fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <asm/unaligned.h>
  32 #include "ctree.h"
  33 #include "disk-io.h"
  34 #include "transaction.h"
  35 #include "btrfs_inode.h"
  36 #include "print-tree.h"
  37 #include "ordered-data.h"
  38 #include "xattr.h"
  39 #include "tree-log.h"
  40 #include "volumes.h"
  41 #include "compression.h"
  42 #include "locking.h"
  43 #include "free-space-cache.h"
  44 #include "inode-map.h"
  45 #include "backref.h"
  46 #include "props.h"
  47 #include "qgroup.h"
  48 #include "dedupe.h"
  49
  50 struct btrfs_iget_args {
  51         struct btrfs_key *location;
  52         struct btrfs_root *root;
  53 };
  54
  55 struct btrfs_dio_data {
  56         u64 reserve;
  57         u64 unsubmitted_oe_range_start;
  58         u64 unsubmitted_oe_range_end;
  59         int overwrite;
  60 };
  61
  62 static const struct inode_operations btrfs_dir_inode_operations;
  63 static const struct inode_operations btrfs_symlink_inode_operations;
  64 static const struct inode_operations btrfs_dir_ro_inode_operations;
  65 static const struct inode_operations btrfs_special_inode_operations;
  66 static const struct inode_operations btrfs_file_inode_operations;
  67 static const struct address_space_operations btrfs_aops;
  68 static const struct file_operations btrfs_dir_file_operations;
  69 static const struct extent_io_ops btrfs_extent_io_ops;
  70
  71 static struct kmem_cache *btrfs_inode_cachep;
  72 struct kmem_cache *btrfs_trans_handle_cachep;
  73 struct kmem_cache *btrfs_path_cachep;
  74 struct kmem_cache *btrfs_free_space_cachep;
  75
  76 #define S_SHIFT 12
  77 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  78         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  79         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  80         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  81         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  82         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  83         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  84         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  85 };
  86
  87 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  88 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  89 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  90 static noinline int cow_file_range(struct inode *inode,
  91                                    struct page *locked_page,
  92                                    u64 start, u64 end, u64 delalloc_end,
  93                                    int *page_started, unsigned long *nr_written,
  94                                    int unlock, struct btrfs_dedupe_hash *hash);
  95 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  96                                        u64 orig_start, u64 block_start,
  97                                        u64 block_len, u64 orig_block_len,
  98                                        u64 ram_bytes, int compress_type,
  99                                        int type);
 100
 101 static void __endio_write_update_ordered(struct inode *inode,
 102                                          const u64 offset, const u64 bytes,
 103                                          const bool uptodate);
 104
 105 /*
 106  * Cleanup all submitted ordered extents in specified range to handle errors
 107  * from the btrfs_run_delalloc_range() callback.
 108  *
 109  * NOTE: caller must ensure that when an error happens, it can not call
 110  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 111  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 112  * to be released, which we want to happen only when finishing the ordered
 113  * extent (btrfs_finish_ordered_io()).
 114  */
 115 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 116                                                  struct page *locked_page,
 117                                                  u64 offset, u64 bytes)
 118 {
 119         unsigned long index = offset >> PAGE_SHIFT;
 120         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 121         u64 page_start = page_offset(locked_page);
 122         u64 page_end = page_start + PAGE_SIZE - 1;
 123
 124         struct page *page;
 125
 126         while (index <= end_index) {
 127                 page = find_get_page(inode->i_mapping, index);
 128                 index++;
 129                 if (!page)
 130                         continue;
 131                 ClearPagePrivate2(page);
 132                 put_page(page);
 133         }
 134
 135         /*
 136          * In case this page belongs to the delalloc range being instantiated
 137          * then skip it, since the first page of a range is going to be
 138          * properly cleaned up by the caller of run_delalloc_range
 139          */
 140         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 141                 offset += PAGE_SIZE;
 142                 bytes -= PAGE_SIZE;
 143         }
 144
 145         return __endio_write_update_ordered(inode, offset, bytes, false);
 146 }
 147
 148 static int btrfs_dirty_inode(struct inode *inode);
 149
 150 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 151 void btrfs_test_inode_set_ops(struct inode *inode)
 152 {
 153         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 154 }
 155 #endif
 156
 157 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 158                                      struct inode *inode,  struct inode *dir,
 159                                      const struct qstr *qstr)
 160 {
 161         int err;
 162
 163         err = btrfs_init_acl(trans, inode, dir);
 164         if (!err)
 165                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 166         return err;
 167 }
 168
 169 /*
 170  * this does all the hard work for inserting an inline extent into
 171  * the btree.  The caller should have done a btrfs_drop_extents so that
 172  * no overlapping inline items exist in the btree
 173  */
 174 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 175                                 struct btrfs_path *path, int extent_inserted,
 176                                 struct btrfs_root *root, struct inode *inode,
 177                                 u64 start, size_t size, size_t compressed_size,
 178                                 int compress_type,
 179                                 struct page **compressed_pages)
 180 {
 181         struct extent_buffer *leaf;
 182         struct page *page = NULL;
 183         char *kaddr;
 184         unsigned long ptr;
 185         struct btrfs_file_extent_item *ei;
 186         int ret;
 187         size_t cur_size = size;
 188         unsigned long offset;
 189
 190         if (compressed_size && compressed_pages)
 191                 cur_size = compressed_size;
 192
 193         inode_add_bytes(inode, size);
 194
 195         if (!extent_inserted) {
 196                 struct btrfs_key key;
 197                 size_t datasize;
 198
 199                 key.objectid = btrfs_ino(BTRFS_I(inode));
 200                 key.offset = start;
 201                 key.type = BTRFS_EXTENT_DATA_KEY;
 202
 203                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 204                 path->leave_spinning = 1;
 205                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 206                                               datasize);
 207                 if (ret)
 208                         goto fail;
 209         }
 210         leaf = path->nodes[0];
 211         ei = btrfs_item_ptr(leaf, path->slots[0],
 212                             struct btrfs_file_extent_item);
 213         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 214         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 215         btrfs_set_file_extent_encryption(leaf, ei, 0);
 216         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 217         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 218         ptr = btrfs_file_extent_inline_start(ei);
 219
 220         if (compress_type != BTRFS_COMPRESS_NONE) {
 221                 struct page *cpage;
 222                 int i = 0;
 223                 while (compressed_size > 0) {
 224                         cpage = compressed_pages[i];
 225                         cur_size = min_t(unsigned long, compressed_size,
 226                                        PAGE_SIZE);
 227
 228                         kaddr = kmap_atomic(cpage);
 229                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 230                         kunmap_atomic(kaddr);
 231
 232                         i++;
 233                         ptr += cur_size;
 234                         compressed_size -= cur_size;
 235                 }
 236                 btrfs_set_file_extent_compression(leaf, ei,
 237                                                   compress_type);
 238         } else {
 239                 page = find_get_page(inode->i_mapping,
 240                                      start >> PAGE_SHIFT);
 241                 btrfs_set_file_extent_compression(leaf, ei, 0);
 242                 kaddr = kmap_atomic(page);
 243                 offset = offset_in_page(start);
 244                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 245                 kunmap_atomic(kaddr);
 246                 put_page(page);
 247         }
 248         btrfs_mark_buffer_dirty(leaf);
 249         btrfs_release_path(path);
 250
 251         /*
 252          * we're an inline extent, so nobody can
 253          * extend the file past i_size without locking
 254          * a page we already have locked.
 255          *
 256          * We must do any isize and inode updates
 257          * before we unlock the pages.  Otherwise we
 258          * could end up racing with unlink.
 259          */
 260         BTRFS_I(inode)->disk_i_size = inode->i_size;
 261         ret = btrfs_update_inode(trans, root, inode);
 262
 263 fail:
 264         return ret;
 265 }
 266
 267
 268 /*
 269  * conditionally insert an inline extent into the file.  This
 270  * does the checks required to make sure the data is small enough
 271  * to fit as an inline extent.
 272  */
 273 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 274                                           u64 end, size_t compressed_size,
 275                                           int compress_type,
 276                                           struct page **compressed_pages)
 277 {
 278         struct btrfs_root *root = BTRFS_I(inode)->root;
 279         struct btrfs_fs_info *fs_info = root->fs_info;
 280         struct btrfs_trans_handle *trans;
 281         u64 isize = i_size_read(inode);
 282         u64 actual_end = min(end + 1, isize);
 283         u64 inline_len = actual_end - start;
 284         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 285         u64 data_len = inline_len;
 286         int ret;
 287         struct btrfs_path *path;
 288         int extent_inserted = 0;
 289         u32 extent_item_size;
 290
 291         if (compressed_size)
 292                 data_len = compressed_size;
 293
 294         if (start > 0 ||
 295             actual_end > fs_info->sectorsize ||
 296             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 297             (!compressed_size &&
 298             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 299             end + 1 < isize ||
 300             data_len > fs_info->max_inline) {
 301                 return 1;
 302         }
 303
 304         path = btrfs_alloc_path();
 305         if (!path)
 306                 return -ENOMEM;
 307
 308         trans = btrfs_join_transaction(root);
 309         if (IS_ERR(trans)) {
 310                 btrfs_free_path(path);
 311                 return PTR_ERR(trans);
 312         }
 313         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 314
 315         if (compressed_size && compressed_pages)
 316                 extent_item_size = btrfs_file_extent_calc_inline_size(
 317                    compressed_size);
 318         else
 319                 extent_item_size = btrfs_file_extent_calc_inline_size(
 320                     inline_len);
 321
 322         ret = __btrfs_drop_extents(trans, root, inode, path,
 323                                    start, aligned_end, NULL,
 324                                    1, 1, extent_item_size, &extent_inserted);
 325         if (ret) {
 326                 btrfs_abort_transaction(trans, ret);
 327                 goto out;
 328         }
 329
 330         if (isize > actual_end)
 331                 inline_len = min_t(u64, isize, actual_end);
 332         ret = insert_inline_extent(trans, path, extent_inserted,
 333                                    root, inode, start,
 334                                    inline_len, compressed_size,
 335                                    compress_type, compressed_pages);
 336         if (ret && ret != -ENOSPC) {
 337                 btrfs_abort_transaction(trans, ret);
 338                 goto out;
 339         } else if (ret == -ENOSPC) {
 340                 ret = 1;
 341                 goto out;
 342         }
 343
 344         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 345         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 346 out:
 347         /*
 348          * Don't forget to free the reserved space, as for inlined extent
 349          * it won't count as data extent, free them directly here.
 350          * And at reserve time, it's always aligned to page size, so
 351          * just free one page here.
 352          */
 353         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 354         btrfs_free_path(path);
 355         btrfs_end_transaction(trans);
 356         return ret;
 357 }
 358
 359 struct async_extent {
 360         u64 start;
 361         u64 ram_size;
 362         u64 compressed_size;
 363         struct page **pages;
 364         unsigned long nr_pages;
 365         int compress_type;
 366         struct list_head list;
 367 };
 368
 369 struct async_cow {
 370         struct inode *inode;
 371         struct btrfs_fs_info *fs_info;
 372         struct page *locked_page;
 373         u64 start;
 374         u64 end;
 375         unsigned int write_flags;
 376         struct list_head extents;
 377         struct btrfs_work work;
 378 };
 379
 380 static noinline int add_async_extent(struct async_cow *cow,
 381                                      u64 start, u64 ram_size,
 382                                      u64 compressed_size,
 383                                      struct page **pages,
 384                                      unsigned long nr_pages,
 385                                      int compress_type)
 386 {
 387         struct async_extent *async_extent;
 388
 389         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 390         BUG_ON(!async_extent); /* -ENOMEM */
 391         async_extent->start = start;
 392         async_extent->ram_size = ram_size;
 393         async_extent->compressed_size = compressed_size;
 394         async_extent->pages = pages;
 395         async_extent->nr_pages = nr_pages;
 396         async_extent->compress_type = compress_type;
 397         list_add_tail(&async_extent->list, &cow->extents);
 398         return 0;
 399 }
 400
 401 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 402 {
 403         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 404
 405         /* force compress */
 406         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 407                 return 1;
 408         /* defrag ioctl */
 409         if (BTRFS_I(inode)->defrag_compress)
 410                 return 1;
 411         /* bad compression ratios */
 412         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 413                 return 0;
 414         if (btrfs_test_opt(fs_info, COMPRESS) ||
 415             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 416             BTRFS_I(inode)->prop_compress)
 417                 return btrfs_compress_heuristic(inode, start, end);
 418         return 0;
 419 }
 420
 421 static inline void inode_should_defrag(struct btrfs_inode *inode,
 422                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 423 {
 424         /* If this is a small write inside eof, kick off a defrag */
 425         if (num_bytes < small_write &&
 426             (start > 0 || end + 1 < inode->disk_i_size))
 427                 btrfs_add_inode_defrag(NULL, inode);
 428 }
 429
 430 /*
 431  * we create compressed extents in two phases.  The first
 432  * phase compresses a range of pages that have already been
 433  * locked (both pages and state bits are locked).
 434  *
 435  * This is done inside an ordered work queue, and the compression
 436  * is spread across many cpus.  The actual IO submission is step
 437  * two, and the ordered work queue takes care of making sure that
 438  * happens in the same order things were put onto the queue by
 439  * writepages and friends.
 440  *
 441  * If this code finds it can't get good compression, it puts an
 442  * entry onto the work queue to write the uncompressed bytes.  This
 443  * makes sure that both compressed inodes and uncompressed inodes
 444  * are written in the same order that the flusher thread sent them
 445  * down.
 446  */
 447 static noinline void compress_file_range(struct inode *inode,
 448                                         struct page *locked_page,
 449                                         u64 start, u64 end,
 450                                         struct async_cow *async_cow,
 451                                         int *num_added)
 452 {
 453         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 454         u64 blocksize = fs_info->sectorsize;
 455         u64 actual_end;
 456         int ret = 0;
 457         struct page **pages = NULL;
 458         unsigned long nr_pages;
 459         unsigned long total_compressed = 0;
 460         unsigned long total_in = 0;
 461         int i;
 462         int will_compress;
 463         int compress_type = fs_info->compress_type;
 464         int redirty = 0;
 465
 466         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 467                         SZ_16K);
 468
 469         actual_end = min_t(u64, i_size_read(inode), end + 1);
 470 again:
 471         will_compress = 0;
 472         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 473         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 474         nr_pages = min_t(unsigned long, nr_pages,
 475                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 476
 477         /*
 478          * we don't want to send crud past the end of i_size through
 479          * compression, that's just a waste of CPU time.  So, if the
 480          * end of the file is before the start of our current
 481          * requested range of bytes, we bail out to the uncompressed
 482          * cleanup code that can deal with all of this.
 483          *
 484          * It isn't really the fastest way to fix things, but this is a
 485          * very uncommon corner.
 486          */
 487         if (actual_end <= start)
 488                 goto cleanup_and_bail_uncompressed;
 489
 490         total_compressed = actual_end - start;
 491
 492         /*
 493          * skip compression for a small file range(<=blocksize) that
 494          * isn't an inline extent, since it doesn't save disk space at all.
 495          */
 496         if (total_compressed <= blocksize &&
 497            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 498                 goto cleanup_and_bail_uncompressed;
 499
 500         total_compressed = min_t(unsigned long, total_compressed,
 501                         BTRFS_MAX_UNCOMPRESSED);
 502         total_in = 0;
 503         ret = 0;
 504
 505         /*
 506          * we do compression for mount -o compress and when the
 507          * inode has not been flagged as nocompress.  This flag can
 508          * change at any time if we discover bad compression ratios.
 509          */
 510         if (inode_need_compress(inode, start, end)) {
 511                 WARN_ON(pages);
 512                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 513                 if (!pages) {
 514                         /* just bail out to the uncompressed code */
 515                         nr_pages = 0;
 516                         goto cont;
 517                 }
 518
 519                 if (BTRFS_I(inode)->defrag_compress)
 520                         compress_type = BTRFS_I(inode)->defrag_compress;
 521                 else if (BTRFS_I(inode)->prop_compress)
 522                         compress_type = BTRFS_I(inode)->prop_compress;
 523
 524                 /*
 525                  * we need to call clear_page_dirty_for_io on each
 526                  * page in the range.  Otherwise applications with the file
 527                  * mmap'd can wander in and change the page contents while
 528                  * we are compressing them.
 529                  *
 530                  * If the compression fails for any reason, we set the pages
 531                  * dirty again later on.
 532                  *
 533                  * Note that the remaining part is redirtied, the start pointer
 534                  * has moved, the end is the original one.
 535                  */
 536                 if (!redirty) {
 537                         extent_range_clear_dirty_for_io(inode, start, end);
 538                         redirty = 1;
 539                 }
 540
 541                 /* Compression level is applied here and only here */
 542                 ret = btrfs_compress_pages(
 543                         compress_type | (fs_info->compress_level << 4),
 544                                            inode->i_mapping, start,
 545                                            pages,
 546                                            &nr_pages,
 547                                            &total_in,
 548                                            &total_compressed);
 549
 550                 if (!ret) {
 551                         unsigned long offset = offset_in_page(total_compressed);
 552                         struct page *page = pages[nr_pages - 1];
 553                         char *kaddr;
 554
 555                         /* zero the tail end of the last page, we might be
 556                          * sending it down to disk
 557                          */
 558                         if (offset) {
 559                                 kaddr = kmap_atomic(page);
 560                                 memset(kaddr + offset, 0,
 561                                        PAGE_SIZE - offset);
 562                                 kunmap_atomic(kaddr);
 563                         }
 564                         will_compress = 1;
 565                 }
 566         }
 567 cont:
 568         if (start == 0) {
 569                 /* lets try to make an inline extent */
 570                 if (ret || total_in < actual_end) {
 571                         /* we didn't compress the entire range, try
 572                          * to make an uncompressed inline extent.
 573                          */
 574                         ret = cow_file_range_inline(inode, start, end, 0,
 575                                                     BTRFS_COMPRESS_NONE, NULL);
 576                 } else {
 577                         /* try making a compressed inline extent */
 578                         ret = cow_file_range_inline(inode, start, end,
 579                                                     total_compressed,
 580                                                     compress_type, pages);
 581                 }
 582                 if (ret <= 0) {
 583                         unsigned long clear_flags = EXTENT_DELALLOC |
 584                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 585                                 EXTENT_DO_ACCOUNTING;
 586                         unsigned long page_error_op;
 587
 588                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 589
 590                         /*
 591                          * inline extent creation worked or returned error,
 592                          * we don't need to create any more async work items.
 593                          * Unlock and free up our temp pages.
 594                          *
 595                          * We use DO_ACCOUNTING here because we need the
 596                          * delalloc_release_metadata to be done _after_ we drop
 597                          * our outstanding extent for clearing delalloc for this
 598                          * range.
 599                          */
 600                         extent_clear_unlock_delalloc(inode, start, end, end,
 601                                                      NULL, clear_flags,
 602                                                      PAGE_UNLOCK |
 603                                                      PAGE_CLEAR_DIRTY |
 604                                                      PAGE_SET_WRITEBACK |
 605                                                      page_error_op |
 606                                                      PAGE_END_WRITEBACK);
 607                         goto free_pages_out;
 608                 }
 609         }
 610
 611         if (will_compress) {
 612                 /*
 613                  * we aren't doing an inline extent round the compressed size
 614                  * up to a block size boundary so the allocator does sane
 615                  * things
 616                  */
 617                 total_compressed = ALIGN(total_compressed, blocksize);
 618
 619                 /*
 620                  * one last check to make sure the compression is really a
 621                  * win, compare the page count read with the blocks on disk,
 622                  * compression must free at least one sector size
 623                  */
 624                 total_in = ALIGN(total_in, PAGE_SIZE);
 625                 if (total_compressed + blocksize <= total_in) {
 626                         *num_added += 1;
 627
 628                         /*
 629                          * The async work queues will take care of doing actual
 630                          * allocation on disk for these compressed pages, and
 631                          * will submit them to the elevator.
 632                          */
 633                         add_async_extent(async_cow, start, total_in,
 634                                         total_compressed, pages, nr_pages,
 635                                         compress_type);
 636
 637                         if (start + total_in < end) {
 638                                 start += total_in;
 639                                 pages = NULL;
 640                                 cond_resched();
 641                                 goto again;
 642                         }
 643                         return;
 644                 }
 645         }
 646         if (pages) {
 647                 /*
 648                  * the compression code ran but failed to make things smaller,
 649                  * free any pages it allocated and our page pointer array
 650                  */
 651                 for (i = 0; i < nr_pages; i++) {
 652                         WARN_ON(pages[i]->mapping);
 653                         put_page(pages[i]);
 654                 }
 655                 kfree(pages);
 656                 pages = NULL;
 657                 total_compressed = 0;
 658                 nr_pages = 0;
 659
 660                 /* flag the file so we don't compress in the future */
 661                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 662                     !(BTRFS_I(inode)->prop_compress)) {
 663                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 664                 }
 665         }
 666 cleanup_and_bail_uncompressed:
 667         /*
 668          * No compression, but we still need to write the pages in the file
 669          * we've been given so far.  redirty the locked page if it corresponds
 670          * to our extent and set things up for the async work queue to run
 671          * cow_file_range to do the normal delalloc dance.
 672          */
 673         if (page_offset(locked_page) >= start &&
 674             page_offset(locked_page) <= end)
 675                 __set_page_dirty_nobuffers(locked_page);
 676                 /* unlocked later on in the async handlers */
 677
 678         if (redirty)
 679                 extent_range_redirty_for_io(inode, start, end);
 680         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 681                          BTRFS_COMPRESS_NONE);
 682         *num_added += 1;
 683
 684         return;
 685
 686 free_pages_out:
 687         for (i = 0; i < nr_pages; i++) {
 688                 WARN_ON(pages[i]->mapping);
 689                 put_page(pages[i]);
 690         }
 691         kfree(pages);
 692 }
 693
 694 static void free_async_extent_pages(struct async_extent *async_extent)
 695 {
 696         int i;
 697
 698         if (!async_extent->pages)
 699                 return;
 700
 701         for (i = 0; i < async_extent->nr_pages; i++) {
 702                 WARN_ON(async_extent->pages[i]->mapping);
 703                 put_page(async_extent->pages[i]);
 704         }
 705         kfree(async_extent->pages);
 706         async_extent->nr_pages = 0;
 707         async_extent->pages = NULL;
 708 }
 709
 710 /*
 711  * phase two of compressed writeback.  This is the ordered portion
 712  * of the code, which only gets called in the order the work was
 713  * queued.  We walk all the async extents created by compress_file_range
 714  * and send them down to the disk.
 715  */
 716 static noinline void submit_compressed_extents(struct async_cow *async_cow)
 717 {
 718         struct inode *inode = async_cow->inode;
 719         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 720         struct async_extent *async_extent;
 721         u64 alloc_hint = 0;
 722         struct btrfs_key ins;
 723         struct extent_map *em;
 724         struct btrfs_root *root = BTRFS_I(inode)->root;
 725         struct extent_io_tree *io_tree;
 726         int ret = 0;
 727
 728 again:
 729         while (!list_empty(&async_cow->extents)) {
 730                 async_extent = list_entry(async_cow->extents.next,
 731                                           struct async_extent, list);
 732                 list_del(&async_extent->list);
 733
 734                 io_tree = &BTRFS_I(inode)->io_tree;
 735
 736 retry:
 737                 /* did the compression code fall back to uncompressed IO? */
 738                 if (!async_extent->pages) {
 739                         int page_started = 0;
 740                         unsigned long nr_written = 0;
 741
 742                         lock_extent(io_tree, async_extent->start,
 743                                          async_extent->start +
 744                                          async_extent->ram_size - 1);
 745
 746                         /* allocate blocks */
 747                         ret = cow_file_range(inode, async_cow->locked_page,
 748                                              async_extent->start,
 749                                              async_extent->start +
 750                                              async_extent->ram_size - 1,
 751                                              async_extent->start +
 752                                              async_extent->ram_size - 1,
 753                                              &page_started, &nr_written, 0,
 754                                              NULL);
 755
 756                         /* JDM XXX */
 757
 758                         /*
 759                          * if page_started, cow_file_range inserted an
 760                          * inline extent and took care of all the unlocking
 761                          * and IO for us.  Otherwise, we need to submit
 762                          * all those pages down to the drive.
 763                          */
 764                         if (!page_started && !ret)
 765                                 extent_write_locked_range(inode,
 766                                                   async_extent->start,
 767                                                   async_extent->start +
 768                                                   async_extent->ram_size - 1,
 769                                                   WB_SYNC_ALL);
 770                         else if (ret)
 771                                 unlock_page(async_cow->locked_page);
 772                         kfree(async_extent);
 773                         cond_resched();
 774                         continue;
 775                 }
 776
 777                 lock_extent(io_tree, async_extent->start,
 778                             async_extent->start + async_extent->ram_size - 1);
 779
 780                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 781                                            async_extent->compressed_size,
 782                                            async_extent->compressed_size,
 783                                            0, alloc_hint, &ins, 1, 1);
 784                 if (ret) {
 785                         free_async_extent_pages(async_extent);
 786
 787                         if (ret == -ENOSPC) {
 788                                 unlock_extent(io_tree, async_extent->start,
 789                                               async_extent->start +
 790                                               async_extent->ram_size - 1);
 791
 792                                 /*
 793                                  * we need to redirty the pages if we decide to
 794                                  * fallback to uncompressed IO, otherwise we
 795                                  * will not submit these pages down to lower
 796                                  * layers.
 797                                  */
 798                                 extent_range_redirty_for_io(inode,
 799                                                 async_extent->start,
 800                                                 async_extent->start +
 801                                                 async_extent->ram_size - 1);
 802
 803                                 goto retry;
 804                         }
 805                         goto out_free;
 806                 }
 807                 /*
 808                  * here we're doing allocation and writeback of the
 809                  * compressed pages
 810                  */
 811                 em = create_io_em(inode, async_extent->start,
 812                                   async_extent->ram_size, /* len */
 813                                   async_extent->start, /* orig_start */
 814                                   ins.objectid, /* block_start */
 815                                   ins.offset, /* block_len */
 816                                   ins.offset, /* orig_block_len */
 817                                   async_extent->ram_size, /* ram_bytes */
 818                                   async_extent->compress_type,
 819                                   BTRFS_ORDERED_COMPRESSED);
 820                 if (IS_ERR(em))
 821                         /* ret value is not necessary due to void function */
 822                         goto out_free_reserve;
 823                 free_extent_map(em);
 824
 825                 ret = btrfs_add_ordered_extent_compress(inode,
 826                                                 async_extent->start,
 827                                                 ins.objectid,
 828                                                 async_extent->ram_size,
 829                                                 ins.offset,
 830                                                 BTRFS_ORDERED_COMPRESSED,
 831                                                 async_extent->compress_type);
 832                 if (ret) {
 833                         btrfs_drop_extent_cache(BTRFS_I(inode),
 834                                                 async_extent->start,
 835                                                 async_extent->start +
 836                                                 async_extent->ram_size - 1, 0);
 837                         goto out_free_reserve;
 838                 }
 839                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 840
 841                 /*
 842                  * clear dirty, set writeback and unlock the pages.
 843                  */
 844                 extent_clear_unlock_delalloc(inode, async_extent->start,
 845                                 async_extent->start +
 846                                 async_extent->ram_size - 1,
 847                                 async_extent->start +
 848                                 async_extent->ram_size - 1,
 849                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 850                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 851                                 PAGE_SET_WRITEBACK);
 852                 if (btrfs_submit_compressed_write(inode,
 853                                     async_extent->start,
 854                                     async_extent->ram_size,
 855                                     ins.objectid,
 856                                     ins.offset, async_extent->pages,
 857                                     async_extent->nr_pages,
 858                                     async_cow->write_flags)) {
 859                         struct page *p = async_extent->pages[0];
 860                         const u64 start = async_extent->start;
 861                         const u64 end = start + async_extent->ram_size - 1;
 862
 863                         p->mapping = inode->i_mapping;
 864                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 865
 866                         p->mapping = NULL;
 867                         extent_clear_unlock_delalloc(inode, start, end, end,
 868                                                      NULL, 0,
 869                                                      PAGE_END_WRITEBACK |
 870                                                      PAGE_SET_ERROR);
 871                         free_async_extent_pages(async_extent);
 872                 }
 873                 alloc_hint = ins.objectid + ins.offset;
 874                 kfree(async_extent);
 875                 cond_resched();
 876         }
 877         return;
 878 out_free_reserve:
 879         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 880         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 881 out_free:
 882         extent_clear_unlock_delalloc(inode, async_extent->start,
 883                                      async_extent->start +
 884                                      async_extent->ram_size - 1,
 885                                      async_extent->start +
 886                                      async_extent->ram_size - 1,
 887                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 888                                      EXTENT_DELALLOC_NEW |
 889                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 890                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 891                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 892                                      PAGE_SET_ERROR);
 893         free_async_extent_pages(async_extent);
 894         kfree(async_extent);
 895         goto again;
 896 }
 897
 898 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 899                                       u64 num_bytes)
 900 {
 901         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 902         struct extent_map *em;
 903         u64 alloc_hint = 0;
 904
 905         read_lock(&em_tree->lock);
 906         em = search_extent_mapping(em_tree, start, num_bytes);
 907         if (em) {
 908                 /*
 909                  * if block start isn't an actual block number then find the
 910                  * first block in this inode and use that as a hint.  If that
 911                  * block is also bogus then just don't worry about it.
 912                  */
 913                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 914                         free_extent_map(em);
 915                         em = search_extent_mapping(em_tree, 0, 0);
 916                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 917                                 alloc_hint = em->block_start;
 918                         if (em)
 919                                 free_extent_map(em);
 920                 } else {
 921                         alloc_hint = em->block_start;
 922                         free_extent_map(em);
 923                 }
 924         }
 925         read_unlock(&em_tree->lock);
 926
 927         return alloc_hint;
 928 }
 929
 930 /*
 931  * when extent_io.c finds a delayed allocation range in the file,
 932  * the call backs end up in this code.  The basic idea is to
 933  * allocate extents on disk for the range, and create ordered data structs
 934  * in ram to track those extents.
 935  *
 936  * locked_page is the page that writepage had locked already.  We use
 937  * it to make sure we don't do extra locks or unlocks.
 938  *
 939  * *page_started is set to one if we unlock locked_page and do everything
 940  * required to start IO on it.  It may be clean and already done with
 941  * IO when we return.
 942  */
 943 static noinline int cow_file_range(struct inode *inode,
 944                                    struct page *locked_page,
 945                                    u64 start, u64 end, u64 delalloc_end,
 946                                    int *page_started, unsigned long *nr_written,
 947                                    int unlock, struct btrfs_dedupe_hash *hash)
 948 {
 949         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 950         struct btrfs_root *root = BTRFS_I(inode)->root;
 951         u64 alloc_hint = 0;
 952         u64 num_bytes;
 953         unsigned long ram_size;
 954         u64 cur_alloc_size = 0;
 955         u64 blocksize = fs_info->sectorsize;
 956         struct btrfs_key ins;
 957         struct extent_map *em;
 958         unsigned clear_bits;
 959         unsigned long page_ops;
 960         bool extent_reserved = false;
 961         int ret = 0;
 962
 963         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 964                 WARN_ON_ONCE(1);
 965                 ret = -EINVAL;
 966                 goto out_unlock;
 967         }
 968
 969         num_bytes = ALIGN(end - start + 1, blocksize);
 970         num_bytes = max(blocksize,  num_bytes);
 971         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 972
 973         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 974
 975         if (start == 0) {
 976                 /* lets try to make an inline extent */
 977                 ret = cow_file_range_inline(inode, start, end, 0,
 978                                             BTRFS_COMPRESS_NONE, NULL);
 979                 if (ret == 0) {
 980                         /*
 981                          * We use DO_ACCOUNTING here because we need the
 982                          * delalloc_release_metadata to be run _after_ we drop
 983                          * our outstanding extent for clearing delalloc for this
 984                          * range.
 985                          */
 986                         extent_clear_unlock_delalloc(inode, start, end,
 987                                      delalloc_end, NULL,
 988                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 989                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 990                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 991                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 992                                      PAGE_END_WRITEBACK);
 993                         *nr_written = *nr_written +
 994                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 995                         *page_started = 1;
 996                         goto out;
 997                 } else if (ret < 0) {
 998                         goto out_unlock;
 999                 }
1000         }
1001
1002         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1003         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1004                         start + num_bytes - 1, 0);
1005
1006         while (num_bytes > 0) {
1007                 cur_alloc_size = num_bytes;
1008                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1009                                            fs_info->sectorsize, 0, alloc_hint,
1010                                            &ins, 1, 1);
1011                 if (ret < 0)
1012                         goto out_unlock;
1013                 cur_alloc_size = ins.offset;
1014                 extent_reserved = true;
1015
1016                 ram_size = ins.offset;
1017                 em = create_io_em(inode, start, ins.offset, /* len */
1018                                   start, /* orig_start */
1019                                   ins.objectid, /* block_start */
1020                                   ins.offset, /* block_len */
1021                                   ins.offset, /* orig_block_len */
1022                                   ram_size, /* ram_bytes */
1023                                   BTRFS_COMPRESS_NONE, /* compress_type */
1024                                   BTRFS_ORDERED_REGULAR /* type */);
1025                 if (IS_ERR(em)) {
1026                         ret = PTR_ERR(em);
1027                         goto out_reserve;
1028                 }
1029                 free_extent_map(em);
1030
1031                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1032                                                ram_size, cur_alloc_size, 0);
1033                 if (ret)
1034                         goto out_drop_extent_cache;
1035
1036                 if (root->root_key.objectid ==
1037                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1038                         ret = btrfs_reloc_clone_csums(inode, start,
1039                                                       cur_alloc_size);
1040                         /*
1041                          * Only drop cache here, and process as normal.
1042                          *
1043                          * We must not allow extent_clear_unlock_delalloc()
1044                          * at out_unlock label to free meta of this ordered
1045                          * extent, as its meta should be freed by
1046                          * btrfs_finish_ordered_io().
1047                          *
1048                          * So we must continue until @start is increased to
1049                          * skip current ordered extent.
1050                          */
1051                         if (ret)
1052                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1053                                                 start + ram_size - 1, 0);
1054                 }
1055
1056                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1057
1058                 /* we're not doing compressed IO, don't unlock the first
1059                  * page (which the caller expects to stay locked), don't
1060                  * clear any dirty bits and don't set any writeback bits
1061                  *
1062                  * Do set the Private2 bit so we know this page was properly
1063                  * setup for writepage
1064                  */
1065                 page_ops = unlock ? PAGE_UNLOCK : 0;
1066                 page_ops |= PAGE_SET_PRIVATE2;
1067
1068                 extent_clear_unlock_delalloc(inode, start,
1069                                              start + ram_size - 1,
1070                                              delalloc_end, locked_page,
1071                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1072                                              page_ops);
1073                 if (num_bytes < cur_alloc_size)
1074                         num_bytes = 0;
1075                 else
1076                         num_bytes -= cur_alloc_size;
1077                 alloc_hint = ins.objectid + ins.offset;
1078                 start += cur_alloc_size;
1079                 extent_reserved = false;
1080
1081                 /*
1082                  * btrfs_reloc_clone_csums() error, since start is increased
1083                  * extent_clear_unlock_delalloc() at out_unlock label won't
1084                  * free metadata of current ordered extent, we're OK to exit.
1085                  */
1086                 if (ret)
1087                         goto out_unlock;
1088         }
1089 out:
1090         return ret;
1091
1092 out_drop_extent_cache:
1093         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1094 out_reserve:
1095         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1096         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1097 out_unlock:
1098         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1099                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1100         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1101                 PAGE_END_WRITEBACK;
1102         /*
1103          * If we reserved an extent for our delalloc range (or a subrange) and
1104          * failed to create the respective ordered extent, then it means that
1105          * when we reserved the extent we decremented the extent's size from
1106          * the data space_info's bytes_may_use counter and incremented the
1107          * space_info's bytes_reserved counter by the same amount. We must make
1108          * sure extent_clear_unlock_delalloc() does not try to decrement again
1109          * the data space_info's bytes_may_use counter, therefore we do not pass
1110          * it the flag EXTENT_CLEAR_DATA_RESV.
1111          */
1112         if (extent_reserved) {
1113                 extent_clear_unlock_delalloc(inode, start,
1114                                              start + cur_alloc_size,
1115                                              start + cur_alloc_size,
1116                                              locked_page,
1117                                              clear_bits,
1118                                              page_ops);
1119                 start += cur_alloc_size;
1120                 if (start >= end)
1121                         goto out;
1122         }
1123         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1124                                      locked_page,
1125                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1126                                      page_ops);
1127         goto out;
1128 }
1129
1130 /*
1131  * work queue call back to started compression on a file and pages
1132  */
1133 static noinline void async_cow_start(struct btrfs_work *work)
1134 {
1135         struct async_cow *async_cow;
1136         int num_added = 0;
1137         async_cow = container_of(work, struct async_cow, work);
1138
1139         compress_file_range(async_cow->inode, async_cow->locked_page,
1140                             async_cow->start, async_cow->end, async_cow,
1141                             &num_added);
1142         if (num_added == 0) {
1143                 btrfs_add_delayed_iput(async_cow->inode);
1144                 async_cow->inode = NULL;
1145         }
1146 }
1147
1148 /*
1149  * work queue call back to submit previously compressed pages
1150  */
1151 static noinline void async_cow_submit(struct btrfs_work *work)
1152 {
1153         struct btrfs_fs_info *fs_info;
1154         struct async_cow *async_cow;
1155         unsigned long nr_pages;
1156
1157         async_cow = container_of(work, struct async_cow, work);
1158
1159         fs_info = async_cow->fs_info;
1160         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1161                 PAGE_SHIFT;
1162
1163         /* atomic_sub_return implies a barrier */
1164         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1165             5 * SZ_1M)
1166                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1167
1168         /*
1169          * ->inode could be NULL if async_cow_start has failed to compress,
1170          * in which case we don't have anything to submit, yet we need to
1171          * always adjust ->async_delalloc_pages as its paired with the init
1172          * happening in cow_file_range_async
1173          */
1174         if (async_cow->inode)
1175                 submit_compressed_extents(async_cow);
1176 }
1177
1178 static noinline void async_cow_free(struct btrfs_work *work)
1179 {
1180         struct async_cow *async_cow;
1181         async_cow = container_of(work, struct async_cow, work);
1182         if (async_cow->inode)
1183                 btrfs_add_delayed_iput(async_cow->inode);
1184         kfree(async_cow);
1185 }
1186
1187 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1188                                 u64 start, u64 end, int *page_started,
1189                                 unsigned long *nr_written,
1190                                 unsigned int write_flags)
1191 {
1192         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1193         struct async_cow *async_cow;
1194         unsigned long nr_pages;
1195         u64 cur_end;
1196
1197         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1198                          1, 0, NULL);
1199         while (start < end) {
1200                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1201                 BUG_ON(!async_cow); /* -ENOMEM */
1202                 /*
1203                  * igrab is called higher up in the call chain, take only the
1204                  * lightweight reference for the callback lifetime
1205                  */
1206                 ihold(inode);
1207                 async_cow->inode = inode;
1208                 async_cow->fs_info = fs_info;
1209                 async_cow->locked_page = locked_page;
1210                 async_cow->start = start;
1211                 async_cow->write_flags = write_flags;
1212
1213                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1214                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1215                         cur_end = end;
1216                 else
1217                         cur_end = min(end, start + SZ_512K - 1);
1218
1219                 async_cow->end = cur_end;
1220                 INIT_LIST_HEAD(&async_cow->extents);
1221
1222                 btrfs_init_work(&async_cow->work,
1223                                 btrfs_delalloc_helper,
1224                                 async_cow_start, async_cow_submit,
1225                                 async_cow_free);
1226
1227                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1228                         PAGE_SHIFT;
1229                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1230
1231                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1232
1233                 *nr_written += nr_pages;
1234                 start = cur_end + 1;
1235         }
1236         *page_started = 1;
1237         return 0;
1238 }
1239
1240 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1241                                         u64 bytenr, u64 num_bytes)
1242 {
1243         int ret;
1244         struct btrfs_ordered_sum *sums;
1245         LIST_HEAD(list);
1246
1247         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1248                                        bytenr + num_bytes - 1, &list, 0);
1249         if (ret == 0 && list_empty(&list))
1250                 return 0;
1251
1252         while (!list_empty(&list)) {
1253                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1254                 list_del(&sums->list);
1255                 kfree(sums);
1256         }
1257         if (ret < 0)
1258                 return ret;
1259         return 1;
1260 }
1261
1262 /*
1263  * when nowcow writeback call back.  This checks for snapshots or COW copies
1264  * of the extents that exist in the file, and COWs the file as required.
1265  *
1266  * If no cow copies or snapshots exist, we write directly to the existing
1267  * blocks on disk
1268  */
1269 static noinline int run_delalloc_nocow(struct inode *inode,
1270                                        struct page *locked_page,
1271                               u64 start, u64 end, int *page_started, int force,
1272                               unsigned long *nr_written)
1273 {
1274         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1275         struct btrfs_root *root = BTRFS_I(inode)->root;
1276         struct extent_buffer *leaf;
1277         struct btrfs_path *path;
1278         struct btrfs_file_extent_item *fi;
1279         struct btrfs_key found_key;
1280         struct extent_map *em;
1281         u64 cow_start;
1282         u64 cur_offset;
1283         u64 extent_end;
1284         u64 extent_offset;
1285         u64 disk_bytenr;
1286         u64 num_bytes;
1287         u64 disk_num_bytes;
1288         u64 ram_bytes;
1289         int extent_type;
1290         int ret;
1291         int type;
1292         int nocow;
1293         int check_prev = 1;
1294         bool nolock;
1295         u64 ino = btrfs_ino(BTRFS_I(inode));
1296
1297         path = btrfs_alloc_path();
1298         if (!path) {
1299                 extent_clear_unlock_delalloc(inode, start, end, end,
1300                                              locked_page,
1301                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1302                                              EXTENT_DO_ACCOUNTING |
1303                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1304                                              PAGE_CLEAR_DIRTY |
1305                                              PAGE_SET_WRITEBACK |
1306                                              PAGE_END_WRITEBACK);
1307                 return -ENOMEM;
1308         }
1309
1310         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1311
1312         cow_start = (u64)-1;
1313         cur_offset = start;
1314         while (1) {
1315                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1316                                                cur_offset, 0);
1317                 if (ret < 0)
1318                         goto error;
1319                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1320                         leaf = path->nodes[0];
1321                         btrfs_item_key_to_cpu(leaf, &found_key,
1322                                               path->slots[0] - 1);
1323                         if (found_key.objectid == ino &&
1324                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1325                                 path->slots[0]--;
1326                 }
1327                 check_prev = 0;
1328 next_slot:
1329                 leaf = path->nodes[0];
1330                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1331                         ret = btrfs_next_leaf(root, path);
1332                         if (ret < 0) {
1333                                 if (cow_start != (u64)-1)
1334                                         cur_offset = cow_start;
1335                                 goto error;
1336                         }
1337                         if (ret > 0)
1338                                 break;
1339                         leaf = path->nodes[0];
1340                 }
1341
1342                 nocow = 0;
1343                 disk_bytenr = 0;
1344                 num_bytes = 0;
1345                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1346
1347                 if (found_key.objectid > ino)
1348                         break;
1349                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1350                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1351                         path->slots[0]++;
1352                         goto next_slot;
1353                 }
1354                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1355                     found_key.offset > end)
1356                         break;
1357
1358                 if (found_key.offset > cur_offset) {
1359                         extent_end = found_key.offset;
1360                         extent_type = 0;
1361                         goto out_check;
1362                 }
1363
1364                 fi = btrfs_item_ptr(leaf, path->slots[0],
1365                                     struct btrfs_file_extent_item);
1366                 extent_type = btrfs_file_extent_type(leaf, fi);
1367
1368                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1369                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1370                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1371                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1372                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1373                         extent_end = found_key.offset +
1374                                 btrfs_file_extent_num_bytes(leaf, fi);
1375                         disk_num_bytes =
1376                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1377                         if (extent_end <= start) {
1378                                 path->slots[0]++;
1379                                 goto next_slot;
1380                         }
1381                         if (disk_bytenr == 0)
1382                                 goto out_check;
1383                         if (btrfs_file_extent_compression(leaf, fi) ||
1384                             btrfs_file_extent_encryption(leaf, fi) ||
1385                             btrfs_file_extent_other_encoding(leaf, fi))
1386                                 goto out_check;
1387                         /*
1388                          * Do the same check as in btrfs_cross_ref_exist but
1389                          * without the unnecessary search.
1390                          */
1391                         if (!nolock &&
1392                             btrfs_file_extent_generation(leaf, fi) <=
1393                             btrfs_root_last_snapshot(&root->root_item))
1394                                 goto out_check;
1395                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1396                                 goto out_check;
1397                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1398                                 goto out_check;
1399                         ret = btrfs_cross_ref_exist(root, ino,
1400                                                     found_key.offset -
1401                                                     extent_offset, disk_bytenr);
1402                         if (ret) {
1403                                 /*
1404                                  * ret could be -EIO if the above fails to read
1405                                  * metadata.
1406                                  */
1407                                 if (ret < 0) {
1408                                         if (cow_start != (u64)-1)
1409                                                 cur_offset = cow_start;
1410                                         goto error;
1411                                 }
1412
1413                                 WARN_ON_ONCE(nolock);
1414                                 goto out_check;
1415                         }
1416                         disk_bytenr += extent_offset;
1417                         disk_bytenr += cur_offset - found_key.offset;
1418                         num_bytes = min(end + 1, extent_end) - cur_offset;
1419                         /*
1420                          * if there are pending snapshots for this root,
1421                          * we fall into common COW way.
1422                          */
1423                         if (!nolock && atomic_read(&root->snapshot_force_cow))
1424                                 goto out_check;
1425                         /*
1426                          * force cow if csum exists in the range.
1427                          * this ensure that csum for a given extent are
1428                          * either valid or do not exist.
1429                          */
1430                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1431                                                   num_bytes);
1432                         if (ret) {
1433                                 /*
1434                                  * ret could be -EIO if the above fails to read
1435                                  * metadata.
1436                                  */
1437                                 if (ret < 0) {
1438                                         if (cow_start != (u64)-1)
1439                                                 cur_offset = cow_start;
1440                                         goto error;
1441                                 }
1442                                 WARN_ON_ONCE(nolock);
1443                                 goto out_check;
1444                         }
1445                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1446                                 goto out_check;
1447                         nocow = 1;
1448                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1449                         extent_end = found_key.offset +
1450                                 btrfs_file_extent_ram_bytes(leaf, fi);
1451                         extent_end = ALIGN(extent_end,
1452                                            fs_info->sectorsize);
1453                 } else {
1454                         BUG_ON(1);
1455                 }
1456 out_check:
1457                 if (extent_end <= start) {
1458                         path->slots[0]++;
1459                         if (nocow)
1460                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1461                         goto next_slot;
1462                 }
1463                 if (!nocow) {
1464                         if (cow_start == (u64)-1)
1465                                 cow_start = cur_offset;
1466                         cur_offset = extent_end;
1467                         if (cur_offset > end)
1468                                 break;
1469                         path->slots[0]++;
1470                         goto next_slot;
1471                 }
1472
1473                 btrfs_release_path(path);
1474                 if (cow_start != (u64)-1) {
1475                         ret = cow_file_range(inode, locked_page,
1476                                              cow_start, found_key.offset - 1,
1477                                              end, page_started, nr_written, 1,
1478                                              NULL);
1479                         if (ret) {
1480                                 if (nocow)
1481                                         btrfs_dec_nocow_writers(fs_info,
1482                                                                 disk_bytenr);
1483                                 goto error;
1484                         }
1485                         cow_start = (u64)-1;
1486                 }
1487
1488                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1489                         u64 orig_start = found_key.offset - extent_offset;
1490
1491                         em = create_io_em(inode, cur_offset, num_bytes,
1492                                           orig_start,
1493                                           disk_bytenr, /* block_start */
1494                                           num_bytes, /* block_len */
1495                                           disk_num_bytes, /* orig_block_len */
1496                                           ram_bytes, BTRFS_COMPRESS_NONE,
1497                                           BTRFS_ORDERED_PREALLOC);
1498                         if (IS_ERR(em)) {
1499                                 if (nocow)
1500                                         btrfs_dec_nocow_writers(fs_info,
1501                                                                 disk_bytenr);
1502                                 ret = PTR_ERR(em);
1503                                 goto error;
1504                         }
1505                         free_extent_map(em);
1506                 }
1507
1508                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1509                         type = BTRFS_ORDERED_PREALLOC;
1510                 } else {
1511                         type = BTRFS_ORDERED_NOCOW;
1512                 }
1513
1514                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1515                                                num_bytes, num_bytes, type);
1516                 if (nocow)
1517                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1518                 BUG_ON(ret); /* -ENOMEM */
1519
1520                 if (root->root_key.objectid ==
1521                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1522                         /*
1523                          * Error handled later, as we must prevent
1524                          * extent_clear_unlock_delalloc() in error handler
1525                          * from freeing metadata of created ordered extent.
1526                          */
1527                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1528                                                       num_bytes);
1529
1530                 extent_clear_unlock_delalloc(inode, cur_offset,
1531                                              cur_offset + num_bytes - 1, end,
1532                                              locked_page, EXTENT_LOCKED |
1533                                              EXTENT_DELALLOC |
1534                                              EXTENT_CLEAR_DATA_RESV,
1535                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1536
1537                 cur_offset = extent_end;
1538
1539                 /*
1540                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1541                  * handler, as metadata for created ordered extent will only
1542                  * be freed by btrfs_finish_ordered_io().
1543                  */
1544                 if (ret)
1545                         goto error;
1546                 if (cur_offset > end)
1547                         break;
1548         }
1549         btrfs_release_path(path);
1550
1551         if (cur_offset <= end && cow_start == (u64)-1)
1552                 cow_start = cur_offset;
1553
1554         if (cow_start != (u64)-1) {
1555                 cur_offset = end;
1556                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1557                                      page_started, nr_written, 1, NULL);
1558                 if (ret)
1559                         goto error;
1560         }
1561
1562 error:
1563         if (ret && cur_offset < end)
1564                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1565                                              locked_page, EXTENT_LOCKED |
1566                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1567                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1568                                              PAGE_CLEAR_DIRTY |
1569                                              PAGE_SET_WRITEBACK |
1570                                              PAGE_END_WRITEBACK);
1571         btrfs_free_path(path);
1572         return ret;
1573 }
1574
1575 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1576 {
1577
1578         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1579             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1580                 return 0;
1581
1582         /*
1583          * @defrag_bytes is a hint value, no spinlock held here,
1584          * if is not zero, it means the file is defragging.
1585          * Force cow if given extent needs to be defragged.
1586          */
1587         if (BTRFS_I(inode)->defrag_bytes &&
1588             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1589                            EXTENT_DEFRAG, 0, NULL))
1590                 return 1;
1591
1592         return 0;
1593 }
1594
1595 /*
1596  * Function to process delayed allocation (create CoW) for ranges which are
1597  * being touched for the first time.
1598  */
1599 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1600                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1601                 struct writeback_control *wbc)
1602 {
1603         int ret;
1604         int force_cow = need_force_cow(inode, start, end);
1605         unsigned int write_flags = wbc_to_write_flags(wbc);
1606
1607         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1608                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1609                                          page_started, 1, nr_written);
1610         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1611                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1612                                          page_started, 0, nr_written);
1613         } else if (!inode_need_compress(inode, start, end)) {
1614                 ret = cow_file_range(inode, locked_page, start, end, end,
1615                                       page_started, nr_written, 1, NULL);
1616         } else {
1617                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1618                         &BTRFS_I(inode)->runtime_flags);
1619                 ret = cow_file_range_async(inode, locked_page, start, end,
1620                                            page_started, nr_written,
1621                                            write_flags);
1622         }
1623         if (ret)
1624                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1625                                               end - start + 1);
1626         return ret;
1627 }
1628
1629 void btrfs_split_delalloc_extent(struct inode *inode,
1630                                  struct extent_state *orig, u64 split)
1631 {
1632         u64 size;
1633
1634         /* not delalloc, ignore it */
1635         if (!(orig->state & EXTENT_DELALLOC))
1636                 return;
1637
1638         size = orig->end - orig->start + 1;
1639         if (size > BTRFS_MAX_EXTENT_SIZE) {
1640                 u32 num_extents;
1641                 u64 new_size;
1642
1643                 /*
1644                  * See the explanation in btrfs_merge_delalloc_extent, the same
1645                  * applies here, just in reverse.
1646                  */
1647                 new_size = orig->end - split + 1;
1648                 num_extents = count_max_extents(new_size);
1649                 new_size = split - orig->start;
1650                 num_extents += count_max_extents(new_size);
1651                 if (count_max_extents(size) >= num_extents)
1652                         return;
1653         }
1654
1655         spin_lock(&BTRFS_I(inode)->lock);
1656         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1657         spin_unlock(&BTRFS_I(inode)->lock);
1658 }
1659
1660 /*
1661  * Handle merged delayed allocation extents so we can keep track of new extents
1662  * that are just merged onto old extents, such as when we are doing sequential
1663  * writes, so we can properly account for the metadata space we'll need.
1664  */
1665 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1666                                  struct extent_state *other)
1667 {
1668         u64 new_size, old_size;
1669         u32 num_extents;
1670
1671         /* not delalloc, ignore it */
1672         if (!(other->state & EXTENT_DELALLOC))
1673                 return;
1674
1675         if (new->start > other->start)
1676                 new_size = new->end - other->start + 1;
1677         else
1678                 new_size = other->end - new->start + 1;
1679
1680         /* we're not bigger than the max, unreserve the space and go */
1681         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1682                 spin_lock(&BTRFS_I(inode)->lock);
1683                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1684                 spin_unlock(&BTRFS_I(inode)->lock);
1685                 return;
1686         }
1687
1688         /*
1689          * We have to add up either side to figure out how many extents were
1690          * accounted for before we merged into one big extent.  If the number of
1691          * extents we accounted for is <= the amount we need for the new range
1692          * then we can return, otherwise drop.  Think of it like this
1693          *
1694          * [ 4k][MAX_SIZE]
1695          *
1696          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1697          * need 2 outstanding extents, on one side we have 1 and the other side
1698          * we have 1 so they are == and we can return.  But in this case
1699          *
1700          * [MAX_SIZE+4k][MAX_SIZE+4k]
1701          *
1702          * Each range on their own accounts for 2 extents, but merged together
1703          * they are only 3 extents worth of accounting, so we need to drop in
1704          * this case.
1705          */
1706         old_size = other->end - other->start + 1;
1707         num_extents = count_max_extents(old_size);
1708         old_size = new->end - new->start + 1;
1709         num_extents += count_max_extents(old_size);
1710         if (count_max_extents(new_size) >= num_extents)
1711                 return;
1712
1713         spin_lock(&BTRFS_I(inode)->lock);
1714         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1715         spin_unlock(&BTRFS_I(inode)->lock);
1716 }
1717
1718 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1719                                       struct inode *inode)
1720 {
1721         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1722
1723         spin_lock(&root->delalloc_lock);
1724         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1725                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1726                               &root->delalloc_inodes);
1727                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1728                         &BTRFS_I(inode)->runtime_flags);
1729                 root->nr_delalloc_inodes++;
1730                 if (root->nr_delalloc_inodes == 1) {
1731                         spin_lock(&fs_info->delalloc_root_lock);
1732                         BUG_ON(!list_empty(&root->delalloc_root));
1733                         list_add_tail(&root->delalloc_root,
1734                                       &fs_info->delalloc_roots);
1735                         spin_unlock(&fs_info->delalloc_root_lock);
1736                 }
1737         }
1738         spin_unlock(&root->delalloc_lock);
1739 }
1740
1741
1742 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1743                                 struct btrfs_inode *inode)
1744 {
1745         struct btrfs_fs_info *fs_info = root->fs_info;
1746
1747         if (!list_empty(&inode->delalloc_inodes)) {
1748                 list_del_init(&inode->delalloc_inodes);
1749                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1750                           &inode->runtime_flags);
1751                 root->nr_delalloc_inodes--;
1752                 if (!root->nr_delalloc_inodes) {
1753                         ASSERT(list_empty(&root->delalloc_inodes));
1754                         spin_lock(&fs_info->delalloc_root_lock);
1755                         BUG_ON(list_empty(&root->delalloc_root));
1756                         list_del_init(&root->delalloc_root);
1757                         spin_unlock(&fs_info->delalloc_root_lock);
1758                 }
1759         }
1760 }
1761
1762 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1763                                      struct btrfs_inode *inode)
1764 {
1765         spin_lock(&root->delalloc_lock);
1766         __btrfs_del_delalloc_inode(root, inode);
1767         spin_unlock(&root->delalloc_lock);
1768 }
1769
1770 /*
1771  * Properly track delayed allocation bytes in the inode and to maintain the
1772  * list of inodes that have pending delalloc work to be done.
1773  */
1774 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1775                                unsigned *bits)
1776 {
1777         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1778
1779         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1780                 WARN_ON(1);
1781         /*
1782          * set_bit and clear bit hooks normally require _irqsave/restore
1783          * but in this case, we are only testing for the DELALLOC
1784          * bit, which is only set or cleared with irqs on
1785          */
1786         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1787                 struct btrfs_root *root = BTRFS_I(inode)->root;
1788                 u64 len = state->end + 1 - state->start;
1789                 u32 num_extents = count_max_extents(len);
1790                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1791
1792                 spin_lock(&BTRFS_I(inode)->lock);
1793                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1794                 spin_unlock(&BTRFS_I(inode)->lock);
1795
1796                 /* For sanity tests */
1797                 if (btrfs_is_testing(fs_info))
1798                         return;
1799
1800                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1801                                          fs_info->delalloc_batch);
1802                 spin_lock(&BTRFS_I(inode)->lock);
1803                 BTRFS_I(inode)->delalloc_bytes += len;
1804                 if (*bits & EXTENT_DEFRAG)
1805                         BTRFS_I(inode)->defrag_bytes += len;
1806                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1807                                          &BTRFS_I(inode)->runtime_flags))
1808                         btrfs_add_delalloc_inodes(root, inode);
1809                 spin_unlock(&BTRFS_I(inode)->lock);
1810         }
1811
1812         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1813             (*bits & EXTENT_DELALLOC_NEW)) {
1814                 spin_lock(&BTRFS_I(inode)->lock);
1815                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1816                         state->start;
1817                 spin_unlock(&BTRFS_I(inode)->lock);
1818         }
1819 }
1820
1821 /*
1822  * Once a range is no longer delalloc this function ensures that proper
1823  * accounting happens.
1824  */
1825 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1826                                  struct extent_state *state, unsigned *bits)
1827 {
1828         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1829         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1830         u64 len = state->end + 1 - state->start;
1831         u32 num_extents = count_max_extents(len);
1832
1833         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1834                 spin_lock(&inode->lock);
1835                 inode->defrag_bytes -= len;
1836                 spin_unlock(&inode->lock);
1837         }
1838
1839         /*
1840          * set_bit and clear bit hooks normally require _irqsave/restore
1841          * but in this case, we are only testing for the DELALLOC
1842          * bit, which is only set or cleared with irqs on
1843          */
1844         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1845                 struct btrfs_root *root = inode->root;
1846                 bool do_list = !btrfs_is_free_space_inode(inode);
1847
1848                 spin_lock(&inode->lock);
1849                 btrfs_mod_outstanding_extents(inode, -num_extents);
1850                 spin_unlock(&inode->lock);
1851
1852                 /*
1853                  * We don't reserve metadata space for space cache inodes so we
1854                  * don't need to call delalloc_release_metadata if there is an
1855                  * error.
1856                  */
1857                 if (*bits & EXTENT_CLEAR_META_RESV &&
1858                     root != fs_info->tree_root)
1859                         btrfs_delalloc_release_metadata(inode, len, false);
1860
1861                 /* For sanity tests. */
1862                 if (btrfs_is_testing(fs_info))
1863                         return;
1864
1865                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1866                     do_list && !(state->state & EXTENT_NORESERVE) &&
1867                     (*bits & EXTENT_CLEAR_DATA_RESV))
1868                         btrfs_free_reserved_data_space_noquota(
1869                                         &inode->vfs_inode,
1870                                         state->start, len);
1871
1872                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1873                                          fs_info->delalloc_batch);
1874                 spin_lock(&inode->lock);
1875                 inode->delalloc_bytes -= len;
1876                 if (do_list && inode->delalloc_bytes == 0 &&
1877                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1878                                         &inode->runtime_flags))
1879                         btrfs_del_delalloc_inode(root, inode);
1880                 spin_unlock(&inode->lock);
1881         }
1882
1883         if ((state->state & EXTENT_DELALLOC_NEW) &&
1884             (*bits & EXTENT_DELALLOC_NEW)) {
1885                 spin_lock(&inode->lock);
1886                 ASSERT(inode->new_delalloc_bytes >= len);
1887                 inode->new_delalloc_bytes -= len;
1888                 spin_unlock(&inode->lock);
1889         }
1890 }
1891
1892 /*
1893  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
1894  * in a chunk's stripe. This function ensures that bios do not span a
1895  * stripe/chunk
1896  *
1897  * @page - The page we are about to add to the bio
1898  * @size - size we want to add to the bio
1899  * @bio - bio we want to ensure is smaller than a stripe
1900  * @bio_flags - flags of the bio
1901  *
1902  * return 1 if page cannot be added to the bio
1903  * return 0 if page can be added to the bio
1904  * return error otherwise
1905  */
1906 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
1907                              unsigned long bio_flags)
1908 {
1909         struct inode *inode = page->mapping->host;
1910         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1911         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1912         u64 length = 0;
1913         u64 map_length;
1914         int ret;
1915
1916         if (bio_flags & EXTENT_BIO_COMPRESSED)
1917                 return 0;
1918
1919         length = bio->bi_iter.bi_size;
1920         map_length = length;
1921         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1922                               NULL, 0);
1923         if (ret < 0)
1924                 return ret;
1925         if (map_length < length + size)
1926                 return 1;
1927         return 0;
1928 }
1929
1930 /*
1931  * in order to insert checksums into the metadata in large chunks,
1932  * we wait until bio submission time.   All the pages in the bio are
1933  * checksummed and sums are attached onto the ordered extent record.
1934  *
1935  * At IO completion time the cums attached on the ordered extent record
1936  * are inserted into the btree
1937  */
1938 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1939                                     u64 bio_offset)
1940 {
1941         struct inode *inode = private_data;
1942         blk_status_t ret = 0;
1943
1944         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1945         BUG_ON(ret); /* -ENOMEM */
1946         return 0;
1947 }
1948
1949 /*
1950  * extent_io.c submission hook. This does the right thing for csum calculation
1951  * on write, or reading the csums from the tree before a read.
1952  *
1953  * Rules about async/sync submit,
1954  * a) read:                             sync submit
1955  *
1956  * b) write without checksum:           sync submit
1957  *
1958  * c) write with checksum:
1959  *    c-1) if bio is issued by fsync:   sync submit
1960  *         (sync_writers != 0)
1961  *
1962  *    c-2) if root is reloc root:       sync submit
1963  *         (only in case of buffered IO)
1964  *
1965  *    c-3) otherwise:                   async submit
1966  */
1967 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1968                                  int mirror_num, unsigned long bio_flags,
1969                                  u64 bio_offset)
1970 {
1971         struct inode *inode = private_data;
1972         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1973         struct btrfs_root *root = BTRFS_I(inode)->root;
1974         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1975         blk_status_t ret = 0;
1976         int skip_sum;
1977         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1978
1979         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1980
1981         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1982                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1983
1984         if (bio_op(bio) != REQ_OP_WRITE) {
1985                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1986                 if (ret)
1987                         goto out;
1988
1989                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1990                         ret = btrfs_submit_compressed_read(inode, bio,
1991                                                            mirror_num,
1992                                                            bio_flags);
1993                         goto out;
1994                 } else if (!skip_sum) {
1995                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1996                         if (ret)
1997                                 goto out;
1998                 }
1999                 goto mapit;
2000         } else if (async && !skip_sum) {
2001                 /* csum items have already been cloned */
2002                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2003                         goto mapit;
2004                 /* we're doing a write, do the async checksumming */
2005                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2006                                           bio_offset, inode,
2007                                           btrfs_submit_bio_start);
2008                 goto out;
2009         } else if (!skip_sum) {
2010                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2011                 if (ret)
2012                         goto out;
2013         }
2014
2015 mapit:
2016         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2017
2018 out:
2019         if (ret) {
2020                 bio->bi_status = ret;
2021                 bio_endio(bio);
2022         }
2023         return ret;
2024 }
2025
2026 /*
2027  * given a list of ordered sums record them in the inode.  This happens
2028  * at IO completion time based on sums calculated at bio submission time.
2029  */
2030 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2031                              struct inode *inode, struct list_head *list)
2032 {
2033         struct btrfs_ordered_sum *sum;
2034         int ret;
2035
2036         list_for_each_entry(sum, list, list) {
2037                 trans->adding_csums = true;
2038                 ret = btrfs_csum_file_blocks(trans,
2039                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2040                 trans->adding_csums = false;
2041                 if (ret)
2042                         return ret;
2043         }
2044         return 0;
2045 }
2046
2047 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2048                               unsigned int extra_bits,
2049                               struct extent_state **cached_state, int dedupe)
2050 {
2051         WARN_ON(PAGE_ALIGNED(end));
2052         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2053                                    extra_bits, cached_state);
2054 }
2055
2056 /* see btrfs_writepage_start_hook for details on why this is required */
2057 struct btrfs_writepage_fixup {
2058         struct page *page;
2059         struct btrfs_work work;
2060 };
2061
2062 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2063 {
2064         struct btrfs_writepage_fixup *fixup;
2065         struct btrfs_ordered_extent *ordered;
2066         struct extent_state *cached_state = NULL;
2067         struct extent_changeset *data_reserved = NULL;
2068         struct page *page;
2069         struct inode *inode;
2070         u64 page_start;
2071         u64 page_end;
2072         int ret;
2073
2074         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2075         page = fixup->page;
2076 again:
2077         lock_page(page);
2078         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2079                 ClearPageChecked(page);
2080                 goto out_page;
2081         }
2082
2083         inode = page->mapping->host;
2084         page_start = page_offset(page);
2085         page_end = page_offset(page) + PAGE_SIZE - 1;
2086
2087         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2088                          &cached_state);
2089
2090         /* already ordered? We're done */
2091         if (PagePrivate2(page))
2092                 goto out;
2093
2094         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2095                                         PAGE_SIZE);
2096         if (ordered) {
2097                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2098                                      page_end, &cached_state);
2099                 unlock_page(page);
2100                 btrfs_start_ordered_extent(inode, ordered, 1);
2101                 btrfs_put_ordered_extent(ordered);
2102                 goto again;
2103         }
2104
2105         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2106                                            PAGE_SIZE);
2107         if (ret) {
2108                 mapping_set_error(page->mapping, ret);
2109                 end_extent_writepage(page, ret, page_start, page_end);
2110                 ClearPageChecked(page);
2111                 goto out;
2112          }
2113
2114         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2115                                         &cached_state, 0);
2116         if (ret) {
2117                 mapping_set_error(page->mapping, ret);
2118                 end_extent_writepage(page, ret, page_start, page_end);
2119                 ClearPageChecked(page);
2120                 goto out;
2121         }
2122
2123         ClearPageChecked(page);
2124         set_page_dirty(page);
2125         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2126 out:
2127         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2128                              &cached_state);
2129 out_page:
2130         unlock_page(page);
2131         put_page(page);
2132         kfree(fixup);
2133         extent_changeset_free(data_reserved);
2134 }
2135
2136 /*
2137  * There are a few paths in the higher layers of the kernel that directly
2138  * set the page dirty bit without asking the filesystem if it is a
2139  * good idea.  This causes problems because we want to make sure COW
2140  * properly happens and the data=ordered rules are followed.
2141  *
2142  * In our case any range that doesn't have the ORDERED bit set
2143  * hasn't been properly setup for IO.  We kick off an async process
2144  * to fix it up.  The async helper will wait for ordered extents, set
2145  * the delalloc bit and make it safe to write the page.
2146  */
2147 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2148 {
2149         struct inode *inode = page->mapping->host;
2150         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2151         struct btrfs_writepage_fixup *fixup;
2152
2153         /* this page is properly in the ordered list */
2154         if (TestClearPagePrivate2(page))
2155                 return 0;
2156
2157         if (PageChecked(page))
2158                 return -EAGAIN;
2159
2160         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2161         if (!fixup)
2162                 return -EAGAIN;
2163
2164         SetPageChecked(page);
2165         get_page(page);
2166         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2167                         btrfs_writepage_fixup_worker, NULL, NULL);
2168         fixup->page = page;
2169         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2170         return -EBUSY;
2171 }
2172
2173 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2174                                        struct inode *inode, u64 file_pos,
2175                                        u64 disk_bytenr, u64 disk_num_bytes,
2176                                        u64 num_bytes, u64 ram_bytes,
2177                                        u8 compression, u8 encryption,
2178                                        u16 other_encoding, int extent_type)
2179 {
2180         struct btrfs_root *root = BTRFS_I(inode)->root;
2181         struct btrfs_file_extent_item *fi;
2182         struct btrfs_path *path;
2183         struct extent_buffer *leaf;
2184         struct btrfs_key ins;
2185         u64 qg_released;
2186         int extent_inserted = 0;
2187         int ret;
2188
2189         path = btrfs_alloc_path();
2190         if (!path)
2191                 return -ENOMEM;
2192
2193         /*
2194          * we may be replacing one extent in the tree with another.
2195          * The new extent is pinned in the extent map, and we don't want
2196          * to drop it from the cache until it is completely in the btree.
2197          *
2198          * So, tell btrfs_drop_extents to leave this extent in the cache.
2199          * the caller is expected to unpin it and allow it to be merged
2200          * with the others.
2201          */
2202         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2203                                    file_pos + num_bytes, NULL, 0,
2204                                    1, sizeof(*fi), &extent_inserted);
2205         if (ret)
2206                 goto out;
2207
2208         if (!extent_inserted) {
2209                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2210                 ins.offset = file_pos;
2211                 ins.type = BTRFS_EXTENT_DATA_KEY;
2212
2213                 path->leave_spinning = 1;
2214                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2215                                               sizeof(*fi));
2216                 if (ret)
2217                         goto out;
2218         }
2219         leaf = path->nodes[0];
2220         fi = btrfs_item_ptr(leaf, path->slots[0],
2221                             struct btrfs_file_extent_item);
2222         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2223         btrfs_set_file_extent_type(leaf, fi, extent_type);
2224         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2225         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2226         btrfs_set_file_extent_offset(leaf, fi, 0);
2227         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2228         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2229         btrfs_set_file_extent_compression(leaf, fi, compression);
2230         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2231         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2232
2233         btrfs_mark_buffer_dirty(leaf);
2234         btrfs_release_path(path);
2235
2236         inode_add_bytes(inode, num_bytes);
2237
2238         ins.objectid = disk_bytenr;
2239         ins.offset = disk_num_bytes;
2240         ins.type = BTRFS_EXTENT_ITEM_KEY;
2241
2242         /*
2243          * Release the reserved range from inode dirty range map, as it is
2244          * already moved into delayed_ref_head
2245          */
2246         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2247         if (ret < 0)
2248                 goto out;
2249         qg_released = ret;
2250         ret = btrfs_alloc_reserved_file_extent(trans, root,
2251                                                btrfs_ino(BTRFS_I(inode)),
2252                                                file_pos, qg_released, &ins);
2253 out:
2254         btrfs_free_path(path);
2255
2256         return ret;
2257 }
2258
2259 /* snapshot-aware defrag */
2260 struct sa_defrag_extent_backref {
2261         struct rb_node node;
2262         struct old_sa_defrag_extent *old;
2263         u64 root_id;
2264         u64 inum;
2265         u64 file_pos;
2266         u64 extent_offset;
2267         u64 num_bytes;
2268         u64 generation;
2269 };
2270
2271 struct old_sa_defrag_extent {
2272         struct list_head list;
2273         struct new_sa_defrag_extent *new;
2274
2275         u64 extent_offset;
2276         u64 bytenr;
2277         u64 offset;
2278         u64 len;
2279         int count;
2280 };
2281
2282 struct new_sa_defrag_extent {
2283         struct rb_root root;
2284         struct list_head head;
2285         struct btrfs_path *path;
2286         struct inode *inode;
2287         u64 file_pos;
2288         u64 len;
2289         u64 bytenr;
2290         u64 disk_len;
2291         u8 compress_type;
2292 };
2293
2294 static int backref_comp(struct sa_defrag_extent_backref *b1,
2295                         struct sa_defrag_extent_backref *b2)
2296 {
2297         if (b1->root_id < b2->root_id)
2298                 return -1;
2299         else if (b1->root_id > b2->root_id)
2300                 return 1;
2301
2302         if (b1->inum < b2->inum)
2303                 return -1;
2304         else if (b1->inum > b2->inum)
2305                 return 1;
2306
2307         if (b1->file_pos < b2->file_pos)
2308                 return -1;
2309         else if (b1->file_pos > b2->file_pos)
2310                 return 1;
2311
2312         /*
2313          * [------------------------------] ===> (a range of space)
2314          *     |<--->|   |<---->| =============> (fs/file tree A)
2315          * |<---------------------------->| ===> (fs/file tree B)
2316          *
2317          * A range of space can refer to two file extents in one tree while
2318          * refer to only one file extent in another tree.
2319          *
2320          * So we may process a disk offset more than one time(two extents in A)
2321          * and locate at the same extent(one extent in B), then insert two same
2322          * backrefs(both refer to the extent in B).
2323          */
2324         return 0;
2325 }
2326
2327 static void backref_insert(struct rb_root *root,
2328                            struct sa_defrag_extent_backref *backref)
2329 {
2330         struct rb_node **p = &root->rb_node;
2331         struct rb_node *parent = NULL;
2332         struct sa_defrag_extent_backref *entry;
2333         int ret;
2334
2335         while (*p) {
2336                 parent = *p;
2337                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2338
2339                 ret = backref_comp(backref, entry);
2340                 if (ret < 0)
2341                         p = &(*p)->rb_left;
2342                 else
2343                         p = &(*p)->rb_right;
2344         }
2345
2346         rb_link_node(&backref->node, parent, p);
2347         rb_insert_color(&backref->node, root);
2348 }
2349
2350 /*
2351  * Note the backref might has changed, and in this case we just return 0.
2352  */
2353 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2354                                        void *ctx)
2355 {
2356         struct btrfs_file_extent_item *extent;
2357         struct old_sa_defrag_extent *old = ctx;
2358         struct new_sa_defrag_extent *new = old->new;
2359         struct btrfs_path *path = new->path;
2360         struct btrfs_key key;
2361         struct btrfs_root *root;
2362         struct sa_defrag_extent_backref *backref;
2363         struct extent_buffer *leaf;
2364         struct inode *inode = new->inode;
2365         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2366         int slot;
2367         int ret;
2368         u64 extent_offset;
2369         u64 num_bytes;
2370
2371         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2372             inum == btrfs_ino(BTRFS_I(inode)))
2373                 return 0;
2374
2375         key.objectid = root_id;
2376         key.type = BTRFS_ROOT_ITEM_KEY;
2377         key.offset = (u64)-1;
2378
2379         root = btrfs_read_fs_root_no_name(fs_info, &key);
2380         if (IS_ERR(root)) {
2381                 if (PTR_ERR(root) == -ENOENT)
2382                         return 0;
2383                 WARN_ON(1);
2384                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2385                          inum, offset, root_id);
2386                 return PTR_ERR(root);
2387         }
2388
2389         key.objectid = inum;
2390         key.type = BTRFS_EXTENT_DATA_KEY;
2391         if (offset > (u64)-1 << 32)
2392                 key.offset = 0;
2393         else
2394                 key.offset = offset;
2395
2396         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2397         if (WARN_ON(ret < 0))
2398                 return ret;
2399         ret = 0;
2400
2401         while (1) {
2402                 cond_resched();
2403
2404                 leaf = path->nodes[0];
2405                 slot = path->slots[0];
2406
2407                 if (slot >= btrfs_header_nritems(leaf)) {
2408                         ret = btrfs_next_leaf(root, path);
2409                         if (ret < 0) {
2410                                 goto out;
2411                         } else if (ret > 0) {
2412                                 ret = 0;
2413                                 goto out;
2414                         }
2415                         continue;
2416                 }
2417
2418                 path->slots[0]++;
2419
2420                 btrfs_item_key_to_cpu(leaf, &key, slot);
2421
2422                 if (key.objectid > inum)
2423                         goto out;
2424
2425                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2426                         continue;
2427
2428                 extent = btrfs_item_ptr(leaf, slot,
2429                                         struct btrfs_file_extent_item);
2430
2431                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2432                         continue;
2433
2434                 /*
2435                  * 'offset' refers to the exact key.offset,
2436                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2437                  * (key.offset - extent_offset).
2438                  */
2439                 if (key.offset != offset)
2440                         continue;
2441
2442                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2443                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2444
2445                 if (extent_offset >= old->extent_offset + old->offset +
2446                     old->len || extent_offset + num_bytes <=
2447                     old->extent_offset + old->offset)
2448                         continue;
2449                 break;
2450         }
2451
2452         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2453         if (!backref) {
2454                 ret = -ENOENT;
2455                 goto out;
2456         }
2457
2458         backref->root_id = root_id;
2459         backref->inum = inum;
2460         backref->file_pos = offset;
2461         backref->num_bytes = num_bytes;
2462         backref->extent_offset = extent_offset;
2463         backref->generation = btrfs_file_extent_generation(leaf, extent);
2464         backref->old = old;
2465         backref_insert(&new->root, backref);
2466         old->count++;
2467 out:
2468         btrfs_release_path(path);
2469         WARN_ON(ret);
2470         return ret;
2471 }
2472
2473 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2474                                    struct new_sa_defrag_extent *new)
2475 {
2476         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2477         struct old_sa_defrag_extent *old, *tmp;
2478         int ret;
2479
2480         new->path = path;
2481
2482         list_for_each_entry_safe(old, tmp, &new->head, list) {
2483                 ret = iterate_inodes_from_logical(old->bytenr +
2484                                                   old->extent_offset, fs_info,
2485                                                   path, record_one_backref,
2486                                                   old, false);
2487                 if (ret < 0 && ret != -ENOENT)
2488                         return false;
2489
2490                 /* no backref to be processed for this extent */
2491                 if (!old->count) {
2492                         list_del(&old->list);
2493                         kfree(old);
2494                 }
2495         }
2496
2497         if (list_empty(&new->head))
2498                 return false;
2499
2500         return true;
2501 }
2502
2503 static int relink_is_mergable(struct extent_buffer *leaf,
2504                               struct btrfs_file_extent_item *fi,
2505                               struct new_sa_defrag_extent *new)
2506 {
2507         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2508                 return 0;
2509
2510         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2511                 return 0;
2512
2513         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2514                 return 0;
2515
2516         if (btrfs_file_extent_encryption(leaf, fi) ||
2517             btrfs_file_extent_other_encoding(leaf, fi))
2518                 return 0;
2519
2520         return 1;
2521 }
2522
2523 /*
2524  * Note the backref might has changed, and in this case we just return 0.
2525  */
2526 static noinline int relink_extent_backref(struct btrfs_path *path,
2527                                  struct sa_defrag_extent_backref *prev,
2528                                  struct sa_defrag_extent_backref *backref)
2529 {
2530         struct btrfs_file_extent_item *extent;
2531         struct btrfs_file_extent_item *item;
2532         struct btrfs_ordered_extent *ordered;
2533         struct btrfs_trans_handle *trans;
2534         struct btrfs_root *root;
2535         struct btrfs_key key;
2536         struct extent_buffer *leaf;
2537         struct old_sa_defrag_extent *old = backref->old;
2538         struct new_sa_defrag_extent *new = old->new;
2539         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2540         struct inode *inode;
2541         struct extent_state *cached = NULL;
2542         int ret = 0;
2543         u64 start;
2544         u64 len;
2545         u64 lock_start;
2546         u64 lock_end;
2547         bool merge = false;
2548         int index;
2549
2550         if (prev && prev->root_id == backref->root_id &&
2551             prev->inum == backref->inum &&
2552             prev->file_pos + prev->num_bytes == backref->file_pos)
2553                 merge = true;
2554
2555         /* step 1: get root */
2556         key.objectid = backref->root_id;
2557         key.type = BTRFS_ROOT_ITEM_KEY;
2558         key.offset = (u64)-1;
2559
2560         index = srcu_read_lock(&fs_info->subvol_srcu);
2561
2562         root = btrfs_read_fs_root_no_name(fs_info, &key);
2563         if (IS_ERR(root)) {
2564                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2565                 if (PTR_ERR(root) == -ENOENT)
2566                         return 0;
2567                 return PTR_ERR(root);
2568         }
2569
2570         if (btrfs_root_readonly(root)) {
2571                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2572                 return 0;
2573         }
2574
2575         /* step 2: get inode */
2576         key.objectid = backref->inum;
2577         key.type = BTRFS_INODE_ITEM_KEY;
2578         key.offset = 0;
2579
2580         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2581         if (IS_ERR(inode)) {
2582                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2583                 return 0;
2584         }
2585
2586         srcu_read_unlock(&fs_info->subvol_srcu, index);
2587
2588         /* step 3: relink backref */
2589         lock_start = backref->file_pos;
2590         lock_end = backref->file_pos + backref->num_bytes - 1;
2591         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2592                          &cached);
2593
2594         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2595         if (ordered) {
2596                 btrfs_put_ordered_extent(ordered);
2597                 goto out_unlock;
2598         }
2599
2600         trans = btrfs_join_transaction(root);
2601         if (IS_ERR(trans)) {
2602                 ret = PTR_ERR(trans);
2603                 goto out_unlock;
2604         }
2605
2606         key.objectid = backref->inum;
2607         key.type = BTRFS_EXTENT_DATA_KEY;
2608         key.offset = backref->file_pos;
2609
2610         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2611         if (ret < 0) {
2612                 goto out_free_path;
2613         } else if (ret > 0) {
2614                 ret = 0;
2615                 goto out_free_path;
2616         }
2617
2618         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2619                                 struct btrfs_file_extent_item);
2620
2621         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2622             backref->generation)
2623                 goto out_free_path;
2624
2625         btrfs_release_path(path);
2626
2627         start = backref->file_pos;
2628         if (backref->extent_offset < old->extent_offset + old->offset)
2629                 start += old->extent_offset + old->offset -
2630                          backref->extent_offset;
2631
2632         len = min(backref->extent_offset + backref->num_bytes,
2633                   old->extent_offset + old->offset + old->len);
2634         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2635
2636         ret = btrfs_drop_extents(trans, root, inode, start,
2637                                  start + len, 1);
2638         if (ret)
2639                 goto out_free_path;
2640 again:
2641         key.objectid = btrfs_ino(BTRFS_I(inode));
2642         key.type = BTRFS_EXTENT_DATA_KEY;
2643         key.offset = start;
2644
2645         path->leave_spinning = 1;
2646         if (merge) {
2647                 struct btrfs_file_extent_item *fi;
2648                 u64 extent_len;
2649                 struct btrfs_key found_key;
2650
2651                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2652                 if (ret < 0)
2653                         goto out_free_path;
2654
2655                 path->slots[0]--;
2656                 leaf = path->nodes[0];
2657                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2658
2659                 fi = btrfs_item_ptr(leaf, path->slots[0],
2660                                     struct btrfs_file_extent_item);
2661                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2662
2663                 if (extent_len + found_key.offset == start &&
2664                     relink_is_mergable(leaf, fi, new)) {
2665                         btrfs_set_file_extent_num_bytes(leaf, fi,
2666                                                         extent_len + len);
2667                         btrfs_mark_buffer_dirty(leaf);
2668                         inode_add_bytes(inode, len);
2669
2670                         ret = 1;
2671                         goto out_free_path;
2672                 } else {
2673                         merge = false;
2674                         btrfs_release_path(path);
2675                         goto again;
2676                 }
2677         }
2678
2679         ret = btrfs_insert_empty_item(trans, root, path, &key,
2680                                         sizeof(*extent));
2681         if (ret) {
2682                 btrfs_abort_transaction(trans, ret);
2683                 goto out_free_path;
2684         }
2685
2686         leaf = path->nodes[0];
2687         item = btrfs_item_ptr(leaf, path->slots[0],
2688                                 struct btrfs_file_extent_item);
2689         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2690         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2691         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2692         btrfs_set_file_extent_num_bytes(leaf, item, len);
2693         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2694         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2695         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2696         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2697         btrfs_set_file_extent_encryption(leaf, item, 0);
2698         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2699
2700         btrfs_mark_buffer_dirty(leaf);
2701         inode_add_bytes(inode, len);
2702         btrfs_release_path(path);
2703
2704         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2705                         new->disk_len, 0,
2706                         backref->root_id, backref->inum,
2707                         new->file_pos); /* start - extent_offset */
2708         if (ret) {
2709                 btrfs_abort_transaction(trans, ret);
2710                 goto out_free_path;
2711         }
2712
2713         ret = 1;
2714 out_free_path:
2715         btrfs_release_path(path);
2716         path->leave_spinning = 0;
2717         btrfs_end_transaction(trans);
2718 out_unlock:
2719         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2720                              &cached);
2721         iput(inode);
2722         return ret;
2723 }
2724
2725 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2726 {
2727         struct old_sa_defrag_extent *old, *tmp;
2728
2729         if (!new)
2730                 return;
2731
2732         list_for_each_entry_safe(old, tmp, &new->head, list) {
2733                 kfree(old);
2734         }
2735         kfree(new);
2736 }
2737
2738 static void relink_file_extents(struct new_sa_defrag_extent *new)
2739 {
2740         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2741         struct btrfs_path *path;
2742         struct sa_defrag_extent_backref *backref;
2743         struct sa_defrag_extent_backref *prev = NULL;
2744         struct rb_node *node;
2745         int ret;
2746
2747         path = btrfs_alloc_path();
2748         if (!path)
2749                 return;
2750
2751         if (!record_extent_backrefs(path, new)) {
2752                 btrfs_free_path(path);
2753                 goto out;
2754         }
2755         btrfs_release_path(path);
2756
2757         while (1) {
2758                 node = rb_first(&new->root);
2759                 if (!node)
2760                         break;
2761                 rb_erase(node, &new->root);
2762
2763                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2764
2765                 ret = relink_extent_backref(path, prev, backref);
2766                 WARN_ON(ret < 0);
2767
2768                 kfree(prev);
2769
2770                 if (ret == 1)
2771                         prev = backref;
2772                 else
2773                         prev = NULL;
2774                 cond_resched();
2775         }
2776         kfree(prev);
2777
2778         btrfs_free_path(path);
2779 out:
2780         free_sa_defrag_extent(new);
2781
2782         atomic_dec(&fs_info->defrag_running);
2783         wake_up(&fs_info->transaction_wait);
2784 }
2785
2786 static struct new_sa_defrag_extent *
2787 record_old_file_extents(struct inode *inode,
2788                         struct btrfs_ordered_extent *ordered)
2789 {
2790         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2791         struct btrfs_root *root = BTRFS_I(inode)->root;
2792         struct btrfs_path *path;
2793         struct btrfs_key key;
2794         struct old_sa_defrag_extent *old;
2795         struct new_sa_defrag_extent *new;
2796         int ret;
2797
2798         new = kmalloc(sizeof(*new), GFP_NOFS);
2799         if (!new)
2800                 return NULL;
2801
2802         new->inode = inode;
2803         new->file_pos = ordered->file_offset;
2804         new->len = ordered->len;
2805         new->bytenr = ordered->start;
2806         new->disk_len = ordered->disk_len;
2807         new->compress_type = ordered->compress_type;
2808         new->root = RB_ROOT;
2809         INIT_LIST_HEAD(&new->head);
2810
2811         path = btrfs_alloc_path();
2812         if (!path)
2813                 goto out_kfree;
2814
2815         key.objectid = btrfs_ino(BTRFS_I(inode));
2816         key.type = BTRFS_EXTENT_DATA_KEY;
2817         key.offset = new->file_pos;
2818
2819         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2820         if (ret < 0)
2821                 goto out_free_path;
2822         if (ret > 0 && path->slots[0] > 0)
2823                 path->slots[0]--;
2824
2825         /* find out all the old extents for the file range */
2826         while (1) {
2827                 struct btrfs_file_extent_item *extent;
2828                 struct extent_buffer *l;
2829                 int slot;
2830                 u64 num_bytes;
2831                 u64 offset;
2832                 u64 end;
2833                 u64 disk_bytenr;
2834                 u64 extent_offset;
2835
2836                 l = path->nodes[0];
2837                 slot = path->slots[0];
2838
2839                 if (slot >= btrfs_header_nritems(l)) {
2840                         ret = btrfs_next_leaf(root, path);
2841                         if (ret < 0)
2842                                 goto out_free_path;
2843                         else if (ret > 0)
2844                                 break;
2845                         continue;
2846                 }
2847
2848                 btrfs_item_key_to_cpu(l, &key, slot);
2849
2850                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2851                         break;
2852                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2853                         break;
2854                 if (key.offset >= new->file_pos + new->len)
2855                         break;
2856
2857                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2858
2859                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2860                 if (key.offset + num_bytes < new->file_pos)
2861                         goto next;
2862
2863                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2864                 if (!disk_bytenr)
2865                         goto next;
2866
2867                 extent_offset = btrfs_file_extent_offset(l, extent);
2868
2869                 old = kmalloc(sizeof(*old), GFP_NOFS);
2870                 if (!old)
2871                         goto out_free_path;
2872
2873                 offset = max(new->file_pos, key.offset);
2874                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2875
2876                 old->bytenr = disk_bytenr;
2877                 old->extent_offset = extent_offset;
2878                 old->offset = offset - key.offset;
2879                 old->len = end - offset;
2880                 old->new = new;
2881                 old->count = 0;
2882                 list_add_tail(&old->list, &new->head);
2883 next:
2884                 path->slots[0]++;
2885                 cond_resched();
2886         }
2887
2888         btrfs_free_path(path);
2889         atomic_inc(&fs_info->defrag_running);
2890
2891         return new;
2892
2893 out_free_path:
2894         btrfs_free_path(path);
2895 out_kfree:
2896         free_sa_defrag_extent(new);
2897         return NULL;
2898 }
2899
2900 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2901                                          u64 start, u64 len)
2902 {
2903         struct btrfs_block_group_cache *cache;
2904
2905         cache = btrfs_lookup_block_group(fs_info, start);
2906         ASSERT(cache);
2907
2908         spin_lock(&cache->lock);
2909         cache->delalloc_bytes -= len;
2910         spin_unlock(&cache->lock);
2911
2912         btrfs_put_block_group(cache);
2913 }
2914
2915 /* as ordered data IO finishes, this gets called so we can finish
2916  * an ordered extent if the range of bytes in the file it covers are
2917  * fully written.
2918  */
2919 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2920 {
2921         struct inode *inode = ordered_extent->inode;
2922         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2923         struct btrfs_root *root = BTRFS_I(inode)->root;
2924         struct btrfs_trans_handle *trans = NULL;
2925         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2926         struct extent_state *cached_state = NULL;
2927         struct new_sa_defrag_extent *new = NULL;
2928         int compress_type = 0;
2929         int ret = 0;
2930         u64 logical_len = ordered_extent->len;
2931         bool nolock;
2932         bool truncated = false;
2933         bool range_locked = false;
2934         bool clear_new_delalloc_bytes = false;
2935         bool clear_reserved_extent = true;
2936
2937         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2938             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2939             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2940                 clear_new_delalloc_bytes = true;
2941
2942         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2943
2944         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2945                 ret = -EIO;
2946                 goto out;
2947         }
2948
2949         btrfs_free_io_failure_record(BTRFS_I(inode),
2950                         ordered_extent->file_offset,
2951                         ordered_extent->file_offset +
2952                         ordered_extent->len - 1);
2953
2954         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2955                 truncated = true;
2956                 logical_len = ordered_extent->truncated_len;
2957                 /* Truncated the entire extent, don't bother adding */
2958                 if (!logical_len)
2959                         goto out;
2960         }
2961
2962         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2963                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2964
2965                 /*
2966                  * For mwrite(mmap + memset to write) case, we still reserve
2967                  * space for NOCOW range.
2968                  * As NOCOW won't cause a new delayed ref, just free the space
2969                  */
2970                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2971                                        ordered_extent->len);
2972                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2973                 if (nolock)
2974                         trans = btrfs_join_transaction_nolock(root);
2975                 else
2976                         trans = btrfs_join_transaction(root);
2977                 if (IS_ERR(trans)) {
2978                         ret = PTR_ERR(trans);
2979                         trans = NULL;
2980                         goto out;
2981                 }
2982                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2983                 ret = btrfs_update_inode_fallback(trans, root, inode);
2984                 if (ret) /* -ENOMEM or corruption */
2985                         btrfs_abort_transaction(trans, ret);
2986                 goto out;
2987         }
2988
2989         range_locked = true;
2990         lock_extent_bits(io_tree, ordered_extent->file_offset,
2991                          ordered_extent->file_offset + ordered_extent->len - 1,
2992                          &cached_state);
2993
2994         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2995                         ordered_extent->file_offset + ordered_extent->len - 1,
2996                         EXTENT_DEFRAG, 0, cached_state);
2997         if (ret) {
2998                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2999                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3000                         /* the inode is shared */
3001                         new = record_old_file_extents(inode, ordered_extent);
3002
3003                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3004                         ordered_extent->file_offset + ordered_extent->len - 1,
3005                         EXTENT_DEFRAG, 0, 0, &cached_state);
3006         }
3007
3008         if (nolock)
3009                 trans = btrfs_join_transaction_nolock(root);
3010         else
3011                 trans = btrfs_join_transaction(root);
3012         if (IS_ERR(trans)) {
3013                 ret = PTR_ERR(trans);
3014                 trans = NULL;
3015                 goto out;
3016         }
3017
3018         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3019
3020         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3021                 compress_type = ordered_extent->compress_type;
3022         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3023                 BUG_ON(compress_type);
3024                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3025                                        ordered_extent->len);
3026                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3027                                                 ordered_extent->file_offset,
3028                                                 ordered_extent->file_offset +
3029                                                 logical_len);
3030         } else {
3031                 BUG_ON(root == fs_info->tree_root);
3032                 ret = insert_reserved_file_extent(trans, inode,
3033                                                 ordered_extent->file_offset,
3034                                                 ordered_extent->start,
3035                                                 ordered_extent->disk_len,
3036                                                 logical_len, logical_len,
3037                                                 compress_type, 0, 0,
3038                                                 BTRFS_FILE_EXTENT_REG);
3039                 if (!ret) {
3040                         clear_reserved_extent = false;
3041                         btrfs_release_delalloc_bytes(fs_info,
3042                                                      ordered_extent->start,
3043                                                      ordered_extent->disk_len);
3044                 }
3045         }
3046         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3047                            ordered_extent->file_offset, ordered_extent->len,
3048                            trans->transid);
3049         if (ret < 0) {
3050                 btrfs_abort_transaction(trans, ret);
3051                 goto out;
3052         }
3053
3054         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3055         if (ret) {
3056                 btrfs_abort_transaction(trans, ret);
3057                 goto out;
3058         }
3059
3060         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3061         ret = btrfs_update_inode_fallback(trans, root, inode);
3062         if (ret) { /* -ENOMEM or corruption */
3063                 btrfs_abort_transaction(trans, ret);
3064                 goto out;
3065         }
3066         ret = 0;
3067 out:
3068         if (range_locked || clear_new_delalloc_bytes) {
3069                 unsigned int clear_bits = 0;
3070
3071                 if (range_locked)
3072                         clear_bits |= EXTENT_LOCKED;
3073                 if (clear_new_delalloc_bytes)
3074                         clear_bits |= EXTENT_DELALLOC_NEW;
3075                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3076                                  ordered_extent->file_offset,
3077                                  ordered_extent->file_offset +
3078                                  ordered_extent->len - 1,
3079                                  clear_bits,
3080                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3081                                  0, &cached_state);
3082         }
3083
3084         if (trans)
3085                 btrfs_end_transaction(trans);
3086
3087         if (ret || truncated) {
3088                 u64 start, end;
3089
3090                 if (truncated)
3091                         start = ordered_extent->file_offset + logical_len;
3092                 else
3093                         start = ordered_extent->file_offset;
3094                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3095                 clear_extent_uptodate(io_tree, start, end, NULL);
3096
3097                 /* Drop the cache for the part of the extent we didn't write. */
3098                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3099
3100                 /*
3101                  * If the ordered extent had an IOERR or something else went
3102                  * wrong we need to return the space for this ordered extent
3103                  * back to the allocator.  We only free the extent in the
3104                  * truncated case if we didn't write out the extent at all.
3105                  *
3106                  * If we made it past insert_reserved_file_extent before we
3107                  * errored out then we don't need to do this as the accounting
3108                  * has already been done.
3109                  */
3110                 if ((ret || !logical_len) &&
3111                     clear_reserved_extent &&
3112                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3113                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3114                         btrfs_free_reserved_extent(fs_info,
3115                                                    ordered_extent->start,
3116                                                    ordered_extent->disk_len, 1);
3117         }
3118
3119
3120         /*
3121          * This needs to be done to make sure anybody waiting knows we are done
3122          * updating everything for this ordered extent.
3123          */
3124         btrfs_remove_ordered_extent(inode, ordered_extent);
3125
3126         /* for snapshot-aware defrag */
3127         if (new) {
3128                 if (ret) {
3129                         free_sa_defrag_extent(new);
3130                         atomic_dec(&fs_info->defrag_running);
3131                 } else {
3132                         relink_file_extents(new);
3133                 }
3134         }
3135
3136         /* once for us */
3137         btrfs_put_ordered_extent(ordered_extent);
3138         /* once for the tree */
3139         btrfs_put_ordered_extent(ordered_extent);
3140
3141         return ret;
3142 }
3143
3144 static void finish_ordered_fn(struct btrfs_work *work)
3145 {
3146         struct btrfs_ordered_extent *ordered_extent;
3147         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3148         btrfs_finish_ordered_io(ordered_extent);
3149 }
3150
3151 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3152                                           u64 end, int uptodate)
3153 {
3154         struct inode *inode = page->mapping->host;
3155         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3156         struct btrfs_ordered_extent *ordered_extent = NULL;
3157         struct btrfs_workqueue *wq;
3158         btrfs_work_func_t func;
3159
3160         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3161
3162         ClearPagePrivate2(page);
3163         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3164                                             end - start + 1, uptodate))
3165                 return;
3166
3167         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3168                 wq = fs_info->endio_freespace_worker;
3169                 func = btrfs_freespace_write_helper;
3170         } else {
3171                 wq = fs_info->endio_write_workers;
3172                 func = btrfs_endio_write_helper;
3173         }
3174
3175         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3176                         NULL);
3177         btrfs_queue_work(wq, &ordered_extent->work);
3178 }
3179
3180 static int __readpage_endio_check(struct inode *inode,
3181                                   struct btrfs_io_bio *io_bio,
3182                                   int icsum, struct page *page,
3183                                   int pgoff, u64 start, size_t len)
3184 {
3185         char *kaddr;
3186         u32 csum_expected;
3187         u32 csum = ~(u32)0;
3188
3189         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3190
3191         kaddr = kmap_atomic(page);
3192         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3193         btrfs_csum_final(csum, (u8 *)&csum);
3194         if (csum != csum_expected)
3195                 goto zeroit;
3196
3197         kunmap_atomic(kaddr);
3198         return 0;
3199 zeroit:
3200         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3201                                     io_bio->mirror_num);
3202         memset(kaddr + pgoff, 1, len);
3203         flush_dcache_page(page);
3204         kunmap_atomic(kaddr);
3205         return -EIO;
3206 }
3207
3208 /*
3209  * when reads are done, we need to check csums to verify the data is correct
3210  * if there's a match, we allow the bio to finish.  If not, the code in
3211  * extent_io.c will try to find good copies for us.
3212  */
3213 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3214                                       u64 phy_offset, struct page *page,
3215                                       u64 start, u64 end, int mirror)
3216 {
3217         size_t offset = start - page_offset(page);
3218         struct inode *inode = page->mapping->host;
3219         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3220         struct btrfs_root *root = BTRFS_I(inode)->root;
3221
3222         if (PageChecked(page)) {
3223                 ClearPageChecked(page);
3224                 return 0;
3225         }
3226
3227         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3228                 return 0;
3229
3230         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3231             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3232                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3233                 return 0;
3234         }
3235
3236         phy_offset >>= inode->i_sb->s_blocksize_bits;
3237         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3238                                       start, (size_t)(end - start + 1));
3239 }
3240
3241 /*
3242  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3243  *
3244  * @inode: The inode we want to perform iput on
3245  *
3246  * This function uses the generic vfs_inode::i_count to track whether we should
3247  * just decrement it (in case it's > 1) or if this is the last iput then link
3248  * the inode to the delayed iput machinery. Delayed iputs are processed at
3249  * transaction commit time/superblock commit/cleaner kthread.
3250  */
3251 void btrfs_add_delayed_iput(struct inode *inode)
3252 {
3253         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3254         struct btrfs_inode *binode = BTRFS_I(inode);
3255
3256         if (atomic_add_unless(&inode->i_count, -1, 1))
3257                 return;
3258
3259         atomic_inc(&fs_info->nr_delayed_iputs);
3260         spin_lock(&fs_info->delayed_iput_lock);
3261         ASSERT(list_empty(&binode->delayed_iput));
3262         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3263         spin_unlock(&fs_info->delayed_iput_lock);
3264         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3265                 wake_up_process(fs_info->cleaner_kthread);
3266 }
3267
3268 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3269 {
3270
3271         spin_lock(&fs_info->delayed_iput_lock);
3272         while (!list_empty(&fs_info->delayed_iputs)) {
3273                 struct btrfs_inode *inode;
3274
3275                 inode = list_first_entry(&fs_info->delayed_iputs,
3276                                 struct btrfs_inode, delayed_iput);
3277                 list_del_init(&inode->delayed_iput);
3278                 spin_unlock(&fs_info->delayed_iput_lock);
3279                 iput(&inode->vfs_inode);
3280                 if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3281                         wake_up(&fs_info->delayed_iputs_wait);
3282                 spin_lock(&fs_info->delayed_iput_lock);
3283         }
3284         spin_unlock(&fs_info->delayed_iput_lock);
3285 }
3286
3287 /**
3288  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3289  * @fs_info - the fs_info for this fs
3290  * @return - EINTR if we were killed, 0 if nothing's pending
3291  *
3292  * This will wait on any delayed iputs that are currently running with KILLABLE
3293  * set.  Once they are all done running we will return, unless we are killed in
3294  * which case we return EINTR. This helps in user operations like fallocate etc
3295  * that might get blocked on the iputs.
3296  */
3297 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3298 {
3299         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3300                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3301         if (ret)
3302                 return -EINTR;
3303         return 0;
3304 }
3305
3306 /*
3307  * This creates an orphan entry for the given inode in case something goes wrong
3308  * in the middle of an unlink.
3309  */
3310 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3311                      struct btrfs_inode *inode)
3312 {
3313         int ret;
3314
3315         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3316         if (ret && ret != -EEXIST) {
3317                 btrfs_abort_transaction(trans, ret);
3318                 return ret;
3319         }
3320
3321         return 0;
3322 }
3323
3324 /*
3325  * We have done the delete so we can go ahead and remove the orphan item for
3326  * this particular inode.
3327  */
3328 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3329                             struct btrfs_inode *inode)
3330 {
3331         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3332 }
3333
3334 /*
3335  * this cleans up any orphans that may be left on the list from the last use
3336  * of this root.
3337  */
3338 int btrfs_orphan_cleanup(struct btrfs_root *root)
3339 {
3340         struct btrfs_fs_info *fs_info = root->fs_info;
3341         struct btrfs_path *path;
3342         struct extent_buffer *leaf;
3343         struct btrfs_key key, found_key;
3344         struct btrfs_trans_handle *trans;
3345         struct inode *inode;
3346         u64 last_objectid = 0;
3347         int ret = 0, nr_unlink = 0;
3348
3349         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3350                 return 0;
3351
3352         path = btrfs_alloc_path();
3353         if (!path) {
3354                 ret = -ENOMEM;
3355                 goto out;
3356         }
3357         path->reada = READA_BACK;
3358
3359         key.objectid = BTRFS_ORPHAN_OBJECTID;
3360         key.type = BTRFS_ORPHAN_ITEM_KEY;
3361         key.offset = (u64)-1;
3362
3363         while (1) {
3364                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3365                 if (ret < 0)
3366                         goto out;
3367
3368                 /*
3369                  * if ret == 0 means we found what we were searching for, which
3370                  * is weird, but possible, so only screw with path if we didn't
3371                  * find the key and see if we have stuff that matches
3372                  */
3373                 if (ret > 0) {
3374                         ret = 0;
3375                         if (path->slots[0] == 0)
3376                                 break;
3377                         path->slots[0]--;
3378                 }
3379
3380                 /* pull out the item */
3381                 leaf = path->nodes[0];
3382                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3383
3384                 /* make sure the item matches what we want */
3385                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3386                         break;
3387                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3388                         break;
3389
3390                 /* release the path since we're done with it */
3391                 btrfs_release_path(path);
3392
3393                 /*
3394                  * this is where we are basically btrfs_lookup, without the
3395                  * crossing root thing.  we store the inode number in the
3396                  * offset of the orphan item.
3397                  */
3398
3399                 if (found_key.offset == last_objectid) {
3400                         btrfs_err(fs_info,
3401                                   "Error removing orphan entry, stopping orphan cleanup");
3402                         ret = -EINVAL;
3403                         goto out;
3404                 }
3405
3406                 last_objectid = found_key.offset;
3407
3408                 found_key.objectid = found_key.offset;
3409                 found_key.type = BTRFS_INODE_ITEM_KEY;
3410                 found_key.offset = 0;
3411                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3412                 ret = PTR_ERR_OR_ZERO(inode);
3413                 if (ret && ret != -ENOENT)
3414                         goto out;
3415
3416                 if (ret == -ENOENT && root == fs_info->tree_root) {
3417                         struct btrfs_root *dead_root;
3418                         struct btrfs_fs_info *fs_info = root->fs_info;
3419                         int is_dead_root = 0;
3420
3421                         /*
3422                          * this is an orphan in the tree root. Currently these
3423                          * could come from 2 sources:
3424                          *  a) a snapshot deletion in progress
3425                          *  b) a free space cache inode
3426                          * We need to distinguish those two, as the snapshot
3427                          * orphan must not get deleted.
3428                          * find_dead_roots already ran before us, so if this
3429                          * is a snapshot deletion, we should find the root
3430                          * in the dead_roots list
3431                          */
3432                         spin_lock(&fs_info->trans_lock);
3433                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3434                                             root_list) {
3435                                 if (dead_root->root_key.objectid ==
3436                                     found_key.objectid) {
3437                                         is_dead_root = 1;
3438                                         break;
3439                                 }
3440                         }
3441                         spin_unlock(&fs_info->trans_lock);
3442                         if (is_dead_root) {
3443                                 /* prevent this orphan from being found again */
3444                                 key.offset = found_key.objectid - 1;
3445                                 continue;
3446                         }
3447
3448                 }
3449
3450                 /*
3451                  * If we have an inode with links, there are a couple of
3452                  * possibilities. Old kernels (before v3.12) used to create an
3453                  * orphan item for truncate indicating that there were possibly
3454                  * extent items past i_size that needed to be deleted. In v3.12,
3455                  * truncate was changed to update i_size in sync with the extent
3456                  * items, but the (useless) orphan item was still created. Since
3457                  * v4.18, we don't create the orphan item for truncate at all.
3458                  *
3459                  * So, this item could mean that we need to do a truncate, but
3460                  * only if this filesystem was last used on a pre-v3.12 kernel
3461                  * and was not cleanly unmounted. The odds of that are quite
3462                  * slim, and it's a pain to do the truncate now, so just delete
3463                  * the orphan item.
3464                  *
3465                  * It's also possible that this orphan item was supposed to be
3466                  * deleted but wasn't. The inode number may have been reused,
3467                  * but either way, we can delete the orphan item.
3468                  */
3469                 if (ret == -ENOENT || inode->i_nlink) {
3470                         if (!ret)
3471                                 iput(inode);
3472                         trans = btrfs_start_transaction(root, 1);
3473                         if (IS_ERR(trans)) {
3474                                 ret = PTR_ERR(trans);
3475                                 goto out;
3476                         }
3477                         btrfs_debug(fs_info, "auto deleting %Lu",
3478                                     found_key.objectid);
3479                         ret = btrfs_del_orphan_item(trans, root,
3480                                                     found_key.objectid);
3481                         btrfs_end_transaction(trans);
3482                         if (ret)
3483                                 goto out;
3484                         continue;
3485                 }
3486
3487                 nr_unlink++;
3488
3489                 /* this will do delete_inode and everything for us */
3490                 iput(inode);
3491         }
3492         /* release the path since we're done with it */
3493         btrfs_release_path(path);
3494
3495         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3496
3497         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3498                 trans = btrfs_join_transaction(root);
3499                 if (!IS_ERR(trans))
3500                         btrfs_end_transaction(trans);
3501         }
3502
3503         if (nr_unlink)
3504                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3505
3506 out:
3507         if (ret)
3508                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3509         btrfs_free_path(path);
3510         return ret;
3511 }
3512
3513 /*
3514  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3515  * don't find any xattrs, we know there can't be any acls.
3516  *
3517  * slot is the slot the inode is in, objectid is the objectid of the inode
3518  */
3519 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3520                                           int slot, u64 objectid,
3521                                           int *first_xattr_slot)
3522 {
3523         u32 nritems = btrfs_header_nritems(leaf);
3524         struct btrfs_key found_key;
3525         static u64 xattr_access = 0;
3526         static u64 xattr_default = 0;
3527         int scanned = 0;
3528
3529         if (!xattr_access) {
3530                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3531                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3532                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3533                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3534         }
3535
3536         slot++;
3537         *first_xattr_slot = -1;
3538         while (slot < nritems) {
3539                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3540
3541                 /* we found a different objectid, there must not be acls */
3542                 if (found_key.objectid != objectid)
3543                         return 0;
3544
3545                 /* we found an xattr, assume we've got an acl */
3546                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3547                         if (*first_xattr_slot == -1)
3548                                 *first_xattr_slot = slot;
3549                         if (found_key.offset == xattr_access ||
3550                             found_key.offset == xattr_default)
3551                                 return 1;
3552                 }
3553
3554                 /*
3555                  * we found a key greater than an xattr key, there can't
3556                  * be any acls later on
3557                  */
3558                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3559                         return 0;
3560
3561                 slot++;
3562                 scanned++;
3563
3564                 /*
3565                  * it goes inode, inode backrefs, xattrs, extents,
3566                  * so if there are a ton of hard links to an inode there can
3567                  * be a lot of backrefs.  Don't waste time searching too hard,
3568                  * this is just an optimization
3569                  */
3570                 if (scanned >= 8)
3571                         break;
3572         }
3573         /* we hit the end of the leaf before we found an xattr or
3574          * something larger than an xattr.  We have to assume the inode
3575          * has acls
3576          */
3577         if (*first_xattr_slot == -1)
3578                 *first_xattr_slot = slot;
3579         return 1;
3580 }
3581
3582 /*
3583  * read an inode from the btree into the in-memory inode
3584  */
3585 static int btrfs_read_locked_inode(struct inode *inode,
3586                                    struct btrfs_path *in_path)
3587 {
3588         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3589         struct btrfs_path *path = in_path;
3590         struct extent_buffer *leaf;
3591         struct btrfs_inode_item *inode_item;
3592         struct btrfs_root *root = BTRFS_I(inode)->root;
3593         struct btrfs_key location;
3594         unsigned long ptr;
3595         int maybe_acls;
3596         u32 rdev;
3597         int ret;
3598         bool filled = false;
3599         int first_xattr_slot;
3600
3601         ret = btrfs_fill_inode(inode, &rdev);
3602         if (!ret)
3603                 filled = true;
3604
3605         if (!path) {
3606                 path = btrfs_alloc_path();
3607                 if (!path)
3608                         return -ENOMEM;
3609         }
3610
3611         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3612
3613         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3614         if (ret) {
3615                 if (path != in_path)
3616                         btrfs_free_path(path);
3617                 return ret;
3618         }
3619
3620         leaf = path->nodes[0];
3621
3622         if (filled)
3623                 goto cache_index;
3624
3625         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3626                                     struct btrfs_inode_item);
3627         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3628         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3629         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3630         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3631         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3632
3633         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3634         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3635
3636         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3637         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3638
3639         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3640         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3641
3642         BTRFS_I(inode)->i_otime.tv_sec =
3643                 btrfs_timespec_sec(leaf, &inode_item->otime);
3644         BTRFS_I(inode)->i_otime.tv_nsec =
3645                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3646
3647         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3648         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3649         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3650
3651         inode_set_iversion_queried(inode,
3652                                    btrfs_inode_sequence(leaf, inode_item));
3653         inode->i_generation = BTRFS_I(inode)->generation;
3654         inode->i_rdev = 0;
3655         rdev = btrfs_inode_rdev(leaf, inode_item);
3656
3657         BTRFS_I(inode)->index_cnt = (u64)-1;
3658         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3659
3660 cache_index:
3661         /*
3662          * If we were modified in the current generation and evicted from memory
3663          * and then re-read we need to do a full sync since we don't have any
3664          * idea about which extents were modified before we were evicted from
3665          * cache.
3666          *
3667          * This is required for both inode re-read from disk and delayed inode
3668          * in delayed_nodes_tree.
3669          */
3670         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3671                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3672                         &BTRFS_I(inode)->runtime_flags);
3673
3674         /*
3675          * We don't persist the id of the transaction where an unlink operation
3676          * against the inode was last made. So here we assume the inode might
3677          * have been evicted, and therefore the exact value of last_unlink_trans
3678          * lost, and set it to last_trans to avoid metadata inconsistencies
3679          * between the inode and its parent if the inode is fsync'ed and the log
3680          * replayed. For example, in the scenario:
3681          *
3682          * touch mydir/foo
3683          * ln mydir/foo mydir/bar
3684          * sync
3685          * unlink mydir/bar
3686          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3687          * xfs_io -c fsync mydir/foo
3688          * <power failure>
3689          * mount fs, triggers fsync log replay
3690          *
3691          * We must make sure that when we fsync our inode foo we also log its
3692          * parent inode, otherwise after log replay the parent still has the
3693          * dentry with the "bar" name but our inode foo has a link count of 1
3694          * and doesn't have an inode ref with the name "bar" anymore.
3695          *
3696          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3697          * but it guarantees correctness at the expense of occasional full
3698          * transaction commits on fsync if our inode is a directory, or if our
3699          * inode is not a directory, logging its parent unnecessarily.
3700          */
3701         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3702         /*
3703          * Similar reasoning for last_link_trans, needs to be set otherwise
3704          * for a case like the following:
3705          *
3706          * mkdir A
3707          * touch foo
3708          * ln foo A/bar
3709          * echo 2 > /proc/sys/vm/drop_caches
3710          * fsync foo
3711          * <power failure>
3712          *
3713          * Would result in link bar and directory A not existing after the power
3714          * failure.
3715          */
3716         BTRFS_I(inode)->last_link_trans = BTRFS_I(inode)->last_trans;
3717
3718         path->slots[0]++;
3719         if (inode->i_nlink != 1 ||
3720             path->slots[0] >= btrfs_header_nritems(leaf))
3721                 goto cache_acl;
3722
3723         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3724         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3725                 goto cache_acl;
3726
3727         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3728         if (location.type == BTRFS_INODE_REF_KEY) {
3729                 struct btrfs_inode_ref *ref;
3730
3731                 ref = (struct btrfs_inode_ref *)ptr;
3732                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3733         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3734                 struct btrfs_inode_extref *extref;
3735
3736                 extref = (struct btrfs_inode_extref *)ptr;
3737                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3738                                                                      extref);
3739         }
3740 cache_acl:
3741         /*
3742          * try to precache a NULL acl entry for files that don't have
3743          * any xattrs or acls
3744          */
3745         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3746                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3747         if (first_xattr_slot != -1) {
3748                 path->slots[0] = first_xattr_slot;
3749                 ret = btrfs_load_inode_props(inode, path);
3750                 if (ret)
3751                         btrfs_err(fs_info,
3752                                   "error loading props for ino %llu (root %llu): %d",
3753                                   btrfs_ino(BTRFS_I(inode)),
3754                                   root->root_key.objectid, ret);
3755         }
3756         if (path != in_path)
3757                 btrfs_free_path(path);
3758
3759         if (!maybe_acls)
3760                 cache_no_acl(inode);
3761
3762         switch (inode->i_mode & S_IFMT) {
3763         case S_IFREG:
3764                 inode->i_mapping->a_ops = &btrfs_aops;
3765                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3766                 inode->i_fop = &btrfs_file_operations;
3767                 inode->i_op = &btrfs_file_inode_operations;
3768                 break;
3769         case S_IFDIR:
3770                 inode->i_fop = &btrfs_dir_file_operations;
3771                 inode->i_op = &btrfs_dir_inode_operations;
3772                 break;
3773         case S_IFLNK:
3774                 inode->i_op = &btrfs_symlink_inode_operations;
3775                 inode_nohighmem(inode);
3776                 inode->i_mapping->a_ops = &btrfs_aops;
3777                 break;
3778         default:
3779                 inode->i_op = &btrfs_special_inode_operations;
3780                 init_special_inode(inode, inode->i_mode, rdev);
3781                 break;
3782         }
3783
3784         btrfs_sync_inode_flags_to_i_flags(inode);
3785         return 0;
3786 }
3787
3788 /*
3789  * given a leaf and an inode, copy the inode fields into the leaf
3790  */
3791 static void fill_inode_item(struct btrfs_trans_handle *trans,
3792                             struct extent_buffer *leaf,
3793                             struct btrfs_inode_item *item,
3794                             struct inode *inode)
3795 {
3796         struct btrfs_map_token token;
3797
3798         btrfs_init_map_token(&token);
3799
3800         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3801         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3802         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3803                                    &token);
3804         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3805         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3806
3807         btrfs_set_token_timespec_sec(leaf, &item->atime,
3808                                      inode->i_atime.tv_sec, &token);
3809         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3810                                       inode->i_atime.tv_nsec, &token);
3811
3812         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3813                                      inode->i_mtime.tv_sec, &token);
3814         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3815                                       inode->i_mtime.tv_nsec, &token);
3816
3817         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3818                                      inode->i_ctime.tv_sec, &token);
3819         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3820                                       inode->i_ctime.tv_nsec, &token);
3821
3822         btrfs_set_token_timespec_sec(leaf, &item->otime,
3823                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3824         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3825                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3826
3827         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3828                                      &token);
3829         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3830                                          &token);
3831         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3832                                        &token);
3833         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3834         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3835         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3836         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3837 }
3838
3839 /*
3840  * copy everything in the in-memory inode into the btree.
3841  */
3842 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3843                                 struct btrfs_root *root, struct inode *inode)
3844 {
3845         struct btrfs_inode_item *inode_item;
3846         struct btrfs_path *path;
3847         struct extent_buffer *leaf;
3848         int ret;
3849
3850         path = btrfs_alloc_path();
3851         if (!path)
3852                 return -ENOMEM;
3853
3854         path->leave_spinning = 1;
3855         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3856                                  1);
3857         if (ret) {
3858                 if (ret > 0)
3859                         ret = -ENOENT;
3860                 goto failed;
3861         }
3862
3863         leaf = path->nodes[0];
3864         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3865                                     struct btrfs_inode_item);
3866
3867         fill_inode_item(trans, leaf, inode_item, inode);
3868         btrfs_mark_buffer_dirty(leaf);
3869         btrfs_set_inode_last_trans(trans, inode);
3870         ret = 0;
3871 failed:
3872         btrfs_free_path(path);
3873         return ret;
3874 }
3875
3876 /*
3877  * copy everything in the in-memory inode into the btree.
3878  */
3879 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3880                                 struct btrfs_root *root, struct inode *inode)
3881 {
3882         struct btrfs_fs_info *fs_info = root->fs_info;
3883         int ret;
3884
3885         /*
3886          * If the inode is a free space inode, we can deadlock during commit
3887          * if we put it into the delayed code.
3888          *
3889          * The data relocation inode should also be directly updated
3890          * without delay
3891          */
3892         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3893             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3894             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3895                 btrfs_update_root_times(trans, root);
3896
3897                 ret = btrfs_delayed_update_inode(trans, root, inode);
3898                 if (!ret)
3899                         btrfs_set_inode_last_trans(trans, inode);
3900                 return ret;
3901         }
3902
3903         return btrfs_update_inode_item(trans, root, inode);
3904 }
3905
3906 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3907                                          struct btrfs_root *root,
3908                                          struct inode *inode)
3909 {
3910         int ret;
3911
3912         ret = btrfs_update_inode(trans, root, inode);
3913         if (ret == -ENOSPC)
3914                 return btrfs_update_inode_item(trans, root, inode);
3915         return ret;
3916 }
3917
3918 /*
3919  * unlink helper that gets used here in inode.c and in the tree logging
3920  * recovery code.  It remove a link in a directory with a given name, and
3921  * also drops the back refs in the inode to the directory
3922  */
3923 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3924                                 struct btrfs_root *root,
3925                                 struct btrfs_inode *dir,
3926                                 struct btrfs_inode *inode,
3927                                 const char *name, int name_len)
3928 {
3929         struct btrfs_fs_info *fs_info = root->fs_info;
3930         struct btrfs_path *path;
3931         int ret = 0;
3932         struct extent_buffer *leaf;
3933         struct btrfs_dir_item *di;
3934         struct btrfs_key key;
3935         u64 index;
3936         u64 ino = btrfs_ino(inode);
3937         u64 dir_ino = btrfs_ino(dir);
3938
3939         path = btrfs_alloc_path();
3940         if (!path) {
3941                 ret = -ENOMEM;
3942                 goto out;
3943         }
3944
3945         path->leave_spinning = 1;
3946         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3947                                     name, name_len, -1);
3948         if (IS_ERR_OR_NULL(di)) {
3949                 ret = di ? PTR_ERR(di) : -ENOENT;
3950                 goto err;
3951         }
3952         leaf = path->nodes[0];
3953         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3954         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3955         if (ret)
3956                 goto err;
3957         btrfs_release_path(path);
3958
3959         /*
3960          * If we don't have dir index, we have to get it by looking up
3961          * the inode ref, since we get the inode ref, remove it directly,
3962          * it is unnecessary to do delayed deletion.
3963          *
3964          * But if we have dir index, needn't search inode ref to get it.
3965          * Since the inode ref is close to the inode item, it is better
3966          * that we delay to delete it, and just do this deletion when
3967          * we update the inode item.
3968          */
3969         if (inode->dir_index) {
3970                 ret = btrfs_delayed_delete_inode_ref(inode);
3971                 if (!ret) {
3972                         index = inode->dir_index;
3973                         goto skip_backref;
3974                 }
3975         }
3976
3977         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3978                                   dir_ino, &index);
3979         if (ret) {
3980                 btrfs_info(fs_info,
3981                         "failed to delete reference to %.*s, inode %llu parent %llu",
3982                         name_len, name, ino, dir_ino);
3983                 btrfs_abort_transaction(trans, ret);
3984                 goto err;
3985         }
3986 skip_backref:
3987         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3988         if (ret) {
3989                 btrfs_abort_transaction(trans, ret);
3990                 goto err;
3991         }
3992
3993         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3994                         dir_ino);
3995         if (ret != 0 && ret != -ENOENT) {
3996                 btrfs_abort_transaction(trans, ret);
3997                 goto err;
3998         }
3999
4000         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4001                         index);
4002         if (ret == -ENOENT)
4003                 ret = 0;
4004         else if (ret)
4005                 btrfs_abort_transaction(trans, ret);
4006 err:
4007         btrfs_free_path(path);
4008         if (ret)
4009                 goto out;
4010
4011         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4012         inode_inc_iversion(&inode->vfs_inode);
4013         inode_inc_iversion(&dir->vfs_inode);
4014         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4015                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4016         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4017 out:
4018         return ret;
4019 }
4020
4021 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4022                        struct btrfs_root *root,
4023                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4024                        const char *name, int name_len)
4025 {
4026         int ret;
4027         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4028         if (!ret) {
4029                 drop_nlink(&inode->vfs_inode);
4030                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4031         }
4032         return ret;
4033 }
4034
4035 /*
4036  * helper to start transaction for unlink and rmdir.
4037  *
4038  * unlink and rmdir are special in btrfs, they do not always free space, so
4039  * if we cannot make our reservations the normal way try and see if there is
4040  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4041  * allow the unlink to occur.
4042  */
4043 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4044 {
4045         struct btrfs_root *root = BTRFS_I(dir)->root;
4046
4047         /*
4048          * 1 for the possible orphan item
4049          * 1 for the dir item
4050          * 1 for the dir index
4051          * 1 for the inode ref
4052          * 1 for the inode
4053          */
4054         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4055 }
4056
4057 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4058 {
4059         struct btrfs_root *root = BTRFS_I(dir)->root;
4060         struct btrfs_trans_handle *trans;
4061         struct inode *inode = d_inode(dentry);
4062         int ret;
4063
4064         trans = __unlink_start_trans(dir);
4065         if (IS_ERR(trans))
4066                 return PTR_ERR(trans);
4067
4068         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4069                         0);
4070
4071         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4072                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4073                         dentry->d_name.len);
4074         if (ret)
4075                 goto out;
4076
4077         if (inode->i_nlink == 0) {
4078                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4079                 if (ret)
4080                         goto out;
4081         }
4082
4083 out:
4084         btrfs_end_transaction(trans);
4085         btrfs_btree_balance_dirty(root->fs_info);
4086         return ret;
4087 }
4088
4089 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4090                                struct inode *dir, u64 objectid,
4091                                const char *name, int name_len)
4092 {
4093         struct btrfs_root *root = BTRFS_I(dir)->root;
4094         struct btrfs_path *path;
4095         struct extent_buffer *leaf;
4096         struct btrfs_dir_item *di;
4097         struct btrfs_key key;
4098         u64 index;
4099         int ret;
4100         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4101
4102         path = btrfs_alloc_path();
4103         if (!path)
4104                 return -ENOMEM;
4105
4106         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4107                                    name, name_len, -1);
4108         if (IS_ERR_OR_NULL(di)) {
4109                 ret = di ? PTR_ERR(di) : -ENOENT;
4110                 goto out;
4111         }
4112
4113         leaf = path->nodes[0];
4114         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4115         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4116         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4117         if (ret) {
4118                 btrfs_abort_transaction(trans, ret);
4119                 goto out;
4120         }
4121         btrfs_release_path(path);
4122
4123         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4124                                  dir_ino, &index, name, name_len);
4125         if (ret < 0) {
4126                 if (ret != -ENOENT) {
4127                         btrfs_abort_transaction(trans, ret);
4128                         goto out;
4129                 }
4130                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4131                                                  name, name_len);
4132                 if (IS_ERR_OR_NULL(di)) {
4133                         if (!di)
4134                                 ret = -ENOENT;
4135                         else
4136                                 ret = PTR_ERR(di);
4137                         btrfs_abort_transaction(trans, ret);
4138                         goto out;
4139                 }
4140
4141                 leaf = path->nodes[0];
4142                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4143                 index = key.offset;
4144         }
4145         btrfs_release_path(path);
4146
4147         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4148         if (ret) {
4149                 btrfs_abort_transaction(trans, ret);
4150                 goto out;
4151         }
4152
4153         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4154         inode_inc_iversion(dir);
4155         dir->i_mtime = dir->i_ctime = current_time(dir);
4156         ret = btrfs_update_inode_fallback(trans, root, dir);
4157         if (ret)
4158                 btrfs_abort_transaction(trans, ret);
4159 out:
4160         btrfs_free_path(path);
4161         return ret;
4162 }
4163
4164 /*
4165  * Helper to check if the subvolume references other subvolumes or if it's
4166  * default.
4167  */
4168 static noinline int may_destroy_subvol(struct btrfs_root *root)
4169 {
4170         struct btrfs_fs_info *fs_info = root->fs_info;
4171         struct btrfs_path *path;
4172         struct btrfs_dir_item *di;
4173         struct btrfs_key key;
4174         u64 dir_id;
4175         int ret;
4176
4177         path = btrfs_alloc_path();
4178         if (!path)
4179                 return -ENOMEM;
4180
4181         /* Make sure this root isn't set as the default subvol */
4182         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4183         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4184                                    dir_id, "default", 7, 0);
4185         if (di && !IS_ERR(di)) {
4186                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4187                 if (key.objectid == root->root_key.objectid) {
4188                         ret = -EPERM;
4189                         btrfs_err(fs_info,
4190                                   "deleting default subvolume %llu is not allowed",
4191                                   key.objectid);
4192                         goto out;
4193                 }
4194                 btrfs_release_path(path);
4195         }
4196
4197         key.objectid = root->root_key.objectid;
4198         key.type = BTRFS_ROOT_REF_KEY;
4199         key.offset = (u64)-1;
4200
4201         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4202         if (ret < 0)
4203                 goto out;
4204         BUG_ON(ret == 0);
4205
4206         ret = 0;
4207         if (path->slots[0] > 0) {
4208                 path->slots[0]--;
4209                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4210                 if (key.objectid == root->root_key.objectid &&
4211                     key.type == BTRFS_ROOT_REF_KEY)
4212                         ret = -ENOTEMPTY;
4213         }
4214 out:
4215         btrfs_free_path(path);
4216         return ret;
4217 }
4218
4219 /* Delete all dentries for inodes belonging to the root */
4220 static void btrfs_prune_dentries(struct btrfs_root *root)
4221 {
4222         struct btrfs_fs_info *fs_info = root->fs_info;
4223         struct rb_node *node;
4224         struct rb_node *prev;
4225         struct btrfs_inode *entry;
4226         struct inode *inode;
4227         u64 objectid = 0;
4228
4229         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4230                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4231
4232         spin_lock(&root->inode_lock);
4233 again:
4234         node = root->inode_tree.rb_node;
4235         prev = NULL;
4236         while (node) {
4237                 prev = node;
4238                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4239
4240                 if (objectid < btrfs_ino(entry))
4241                         node = node->rb_left;
4242                 else if (objectid > btrfs_ino(entry))
4243                         node = node->rb_right;
4244                 else
4245                         break;
4246         }
4247         if (!node) {
4248                 while (prev) {
4249                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4250                         if (objectid <= btrfs_ino(entry)) {
4251                                 node = prev;
4252                                 break;
4253                         }
4254                         prev = rb_next(prev);
4255                 }
4256         }
4257         while (node) {
4258                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4259                 objectid = btrfs_ino(entry) + 1;
4260                 inode = igrab(&entry->vfs_inode);
4261                 if (inode) {
4262                         spin_unlock(&root->inode_lock);
4263                         if (atomic_read(&inode->i_count) > 1)
4264                                 d_prune_aliases(inode);
4265                         /*
4266                          * btrfs_drop_inode will have it removed from the inode
4267                          * cache when its usage count hits zero.
4268                          */
4269                         iput(inode);
4270                         cond_resched();
4271                         spin_lock(&root->inode_lock);
4272                         goto again;
4273                 }
4274
4275                 if (cond_resched_lock(&root->inode_lock))
4276                         goto again;
4277
4278                 node = rb_next(node);
4279         }
4280         spin_unlock(&root->inode_lock);
4281 }
4282
4283 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4284 {
4285         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4286         struct btrfs_root *root = BTRFS_I(dir)->root;
4287         struct inode *inode = d_inode(dentry);
4288         struct btrfs_root *dest = BTRFS_I(inode)->root;
4289         struct btrfs_trans_handle *trans;
4290         struct btrfs_block_rsv block_rsv;
4291         u64 root_flags;
4292         int ret;
4293         int err;
4294
4295         /*
4296          * Don't allow to delete a subvolume with send in progress. This is
4297          * inside the inode lock so the error handling that has to drop the bit
4298          * again is not run concurrently.
4299          */
4300         spin_lock(&dest->root_item_lock);
4301         if (dest->send_in_progress) {
4302                 spin_unlock(&dest->root_item_lock);
4303                 btrfs_warn(fs_info,
4304                            "attempt to delete subvolume %llu during send",
4305                            dest->root_key.objectid);
4306                 return -EPERM;
4307         }
4308         root_flags = btrfs_root_flags(&dest->root_item);
4309         btrfs_set_root_flags(&dest->root_item,
4310                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4311         spin_unlock(&dest->root_item_lock);
4312
4313         down_write(&fs_info->subvol_sem);
4314
4315         err = may_destroy_subvol(dest);
4316         if (err)
4317                 goto out_up_write;
4318
4319         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4320         /*
4321          * One for dir inode,
4322          * two for dir entries,
4323          * two for root ref/backref.
4324          */
4325         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4326         if (err)
4327                 goto out_up_write;
4328
4329         trans = btrfs_start_transaction(root, 0);
4330         if (IS_ERR(trans)) {
4331                 err = PTR_ERR(trans);
4332                 goto out_release;
4333         }
4334         trans->block_rsv = &block_rsv;
4335         trans->bytes_reserved = block_rsv.size;
4336
4337         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4338
4339         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4340                                   dentry->d_name.name, dentry->d_name.len);
4341         if (ret) {
4342                 err = ret;
4343                 btrfs_abort_transaction(trans, ret);
4344                 goto out_end_trans;
4345         }
4346
4347         btrfs_record_root_in_trans(trans, dest);
4348
4349         memset(&dest->root_item.drop_progress, 0,
4350                 sizeof(dest->root_item.drop_progress));
4351         dest->root_item.drop_level = 0;
4352         btrfs_set_root_refs(&dest->root_item, 0);
4353
4354         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4355                 ret = btrfs_insert_orphan_item(trans,
4356                                         fs_info->tree_root,
4357                                         dest->root_key.objectid);
4358                 if (ret) {
4359                         btrfs_abort_transaction(trans, ret);
4360                         err = ret;
4361                         goto out_end_trans;
4362                 }
4363         }
4364
4365         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4366                                   BTRFS_UUID_KEY_SUBVOL,
4367                                   dest->root_key.objectid);
4368         if (ret && ret != -ENOENT) {
4369                 btrfs_abort_transaction(trans, ret);
4370                 err = ret;
4371                 goto out_end_trans;
4372         }
4373         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4374                 ret = btrfs_uuid_tree_remove(trans,
4375                                           dest->root_item.received_uuid,
4376                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4377                                           dest->root_key.objectid);
4378                 if (ret && ret != -ENOENT) {
4379                         btrfs_abort_transaction(trans, ret);
4380                         err = ret;
4381                         goto out_end_trans;
4382                 }
4383         }
4384
4385 out_end_trans:
4386         trans->block_rsv = NULL;
4387         trans->bytes_reserved = 0;
4388         ret = btrfs_end_transaction(trans);
4389         if (ret && !err)
4390                 err = ret;
4391         inode->i_flags |= S_DEAD;
4392 out_release:
4393         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4394 out_up_write:
4395         up_write(&fs_info->subvol_sem);
4396         if (err) {
4397                 spin_lock(&dest->root_item_lock);
4398                 root_flags = btrfs_root_flags(&dest->root_item);
4399                 btrfs_set_root_flags(&dest->root_item,
4400                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4401                 spin_unlock(&dest->root_item_lock);
4402         } else {
4403                 d_invalidate(dentry);
4404                 btrfs_prune_dentries(dest);
4405                 ASSERT(dest->send_in_progress == 0);
4406
4407                 /* the last ref */
4408                 if (dest->ino_cache_inode) {
4409                         iput(dest->ino_cache_inode);
4410                         dest->ino_cache_inode = NULL;
4411                 }
4412         }
4413
4414         return err;
4415 }
4416
4417 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4418 {
4419         struct inode *inode = d_inode(dentry);
4420         int err = 0;
4421         struct btrfs_root *root = BTRFS_I(dir)->root;
4422         struct btrfs_trans_handle *trans;
4423         u64 last_unlink_trans;
4424
4425         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4426                 return -ENOTEMPTY;
4427         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4428                 return btrfs_delete_subvolume(dir, dentry);
4429
4430         trans = __unlink_start_trans(dir);
4431         if (IS_ERR(trans))
4432                 return PTR_ERR(trans);
4433
4434         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4435                 err = btrfs_unlink_subvol(trans, dir,
4436                                           BTRFS_I(inode)->location.objectid,
4437                                           dentry->d_name.name,
4438                                           dentry->d_name.len);
4439                 goto out;
4440         }
4441
4442         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4443         if (err)
4444                 goto out;
4445
4446         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4447
4448         /* now the directory is empty */
4449         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4450                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4451                         dentry->d_name.len);
4452         if (!err) {
4453                 btrfs_i_size_write(BTRFS_I(inode), 0);
4454                 /*
4455                  * Propagate the last_unlink_trans value of the deleted dir to
4456                  * its parent directory. This is to prevent an unrecoverable
4457                  * log tree in the case we do something like this:
4458                  * 1) create dir foo
4459                  * 2) create snapshot under dir foo
4460                  * 3) delete the snapshot
4461                  * 4) rmdir foo
4462                  * 5) mkdir foo
4463                  * 6) fsync foo or some file inside foo
4464                  */
4465                 if (last_unlink_trans >= trans->transid)
4466                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4467         }
4468 out:
4469         btrfs_end_transaction(trans);
4470         btrfs_btree_balance_dirty(root->fs_info);
4471
4472         return err;
4473 }
4474
4475 /*
4476  * Return this if we need to call truncate_block for the last bit of the
4477  * truncate.
4478  */
4479 #define NEED_TRUNCATE_BLOCK 1
4480
4481 /*
4482  * this can truncate away extent items, csum items and directory items.
4483  * It starts at a high offset and removes keys until it can't find
4484  * any higher than new_size
4485  *
4486  * csum items that cross the new i_size are truncated to the new size
4487  * as well.
4488  *
4489  * min_type is the minimum key type to truncate down to.  If set to 0, this
4490  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4491  */
4492 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4493                                struct btrfs_root *root,
4494                                struct inode *inode,
4495                                u64 new_size, u32 min_type)
4496 {
4497         struct btrfs_fs_info *fs_info = root->fs_info;
4498         struct btrfs_path *path;
4499         struct extent_buffer *leaf;
4500         struct btrfs_file_extent_item *fi;
4501         struct btrfs_key key;
4502         struct btrfs_key found_key;
4503         u64 extent_start = 0;
4504         u64 extent_num_bytes = 0;
4505         u64 extent_offset = 0;
4506         u64 item_end = 0;
4507         u64 last_size = new_size;
4508         u32 found_type = (u8)-1;
4509         int found_extent;
4510         int del_item;
4511         int pending_del_nr = 0;
4512         int pending_del_slot = 0;
4513         int extent_type = -1;
4514         int ret;
4515         u64 ino = btrfs_ino(BTRFS_I(inode));
4516         u64 bytes_deleted = 0;
4517         bool be_nice = false;
4518         bool should_throttle = false;
4519
4520         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4521
4522         /*
4523          * for non-free space inodes and ref cows, we want to back off from
4524          * time to time
4525          */
4526         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4527             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4528                 be_nice = true;
4529
4530         path = btrfs_alloc_path();
4531         if (!path)
4532                 return -ENOMEM;
4533         path->reada = READA_BACK;
4534
4535         /*
4536          * We want to drop from the next block forward in case this new size is
4537          * not block aligned since we will be keeping the last block of the
4538          * extent just the way it is.
4539          */
4540         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4541             root == fs_info->tree_root)
4542                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4543                                         fs_info->sectorsize),
4544                                         (u64)-1, 0);
4545
4546         /*
4547          * This function is also used to drop the items in the log tree before
4548          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4549          * it is used to drop the logged items. So we shouldn't kill the delayed
4550          * items.
4551          */
4552         if (min_type == 0 && root == BTRFS_I(inode)->root)
4553                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4554
4555         key.objectid = ino;
4556         key.offset = (u64)-1;
4557         key.type = (u8)-1;
4558
4559 search_again:
4560         /*
4561          * with a 16K leaf size and 128MB extents, you can actually queue
4562          * up a huge file in a single leaf.  Most of the time that
4563          * bytes_deleted is > 0, it will be huge by the time we get here
4564          */
4565         if (be_nice && bytes_deleted > SZ_32M &&
4566             btrfs_should_end_transaction(trans)) {
4567                 ret = -EAGAIN;
4568                 goto out;
4569         }
4570
4571         path->leave_spinning = 1;
4572         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4573         if (ret < 0)
4574                 goto out;
4575
4576         if (ret > 0) {
4577                 ret = 0;
4578                 /* there are no items in the tree for us to truncate, we're
4579                  * done
4580                  */
4581                 if (path->slots[0] == 0)
4582                         goto out;
4583                 path->slots[0]--;
4584         }
4585
4586         while (1) {
4587                 fi = NULL;
4588                 leaf = path->nodes[0];
4589                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4590                 found_type = found_key.type;
4591
4592                 if (found_key.objectid != ino)
4593                         break;
4594
4595                 if (found_type < min_type)
4596                         break;
4597
4598                 item_end = found_key.offset;
4599                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4600                         fi = btrfs_item_ptr(leaf, path->slots[0],
4601                                             struct btrfs_file_extent_item);
4602                         extent_type = btrfs_file_extent_type(leaf, fi);
4603                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4604                                 item_end +=
4605                                     btrfs_file_extent_num_bytes(leaf, fi);
4606
4607                                 trace_btrfs_truncate_show_fi_regular(
4608                                         BTRFS_I(inode), leaf, fi,
4609                                         found_key.offset);
4610                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4611                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4612                                                                         fi);
4613
4614                                 trace_btrfs_truncate_show_fi_inline(
4615                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4616                                         found_key.offset);
4617                         }
4618                         item_end--;
4619                 }
4620                 if (found_type > min_type) {
4621                         del_item = 1;
4622                 } else {
4623                         if (item_end < new_size)
4624                                 break;
4625                         if (found_key.offset >= new_size)
4626                                 del_item = 1;
4627                         else
4628                                 del_item = 0;
4629                 }
4630                 found_extent = 0;
4631                 /* FIXME, shrink the extent if the ref count is only 1 */
4632                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4633                         goto delete;
4634
4635                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4636                         u64 num_dec;
4637                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4638                         if (!del_item) {
4639                                 u64 orig_num_bytes =
4640                                         btrfs_file_extent_num_bytes(leaf, fi);
4641                                 extent_num_bytes = ALIGN(new_size -
4642                                                 found_key.offset,
4643                                                 fs_info->sectorsize);
4644                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4645                                                          extent_num_bytes);
4646                                 num_dec = (orig_num_bytes -
4647                                            extent_num_bytes);
4648                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4649                                              &root->state) &&
4650                                     extent_start != 0)
4651                                         inode_sub_bytes(inode, num_dec);
4652                                 btrfs_mark_buffer_dirty(leaf);
4653                         } else {
4654                                 extent_num_bytes =
4655                                         btrfs_file_extent_disk_num_bytes(leaf,
4656                                                                          fi);
4657                                 extent_offset = found_key.offset -
4658                                         btrfs_file_extent_offset(leaf, fi);
4659
4660                                 /* FIXME blocksize != 4096 */
4661                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4662                                 if (extent_start != 0) {
4663                                         found_extent = 1;
4664                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4665                                                      &root->state))
4666                                                 inode_sub_bytes(inode, num_dec);
4667                                 }
4668                         }
4669                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4670                         /*
4671                          * we can't truncate inline items that have had
4672                          * special encodings
4673                          */
4674                         if (!del_item &&
4675                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4676                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4677                             btrfs_file_extent_compression(leaf, fi) == 0) {
4678                                 u32 size = (u32)(new_size - found_key.offset);
4679
4680                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4681                                 size = btrfs_file_extent_calc_inline_size(size);
4682                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4683                         } else if (!del_item) {
4684                                 /*
4685                                  * We have to bail so the last_size is set to
4686                                  * just before this extent.
4687                                  */
4688                                 ret = NEED_TRUNCATE_BLOCK;
4689                                 break;
4690                         }
4691
4692                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4693                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4694                 }
4695 delete:
4696                 if (del_item)
4697                         last_size = found_key.offset;
4698                 else
4699                         last_size = new_size;
4700                 if (del_item) {
4701                         if (!pending_del_nr) {
4702                                 /* no pending yet, add ourselves */
4703                                 pending_del_slot = path->slots[0];
4704                                 pending_del_nr = 1;
4705                         } else if (pending_del_nr &&
4706                                    path->slots[0] + 1 == pending_del_slot) {
4707                                 /* hop on the pending chunk */
4708                                 pending_del_nr++;
4709                                 pending_del_slot = path->slots[0];
4710                         } else {
4711                                 BUG();
4712                         }
4713                 } else {
4714                         break;
4715                 }
4716                 should_throttle = false;
4717
4718                 if (found_extent &&
4719                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4720                      root == fs_info->tree_root)) {
4721                         btrfs_set_path_blocking(path);
4722                         bytes_deleted += extent_num_bytes;
4723                         ret = btrfs_free_extent(trans, root, extent_start,
4724                                                 extent_num_bytes, 0,
4725                                                 btrfs_header_owner(leaf),
4726                                                 ino, extent_offset);
4727                         if (ret) {
4728                                 btrfs_abort_transaction(trans, ret);
4729                                 break;
4730                         }
4731                         if (be_nice) {
4732                                 if (btrfs_should_throttle_delayed_refs(trans))
4733                                         should_throttle = true;
4734                         }
4735                 }
4736
4737                 if (found_type == BTRFS_INODE_ITEM_KEY)
4738                         break;
4739
4740                 if (path->slots[0] == 0 ||
4741                     path->slots[0] != pending_del_slot ||
4742                     should_throttle) {
4743                         if (pending_del_nr) {
4744                                 ret = btrfs_del_items(trans, root, path,
4745                                                 pending_del_slot,
4746                                                 pending_del_nr);
4747                                 if (ret) {
4748                                         btrfs_abort_transaction(trans, ret);
4749                                         break;
4750                                 }
4751                                 pending_del_nr = 0;
4752                         }
4753                         btrfs_release_path(path);
4754
4755                         /*
4756                          * We can generate a lot of delayed refs, so we need to
4757                          * throttle every once and a while and make sure we're
4758                          * adding enough space to keep up with the work we are
4759                          * generating.  Since we hold a transaction here we
4760                          * can't flush, and we don't want to FLUSH_LIMIT because
4761                          * we could have generated too many delayed refs to
4762                          * actually allocate, so just bail if we're short and
4763                          * let the normal reservation dance happen higher up.
4764                          */
4765                         if (should_throttle) {
4766                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4767                                                         BTRFS_RESERVE_NO_FLUSH);
4768                                 if (ret) {
4769                                         ret = -EAGAIN;
4770                                         break;
4771                                 }
4772                         }
4773                         goto search_again;
4774                 } else {
4775                         path->slots[0]--;
4776                 }
4777         }
4778 out:
4779         if (ret >= 0 && pending_del_nr) {
4780                 int err;
4781
4782                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4783                                       pending_del_nr);
4784                 if (err) {
4785                         btrfs_abort_transaction(trans, err);
4786                         ret = err;
4787                 }
4788         }
4789         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4790                 ASSERT(last_size >= new_size);
4791                 if (!ret && last_size > new_size)
4792                         last_size = new_size;
4793                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4794         }
4795
4796         btrfs_free_path(path);
4797         return ret;
4798 }
4799
4800 /*
4801  * btrfs_truncate_block - read, zero a chunk and write a block
4802  * @inode - inode that we're zeroing
4803  * @from - the offset to start zeroing
4804  * @len - the length to zero, 0 to zero the entire range respective to the
4805  *      offset
4806  * @front - zero up to the offset instead of from the offset on
4807  *
4808  * This will find the block for the "from" offset and cow the block and zero the
4809  * part we want to zero.  This is used with truncate and hole punching.
4810  */
4811 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4812                         int front)
4813 {
4814         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4815         struct address_space *mapping = inode->i_mapping;
4816         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4817         struct btrfs_ordered_extent *ordered;
4818         struct extent_state *cached_state = NULL;
4819         struct extent_changeset *data_reserved = NULL;
4820         char *kaddr;
4821         u32 blocksize = fs_info->sectorsize;
4822         pgoff_t index = from >> PAGE_SHIFT;
4823         unsigned offset = from & (blocksize - 1);
4824         struct page *page;
4825         gfp_t mask = btrfs_alloc_write_mask(mapping);
4826         int ret = 0;
4827         u64 block_start;
4828         u64 block_end;
4829
4830         if (IS_ALIGNED(offset, blocksize) &&
4831             (!len || IS_ALIGNED(len, blocksize)))
4832                 goto out;
4833
4834         block_start = round_down(from, blocksize);
4835         block_end = block_start + blocksize - 1;
4836
4837         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4838                                            block_start, blocksize);
4839         if (ret)
4840                 goto out;
4841
4842 again:
4843         page = find_or_create_page(mapping, index, mask);
4844         if (!page) {
4845                 btrfs_delalloc_release_space(inode, data_reserved,
4846                                              block_start, blocksize, true);
4847                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4848                 ret = -ENOMEM;
4849                 goto out;
4850         }
4851
4852         if (!PageUptodate(page)) {
4853                 ret = btrfs_readpage(NULL, page);
4854                 lock_page(page);
4855                 if (page->mapping != mapping) {
4856                         unlock_page(page);
4857                         put_page(page);
4858                         goto again;
4859                 }
4860                 if (!PageUptodate(page)) {
4861                         ret = -EIO;
4862                         goto out_unlock;
4863                 }
4864         }
4865         wait_on_page_writeback(page);
4866
4867         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4868         set_page_extent_mapped(page);
4869
4870         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4871         if (ordered) {
4872                 unlock_extent_cached(io_tree, block_start, block_end,
4873                                      &cached_state);
4874                 unlock_page(page);
4875                 put_page(page);
4876                 btrfs_start_ordered_extent(inode, ordered, 1);
4877                 btrfs_put_ordered_extent(ordered);
4878                 goto again;
4879         }
4880
4881         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4882                           EXTENT_DIRTY | EXTENT_DELALLOC |
4883                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4884                           0, 0, &cached_state);
4885
4886         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4887                                         &cached_state, 0);
4888         if (ret) {
4889                 unlock_extent_cached(io_tree, block_start, block_end,
4890                                      &cached_state);
4891                 goto out_unlock;
4892         }
4893
4894         if (offset != blocksize) {
4895                 if (!len)
4896                         len = blocksize - offset;
4897                 kaddr = kmap(page);
4898                 if (front)
4899                         memset(kaddr + (block_start - page_offset(page)),
4900                                 0, offset);
4901                 else
4902                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4903                                 0, len);
4904                 flush_dcache_page(page);
4905                 kunmap(page);
4906         }
4907         ClearPageChecked(page);
4908         set_page_dirty(page);
4909         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4910
4911 out_unlock:
4912         if (ret)
4913                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4914                                              blocksize, true);
4915         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
4916         unlock_page(page);
4917         put_page(page);
4918 out:
4919         extent_changeset_free(data_reserved);
4920         return ret;
4921 }
4922
4923 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4924                              u64 offset, u64 len)
4925 {
4926         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4927         struct btrfs_trans_handle *trans;
4928         int ret;
4929
4930         /*
4931          * Still need to make sure the inode looks like it's been updated so
4932          * that any holes get logged if we fsync.
4933          */
4934         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4935                 BTRFS_I(inode)->last_trans = fs_info->generation;
4936                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4937                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4938                 return 0;
4939         }
4940
4941         /*
4942          * 1 - for the one we're dropping
4943          * 1 - for the one we're adding
4944          * 1 - for updating the inode.
4945          */
4946         trans = btrfs_start_transaction(root, 3);
4947         if (IS_ERR(trans))
4948                 return PTR_ERR(trans);
4949
4950         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4951         if (ret) {
4952                 btrfs_abort_transaction(trans, ret);
4953                 btrfs_end_transaction(trans);
4954                 return ret;
4955         }
4956
4957         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4958                         offset, 0, 0, len, 0, len, 0, 0, 0);
4959         if (ret)
4960                 btrfs_abort_transaction(trans, ret);
4961         else
4962                 btrfs_update_inode(trans, root, inode);
4963         btrfs_end_transaction(trans);
4964         return ret;
4965 }
4966
4967 /*
4968  * This function puts in dummy file extents for the area we're creating a hole
4969  * for.  So if we are truncating this file to a larger size we need to insert
4970  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4971  * the range between oldsize and size
4972  */
4973 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4974 {
4975         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4976         struct btrfs_root *root = BTRFS_I(inode)->root;
4977         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4978         struct extent_map *em = NULL;
4979         struct extent_state *cached_state = NULL;
4980         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4981         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4982         u64 block_end = ALIGN(size, fs_info->sectorsize);
4983         u64 last_byte;
4984         u64 cur_offset;
4985         u64 hole_size;
4986         int err = 0;
4987
4988         /*
4989          * If our size started in the middle of a block we need to zero out the
4990          * rest of the block before we expand the i_size, otherwise we could
4991          * expose stale data.
4992          */
4993         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4994         if (err)
4995                 return err;
4996
4997         if (size <= hole_start)
4998                 return 0;
4999
5000         while (1) {
5001                 struct btrfs_ordered_extent *ordered;
5002
5003                 lock_extent_bits(io_tree, hole_start, block_end - 1,
5004                                  &cached_state);
5005                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
5006                                                      block_end - hole_start);
5007                 if (!ordered)
5008                         break;
5009                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
5010                                      &cached_state);
5011                 btrfs_start_ordered_extent(inode, ordered, 1);
5012                 btrfs_put_ordered_extent(ordered);
5013         }
5014
5015         cur_offset = hole_start;
5016         while (1) {
5017                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5018                                 block_end - cur_offset, 0);
5019                 if (IS_ERR(em)) {
5020                         err = PTR_ERR(em);
5021                         em = NULL;
5022                         break;
5023                 }
5024                 last_byte = min(extent_map_end(em), block_end);
5025                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5026                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5027                         struct extent_map *hole_em;
5028                         hole_size = last_byte - cur_offset;
5029
5030                         err = maybe_insert_hole(root, inode, cur_offset,
5031                                                 hole_size);
5032                         if (err)
5033                                 break;
5034                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5035                                                 cur_offset + hole_size - 1, 0);
5036                         hole_em = alloc_extent_map();
5037                         if (!hole_em) {
5038                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5039                                         &BTRFS_I(inode)->runtime_flags);
5040                                 goto next;
5041                         }
5042                         hole_em->start = cur_offset;
5043                         hole_em->len = hole_size;
5044                         hole_em->orig_start = cur_offset;
5045
5046                         hole_em->block_start = EXTENT_MAP_HOLE;
5047                         hole_em->block_len = 0;
5048                         hole_em->orig_block_len = 0;
5049                         hole_em->ram_bytes = hole_size;
5050                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5051                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5052                         hole_em->generation = fs_info->generation;
5053
5054                         while (1) {
5055                                 write_lock(&em_tree->lock);
5056                                 err = add_extent_mapping(em_tree, hole_em, 1);
5057                                 write_unlock(&em_tree->lock);
5058                                 if (err != -EEXIST)
5059                                         break;
5060                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5061                                                         cur_offset,
5062                                                         cur_offset +
5063                                                         hole_size - 1, 0);
5064                         }
5065                         free_extent_map(hole_em);
5066                 }
5067 next:
5068                 free_extent_map(em);
5069                 em = NULL;
5070                 cur_offset = last_byte;
5071                 if (cur_offset >= block_end)
5072                         break;
5073         }
5074         free_extent_map(em);
5075         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5076         return err;
5077 }
5078
5079 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5080 {
5081         struct btrfs_root *root = BTRFS_I(inode)->root;
5082         struct btrfs_trans_handle *trans;
5083         loff_t oldsize = i_size_read(inode);
5084         loff_t newsize = attr->ia_size;
5085         int mask = attr->ia_valid;
5086         int ret;
5087
5088         /*
5089          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5090          * special case where we need to update the times despite not having
5091          * these flags set.  For all other operations the VFS set these flags
5092          * explicitly if it wants a timestamp update.
5093          */
5094         if (newsize != oldsize) {
5095                 inode_inc_iversion(inode);
5096                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5097                         inode->i_ctime = inode->i_mtime =
5098                                 current_time(inode);
5099         }
5100
5101         if (newsize > oldsize) {
5102                 /*
5103                  * Don't do an expanding truncate while snapshotting is ongoing.
5104                  * This is to ensure the snapshot captures a fully consistent
5105                  * state of this file - if the snapshot captures this expanding
5106                  * truncation, it must capture all writes that happened before
5107                  * this truncation.
5108                  */
5109                 btrfs_wait_for_snapshot_creation(root);
5110                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5111                 if (ret) {
5112                         btrfs_end_write_no_snapshotting(root);
5113                         return ret;
5114                 }
5115
5116                 trans = btrfs_start_transaction(root, 1);
5117                 if (IS_ERR(trans)) {
5118                         btrfs_end_write_no_snapshotting(root);
5119                         return PTR_ERR(trans);
5120                 }
5121
5122                 i_size_write(inode, newsize);
5123                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5124                 pagecache_isize_extended(inode, oldsize, newsize);
5125                 ret = btrfs_update_inode(trans, root, inode);
5126                 btrfs_end_write_no_snapshotting(root);
5127                 btrfs_end_transaction(trans);
5128         } else {
5129
5130                 /*
5131                  * We're truncating a file that used to have good data down to
5132                  * zero. Make sure it gets into the ordered flush list so that
5133                  * any new writes get down to disk quickly.
5134                  */
5135                 if (newsize == 0)
5136                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5137                                 &BTRFS_I(inode)->runtime_flags);
5138
5139                 truncate_setsize(inode, newsize);
5140
5141                 /* Disable nonlocked read DIO to avoid the endless truncate */
5142                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5143                 inode_dio_wait(inode);
5144                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5145
5146                 ret = btrfs_truncate(inode, newsize == oldsize);
5147                 if (ret && inode->i_nlink) {
5148                         int err;
5149
5150                         /*
5151                          * Truncate failed, so fix up the in-memory size. We
5152                          * adjusted disk_i_size down as we removed extents, so
5153                          * wait for disk_i_size to be stable and then update the
5154                          * in-memory size to match.
5155                          */
5156                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5157                         if (err)
5158                                 return err;
5159                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5160                 }
5161         }
5162
5163         return ret;
5164 }
5165
5166 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5167 {
5168         struct inode *inode = d_inode(dentry);
5169         struct btrfs_root *root = BTRFS_I(inode)->root;
5170         int err;
5171
5172         if (btrfs_root_readonly(root))
5173                 return -EROFS;
5174
5175         err = setattr_prepare(dentry, attr);
5176         if (err)
5177                 return err;
5178
5179         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5180                 err = btrfs_setsize(inode, attr);
5181                 if (err)
5182                         return err;
5183         }
5184
5185         if (attr->ia_valid) {
5186                 setattr_copy(inode, attr);
5187                 inode_inc_iversion(inode);
5188                 err = btrfs_dirty_inode(inode);
5189
5190                 if (!err && attr->ia_valid & ATTR_MODE)
5191                         err = posix_acl_chmod(inode, inode->i_mode);
5192         }
5193
5194         return err;
5195 }
5196
5197 /*
5198  * While truncating the inode pages during eviction, we get the VFS calling
5199  * btrfs_invalidatepage() against each page of the inode. This is slow because
5200  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5201  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5202  * extent_state structures over and over, wasting lots of time.
5203  *
5204  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5205  * those expensive operations on a per page basis and do only the ordered io
5206  * finishing, while we release here the extent_map and extent_state structures,
5207  * without the excessive merging and splitting.
5208  */
5209 static void evict_inode_truncate_pages(struct inode *inode)
5210 {
5211         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5212         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5213         struct rb_node *node;
5214
5215         ASSERT(inode->i_state & I_FREEING);
5216         truncate_inode_pages_final(&inode->i_data);
5217
5218         write_lock(&map_tree->lock);
5219         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5220                 struct extent_map *em;
5221
5222                 node = rb_first_cached(&map_tree->map);
5223                 em = rb_entry(node, struct extent_map, rb_node);
5224                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5225                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5226                 remove_extent_mapping(map_tree, em);
5227                 free_extent_map(em);
5228                 if (need_resched()) {
5229                         write_unlock(&map_tree->lock);
5230                         cond_resched();
5231                         write_lock(&map_tree->lock);
5232                 }
5233         }
5234         write_unlock(&map_tree->lock);
5235
5236         /*
5237          * Keep looping until we have no more ranges in the io tree.
5238          * We can have ongoing bios started by readpages (called from readahead)
5239          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5240          * still in progress (unlocked the pages in the bio but did not yet
5241          * unlocked the ranges in the io tree). Therefore this means some
5242          * ranges can still be locked and eviction started because before
5243          * submitting those bios, which are executed by a separate task (work
5244          * queue kthread), inode references (inode->i_count) were not taken
5245          * (which would be dropped in the end io callback of each bio).
5246          * Therefore here we effectively end up waiting for those bios and
5247          * anyone else holding locked ranges without having bumped the inode's
5248          * reference count - if we don't do it, when they access the inode's
5249          * io_tree to unlock a range it may be too late, leading to an
5250          * use-after-free issue.
5251          */
5252         spin_lock(&io_tree->lock);
5253         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5254                 struct extent_state *state;
5255                 struct extent_state *cached_state = NULL;
5256                 u64 start;
5257                 u64 end;
5258                 unsigned state_flags;
5259
5260                 node = rb_first(&io_tree->state);
5261                 state = rb_entry(node, struct extent_state, rb_node);
5262                 start = state->start;
5263                 end = state->end;
5264                 state_flags = state->state;
5265                 spin_unlock(&io_tree->lock);
5266
5267                 lock_extent_bits(io_tree, start, end, &cached_state);
5268
5269                 /*
5270                  * If still has DELALLOC flag, the extent didn't reach disk,
5271                  * and its reserved space won't be freed by delayed_ref.
5272                  * So we need to free its reserved space here.
5273                  * (Refer to comment in btrfs_invalidatepage, case 2)
5274                  *
5275                  * Note, end is the bytenr of last byte, so we need + 1 here.
5276                  */
5277                 if (state_flags & EXTENT_DELALLOC)
5278                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5279
5280                 clear_extent_bit(io_tree, start, end,
5281                                  EXTENT_LOCKED | EXTENT_DIRTY |
5282                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5283                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5284
5285                 cond_resched();
5286                 spin_lock(&io_tree->lock);
5287         }
5288         spin_unlock(&io_tree->lock);
5289 }
5290
5291 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5292                                                         struct btrfs_block_rsv *rsv)
5293 {
5294         struct btrfs_fs_info *fs_info = root->fs_info;
5295         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5296         u64 delayed_refs_extra = btrfs_calc_trans_metadata_size(fs_info, 1);
5297         int failures = 0;
5298
5299         for (;;) {
5300                 struct btrfs_trans_handle *trans;
5301                 int ret;
5302
5303                 ret = btrfs_block_rsv_refill(root, rsv,
5304                                              rsv->size + delayed_refs_extra,
5305                                              BTRFS_RESERVE_FLUSH_LIMIT);
5306
5307                 if (ret && ++failures > 2) {
5308                         btrfs_warn(fs_info,
5309                                    "could not allocate space for a delete; will truncate on mount");
5310                         return ERR_PTR(-ENOSPC);
5311                 }
5312
5313                 /*
5314                  * Evict can generate a large amount of delayed refs without
5315                  * having a way to add space back since we exhaust our temporary
5316                  * block rsv.  We aren't allowed to do FLUSH_ALL in this case
5317                  * because we could deadlock with so many things in the flushing
5318                  * code, so we have to try and hold some extra space to
5319                  * compensate for our delayed ref generation.  If we can't get
5320                  * that space then we need see if we can steal our minimum from
5321                  * the global reserve.  We will be ratelimited by the amount of
5322                  * space we have for the delayed refs rsv, so we'll end up
5323                  * committing and trying again.
5324                  */
5325                 trans = btrfs_join_transaction(root);
5326                 if (IS_ERR(trans) || !ret) {
5327                         if (!IS_ERR(trans)) {
5328                                 trans->block_rsv = &fs_info->trans_block_rsv;
5329                                 trans->bytes_reserved = delayed_refs_extra;
5330                                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5331                                                         delayed_refs_extra, 1);
5332                         }
5333                         return trans;
5334                 }
5335
5336                 /*
5337                  * Try to steal from the global reserve if there is space for
5338                  * it.
5339                  */
5340                 if (!btrfs_check_space_for_delayed_refs(fs_info) &&
5341                     !btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0))
5342                         return trans;
5343
5344                 /* If not, commit and try again. */
5345                 ret = btrfs_commit_transaction(trans);
5346                 if (ret)
5347                         return ERR_PTR(ret);
5348         }
5349 }
5350
5351 void btrfs_evict_inode(struct inode *inode)
5352 {
5353         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5354         struct btrfs_trans_handle *trans;
5355         struct btrfs_root *root = BTRFS_I(inode)->root;
5356         struct btrfs_block_rsv *rsv;
5357         int ret;
5358
5359         trace_btrfs_inode_evict(inode);
5360
5361         if (!root) {
5362                 clear_inode(inode);
5363                 return;
5364         }
5365
5366         evict_inode_truncate_pages(inode);
5367
5368         if (inode->i_nlink &&
5369             ((btrfs_root_refs(&root->root_item) != 0 &&
5370               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5371              btrfs_is_free_space_inode(BTRFS_I(inode))))
5372                 goto no_delete;
5373
5374         if (is_bad_inode(inode))
5375                 goto no_delete;
5376
5377         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5378
5379         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5380                 goto no_delete;
5381
5382         if (inode->i_nlink > 0) {
5383                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5384                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5385                 goto no_delete;
5386         }
5387
5388         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5389         if (ret)
5390                 goto no_delete;
5391
5392         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5393         if (!rsv)
5394                 goto no_delete;
5395         rsv->size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5396         rsv->failfast = 1;
5397
5398         btrfs_i_size_write(BTRFS_I(inode), 0);
5399
5400         while (1) {
5401                 trans = evict_refill_and_join(root, rsv);
5402                 if (IS_ERR(trans))
5403                         goto free_rsv;
5404
5405                 trans->block_rsv = rsv;
5406
5407                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5408                 trans->block_rsv = &fs_info->trans_block_rsv;
5409                 btrfs_end_transaction(trans);
5410                 btrfs_btree_balance_dirty(fs_info);
5411                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5412                         goto free_rsv;
5413                 else if (!ret)
5414                         break;
5415         }
5416
5417         /*
5418          * Errors here aren't a big deal, it just means we leave orphan items in
5419          * the tree. They will be cleaned up on the next mount. If the inode
5420          * number gets reused, cleanup deletes the orphan item without doing
5421          * anything, and unlink reuses the existing orphan item.
5422          *
5423          * If it turns out that we are dropping too many of these, we might want
5424          * to add a mechanism for retrying these after a commit.
5425          */
5426         trans = evict_refill_and_join(root, rsv);
5427         if (!IS_ERR(trans)) {
5428                 trans->block_rsv = rsv;
5429                 btrfs_orphan_del(trans, BTRFS_I(inode));
5430                 trans->block_rsv = &fs_info->trans_block_rsv;
5431                 btrfs_end_transaction(trans);
5432         }
5433
5434         if (!(root == fs_info->tree_root ||
5435               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5436                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5437
5438 free_rsv:
5439         btrfs_free_block_rsv(fs_info, rsv);
5440 no_delete:
5441         /*
5442          * If we didn't successfully delete, the orphan item will still be in
5443          * the tree and we'll retry on the next mount. Again, we might also want
5444          * to retry these periodically in the future.
5445          */
5446         btrfs_remove_delayed_node(BTRFS_I(inode));
5447         clear_inode(inode);
5448 }
5449
5450 /*
5451  * this returns the key found in the dir entry in the location pointer.
5452  * If no dir entries were found, returns -ENOENT.
5453  * If found a corrupted location in dir entry, returns -EUCLEAN.
5454  */
5455 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5456                                struct btrfs_key *location)
5457 {
5458         const char *name = dentry->d_name.name;
5459         int namelen = dentry->d_name.len;
5460         struct btrfs_dir_item *di;
5461         struct btrfs_path *path;
5462         struct btrfs_root *root = BTRFS_I(dir)->root;
5463         int ret = 0;
5464
5465         path = btrfs_alloc_path();
5466         if (!path)
5467                 return -ENOMEM;
5468
5469         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5470                         name, namelen, 0);
5471         if (IS_ERR_OR_NULL(di)) {
5472                 ret = di ? PTR_ERR(di) : -ENOENT;
5473                 goto out;
5474         }
5475
5476         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5477         if (location->type != BTRFS_INODE_ITEM_KEY &&
5478             location->type != BTRFS_ROOT_ITEM_KEY) {
5479                 ret = -EUCLEAN;
5480                 btrfs_warn(root->fs_info,
5481 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5482                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5483                            location->objectid, location->type, location->offset);
5484         }
5485 out:
5486         btrfs_free_path(path);
5487         return ret;
5488 }
5489
5490 /*
5491  * when we hit a tree root in a directory, the btrfs part of the inode
5492  * needs to be changed to reflect the root directory of the tree root.  This
5493  * is kind of like crossing a mount point.
5494  */
5495 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5496                                     struct inode *dir,
5497                                     struct dentry *dentry,
5498                                     struct btrfs_key *location,
5499                                     struct btrfs_root **sub_root)
5500 {
5501         struct btrfs_path *path;
5502         struct btrfs_root *new_root;
5503         struct btrfs_root_ref *ref;
5504         struct extent_buffer *leaf;
5505         struct btrfs_key key;
5506         int ret;
5507         int err = 0;
5508
5509         path = btrfs_alloc_path();
5510         if (!path) {
5511                 err = -ENOMEM;
5512                 goto out;
5513         }
5514
5515         err = -ENOENT;
5516         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5517         key.type = BTRFS_ROOT_REF_KEY;
5518         key.offset = location->objectid;
5519
5520         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5521         if (ret) {
5522                 if (ret < 0)
5523                         err = ret;
5524                 goto out;
5525         }
5526
5527         leaf = path->nodes[0];
5528         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5529         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5530             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5531                 goto out;
5532
5533         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5534                                    (unsigned long)(ref + 1),
5535                                    dentry->d_name.len);
5536         if (ret)
5537                 goto out;
5538
5539         btrfs_release_path(path);
5540
5541         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5542         if (IS_ERR(new_root)) {
5543                 err = PTR_ERR(new_root);
5544                 goto out;
5545         }
5546
5547         *sub_root = new_root;
5548         location->objectid = btrfs_root_dirid(&new_root->root_item);
5549         location->type = BTRFS_INODE_ITEM_KEY;
5550         location->offset = 0;
5551         err = 0;
5552 out:
5553         btrfs_free_path(path);
5554         return err;
5555 }
5556
5557 static void inode_tree_add(struct inode *inode)
5558 {
5559         struct btrfs_root *root = BTRFS_I(inode)->root;
5560         struct btrfs_inode *entry;
5561         struct rb_node **p;
5562         struct rb_node *parent;
5563         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5564         u64 ino = btrfs_ino(BTRFS_I(inode));
5565
5566         if (inode_unhashed(inode))
5567                 return;
5568         parent = NULL;
5569         spin_lock(&root->inode_lock);
5570         p = &root->inode_tree.rb_node;
5571         while (*p) {
5572                 parent = *p;
5573                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5574
5575                 if (ino < btrfs_ino(entry))
5576                         p = &parent->rb_left;
5577                 else if (ino > btrfs_ino(entry))
5578                         p = &parent->rb_right;
5579                 else {
5580                         WARN_ON(!(entry->vfs_inode.i_state &
5581                                   (I_WILL_FREE | I_FREEING)));
5582                         rb_replace_node(parent, new, &root->inode_tree);
5583                         RB_CLEAR_NODE(parent);
5584                         spin_unlock(&root->inode_lock);
5585                         return;
5586                 }
5587         }
5588         rb_link_node(new, parent, p);
5589         rb_insert_color(new, &root->inode_tree);
5590         spin_unlock(&root->inode_lock);
5591 }
5592
5593 static void inode_tree_del(struct inode *inode)
5594 {
5595         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5596         struct btrfs_root *root = BTRFS_I(inode)->root;
5597         int empty = 0;
5598
5599         spin_lock(&root->inode_lock);
5600         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5601                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5602                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5603                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5604         }
5605         spin_unlock(&root->inode_lock);
5606
5607         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5608                 synchronize_srcu(&fs_info->subvol_srcu);
5609                 spin_lock(&root->inode_lock);
5610                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5611                 spin_unlock(&root->inode_lock);
5612                 if (empty)
5613                         btrfs_add_dead_root(root);
5614         }
5615 }
5616
5617
5618 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5619 {
5620         struct btrfs_iget_args *args = p;
5621         inode->i_ino = args->location->objectid;
5622         memcpy(&BTRFS_I(inode)->location, args->location,
5623                sizeof(*args->location));
5624         BTRFS_I(inode)->root = args->root;
5625         return 0;
5626 }
5627
5628 static int btrfs_find_actor(struct inode *inode, void *opaque)
5629 {
5630         struct btrfs_iget_args *args = opaque;
5631         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5632                 args->root == BTRFS_I(inode)->root;
5633 }
5634
5635 static struct inode *btrfs_iget_locked(struct super_block *s,
5636                                        struct btrfs_key *location,
5637                                        struct btrfs_root *root)
5638 {
5639         struct inode *inode;
5640         struct btrfs_iget_args args;
5641         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5642
5643         args.location = location;
5644         args.root = root;
5645
5646         inode = iget5_locked(s, hashval, btrfs_find_actor,
5647                              btrfs_init_locked_inode,
5648                              (void *)&args);
5649         return inode;
5650 }
5651
5652 /* Get an inode object given its location and corresponding root.
5653  * Returns in *is_new if the inode was read from disk
5654  */
5655 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5656                               struct btrfs_root *root, int *new,
5657                               struct btrfs_path *path)
5658 {
5659         struct inode *inode;
5660
5661         inode = btrfs_iget_locked(s, location, root);
5662         if (!inode)
5663                 return ERR_PTR(-ENOMEM);
5664
5665         if (inode->i_state & I_NEW) {
5666                 int ret;
5667
5668                 ret = btrfs_read_locked_inode(inode, path);
5669                 if (!ret) {
5670                         inode_tree_add(inode);
5671                         unlock_new_inode(inode);
5672                         if (new)
5673                                 *new = 1;
5674                 } else {
5675                         iget_failed(inode);
5676                         /*
5677                          * ret > 0 can come from btrfs_search_slot called by
5678                          * btrfs_read_locked_inode, this means the inode item
5679                          * was not found.
5680                          */
5681                         if (ret > 0)
5682                                 ret = -ENOENT;
5683                         inode = ERR_PTR(ret);
5684                 }
5685         }
5686
5687         return inode;
5688 }
5689
5690 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5691                          struct btrfs_root *root, int *new)
5692 {
5693         return btrfs_iget_path(s, location, root, new, NULL);
5694 }
5695
5696 static struct inode *new_simple_dir(struct super_block *s,
5697                                     struct btrfs_key *key,
5698                                     struct btrfs_root *root)
5699 {
5700         struct inode *inode = new_inode(s);
5701
5702         if (!inode)
5703                 return ERR_PTR(-ENOMEM);
5704
5705         BTRFS_I(inode)->root = root;
5706         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5707         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5708
5709         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5710         inode->i_op = &btrfs_dir_ro_inode_operations;
5711         inode->i_opflags &= ~IOP_XATTR;
5712         inode->i_fop = &simple_dir_operations;
5713         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5714         inode->i_mtime = current_time(inode);
5715         inode->i_atime = inode->i_mtime;
5716         inode->i_ctime = inode->i_mtime;
5717         BTRFS_I(inode)->i_otime = inode->i_mtime;
5718
5719         return inode;
5720 }
5721
5722 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5723 {
5724         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5725         struct inode *inode;
5726         struct btrfs_root *root = BTRFS_I(dir)->root;
5727         struct btrfs_root *sub_root = root;
5728         struct btrfs_key location;
5729         int index;
5730         int ret = 0;
5731
5732         if (dentry->d_name.len > BTRFS_NAME_LEN)
5733                 return ERR_PTR(-ENAMETOOLONG);
5734
5735         ret = btrfs_inode_by_name(dir, dentry, &location);
5736         if (ret < 0)
5737                 return ERR_PTR(ret);
5738
5739         if (location.type == BTRFS_INODE_ITEM_KEY) {
5740                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5741                 return inode;
5742         }
5743
5744         index = srcu_read_lock(&fs_info->subvol_srcu);
5745         ret = fixup_tree_root_location(fs_info, dir, dentry,
5746                                        &location, &sub_root);
5747         if (ret < 0) {
5748                 if (ret != -ENOENT)
5749                         inode = ERR_PTR(ret);
5750                 else
5751                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5752         } else {
5753                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5754         }
5755         srcu_read_unlock(&fs_info->subvol_srcu, index);
5756
5757         if (!IS_ERR(inode) && root != sub_root) {
5758                 down_read(&fs_info->cleanup_work_sem);
5759                 if (!sb_rdonly(inode->i_sb))
5760                         ret = btrfs_orphan_cleanup(sub_root);
5761                 up_read(&fs_info->cleanup_work_sem);
5762                 if (ret) {
5763                         iput(inode);
5764                         inode = ERR_PTR(ret);
5765                 }
5766         }
5767
5768         return inode;
5769 }
5770
5771 static int btrfs_dentry_delete(const struct dentry *dentry)
5772 {
5773         struct btrfs_root *root;
5774         struct inode *inode = d_inode(dentry);
5775
5776         if (!inode && !IS_ROOT(dentry))
5777                 inode = d_inode(dentry->d_parent);
5778
5779         if (inode) {
5780                 root = BTRFS_I(inode)->root;
5781                 if (btrfs_root_refs(&root->root_item) == 0)
5782                         return 1;
5783
5784                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5785                         return 1;
5786         }
5787         return 0;
5788 }
5789
5790 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5791                                    unsigned int flags)
5792 {
5793         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5794
5795         if (inode == ERR_PTR(-ENOENT))
5796                 inode = NULL;
5797         return d_splice_alias(inode, dentry);
5798 }
5799
5800 unsigned char btrfs_filetype_table[] = {
5801         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5802 };
5803
5804 /*
5805  * All this infrastructure exists because dir_emit can fault, and we are holding
5806  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5807  * our information into that, and then dir_emit from the buffer.  This is
5808  * similar to what NFS does, only we don't keep the buffer around in pagecache
5809  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5810  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5811  * tree lock.
5812  */
5813 static int btrfs_opendir(struct inode *inode, struct file *file)
5814 {
5815         struct btrfs_file_private *private;
5816
5817         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5818         if (!private)
5819                 return -ENOMEM;
5820         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5821         if (!private->filldir_buf) {
5822                 kfree(private);
5823                 return -ENOMEM;
5824         }
5825         file->private_data = private;
5826         return 0;
5827 }
5828
5829 struct dir_entry {
5830         u64 ino;
5831         u64 offset;
5832         unsigned type;
5833         int name_len;
5834 };
5835
5836 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5837 {
5838         while (entries--) {
5839                 struct dir_entry *entry = addr;
5840                 char *name = (char *)(entry + 1);
5841
5842                 ctx->pos = get_unaligned(&entry->offset);
5843                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5844                                          get_unaligned(&entry->ino),
5845                                          get_unaligned(&entry->type)))
5846                         return 1;
5847                 addr += sizeof(struct dir_entry) +
5848                         get_unaligned(&entry->name_len);
5849                 ctx->pos++;
5850         }
5851         return 0;
5852 }
5853
5854 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5855 {
5856         struct inode *inode = file_inode(file);
5857         struct btrfs_root *root = BTRFS_I(inode)->root;
5858         struct btrfs_file_private *private = file->private_data;
5859         struct btrfs_dir_item *di;
5860         struct btrfs_key key;
5861         struct btrfs_key found_key;
5862         struct btrfs_path *path;
5863         void *addr;
5864         struct list_head ins_list;
5865         struct list_head del_list;
5866         int ret;
5867         struct extent_buffer *leaf;
5868         int slot;
5869         char *name_ptr;
5870         int name_len;
5871         int entries = 0;
5872         int total_len = 0;
5873         bool put = false;
5874         struct btrfs_key location;
5875
5876         if (!dir_emit_dots(file, ctx))
5877                 return 0;
5878
5879         path = btrfs_alloc_path();
5880         if (!path)
5881                 return -ENOMEM;
5882
5883         addr = private->filldir_buf;
5884         path->reada = READA_FORWARD;
5885
5886         INIT_LIST_HEAD(&ins_list);
5887         INIT_LIST_HEAD(&del_list);
5888         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5889
5890 again:
5891         key.type = BTRFS_DIR_INDEX_KEY;
5892         key.offset = ctx->pos;
5893         key.objectid = btrfs_ino(BTRFS_I(inode));
5894
5895         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5896         if (ret < 0)
5897                 goto err;
5898
5899         while (1) {
5900                 struct dir_entry *entry;
5901
5902                 leaf = path->nodes[0];
5903                 slot = path->slots[0];
5904                 if (slot >= btrfs_header_nritems(leaf)) {
5905                         ret = btrfs_next_leaf(root, path);
5906                         if (ret < 0)
5907                                 goto err;
5908                         else if (ret > 0)
5909                                 break;
5910                         continue;
5911                 }
5912
5913                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5914
5915                 if (found_key.objectid != key.objectid)
5916                         break;
5917                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5918                         break;
5919                 if (found_key.offset < ctx->pos)
5920                         goto next;
5921                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5922                         goto next;
5923                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5924                 name_len = btrfs_dir_name_len(leaf, di);
5925                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5926                     PAGE_SIZE) {
5927                         btrfs_release_path(path);
5928                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5929                         if (ret)
5930                                 goto nopos;
5931                         addr = private->filldir_buf;
5932                         entries = 0;
5933                         total_len = 0;
5934                         goto again;
5935                 }
5936
5937                 entry = addr;
5938                 put_unaligned(name_len, &entry->name_len);
5939                 name_ptr = (char *)(entry + 1);
5940                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5941                                    name_len);
5942                 put_unaligned(btrfs_filetype_table[btrfs_dir_type(leaf, di)],
5943                                 &entry->type);
5944                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5945                 put_unaligned(location.objectid, &entry->ino);
5946                 put_unaligned(found_key.offset, &entry->offset);
5947                 entries++;
5948                 addr += sizeof(struct dir_entry) + name_len;
5949                 total_len += sizeof(struct dir_entry) + name_len;
5950 next:
5951                 path->slots[0]++;
5952         }
5953         btrfs_release_path(path);
5954
5955         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5956         if (ret)
5957                 goto nopos;
5958
5959         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5960         if (ret)
5961                 goto nopos;
5962
5963         /*
5964          * Stop new entries from being returned after we return the last
5965          * entry.
5966          *
5967          * New directory entries are assigned a strictly increasing
5968          * offset.  This means that new entries created during readdir
5969          * are *guaranteed* to be seen in the future by that readdir.
5970          * This has broken buggy programs which operate on names as
5971          * they're returned by readdir.  Until we re-use freed offsets
5972          * we have this hack to stop new entries from being returned
5973          * under the assumption that they'll never reach this huge
5974          * offset.
5975          *
5976          * This is being careful not to overflow 32bit loff_t unless the
5977          * last entry requires it because doing so has broken 32bit apps
5978          * in the past.
5979          */
5980         if (ctx->pos >= INT_MAX)
5981                 ctx->pos = LLONG_MAX;
5982         else
5983                 ctx->pos = INT_MAX;
5984 nopos:
5985         ret = 0;
5986 err:
5987         if (put)
5988                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5989         btrfs_free_path(path);
5990         return ret;
5991 }
5992
5993 /*
5994  * This is somewhat expensive, updating the tree every time the
5995  * inode changes.  But, it is most likely to find the inode in cache.
5996  * FIXME, needs more benchmarking...there are no reasons other than performance
5997  * to keep or drop this code.
5998  */
5999 static int btrfs_dirty_inode(struct inode *inode)
6000 {
6001         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6002         struct btrfs_root *root = BTRFS_I(inode)->root;
6003         struct btrfs_trans_handle *trans;
6004         int ret;
6005
6006         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6007                 return 0;
6008
6009         trans = btrfs_join_transaction(root);
6010         if (IS_ERR(trans))
6011                 return PTR_ERR(trans);
6012
6013         ret = btrfs_update_inode(trans, root, inode);
6014         if (ret && ret == -ENOSPC) {
6015                 /* whoops, lets try again with the full transaction */
6016                 btrfs_end_transaction(trans);
6017                 trans = btrfs_start_transaction(root, 1);
6018                 if (IS_ERR(trans))
6019                         return PTR_ERR(trans);
6020
6021                 ret = btrfs_update_inode(trans, root, inode);
6022         }
6023         btrfs_end_transaction(trans);
6024         if (BTRFS_I(inode)->delayed_node)
6025                 btrfs_balance_delayed_items(fs_info);
6026
6027         return ret;
6028 }
6029
6030 /*
6031  * This is a copy of file_update_time.  We need this so we can return error on
6032  * ENOSPC for updating the inode in the case of file write and mmap writes.
6033  */
6034 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6035                              int flags)
6036 {
6037         struct btrfs_root *root = BTRFS_I(inode)->root;
6038         bool dirty = flags & ~S_VERSION;
6039
6040         if (btrfs_root_readonly(root))
6041                 return -EROFS;
6042
6043         if (flags & S_VERSION)
6044                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6045         if (flags & S_CTIME)
6046                 inode->i_ctime = *now;
6047         if (flags & S_MTIME)
6048                 inode->i_mtime = *now;
6049         if (flags & S_ATIME)
6050                 inode->i_atime = *now;
6051         return dirty ? btrfs_dirty_inode(inode) : 0;
6052 }
6053
6054 /*
6055  * find the highest existing sequence number in a directory
6056  * and then set the in-memory index_cnt variable to reflect
6057  * free sequence numbers
6058  */
6059 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6060 {
6061         struct btrfs_root *root = inode->root;
6062         struct btrfs_key key, found_key;
6063         struct btrfs_path *path;
6064         struct extent_buffer *leaf;
6065         int ret;
6066
6067         key.objectid = btrfs_ino(inode);
6068         key.type = BTRFS_DIR_INDEX_KEY;
6069         key.offset = (u64)-1;
6070
6071         path = btrfs_alloc_path();
6072         if (!path)
6073                 return -ENOMEM;
6074
6075         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6076         if (ret < 0)
6077                 goto out;
6078         /* FIXME: we should be able to handle this */
6079         if (ret == 0)
6080                 goto out;
6081         ret = 0;
6082
6083         /*
6084          * MAGIC NUMBER EXPLANATION:
6085          * since we search a directory based on f_pos we have to start at 2
6086          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6087          * else has to start at 2
6088          */
6089         if (path->slots[0] == 0) {
6090                 inode->index_cnt = 2;
6091                 goto out;
6092         }
6093
6094         path->slots[0]--;
6095
6096         leaf = path->nodes[0];
6097         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6098
6099         if (found_key.objectid != btrfs_ino(inode) ||
6100             found_key.type != BTRFS_DIR_INDEX_KEY) {
6101                 inode->index_cnt = 2;
6102                 goto out;
6103         }
6104
6105         inode->index_cnt = found_key.offset + 1;
6106 out:
6107         btrfs_free_path(path);
6108         return ret;
6109 }
6110
6111 /*
6112  * helper to find a free sequence number in a given directory.  This current
6113  * code is very simple, later versions will do smarter things in the btree
6114  */
6115 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6116 {
6117         int ret = 0;
6118
6119         if (dir->index_cnt == (u64)-1) {
6120                 ret = btrfs_inode_delayed_dir_index_count(dir);
6121                 if (ret) {
6122                         ret = btrfs_set_inode_index_count(dir);
6123                         if (ret)
6124                                 return ret;
6125                 }
6126         }
6127
6128         *index = dir->index_cnt;
6129         dir->index_cnt++;
6130
6131         return ret;
6132 }
6133
6134 static int btrfs_insert_inode_locked(struct inode *inode)
6135 {
6136         struct btrfs_iget_args args;
6137         args.location = &BTRFS_I(inode)->location;
6138         args.root = BTRFS_I(inode)->root;
6139
6140         return insert_inode_locked4(inode,
6141                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6142                    btrfs_find_actor, &args);
6143 }
6144
6145 /*
6146  * Inherit flags from the parent inode.
6147  *
6148  * Currently only the compression flags and the cow flags are inherited.
6149  */
6150 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6151 {
6152         unsigned int flags;
6153
6154         if (!dir)
6155                 return;
6156
6157         flags = BTRFS_I(dir)->flags;
6158
6159         if (flags & BTRFS_INODE_NOCOMPRESS) {
6160                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6161                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6162         } else if (flags & BTRFS_INODE_COMPRESS) {
6163                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6164                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6165         }
6166
6167         if (flags & BTRFS_INODE_NODATACOW) {
6168                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6169                 if (S_ISREG(inode->i_mode))
6170                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6171         }
6172
6173         btrfs_sync_inode_flags_to_i_flags(inode);
6174 }
6175
6176 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6177                                      struct btrfs_root *root,
6178                                      struct inode *dir,
6179                                      const char *name, int name_len,
6180                                      u64 ref_objectid, u64 objectid,
6181                                      umode_t mode, u64 *index)
6182 {
6183         struct btrfs_fs_info *fs_info = root->fs_info;
6184         struct inode *inode;
6185         struct btrfs_inode_item *inode_item;
6186         struct btrfs_key *location;
6187         struct btrfs_path *path;
6188         struct btrfs_inode_ref *ref;
6189         struct btrfs_key key[2];
6190         u32 sizes[2];
6191         int nitems = name ? 2 : 1;
6192         unsigned long ptr;
6193         int ret;
6194
6195         path = btrfs_alloc_path();
6196         if (!path)
6197                 return ERR_PTR(-ENOMEM);
6198
6199         inode = new_inode(fs_info->sb);
6200         if (!inode) {
6201                 btrfs_free_path(path);
6202                 return ERR_PTR(-ENOMEM);
6203         }
6204
6205         /*
6206          * O_TMPFILE, set link count to 0, so that after this point,
6207          * we fill in an inode item with the correct link count.
6208          */
6209         if (!name)
6210                 set_nlink(inode, 0);
6211
6212         /*
6213          * we have to initialize this early, so we can reclaim the inode
6214          * number if we fail afterwards in this function.
6215          */
6216         inode->i_ino = objectid;
6217
6218         if (dir && name) {
6219                 trace_btrfs_inode_request(dir);
6220
6221                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6222                 if (ret) {
6223                         btrfs_free_path(path);
6224                         iput(inode);
6225                         return ERR_PTR(ret);
6226                 }
6227         } else if (dir) {
6228                 *index = 0;
6229         }
6230         /*
6231          * index_cnt is ignored for everything but a dir,
6232          * btrfs_set_inode_index_count has an explanation for the magic
6233          * number
6234          */
6235         BTRFS_I(inode)->index_cnt = 2;
6236         BTRFS_I(inode)->dir_index = *index;
6237         BTRFS_I(inode)->root = root;
6238         BTRFS_I(inode)->generation = trans->transid;
6239         inode->i_generation = BTRFS_I(inode)->generation;
6240
6241         /*
6242          * We could have gotten an inode number from somebody who was fsynced
6243          * and then removed in this same transaction, so let's just set full
6244          * sync since it will be a full sync anyway and this will blow away the
6245          * old info in the log.
6246          */
6247         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6248
6249         key[0].objectid = objectid;
6250         key[0].type = BTRFS_INODE_ITEM_KEY;
6251         key[0].offset = 0;
6252
6253         sizes[0] = sizeof(struct btrfs_inode_item);
6254
6255         if (name) {
6256                 /*
6257                  * Start new inodes with an inode_ref. This is slightly more
6258                  * efficient for small numbers of hard links since they will
6259                  * be packed into one item. Extended refs will kick in if we
6260                  * add more hard links than can fit in the ref item.
6261                  */
6262                 key[1].objectid = objectid;
6263                 key[1].type = BTRFS_INODE_REF_KEY;
6264                 key[1].offset = ref_objectid;
6265
6266                 sizes[1] = name_len + sizeof(*ref);
6267         }
6268
6269         location = &BTRFS_I(inode)->location;
6270         location->objectid = objectid;
6271         location->offset = 0;
6272         location->type = BTRFS_INODE_ITEM_KEY;
6273
6274         ret = btrfs_insert_inode_locked(inode);
6275         if (ret < 0) {
6276                 iput(inode);
6277                 goto fail;
6278         }
6279
6280         path->leave_spinning = 1;
6281         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6282         if (ret != 0)
6283                 goto fail_unlock;
6284
6285         inode_init_owner(inode, dir, mode);
6286         inode_set_bytes(inode, 0);
6287
6288         inode->i_mtime = current_time(inode);
6289         inode->i_atime = inode->i_mtime;
6290         inode->i_ctime = inode->i_mtime;
6291         BTRFS_I(inode)->i_otime = inode->i_mtime;
6292
6293         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6294                                   struct btrfs_inode_item);
6295         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6296                              sizeof(*inode_item));
6297         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6298
6299         if (name) {
6300                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6301                                      struct btrfs_inode_ref);
6302                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6303                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6304                 ptr = (unsigned long)(ref + 1);
6305                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6306         }
6307
6308         btrfs_mark_buffer_dirty(path->nodes[0]);
6309         btrfs_free_path(path);
6310
6311         btrfs_inherit_iflags(inode, dir);
6312
6313         if (S_ISREG(mode)) {
6314                 if (btrfs_test_opt(fs_info, NODATASUM))
6315                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6316                 if (btrfs_test_opt(fs_info, NODATACOW))
6317                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6318                                 BTRFS_INODE_NODATASUM;
6319         }
6320
6321         inode_tree_add(inode);
6322
6323         trace_btrfs_inode_new(inode);
6324         btrfs_set_inode_last_trans(trans, inode);
6325
6326         btrfs_update_root_times(trans, root);
6327
6328         ret = btrfs_inode_inherit_props(trans, inode, dir);
6329         if (ret)
6330                 btrfs_err(fs_info,
6331                           "error inheriting props for ino %llu (root %llu): %d",
6332                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6333
6334         return inode;
6335
6336 fail_unlock:
6337         discard_new_inode(inode);
6338 fail:
6339         if (dir && name)
6340                 BTRFS_I(dir)->index_cnt--;
6341         btrfs_free_path(path);
6342         return ERR_PTR(ret);
6343 }
6344
6345 static inline u8 btrfs_inode_type(struct inode *inode)
6346 {
6347         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6348 }
6349
6350 /*
6351  * utility function to add 'inode' into 'parent_inode' with
6352  * a give name and a given sequence number.
6353  * if 'add_backref' is true, also insert a backref from the
6354  * inode to the parent directory.
6355  */
6356 int btrfs_add_link(struct btrfs_trans_handle *trans,
6357                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6358                    const char *name, int name_len, int add_backref, u64 index)
6359 {
6360         int ret = 0;
6361         struct btrfs_key key;
6362         struct btrfs_root *root = parent_inode->root;
6363         u64 ino = btrfs_ino(inode);
6364         u64 parent_ino = btrfs_ino(parent_inode);
6365
6366         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6367                 memcpy(&key, &inode->root->root_key, sizeof(key));
6368         } else {
6369                 key.objectid = ino;
6370                 key.type = BTRFS_INODE_ITEM_KEY;
6371                 key.offset = 0;
6372         }
6373
6374         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6375                 ret = btrfs_add_root_ref(trans, key.objectid,
6376                                          root->root_key.objectid, parent_ino,
6377                                          index, name, name_len);
6378         } else if (add_backref) {
6379                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6380                                              parent_ino, index);
6381         }
6382
6383         /* Nothing to clean up yet */
6384         if (ret)
6385                 return ret;
6386
6387         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6388                                     btrfs_inode_type(&inode->vfs_inode), index);
6389         if (ret == -EEXIST || ret == -EOVERFLOW)
6390                 goto fail_dir_item;
6391         else if (ret) {
6392                 btrfs_abort_transaction(trans, ret);
6393                 return ret;
6394         }
6395
6396         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6397                            name_len * 2);
6398         inode_inc_iversion(&parent_inode->vfs_inode);
6399         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6400                 current_time(&parent_inode->vfs_inode);
6401         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6402         if (ret)
6403                 btrfs_abort_transaction(trans, ret);
6404         return ret;
6405
6406 fail_dir_item:
6407         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6408                 u64 local_index;
6409                 int err;
6410                 err = btrfs_del_root_ref(trans, key.objectid,
6411                                          root->root_key.objectid, parent_ino,
6412                                          &local_index, name, name_len);
6413                 if (err)
6414                         btrfs_abort_transaction(trans, err);
6415         } else if (add_backref) {
6416                 u64 local_index;
6417                 int err;
6418
6419                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6420                                           ino, parent_ino, &local_index);
6421                 if (err)
6422                         btrfs_abort_transaction(trans, err);
6423         }
6424
6425         /* Return the original error code */
6426         return ret;
6427 }
6428
6429 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6430                             struct btrfs_inode *dir, struct dentry *dentry,
6431                             struct btrfs_inode *inode, int backref, u64 index)
6432 {
6433         int err = btrfs_add_link(trans, dir, inode,
6434                                  dentry->d_name.name, dentry->d_name.len,
6435                                  backref, index);
6436         if (err > 0)
6437                 err = -EEXIST;
6438         return err;
6439 }
6440
6441 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6442                         umode_t mode, dev_t rdev)
6443 {
6444         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6445         struct btrfs_trans_handle *trans;
6446         struct btrfs_root *root = BTRFS_I(dir)->root;
6447         struct inode *inode = NULL;
6448         int err;
6449         u64 objectid;
6450         u64 index = 0;
6451
6452         /*
6453          * 2 for inode item and ref
6454          * 2 for dir items
6455          * 1 for xattr if selinux is on
6456          */
6457         trans = btrfs_start_transaction(root, 5);
6458         if (IS_ERR(trans))
6459                 return PTR_ERR(trans);
6460
6461         err = btrfs_find_free_ino(root, &objectid);
6462         if (err)
6463                 goto out_unlock;
6464
6465         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6466                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6467                         mode, &index);
6468         if (IS_ERR(inode)) {
6469                 err = PTR_ERR(inode);
6470                 inode = NULL;
6471                 goto out_unlock;
6472         }
6473
6474         /*
6475         * If the active LSM wants to access the inode during
6476         * d_instantiate it needs these. Smack checks to see
6477         * if the filesystem supports xattrs by looking at the
6478         * ops vector.
6479         */
6480         inode->i_op = &btrfs_special_inode_operations;
6481         init_special_inode(inode, inode->i_mode, rdev);
6482
6483         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6484         if (err)
6485                 goto out_unlock;
6486
6487         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6488                         0, index);
6489         if (err)
6490                 goto out_unlock;
6491
6492         btrfs_update_inode(trans, root, inode);
6493         d_instantiate_new(dentry, inode);
6494
6495 out_unlock:
6496         btrfs_end_transaction(trans);
6497         btrfs_btree_balance_dirty(fs_info);
6498         if (err && inode) {
6499                 inode_dec_link_count(inode);
6500                 discard_new_inode(inode);
6501         }
6502         return err;
6503 }
6504
6505 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6506                         umode_t mode, bool excl)
6507 {
6508         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6509         struct btrfs_trans_handle *trans;
6510         struct btrfs_root *root = BTRFS_I(dir)->root;
6511         struct inode *inode = NULL;
6512         int err;
6513         u64 objectid;
6514         u64 index = 0;
6515
6516         /*
6517          * 2 for inode item and ref
6518          * 2 for dir items
6519          * 1 for xattr if selinux is on
6520          */
6521         trans = btrfs_start_transaction(root, 5);
6522         if (IS_ERR(trans))
6523                 return PTR_ERR(trans);
6524
6525         err = btrfs_find_free_ino(root, &objectid);
6526         if (err)
6527                 goto out_unlock;
6528
6529         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6530                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6531                         mode, &index);
6532         if (IS_ERR(inode)) {
6533                 err = PTR_ERR(inode);
6534                 inode = NULL;
6535                 goto out_unlock;
6536         }
6537         /*
6538         * If the active LSM wants to access the inode during
6539         * d_instantiate it needs these. Smack checks to see
6540         * if the filesystem supports xattrs by looking at the
6541         * ops vector.
6542         */
6543         inode->i_fop = &btrfs_file_operations;
6544         inode->i_op = &btrfs_file_inode_operations;
6545         inode->i_mapping->a_ops = &btrfs_aops;
6546
6547         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6548         if (err)
6549                 goto out_unlock;
6550
6551         err = btrfs_update_inode(trans, root, inode);
6552         if (err)
6553                 goto out_unlock;
6554
6555         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6556                         0, index);
6557         if (err)
6558                 goto out_unlock;
6559
6560         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6561         d_instantiate_new(dentry, inode);
6562
6563 out_unlock:
6564         btrfs_end_transaction(trans);
6565         if (err && inode) {
6566                 inode_dec_link_count(inode);
6567                 discard_new_inode(inode);
6568         }
6569         btrfs_btree_balance_dirty(fs_info);
6570         return err;
6571 }
6572
6573 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6574                       struct dentry *dentry)
6575 {
6576         struct btrfs_trans_handle *trans = NULL;
6577         struct btrfs_root *root = BTRFS_I(dir)->root;
6578         struct inode *inode = d_inode(old_dentry);
6579         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6580         u64 index;
6581         int err;
6582         int drop_inode = 0;
6583
6584         /* do not allow sys_link's with other subvols of the same device */
6585         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6586                 return -EXDEV;
6587
6588         if (inode->i_nlink >= BTRFS_LINK_MAX)
6589                 return -EMLINK;
6590
6591         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6592         if (err)
6593                 goto fail;
6594
6595         /*
6596          * 2 items for inode and inode ref
6597          * 2 items for dir items
6598          * 1 item for parent inode
6599          * 1 item for orphan item deletion if O_TMPFILE
6600          */
6601         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6602         if (IS_ERR(trans)) {
6603                 err = PTR_ERR(trans);
6604                 trans = NULL;
6605                 goto fail;
6606         }
6607
6608         /* There are several dir indexes for this inode, clear the cache. */
6609         BTRFS_I(inode)->dir_index = 0ULL;
6610         inc_nlink(inode);
6611         inode_inc_iversion(inode);
6612         inode->i_ctime = current_time(inode);
6613         ihold(inode);
6614         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6615
6616         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6617                         1, index);
6618
6619         if (err) {
6620                 drop_inode = 1;
6621         } else {
6622                 struct dentry *parent = dentry->d_parent;
6623                 int ret;
6624
6625                 err = btrfs_update_inode(trans, root, inode);
6626                 if (err)
6627                         goto fail;
6628                 if (inode->i_nlink == 1) {
6629                         /*
6630                          * If new hard link count is 1, it's a file created
6631                          * with open(2) O_TMPFILE flag.
6632                          */
6633                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6634                         if (err)
6635                                 goto fail;
6636                 }
6637                 BTRFS_I(inode)->last_link_trans = trans->transid;
6638                 d_instantiate(dentry, inode);
6639                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6640                                          true, NULL);
6641                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6642                         err = btrfs_commit_transaction(trans);
6643                         trans = NULL;
6644                 }
6645         }
6646
6647 fail:
6648         if (trans)
6649                 btrfs_end_transaction(trans);
6650         if (drop_inode) {
6651                 inode_dec_link_count(inode);
6652                 iput(inode);
6653         }
6654         btrfs_btree_balance_dirty(fs_info);
6655         return err;
6656 }
6657
6658 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6659 {
6660         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6661         struct inode *inode = NULL;
6662         struct btrfs_trans_handle *trans;
6663         struct btrfs_root *root = BTRFS_I(dir)->root;
6664         int err = 0;
6665         u64 objectid = 0;
6666         u64 index = 0;
6667
6668         /*
6669          * 2 items for inode and ref
6670          * 2 items for dir items
6671          * 1 for xattr if selinux is on
6672          */
6673         trans = btrfs_start_transaction(root, 5);
6674         if (IS_ERR(trans))
6675                 return PTR_ERR(trans);
6676
6677         err = btrfs_find_free_ino(root, &objectid);
6678         if (err)
6679                 goto out_fail;
6680
6681         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6682                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6683                         S_IFDIR | mode, &index);
6684         if (IS_ERR(inode)) {
6685                 err = PTR_ERR(inode);
6686                 inode = NULL;
6687                 goto out_fail;
6688         }
6689
6690         /* these must be set before we unlock the inode */
6691         inode->i_op = &btrfs_dir_inode_operations;
6692         inode->i_fop = &btrfs_dir_file_operations;
6693
6694         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6695         if (err)
6696                 goto out_fail;
6697
6698         btrfs_i_size_write(BTRFS_I(inode), 0);
6699         err = btrfs_update_inode(trans, root, inode);
6700         if (err)
6701                 goto out_fail;
6702
6703         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6704                         dentry->d_name.name,
6705                         dentry->d_name.len, 0, index);
6706         if (err)
6707                 goto out_fail;
6708
6709         d_instantiate_new(dentry, inode);
6710
6711 out_fail:
6712         btrfs_end_transaction(trans);
6713         if (err && inode) {
6714                 inode_dec_link_count(inode);
6715                 discard_new_inode(inode);
6716         }
6717         btrfs_btree_balance_dirty(fs_info);
6718         return err;
6719 }
6720
6721 static noinline int uncompress_inline(struct btrfs_path *path,
6722                                       struct page *page,
6723                                       size_t pg_offset, u64 extent_offset,
6724                                       struct btrfs_file_extent_item *item)
6725 {
6726         int ret;
6727         struct extent_buffer *leaf = path->nodes[0];
6728         char *tmp;
6729         size_t max_size;
6730         unsigned long inline_size;
6731         unsigned long ptr;
6732         int compress_type;
6733
6734         WARN_ON(pg_offset != 0);
6735         compress_type = btrfs_file_extent_compression(leaf, item);
6736         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6737         inline_size = btrfs_file_extent_inline_item_len(leaf,
6738                                         btrfs_item_nr(path->slots[0]));
6739         tmp = kmalloc(inline_size, GFP_NOFS);
6740         if (!tmp)
6741                 return -ENOMEM;
6742         ptr = btrfs_file_extent_inline_start(item);
6743
6744         read_extent_buffer(leaf, tmp, ptr, inline_size);
6745
6746         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6747         ret = btrfs_decompress(compress_type, tmp, page,
6748                                extent_offset, inline_size, max_size);
6749
6750         /*
6751          * decompression code contains a memset to fill in any space between the end
6752          * of the uncompressed data and the end of max_size in case the decompressed
6753          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6754          * the end of an inline extent and the beginning of the next block, so we
6755          * cover that region here.
6756          */
6757
6758         if (max_size + pg_offset < PAGE_SIZE) {
6759                 char *map = kmap(page);
6760                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6761                 kunmap(page);
6762         }
6763         kfree(tmp);
6764         return ret;
6765 }
6766
6767 /*
6768  * a bit scary, this does extent mapping from logical file offset to the disk.
6769  * the ugly parts come from merging extents from the disk with the in-ram
6770  * representation.  This gets more complex because of the data=ordered code,
6771  * where the in-ram extents might be locked pending data=ordered completion.
6772  *
6773  * This also copies inline extents directly into the page.
6774  */
6775 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6776                                     struct page *page,
6777                                     size_t pg_offset, u64 start, u64 len,
6778                                     int create)
6779 {
6780         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6781         int ret;
6782         int err = 0;
6783         u64 extent_start = 0;
6784         u64 extent_end = 0;
6785         u64 objectid = btrfs_ino(inode);
6786         u8 extent_type;
6787         struct btrfs_path *path = NULL;
6788         struct btrfs_root *root = inode->root;
6789         struct btrfs_file_extent_item *item;
6790         struct extent_buffer *leaf;
6791         struct btrfs_key found_key;
6792         struct extent_map *em = NULL;
6793         struct extent_map_tree *em_tree = &inode->extent_tree;
6794         struct extent_io_tree *io_tree = &inode->io_tree;
6795         const bool new_inline = !page || create;
6796
6797         read_lock(&em_tree->lock);
6798         em = lookup_extent_mapping(em_tree, start, len);
6799         if (em)
6800                 em->bdev = fs_info->fs_devices->latest_bdev;
6801         read_unlock(&em_tree->lock);
6802
6803         if (em) {
6804                 if (em->start > start || em->start + em->len <= start)
6805                         free_extent_map(em);
6806                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6807                         free_extent_map(em);
6808                 else
6809                         goto out;
6810         }
6811         em = alloc_extent_map();
6812         if (!em) {
6813                 err = -ENOMEM;
6814                 goto out;
6815         }
6816         em->bdev = fs_info->fs_devices->latest_bdev;
6817         em->start = EXTENT_MAP_HOLE;
6818         em->orig_start = EXTENT_MAP_HOLE;
6819         em->len = (u64)-1;
6820         em->block_len = (u64)-1;
6821
6822         path = btrfs_alloc_path();
6823         if (!path) {
6824                 err = -ENOMEM;
6825                 goto out;
6826         }
6827
6828         /* Chances are we'll be called again, so go ahead and do readahead */
6829         path->reada = READA_FORWARD;
6830
6831         /*
6832          * Unless we're going to uncompress the inline extent, no sleep would
6833          * happen.
6834          */
6835         path->leave_spinning = 1;
6836
6837         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6838         if (ret < 0) {
6839                 err = ret;
6840                 goto out;
6841         } else if (ret > 0) {
6842                 if (path->slots[0] == 0)
6843                         goto not_found;
6844                 path->slots[0]--;
6845         }
6846
6847         leaf = path->nodes[0];
6848         item = btrfs_item_ptr(leaf, path->slots[0],
6849                               struct btrfs_file_extent_item);
6850         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6851         if (found_key.objectid != objectid ||
6852             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6853                 /*
6854                  * If we backup past the first extent we want to move forward
6855                  * and see if there is an extent in front of us, otherwise we'll
6856                  * say there is a hole for our whole search range which can
6857                  * cause problems.
6858                  */
6859                 extent_end = start;
6860                 goto next;
6861         }
6862
6863         extent_type = btrfs_file_extent_type(leaf, item);
6864         extent_start = found_key.offset;
6865         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6866             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6867                 extent_end = extent_start +
6868                        btrfs_file_extent_num_bytes(leaf, item);
6869
6870                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6871                                                        extent_start);
6872         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6873                 size_t size;
6874
6875                 size = btrfs_file_extent_ram_bytes(leaf, item);
6876                 extent_end = ALIGN(extent_start + size,
6877                                    fs_info->sectorsize);
6878
6879                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6880                                                       path->slots[0],
6881                                                       extent_start);
6882         }
6883 next:
6884         if (start >= extent_end) {
6885                 path->slots[0]++;
6886                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6887                         ret = btrfs_next_leaf(root, path);
6888                         if (ret < 0) {
6889                                 err = ret;
6890                                 goto out;
6891                         } else if (ret > 0) {
6892                                 goto not_found;
6893                         }
6894                         leaf = path->nodes[0];
6895                 }
6896                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6897                 if (found_key.objectid != objectid ||
6898                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6899                         goto not_found;
6900                 if (start + len <= found_key.offset)
6901                         goto not_found;
6902                 if (start > found_key.offset)
6903                         goto next;
6904
6905                 /* New extent overlaps with existing one */
6906                 em->start = start;
6907                 em->orig_start = start;
6908                 em->len = found_key.offset - start;
6909                 em->block_start = EXTENT_MAP_HOLE;
6910                 goto insert;
6911         }
6912
6913         btrfs_extent_item_to_extent_map(inode, path, item,
6914                         new_inline, em);
6915
6916         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6917             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6918                 goto insert;
6919         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6920                 unsigned long ptr;
6921                 char *map;
6922                 size_t size;
6923                 size_t extent_offset;
6924                 size_t copy_size;
6925
6926                 if (new_inline)
6927                         goto out;
6928
6929                 size = btrfs_file_extent_ram_bytes(leaf, item);
6930                 extent_offset = page_offset(page) + pg_offset - extent_start;
6931                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6932                                   size - extent_offset);
6933                 em->start = extent_start + extent_offset;
6934                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6935                 em->orig_block_len = em->len;
6936                 em->orig_start = em->start;
6937                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6938
6939                 btrfs_set_path_blocking(path);
6940                 if (!PageUptodate(page)) {
6941                         if (btrfs_file_extent_compression(leaf, item) !=
6942                             BTRFS_COMPRESS_NONE) {
6943                                 ret = uncompress_inline(path, page, pg_offset,
6944                                                         extent_offset, item);
6945                                 if (ret) {
6946                                         err = ret;
6947                                         goto out;
6948                                 }
6949                         } else {
6950                                 map = kmap(page);
6951                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6952                                                    copy_size);
6953                                 if (pg_offset + copy_size < PAGE_SIZE) {
6954                                         memset(map + pg_offset + copy_size, 0,
6955                                                PAGE_SIZE - pg_offset -
6956                                                copy_size);
6957                                 }
6958                                 kunmap(page);
6959                         }
6960                         flush_dcache_page(page);
6961                 }
6962                 set_extent_uptodate(io_tree, em->start,
6963                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6964                 goto insert;
6965         }
6966 not_found:
6967         em->start = start;
6968         em->orig_start = start;
6969         em->len = len;
6970         em->block_start = EXTENT_MAP_HOLE;
6971 insert:
6972         btrfs_release_path(path);
6973         if (em->start > start || extent_map_end(em) <= start) {
6974                 btrfs_err(fs_info,
6975                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6976                           em->start, em->len, start, len);
6977                 err = -EIO;
6978                 goto out;
6979         }
6980
6981         err = 0;
6982         write_lock(&em_tree->lock);
6983         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6984         write_unlock(&em_tree->lock);
6985 out:
6986         btrfs_free_path(path);
6987
6988         trace_btrfs_get_extent(root, inode, em);
6989
6990         if (err) {
6991                 free_extent_map(em);
6992                 return ERR_PTR(err);
6993         }
6994         BUG_ON(!em); /* Error is always set */
6995         return em;
6996 }
6997
6998 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6999                                            u64 start, u64 len)
7000 {
7001         struct extent_map *em;
7002         struct extent_map *hole_em = NULL;
7003         u64 delalloc_start = start;
7004         u64 end;
7005         u64 delalloc_len;
7006         u64 delalloc_end;
7007         int err = 0;
7008
7009         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7010         if (IS_ERR(em))
7011                 return em;
7012         /*
7013          * If our em maps to:
7014          * - a hole or
7015          * - a pre-alloc extent,
7016          * there might actually be delalloc bytes behind it.
7017          */
7018         if (em->block_start != EXTENT_MAP_HOLE &&
7019             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7020                 return em;
7021         else
7022                 hole_em = em;
7023
7024         /* check to see if we've wrapped (len == -1 or similar) */
7025         end = start + len;
7026         if (end < start)
7027                 end = (u64)-1;
7028         else
7029                 end -= 1;
7030
7031         em = NULL;
7032
7033         /* ok, we didn't find anything, lets look for delalloc */
7034         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7035                                  end, len, EXTENT_DELALLOC, 1);
7036         delalloc_end = delalloc_start + delalloc_len;
7037         if (delalloc_end < delalloc_start)
7038                 delalloc_end = (u64)-1;
7039
7040         /*
7041          * We didn't find anything useful, return the original results from
7042          * get_extent()
7043          */
7044         if (delalloc_start > end || delalloc_end <= start) {
7045                 em = hole_em;
7046                 hole_em = NULL;
7047                 goto out;
7048         }
7049
7050         /*
7051          * Adjust the delalloc_start to make sure it doesn't go backwards from
7052          * the start they passed in
7053          */
7054         delalloc_start = max(start, delalloc_start);
7055         delalloc_len = delalloc_end - delalloc_start;
7056
7057         if (delalloc_len > 0) {
7058                 u64 hole_start;
7059                 u64 hole_len;
7060                 const u64 hole_end = extent_map_end(hole_em);
7061
7062                 em = alloc_extent_map();
7063                 if (!em) {
7064                         err = -ENOMEM;
7065                         goto out;
7066                 }
7067                 em->bdev = NULL;
7068
7069                 ASSERT(hole_em);
7070                 /*
7071                  * When btrfs_get_extent can't find anything it returns one
7072                  * huge hole
7073                  *
7074                  * Make sure what it found really fits our range, and adjust to
7075                  * make sure it is based on the start from the caller
7076                  */
7077                 if (hole_end <= start || hole_em->start > end) {
7078                        free_extent_map(hole_em);
7079                        hole_em = NULL;
7080                 } else {
7081                        hole_start = max(hole_em->start, start);
7082                        hole_len = hole_end - hole_start;
7083                 }
7084
7085                 if (hole_em && delalloc_start > hole_start) {
7086                         /*
7087                          * Our hole starts before our delalloc, so we have to
7088                          * return just the parts of the hole that go until the
7089                          * delalloc starts
7090                          */
7091                         em->len = min(hole_len, delalloc_start - hole_start);
7092                         em->start = hole_start;
7093                         em->orig_start = hole_start;
7094                         /*
7095                          * Don't adjust block start at all, it is fixed at
7096                          * EXTENT_MAP_HOLE
7097                          */
7098                         em->block_start = hole_em->block_start;
7099                         em->block_len = hole_len;
7100                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7101                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7102                 } else {
7103                         /*
7104                          * Hole is out of passed range or it starts after
7105                          * delalloc range
7106                          */
7107                         em->start = delalloc_start;
7108                         em->len = delalloc_len;
7109                         em->orig_start = delalloc_start;
7110                         em->block_start = EXTENT_MAP_DELALLOC;
7111                         em->block_len = delalloc_len;
7112                 }
7113         } else {
7114                 return hole_em;
7115         }
7116 out:
7117
7118         free_extent_map(hole_em);
7119         if (err) {
7120                 free_extent_map(em);
7121                 return ERR_PTR(err);
7122         }
7123         return em;
7124 }
7125
7126 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7127                                                   const u64 start,
7128                                                   const u64 len,
7129                                                   const u64 orig_start,
7130                                                   const u64 block_start,
7131                                                   const u64 block_len,
7132                                                   const u64 orig_block_len,
7133                                                   const u64 ram_bytes,
7134                                                   const int type)
7135 {
7136         struct extent_map *em = NULL;
7137         int ret;
7138
7139         if (type != BTRFS_ORDERED_NOCOW) {
7140                 em = create_io_em(inode, start, len, orig_start,
7141                                   block_start, block_len, orig_block_len,
7142                                   ram_bytes,
7143                                   BTRFS_COMPRESS_NONE, /* compress_type */
7144                                   type);
7145                 if (IS_ERR(em))
7146                         goto out;
7147         }
7148         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7149                                            len, block_len, type);
7150         if (ret) {
7151                 if (em) {
7152                         free_extent_map(em);
7153                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7154                                                 start + len - 1, 0);
7155                 }
7156                 em = ERR_PTR(ret);
7157         }
7158  out:
7159
7160         return em;
7161 }
7162
7163 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7164                                                   u64 start, u64 len)
7165 {
7166         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7167         struct btrfs_root *root = BTRFS_I(inode)->root;
7168         struct extent_map *em;
7169         struct btrfs_key ins;
7170         u64 alloc_hint;
7171         int ret;
7172
7173         alloc_hint = get_extent_allocation_hint(inode, start, len);
7174         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7175                                    0, alloc_hint, &ins, 1, 1);
7176         if (ret)
7177                 return ERR_PTR(ret);
7178
7179         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7180                                      ins.objectid, ins.offset, ins.offset,
7181                                      ins.offset, BTRFS_ORDERED_REGULAR);
7182         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7183         if (IS_ERR(em))
7184                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7185                                            ins.offset, 1);
7186
7187         return em;
7188 }
7189
7190 /*
7191  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7192  * block must be cow'd
7193  */
7194 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7195                               u64 *orig_start, u64 *orig_block_len,
7196                               u64 *ram_bytes)
7197 {
7198         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7199         struct btrfs_path *path;
7200         int ret;
7201         struct extent_buffer *leaf;
7202         struct btrfs_root *root = BTRFS_I(inode)->root;
7203         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7204         struct btrfs_file_extent_item *fi;
7205         struct btrfs_key key;
7206         u64 disk_bytenr;
7207         u64 backref_offset;
7208         u64 extent_end;
7209         u64 num_bytes;
7210         int slot;
7211         int found_type;
7212         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7213
7214         path = btrfs_alloc_path();
7215         if (!path)
7216                 return -ENOMEM;
7217
7218         ret = btrfs_lookup_file_extent(NULL, root, path,
7219                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7220         if (ret < 0)
7221                 goto out;
7222
7223         slot = path->slots[0];
7224         if (ret == 1) {
7225                 if (slot == 0) {
7226                         /* can't find the item, must cow */
7227                         ret = 0;
7228                         goto out;
7229                 }
7230                 slot--;
7231         }
7232         ret = 0;
7233         leaf = path->nodes[0];
7234         btrfs_item_key_to_cpu(leaf, &key, slot);
7235         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7236             key.type != BTRFS_EXTENT_DATA_KEY) {
7237                 /* not our file or wrong item type, must cow */
7238                 goto out;
7239         }
7240
7241         if (key.offset > offset) {
7242                 /* Wrong offset, must cow */
7243                 goto out;
7244         }
7245
7246         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7247         found_type = btrfs_file_extent_type(leaf, fi);
7248         if (found_type != BTRFS_FILE_EXTENT_REG &&
7249             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7250                 /* not a regular extent, must cow */
7251                 goto out;
7252         }
7253
7254         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7255                 goto out;
7256
7257         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7258         if (extent_end <= offset)
7259                 goto out;
7260
7261         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7262         if (disk_bytenr == 0)
7263                 goto out;
7264
7265         if (btrfs_file_extent_compression(leaf, fi) ||
7266             btrfs_file_extent_encryption(leaf, fi) ||
7267             btrfs_file_extent_other_encoding(leaf, fi))
7268                 goto out;
7269
7270         /*
7271          * Do the same check as in btrfs_cross_ref_exist but without the
7272          * unnecessary search.
7273          */
7274         if (btrfs_file_extent_generation(leaf, fi) <=
7275             btrfs_root_last_snapshot(&root->root_item))
7276                 goto out;
7277
7278         backref_offset = btrfs_file_extent_offset(leaf, fi);
7279
7280         if (orig_start) {
7281                 *orig_start = key.offset - backref_offset;
7282                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7283                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7284         }
7285
7286         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7287                 goto out;
7288
7289         num_bytes = min(offset + *len, extent_end) - offset;
7290         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7291                 u64 range_end;
7292
7293                 range_end = round_up(offset + num_bytes,
7294                                      root->fs_info->sectorsize) - 1;
7295                 ret = test_range_bit(io_tree, offset, range_end,
7296                                      EXTENT_DELALLOC, 0, NULL);
7297                 if (ret) {
7298                         ret = -EAGAIN;
7299                         goto out;
7300                 }
7301         }
7302
7303         btrfs_release_path(path);
7304
7305         /*
7306          * look for other files referencing this extent, if we
7307          * find any we must cow
7308          */
7309
7310         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7311                                     key.offset - backref_offset, disk_bytenr);
7312         if (ret) {
7313                 ret = 0;
7314                 goto out;
7315         }
7316
7317         /*
7318          * adjust disk_bytenr and num_bytes to cover just the bytes
7319          * in this extent we are about to write.  If there
7320          * are any csums in that range we have to cow in order
7321          * to keep the csums correct
7322          */
7323         disk_bytenr += backref_offset;
7324         disk_bytenr += offset - key.offset;
7325         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7326                 goto out;
7327         /*
7328          * all of the above have passed, it is safe to overwrite this extent
7329          * without cow
7330          */
7331         *len = num_bytes;
7332         ret = 1;
7333 out:
7334         btrfs_free_path(path);
7335         return ret;
7336 }
7337
7338 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7339                               struct extent_state **cached_state, int writing)
7340 {
7341         struct btrfs_ordered_extent *ordered;
7342         int ret = 0;
7343
7344         while (1) {
7345                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7346                                  cached_state);
7347                 /*
7348                  * We're concerned with the entire range that we're going to be
7349                  * doing DIO to, so we need to make sure there's no ordered
7350                  * extents in this range.
7351                  */
7352                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7353                                                      lockend - lockstart + 1);
7354
7355                 /*
7356                  * We need to make sure there are no buffered pages in this
7357                  * range either, we could have raced between the invalidate in
7358                  * generic_file_direct_write and locking the extent.  The
7359                  * invalidate needs to happen so that reads after a write do not
7360                  * get stale data.
7361                  */
7362                 if (!ordered &&
7363                     (!writing || !filemap_range_has_page(inode->i_mapping,
7364                                                          lockstart, lockend)))
7365                         break;
7366
7367                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7368                                      cached_state);
7369
7370                 if (ordered) {
7371                         /*
7372                          * If we are doing a DIO read and the ordered extent we
7373                          * found is for a buffered write, we can not wait for it
7374                          * to complete and retry, because if we do so we can
7375                          * deadlock with concurrent buffered writes on page
7376                          * locks. This happens only if our DIO read covers more
7377                          * than one extent map, if at this point has already
7378                          * created an ordered extent for a previous extent map
7379                          * and locked its range in the inode's io tree, and a
7380                          * concurrent write against that previous extent map's
7381                          * range and this range started (we unlock the ranges
7382                          * in the io tree only when the bios complete and
7383                          * buffered writes always lock pages before attempting
7384                          * to lock range in the io tree).
7385                          */
7386                         if (writing ||
7387                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7388                                 btrfs_start_ordered_extent(inode, ordered, 1);
7389                         else
7390                                 ret = -ENOTBLK;
7391                         btrfs_put_ordered_extent(ordered);
7392                 } else {
7393                         /*
7394                          * We could trigger writeback for this range (and wait
7395                          * for it to complete) and then invalidate the pages for
7396                          * this range (through invalidate_inode_pages2_range()),
7397                          * but that can lead us to a deadlock with a concurrent
7398                          * call to readpages() (a buffered read or a defrag call
7399                          * triggered a readahead) on a page lock due to an
7400                          * ordered dio extent we created before but did not have
7401                          * yet a corresponding bio submitted (whence it can not
7402                          * complete), which makes readpages() wait for that
7403                          * ordered extent to complete while holding a lock on
7404                          * that page.
7405                          */
7406                         ret = -ENOTBLK;
7407                 }
7408
7409                 if (ret)
7410                         break;
7411
7412                 cond_resched();
7413         }
7414
7415         return ret;
7416 }
7417
7418 /* The callers of this must take lock_extent() */
7419 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7420                                        u64 orig_start, u64 block_start,
7421                                        u64 block_len, u64 orig_block_len,
7422                                        u64 ram_bytes, int compress_type,
7423                                        int type)
7424 {
7425         struct extent_map_tree *em_tree;
7426         struct extent_map *em;
7427         struct btrfs_root *root = BTRFS_I(inode)->root;
7428         int ret;
7429
7430         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7431                type == BTRFS_ORDERED_COMPRESSED ||
7432                type == BTRFS_ORDERED_NOCOW ||
7433                type == BTRFS_ORDERED_REGULAR);
7434
7435         em_tree = &BTRFS_I(inode)->extent_tree;
7436         em = alloc_extent_map();
7437         if (!em)
7438                 return ERR_PTR(-ENOMEM);
7439
7440         em->start = start;
7441         em->orig_start = orig_start;
7442         em->len = len;
7443         em->block_len = block_len;
7444         em->block_start = block_start;
7445         em->bdev = root->fs_info->fs_devices->latest_bdev;
7446         em->orig_block_len = orig_block_len;
7447         em->ram_bytes = ram_bytes;
7448         em->generation = -1;
7449         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7450         if (type == BTRFS_ORDERED_PREALLOC) {
7451                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7452         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7453                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7454                 em->compress_type = compress_type;
7455         }
7456
7457         do {
7458                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7459                                 em->start + em->len - 1, 0);
7460                 write_lock(&em_tree->lock);
7461                 ret = add_extent_mapping(em_tree, em, 1);
7462                 write_unlock(&em_tree->lock);
7463                 /*
7464                  * The caller has taken lock_extent(), who could race with us
7465                  * to add em?
7466                  */
7467         } while (ret == -EEXIST);
7468
7469         if (ret) {
7470                 free_extent_map(em);
7471                 return ERR_PTR(ret);
7472         }
7473
7474         /* em got 2 refs now, callers needs to do free_extent_map once. */
7475         return em;
7476 }
7477
7478
7479 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7480                                         struct buffer_head *bh_result,
7481                                         struct inode *inode,
7482                                         u64 start, u64 len)
7483 {
7484         if (em->block_start == EXTENT_MAP_HOLE ||
7485                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7486                 return -ENOENT;
7487
7488         len = min(len, em->len - (start - em->start));
7489
7490         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7491                 inode->i_blkbits;
7492         bh_result->b_size = len;
7493         bh_result->b_bdev = em->bdev;
7494         set_buffer_mapped(bh_result);
7495
7496         return 0;
7497 }
7498
7499 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7500                                          struct buffer_head *bh_result,
7501                                          struct inode *inode,
7502                                          struct btrfs_dio_data *dio_data,
7503                                          u64 start, u64 len)
7504 {
7505         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7506         struct extent_map *em = *map;
7507         int ret = 0;
7508
7509         /*
7510          * We don't allocate a new extent in the following cases
7511          *
7512          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7513          * existing extent.
7514          * 2) The extent is marked as PREALLOC. We're good to go here and can
7515          * just use the extent.
7516          *
7517          */
7518         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7519             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7520              em->block_start != EXTENT_MAP_HOLE)) {
7521                 int type;
7522                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7523
7524                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7525                         type = BTRFS_ORDERED_PREALLOC;
7526                 else
7527                         type = BTRFS_ORDERED_NOCOW;
7528                 len = min(len, em->len - (start - em->start));
7529                 block_start = em->block_start + (start - em->start);
7530
7531                 if (can_nocow_extent(inode, start, &len, &orig_start,
7532                                      &orig_block_len, &ram_bytes) == 1 &&
7533                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7534                         struct extent_map *em2;
7535
7536                         em2 = btrfs_create_dio_extent(inode, start, len,
7537                                                       orig_start, block_start,
7538                                                       len, orig_block_len,
7539                                                       ram_bytes, type);
7540                         btrfs_dec_nocow_writers(fs_info, block_start);
7541                         if (type == BTRFS_ORDERED_PREALLOC) {
7542                                 free_extent_map(em);
7543                                 *map = em = em2;
7544                         }
7545
7546                         if (em2 && IS_ERR(em2)) {
7547                                 ret = PTR_ERR(em2);
7548                                 goto out;
7549                         }
7550                         /*
7551                          * For inode marked NODATACOW or extent marked PREALLOC,
7552                          * use the existing or preallocated extent, so does not
7553                          * need to adjust btrfs_space_info's bytes_may_use.
7554                          */
7555                         btrfs_free_reserved_data_space_noquota(inode, start,
7556                                                                len);
7557                         goto skip_cow;
7558                 }
7559         }
7560
7561         /* this will cow the extent */
7562         len = bh_result->b_size;
7563         free_extent_map(em);
7564         *map = em = btrfs_new_extent_direct(inode, start, len);
7565         if (IS_ERR(em)) {
7566                 ret = PTR_ERR(em);
7567                 goto out;
7568         }
7569
7570         len = min(len, em->len - (start - em->start));
7571
7572 skip_cow:
7573         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7574                 inode->i_blkbits;
7575         bh_result->b_size = len;
7576         bh_result->b_bdev = em->bdev;
7577         set_buffer_mapped(bh_result);
7578
7579         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7580                 set_buffer_new(bh_result);
7581
7582         /*
7583          * Need to update the i_size under the extent lock so buffered
7584          * readers will get the updated i_size when we unlock.
7585          */
7586         if (!dio_data->overwrite && start + len > i_size_read(inode))
7587                 i_size_write(inode, start + len);
7588
7589         WARN_ON(dio_data->reserve < len);
7590         dio_data->reserve -= len;
7591         dio_data->unsubmitted_oe_range_end = start + len;
7592         current->journal_info = dio_data;
7593 out:
7594         return ret;
7595 }
7596
7597 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7598                                    struct buffer_head *bh_result, int create)
7599 {
7600         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7601         struct extent_map *em;
7602         struct extent_state *cached_state = NULL;
7603         struct btrfs_dio_data *dio_data = NULL;
7604         u64 start = iblock << inode->i_blkbits;
7605         u64 lockstart, lockend;
7606         u64 len = bh_result->b_size;
7607         int unlock_bits = EXTENT_LOCKED;
7608         int ret = 0;
7609
7610         if (create)
7611                 unlock_bits |= EXTENT_DIRTY;
7612         else
7613                 len = min_t(u64, len, fs_info->sectorsize);
7614
7615         lockstart = start;
7616         lockend = start + len - 1;
7617
7618         if (current->journal_info) {
7619                 /*
7620                  * Need to pull our outstanding extents and set journal_info to NULL so
7621                  * that anything that needs to check if there's a transaction doesn't get
7622                  * confused.
7623                  */
7624                 dio_data = current->journal_info;
7625                 current->journal_info = NULL;
7626         }
7627
7628         /*
7629          * If this errors out it's because we couldn't invalidate pagecache for
7630          * this range and we need to fallback to buffered.
7631          */
7632         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7633                                create)) {
7634                 ret = -ENOTBLK;
7635                 goto err;
7636         }
7637
7638         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7639         if (IS_ERR(em)) {
7640                 ret = PTR_ERR(em);
7641                 goto unlock_err;
7642         }
7643
7644         /*
7645          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7646          * io.  INLINE is special, and we could probably kludge it in here, but
7647          * it's still buffered so for safety lets just fall back to the generic
7648          * buffered path.
7649          *
7650          * For COMPRESSED we _have_ to read the entire extent in so we can
7651          * decompress it, so there will be buffering required no matter what we
7652          * do, so go ahead and fallback to buffered.
7653          *
7654          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7655          * to buffered IO.  Don't blame me, this is the price we pay for using
7656          * the generic code.
7657          */
7658         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7659             em->block_start == EXTENT_MAP_INLINE) {
7660                 free_extent_map(em);
7661                 ret = -ENOTBLK;
7662                 goto unlock_err;
7663         }
7664
7665         if (create) {
7666                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7667                                                     dio_data, start, len);
7668                 if (ret < 0)
7669                         goto unlock_err;
7670
7671                 /* clear and unlock the entire range */
7672                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7673                                  unlock_bits, 1, 0, &cached_state);
7674         } else {
7675                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7676                                                    start, len);
7677                 /* Can be negative only if we read from a hole */
7678                 if (ret < 0) {
7679                         ret = 0;
7680                         free_extent_map(em);
7681                         goto unlock_err;
7682                 }
7683                 /*
7684                  * We need to unlock only the end area that we aren't using.
7685                  * The rest is going to be unlocked by the endio routine.
7686                  */
7687                 lockstart = start + bh_result->b_size;
7688                 if (lockstart < lockend) {
7689                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7690                                          lockend, unlock_bits, 1, 0,
7691                                          &cached_state);
7692                 } else {
7693                         free_extent_state(cached_state);
7694                 }
7695         }
7696
7697         free_extent_map(em);
7698
7699         return 0;
7700
7701 unlock_err:
7702         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7703                          unlock_bits, 1, 0, &cached_state);
7704 err:
7705         if (dio_data)
7706                 current->journal_info = dio_data;
7707         return ret;
7708 }
7709
7710 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7711                                                  struct bio *bio,
7712                                                  int mirror_num)
7713 {
7714         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7715         blk_status_t ret;
7716
7717         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7718
7719         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7720         if (ret)
7721                 return ret;
7722
7723         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7724
7725         return ret;
7726 }
7727
7728 static int btrfs_check_dio_repairable(struct inode *inode,
7729                                       struct bio *failed_bio,
7730                                       struct io_failure_record *failrec,
7731                                       int failed_mirror)
7732 {
7733         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7734         int num_copies;
7735
7736         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7737         if (num_copies == 1) {
7738                 /*
7739                  * we only have a single copy of the data, so don't bother with
7740                  * all the retry and error correction code that follows. no
7741                  * matter what the error is, it is very likely to persist.
7742                  */
7743                 btrfs_debug(fs_info,
7744                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7745                         num_copies, failrec->this_mirror, failed_mirror);
7746                 return 0;
7747         }
7748
7749         failrec->failed_mirror = failed_mirror;
7750         failrec->this_mirror++;
7751         if (failrec->this_mirror == failed_mirror)
7752                 failrec->this_mirror++;
7753
7754         if (failrec->this_mirror > num_copies) {
7755                 btrfs_debug(fs_info,
7756                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7757                         num_copies, failrec->this_mirror, failed_mirror);
7758                 return 0;
7759         }
7760
7761         return 1;
7762 }
7763
7764 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7765                                    struct page *page, unsigned int pgoff,
7766                                    u64 start, u64 end, int failed_mirror,
7767                                    bio_end_io_t *repair_endio, void *repair_arg)
7768 {
7769         struct io_failure_record *failrec;
7770         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7771         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7772         struct bio *bio;
7773         int isector;
7774         unsigned int read_mode = 0;
7775         int segs;
7776         int ret;
7777         blk_status_t status;
7778         struct bio_vec bvec;
7779
7780         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7781
7782         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7783         if (ret)
7784                 return errno_to_blk_status(ret);
7785
7786         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7787                                          failed_mirror);
7788         if (!ret) {
7789                 free_io_failure(failure_tree, io_tree, failrec);
7790                 return BLK_STS_IOERR;
7791         }
7792
7793         segs = bio_segments(failed_bio);
7794         bio_get_first_bvec(failed_bio, &bvec);
7795         if (segs > 1 ||
7796             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7797                 read_mode |= REQ_FAILFAST_DEV;
7798
7799         isector = start - btrfs_io_bio(failed_bio)->logical;
7800         isector >>= inode->i_sb->s_blocksize_bits;
7801         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7802                                 pgoff, isector, repair_endio, repair_arg);
7803         bio->bi_opf = REQ_OP_READ | read_mode;
7804
7805         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7806                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7807                     read_mode, failrec->this_mirror, failrec->in_validation);
7808
7809         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7810         if (status) {
7811                 free_io_failure(failure_tree, io_tree, failrec);
7812                 bio_put(bio);
7813         }
7814
7815         return status;
7816 }
7817
7818 struct btrfs_retry_complete {
7819         struct completion done;
7820         struct inode *inode;
7821         u64 start;
7822         int uptodate;
7823 };
7824
7825 static void btrfs_retry_endio_nocsum(struct bio *bio)
7826 {
7827         struct btrfs_retry_complete *done = bio->bi_private;
7828         struct inode *inode = done->inode;
7829         struct bio_vec *bvec;
7830         struct extent_io_tree *io_tree, *failure_tree;
7831         int i;
7832
7833         if (bio->bi_status)
7834                 goto end;
7835
7836         ASSERT(bio->bi_vcnt == 1);
7837         io_tree = &BTRFS_I(inode)->io_tree;
7838         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7839         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7840
7841         done->uptodate = 1;
7842         ASSERT(!bio_flagged(bio, BIO_CLONED));
7843         bio_for_each_segment_all(bvec, bio, i)
7844                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7845                                  io_tree, done->start, bvec->bv_page,
7846                                  btrfs_ino(BTRFS_I(inode)), 0);
7847 end:
7848         complete(&done->done);
7849         bio_put(bio);
7850 }
7851
7852 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7853                                                 struct btrfs_io_bio *io_bio)
7854 {
7855         struct btrfs_fs_info *fs_info;
7856         struct bio_vec bvec;
7857         struct bvec_iter iter;
7858         struct btrfs_retry_complete done;
7859         u64 start;
7860         unsigned int pgoff;
7861         u32 sectorsize;
7862         int nr_sectors;
7863         blk_status_t ret;
7864         blk_status_t err = BLK_STS_OK;
7865
7866         fs_info = BTRFS_I(inode)->root->fs_info;
7867         sectorsize = fs_info->sectorsize;
7868
7869         start = io_bio->logical;
7870         done.inode = inode;
7871         io_bio->bio.bi_iter = io_bio->iter;
7872
7873         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7874                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7875                 pgoff = bvec.bv_offset;
7876
7877 next_block_or_try_again:
7878                 done.uptodate = 0;
7879                 done.start = start;
7880                 init_completion(&done.done);
7881
7882                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7883                                 pgoff, start, start + sectorsize - 1,
7884                                 io_bio->mirror_num,
7885                                 btrfs_retry_endio_nocsum, &done);
7886                 if (ret) {
7887                         err = ret;
7888                         goto next;
7889                 }
7890
7891                 wait_for_completion_io(&done.done);
7892
7893                 if (!done.uptodate) {
7894                         /* We might have another mirror, so try again */
7895                         goto next_block_or_try_again;
7896                 }
7897
7898 next:
7899                 start += sectorsize;
7900
7901                 nr_sectors--;
7902                 if (nr_sectors) {
7903                         pgoff += sectorsize;
7904                         ASSERT(pgoff < PAGE_SIZE);
7905                         goto next_block_or_try_again;
7906                 }
7907         }
7908
7909         return err;
7910 }
7911
7912 static void btrfs_retry_endio(struct bio *bio)
7913 {
7914         struct btrfs_retry_complete *done = bio->bi_private;
7915         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7916         struct extent_io_tree *io_tree, *failure_tree;
7917         struct inode *inode = done->inode;
7918         struct bio_vec *bvec;
7919         int uptodate;
7920         int ret;
7921         int i;
7922
7923         if (bio->bi_status)
7924                 goto end;
7925
7926         uptodate = 1;
7927
7928         ASSERT(bio->bi_vcnt == 1);
7929         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7930
7931         io_tree = &BTRFS_I(inode)->io_tree;
7932         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7933
7934         ASSERT(!bio_flagged(bio, BIO_CLONED));
7935         bio_for_each_segment_all(bvec, bio, i) {
7936                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7937                                              bvec->bv_offset, done->start,
7938                                              bvec->bv_len);
7939                 if (!ret)
7940                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7941                                          failure_tree, io_tree, done->start,
7942                                          bvec->bv_page,
7943                                          btrfs_ino(BTRFS_I(inode)),
7944                                          bvec->bv_offset);
7945                 else
7946                         uptodate = 0;
7947         }
7948
7949         done->uptodate = uptodate;
7950 end:
7951         complete(&done->done);
7952         bio_put(bio);
7953 }
7954
7955 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7956                 struct btrfs_io_bio *io_bio, blk_status_t err)
7957 {
7958         struct btrfs_fs_info *fs_info;
7959         struct bio_vec bvec;
7960         struct bvec_iter iter;
7961         struct btrfs_retry_complete done;
7962         u64 start;
7963         u64 offset = 0;
7964         u32 sectorsize;
7965         int nr_sectors;
7966         unsigned int pgoff;
7967         int csum_pos;
7968         bool uptodate = (err == 0);
7969         int ret;
7970         blk_status_t status;
7971
7972         fs_info = BTRFS_I(inode)->root->fs_info;
7973         sectorsize = fs_info->sectorsize;
7974
7975         err = BLK_STS_OK;
7976         start = io_bio->logical;
7977         done.inode = inode;
7978         io_bio->bio.bi_iter = io_bio->iter;
7979
7980         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7981                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7982
7983                 pgoff = bvec.bv_offset;
7984 next_block:
7985                 if (uptodate) {
7986                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7987                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7988                                         bvec.bv_page, pgoff, start, sectorsize);
7989                         if (likely(!ret))
7990                                 goto next;
7991                 }
7992 try_again:
7993                 done.uptodate = 0;
7994                 done.start = start;
7995                 init_completion(&done.done);
7996
7997                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7998                                         pgoff, start, start + sectorsize - 1,
7999                                         io_bio->mirror_num, btrfs_retry_endio,
8000                                         &done);
8001                 if (status) {
8002                         err = status;
8003                         goto next;
8004                 }
8005
8006                 wait_for_completion_io(&done.done);
8007
8008                 if (!done.uptodate) {
8009                         /* We might have another mirror, so try again */
8010                         goto try_again;
8011                 }
8012 next:
8013                 offset += sectorsize;
8014                 start += sectorsize;
8015
8016                 ASSERT(nr_sectors);
8017
8018                 nr_sectors--;
8019                 if (nr_sectors) {
8020                         pgoff += sectorsize;
8021                         ASSERT(pgoff < PAGE_SIZE);
8022                         goto next_block;
8023                 }
8024         }
8025
8026         return err;
8027 }
8028
8029 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8030                 struct btrfs_io_bio *io_bio, blk_status_t err)
8031 {
8032         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8033
8034         if (skip_csum) {
8035                 if (unlikely(err))
8036                         return __btrfs_correct_data_nocsum(inode, io_bio);
8037                 else
8038                         return BLK_STS_OK;
8039         } else {
8040                 return __btrfs_subio_endio_read(inode, io_bio, err);
8041         }
8042 }
8043
8044 static void btrfs_endio_direct_read(struct bio *bio)
8045 {
8046         struct btrfs_dio_private *dip = bio->bi_private;
8047         struct inode *inode = dip->inode;
8048         struct bio *dio_bio;
8049         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8050         blk_status_t err = bio->bi_status;
8051
8052         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8053                 err = btrfs_subio_endio_read(inode, io_bio, err);
8054
8055         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8056                       dip->logical_offset + dip->bytes - 1);
8057         dio_bio = dip->dio_bio;
8058
8059         kfree(dip);
8060
8061         dio_bio->bi_status = err;
8062         dio_end_io(dio_bio);
8063         btrfs_io_bio_free_csum(io_bio);
8064         bio_put(bio);
8065 }
8066
8067 static void __endio_write_update_ordered(struct inode *inode,
8068                                          const u64 offset, const u64 bytes,
8069                                          const bool uptodate)
8070 {
8071         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8072         struct btrfs_ordered_extent *ordered = NULL;
8073         struct btrfs_workqueue *wq;
8074         btrfs_work_func_t func;
8075         u64 ordered_offset = offset;
8076         u64 ordered_bytes = bytes;
8077         u64 last_offset;
8078
8079         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8080                 wq = fs_info->endio_freespace_worker;
8081                 func = btrfs_freespace_write_helper;
8082         } else {
8083                 wq = fs_info->endio_write_workers;
8084                 func = btrfs_endio_write_helper;
8085         }
8086
8087         while (ordered_offset < offset + bytes) {
8088                 last_offset = ordered_offset;
8089                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8090                                                            &ordered_offset,
8091                                                            ordered_bytes,
8092                                                            uptodate)) {
8093                         btrfs_init_work(&ordered->work, func,
8094                                         finish_ordered_fn,
8095                                         NULL, NULL);
8096                         btrfs_queue_work(wq, &ordered->work);
8097                 }
8098                 /*
8099                  * If btrfs_dec_test_ordered_pending does not find any ordered
8100                  * extent in the range, we can exit.
8101                  */
8102                 if (ordered_offset == last_offset)
8103                         return;
8104                 /*
8105                  * Our bio might span multiple ordered extents. In this case
8106                  * we keep going until we have accounted the whole dio.
8107                  */
8108                 if (ordered_offset < offset + bytes) {
8109                         ordered_bytes = offset + bytes - ordered_offset;
8110                         ordered = NULL;
8111                 }
8112         }
8113 }
8114
8115 static void btrfs_endio_direct_write(struct bio *bio)
8116 {
8117         struct btrfs_dio_private *dip = bio->bi_private;
8118         struct bio *dio_bio = dip->dio_bio;
8119
8120         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8121                                      dip->bytes, !bio->bi_status);
8122
8123         kfree(dip);
8124
8125         dio_bio->bi_status = bio->bi_status;
8126         dio_end_io(dio_bio);
8127         bio_put(bio);
8128 }
8129
8130 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8131                                     struct bio *bio, u64 offset)
8132 {
8133         struct inode *inode = private_data;
8134         blk_status_t ret;
8135         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8136         BUG_ON(ret); /* -ENOMEM */
8137         return 0;
8138 }
8139
8140 static void btrfs_end_dio_bio(struct bio *bio)
8141 {
8142         struct btrfs_dio_private *dip = bio->bi_private;
8143         blk_status_t err = bio->bi_status;
8144
8145         if (err)
8146                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8147                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8148                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8149                            bio->bi_opf,
8150                            (unsigned long long)bio->bi_iter.bi_sector,
8151                            bio->bi_iter.bi_size, err);
8152
8153         if (dip->subio_endio)
8154                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8155
8156         if (err) {
8157                 /*
8158                  * We want to perceive the errors flag being set before
8159                  * decrementing the reference count. We don't need a barrier
8160                  * since atomic operations with a return value are fully
8161                  * ordered as per atomic_t.txt
8162                  */
8163                 dip->errors = 1;
8164         }
8165
8166         /* if there are more bios still pending for this dio, just exit */
8167         if (!atomic_dec_and_test(&dip->pending_bios))
8168                 goto out;
8169
8170         if (dip->errors) {
8171                 bio_io_error(dip->orig_bio);
8172         } else {
8173                 dip->dio_bio->bi_status = BLK_STS_OK;
8174                 bio_endio(dip->orig_bio);
8175         }
8176 out:
8177         bio_put(bio);
8178 }
8179
8180 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8181                                                  struct btrfs_dio_private *dip,
8182                                                  struct bio *bio,
8183                                                  u64 file_offset)
8184 {
8185         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8186         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8187         blk_status_t ret;
8188
8189         /*
8190          * We load all the csum data we need when we submit
8191          * the first bio to reduce the csum tree search and
8192          * contention.
8193          */
8194         if (dip->logical_offset == file_offset) {
8195                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8196                                                 file_offset);
8197                 if (ret)
8198                         return ret;
8199         }
8200
8201         if (bio == dip->orig_bio)
8202                 return 0;
8203
8204         file_offset -= dip->logical_offset;
8205         file_offset >>= inode->i_sb->s_blocksize_bits;
8206         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8207
8208         return 0;
8209 }
8210
8211 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8212                 struct inode *inode, u64 file_offset, int async_submit)
8213 {
8214         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8215         struct btrfs_dio_private *dip = bio->bi_private;
8216         bool write = bio_op(bio) == REQ_OP_WRITE;
8217         blk_status_t ret;
8218
8219         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8220         if (async_submit)
8221                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8222
8223         if (!write) {
8224                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8225                 if (ret)
8226                         goto err;
8227         }
8228
8229         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8230                 goto map;
8231
8232         if (write && async_submit) {
8233                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8234                                           file_offset, inode,
8235                                           btrfs_submit_bio_start_direct_io);
8236                 goto err;
8237         } else if (write) {
8238                 /*
8239                  * If we aren't doing async submit, calculate the csum of the
8240                  * bio now.
8241                  */
8242                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8243                 if (ret)
8244                         goto err;
8245         } else {
8246                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8247                                                      file_offset);
8248                 if (ret)
8249                         goto err;
8250         }
8251 map:
8252         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8253 err:
8254         return ret;
8255 }
8256
8257 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8258 {
8259         struct inode *inode = dip->inode;
8260         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8261         struct bio *bio;
8262         struct bio *orig_bio = dip->orig_bio;
8263         u64 start_sector = orig_bio->bi_iter.bi_sector;
8264         u64 file_offset = dip->logical_offset;
8265         u64 map_length;
8266         int async_submit = 0;
8267         u64 submit_len;
8268         int clone_offset = 0;
8269         int clone_len;
8270         int ret;
8271         blk_status_t status;
8272
8273         map_length = orig_bio->bi_iter.bi_size;
8274         submit_len = map_length;
8275         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8276                               &map_length, NULL, 0);
8277         if (ret)
8278                 return -EIO;
8279
8280         if (map_length >= submit_len) {
8281                 bio = orig_bio;
8282                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8283                 goto submit;
8284         }
8285
8286         /* async crcs make it difficult to collect full stripe writes. */
8287         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8288                 async_submit = 0;
8289         else
8290                 async_submit = 1;
8291
8292         /* bio split */
8293         ASSERT(map_length <= INT_MAX);
8294         atomic_inc(&dip->pending_bios);
8295         do {
8296                 clone_len = min_t(int, submit_len, map_length);
8297
8298                 /*
8299                  * This will never fail as it's passing GPF_NOFS and
8300                  * the allocation is backed by btrfs_bioset.
8301                  */
8302                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8303                                               clone_len);
8304                 bio->bi_private = dip;
8305                 bio->bi_end_io = btrfs_end_dio_bio;
8306                 btrfs_io_bio(bio)->logical = file_offset;
8307
8308                 ASSERT(submit_len >= clone_len);
8309                 submit_len -= clone_len;
8310                 if (submit_len == 0)
8311                         break;
8312
8313                 /*
8314                  * Increase the count before we submit the bio so we know
8315                  * the end IO handler won't happen before we increase the
8316                  * count. Otherwise, the dip might get freed before we're
8317                  * done setting it up.
8318                  */
8319                 atomic_inc(&dip->pending_bios);
8320
8321                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8322                                                 async_submit);
8323                 if (status) {
8324                         bio_put(bio);
8325                         atomic_dec(&dip->pending_bios);
8326                         goto out_err;
8327                 }
8328
8329                 clone_offset += clone_len;
8330                 start_sector += clone_len >> 9;
8331                 file_offset += clone_len;
8332
8333                 map_length = submit_len;
8334                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8335                                       start_sector << 9, &map_length, NULL, 0);
8336                 if (ret)
8337                         goto out_err;
8338         } while (submit_len > 0);
8339
8340 submit:
8341         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8342         if (!status)
8343                 return 0;
8344
8345         bio_put(bio);
8346 out_err:
8347         dip->errors = 1;
8348         /*
8349          * Before atomic variable goto zero, we must  make sure dip->errors is
8350          * perceived to be set. This ordering is ensured by the fact that an
8351          * atomic operations with a return value are fully ordered as per
8352          * atomic_t.txt
8353          */
8354         if (atomic_dec_and_test(&dip->pending_bios))
8355                 bio_io_error(dip->orig_bio);
8356
8357         /* bio_end_io() will handle error, so we needn't return it */
8358         return 0;
8359 }
8360
8361 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8362                                 loff_t file_offset)
8363 {
8364         struct btrfs_dio_private *dip = NULL;
8365         struct bio *bio = NULL;
8366         struct btrfs_io_bio *io_bio;
8367         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8368         int ret = 0;
8369
8370         bio = btrfs_bio_clone(dio_bio);
8371
8372         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8373         if (!dip) {
8374                 ret = -ENOMEM;
8375                 goto free_ordered;
8376         }
8377
8378         dip->private = dio_bio->bi_private;
8379         dip->inode = inode;
8380         dip->logical_offset = file_offset;
8381         dip->bytes = dio_bio->bi_iter.bi_size;
8382         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8383         bio->bi_private = dip;
8384         dip->orig_bio = bio;
8385         dip->dio_bio = dio_bio;
8386         atomic_set(&dip->pending_bios, 0);
8387         io_bio = btrfs_io_bio(bio);
8388         io_bio->logical = file_offset;
8389
8390         if (write) {
8391                 bio->bi_end_io = btrfs_endio_direct_write;
8392         } else {
8393                 bio->bi_end_io = btrfs_endio_direct_read;
8394                 dip->subio_endio = btrfs_subio_endio_read;
8395         }
8396
8397         /*
8398          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8399          * even if we fail to submit a bio, because in such case we do the
8400          * corresponding error handling below and it must not be done a second
8401          * time by btrfs_direct_IO().
8402          */
8403         if (write) {
8404                 struct btrfs_dio_data *dio_data = current->journal_info;
8405
8406                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8407                         dip->bytes;
8408                 dio_data->unsubmitted_oe_range_start =
8409                         dio_data->unsubmitted_oe_range_end;
8410         }
8411
8412         ret = btrfs_submit_direct_hook(dip);
8413         if (!ret)
8414                 return;
8415
8416         btrfs_io_bio_free_csum(io_bio);
8417
8418 free_ordered:
8419         /*
8420          * If we arrived here it means either we failed to submit the dip
8421          * or we either failed to clone the dio_bio or failed to allocate the
8422          * dip. If we cloned the dio_bio and allocated the dip, we can just
8423          * call bio_endio against our io_bio so that we get proper resource
8424          * cleanup if we fail to submit the dip, otherwise, we must do the
8425          * same as btrfs_endio_direct_[write|read] because we can't call these
8426          * callbacks - they require an allocated dip and a clone of dio_bio.
8427          */
8428         if (bio && dip) {
8429                 bio_io_error(bio);
8430                 /*
8431                  * The end io callbacks free our dip, do the final put on bio
8432                  * and all the cleanup and final put for dio_bio (through
8433                  * dio_end_io()).
8434                  */
8435                 dip = NULL;
8436                 bio = NULL;
8437         } else {
8438                 if (write)
8439                         __endio_write_update_ordered(inode,
8440                                                 file_offset,
8441                                                 dio_bio->bi_iter.bi_size,
8442                                                 false);
8443                 else
8444                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8445                               file_offset + dio_bio->bi_iter.bi_size - 1);
8446
8447                 dio_bio->bi_status = BLK_STS_IOERR;
8448                 /*
8449                  * Releases and cleans up our dio_bio, no need to bio_put()
8450                  * nor bio_endio()/bio_io_error() against dio_bio.
8451                  */
8452                 dio_end_io(dio_bio);
8453         }
8454         if (bio)
8455                 bio_put(bio);
8456         kfree(dip);
8457 }
8458
8459 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8460                                const struct iov_iter *iter, loff_t offset)
8461 {
8462         int seg;
8463         int i;
8464         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8465         ssize_t retval = -EINVAL;
8466
8467         if (offset & blocksize_mask)
8468                 goto out;
8469
8470         if (iov_iter_alignment(iter) & blocksize_mask)
8471                 goto out;
8472
8473         /* If this is a write we don't need to check anymore */
8474         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8475                 return 0;
8476         /*
8477          * Check to make sure we don't have duplicate iov_base's in this
8478          * iovec, if so return EINVAL, otherwise we'll get csum errors
8479          * when reading back.
8480          */
8481         for (seg = 0; seg < iter->nr_segs; seg++) {
8482                 for (i = seg + 1; i < iter->nr_segs; i++) {
8483                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8484                                 goto out;
8485                 }
8486         }
8487         retval = 0;
8488 out:
8489         return retval;
8490 }
8491
8492 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8493 {
8494         struct file *file = iocb->ki_filp;
8495         struct inode *inode = file->f_mapping->host;
8496         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8497         struct btrfs_dio_data dio_data = { 0 };
8498         struct extent_changeset *data_reserved = NULL;
8499         loff_t offset = iocb->ki_pos;
8500         size_t count = 0;
8501         int flags = 0;
8502         bool wakeup = true;
8503         bool relock = false;
8504         ssize_t ret;
8505
8506         if (check_direct_IO(fs_info, iter, offset))
8507                 return 0;
8508
8509         inode_dio_begin(inode);
8510
8511         /*
8512          * The generic stuff only does filemap_write_and_wait_range, which
8513          * isn't enough if we've written compressed pages to this area, so
8514          * we need to flush the dirty pages again to make absolutely sure
8515          * that any outstanding dirty pages are on disk.
8516          */
8517         count = iov_iter_count(iter);
8518         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8519                      &BTRFS_I(inode)->runtime_flags))
8520                 filemap_fdatawrite_range(inode->i_mapping, offset,
8521                                          offset + count - 1);
8522
8523         if (iov_iter_rw(iter) == WRITE) {
8524                 /*
8525                  * If the write DIO is beyond the EOF, we need update
8526                  * the isize, but it is protected by i_mutex. So we can
8527                  * not unlock the i_mutex at this case.
8528                  */
8529                 if (offset + count <= inode->i_size) {
8530                         dio_data.overwrite = 1;
8531                         inode_unlock(inode);
8532                         relock = true;
8533                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8534                         ret = -EAGAIN;
8535                         goto out;
8536                 }
8537                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8538                                                    offset, count);
8539                 if (ret)
8540                         goto out;
8541
8542                 /*
8543                  * We need to know how many extents we reserved so that we can
8544                  * do the accounting properly if we go over the number we
8545                  * originally calculated.  Abuse current->journal_info for this.
8546                  */
8547                 dio_data.reserve = round_up(count,
8548                                             fs_info->sectorsize);
8549                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8550                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8551                 current->journal_info = &dio_data;
8552                 down_read(&BTRFS_I(inode)->dio_sem);
8553         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8554                                      &BTRFS_I(inode)->runtime_flags)) {
8555                 inode_dio_end(inode);
8556                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8557                 wakeup = false;
8558         }
8559
8560         ret = __blockdev_direct_IO(iocb, inode,
8561                                    fs_info->fs_devices->latest_bdev,
8562                                    iter, btrfs_get_blocks_direct, NULL,
8563                                    btrfs_submit_direct, flags);
8564         if (iov_iter_rw(iter) == WRITE) {
8565                 up_read(&BTRFS_I(inode)->dio_sem);
8566                 current->journal_info = NULL;
8567                 if (ret < 0 && ret != -EIOCBQUEUED) {
8568                         if (dio_data.reserve)
8569                                 btrfs_delalloc_release_space(inode, data_reserved,
8570                                         offset, dio_data.reserve, true);
8571                         /*
8572                          * On error we might have left some ordered extents
8573                          * without submitting corresponding bios for them, so
8574                          * cleanup them up to avoid other tasks getting them
8575                          * and waiting for them to complete forever.
8576                          */
8577                         if (dio_data.unsubmitted_oe_range_start <
8578                             dio_data.unsubmitted_oe_range_end)
8579                                 __endio_write_update_ordered(inode,
8580                                         dio_data.unsubmitted_oe_range_start,
8581                                         dio_data.unsubmitted_oe_range_end -
8582                                         dio_data.unsubmitted_oe_range_start,
8583                                         false);
8584                 } else if (ret >= 0 && (size_t)ret < count)
8585                         btrfs_delalloc_release_space(inode, data_reserved,
8586                                         offset, count - (size_t)ret, true);
8587                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8588         }
8589 out:
8590         if (wakeup)
8591                 inode_dio_end(inode);
8592         if (relock)
8593                 inode_lock(inode);
8594
8595         extent_changeset_free(data_reserved);
8596         return ret;
8597 }
8598
8599 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8600
8601 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8602                 __u64 start, __u64 len)
8603 {
8604         int     ret;
8605
8606         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8607         if (ret)
8608                 return ret;
8609
8610         return extent_fiemap(inode, fieinfo, start, len);
8611 }
8612
8613 int btrfs_readpage(struct file *file, struct page *page)
8614 {
8615         struct extent_io_tree *tree;
8616         tree = &BTRFS_I(page->mapping->host)->io_tree;
8617         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8618 }
8619
8620 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8621 {
8622         struct inode *inode = page->mapping->host;
8623         int ret;
8624
8625         if (current->flags & PF_MEMALLOC) {
8626                 redirty_page_for_writepage(wbc, page);
8627                 unlock_page(page);
8628                 return 0;
8629         }
8630
8631         /*
8632          * If we are under memory pressure we will call this directly from the
8633          * VM, we need to make sure we have the inode referenced for the ordered
8634          * extent.  If not just return like we didn't do anything.
8635          */
8636         if (!igrab(inode)) {
8637                 redirty_page_for_writepage(wbc, page);
8638                 return AOP_WRITEPAGE_ACTIVATE;
8639         }
8640         ret = extent_write_full_page(page, wbc);
8641         btrfs_add_delayed_iput(inode);
8642         return ret;
8643 }
8644
8645 static int btrfs_writepages(struct address_space *mapping,
8646                             struct writeback_control *wbc)
8647 {
8648         return extent_writepages(mapping, wbc);
8649 }
8650
8651 static int
8652 btrfs_readpages(struct file *file, struct address_space *mapping,
8653                 struct list_head *pages, unsigned nr_pages)
8654 {
8655         return extent_readpages(mapping, pages, nr_pages);
8656 }
8657
8658 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8659 {
8660         int ret = try_release_extent_mapping(page, gfp_flags);
8661         if (ret == 1) {
8662                 ClearPagePrivate(page);
8663                 set_page_private(page, 0);
8664                 put_page(page);
8665         }
8666         return ret;
8667 }
8668
8669 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8670 {
8671         if (PageWriteback(page) || PageDirty(page))
8672                 return 0;
8673         return __btrfs_releasepage(page, gfp_flags);
8674 }
8675
8676 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8677                                  unsigned int length)
8678 {
8679         struct inode *inode = page->mapping->host;
8680         struct extent_io_tree *tree;
8681         struct btrfs_ordered_extent *ordered;
8682         struct extent_state *cached_state = NULL;
8683         u64 page_start = page_offset(page);
8684         u64 page_end = page_start + PAGE_SIZE - 1;
8685         u64 start;
8686         u64 end;
8687         int inode_evicting = inode->i_state & I_FREEING;
8688
8689         /*
8690          * we have the page locked, so new writeback can't start,
8691          * and the dirty bit won't be cleared while we are here.
8692          *
8693          * Wait for IO on this page so that we can safely clear
8694          * the PagePrivate2 bit and do ordered accounting
8695          */
8696         wait_on_page_writeback(page);
8697
8698         tree = &BTRFS_I(inode)->io_tree;
8699         if (offset) {
8700                 btrfs_releasepage(page, GFP_NOFS);
8701                 return;
8702         }
8703
8704         if (!inode_evicting)
8705                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8706 again:
8707         start = page_start;
8708         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8709                                         page_end - start + 1);
8710         if (ordered) {
8711                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8712                 /*
8713                  * IO on this page will never be started, so we need
8714                  * to account for any ordered extents now
8715                  */
8716                 if (!inode_evicting)
8717                         clear_extent_bit(tree, start, end,
8718                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8719                                          EXTENT_DELALLOC_NEW |
8720                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8721                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8722                 /*
8723                  * whoever cleared the private bit is responsible
8724                  * for the finish_ordered_io
8725                  */
8726                 if (TestClearPagePrivate2(page)) {
8727                         struct btrfs_ordered_inode_tree *tree;
8728                         u64 new_len;
8729
8730                         tree = &BTRFS_I(inode)->ordered_tree;
8731
8732                         spin_lock_irq(&tree->lock);
8733                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8734                         new_len = start - ordered->file_offset;
8735                         if (new_len < ordered->truncated_len)
8736                                 ordered->truncated_len = new_len;
8737                         spin_unlock_irq(&tree->lock);
8738
8739                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8740                                                            start,
8741                                                            end - start + 1, 1))
8742                                 btrfs_finish_ordered_io(ordered);
8743                 }
8744                 btrfs_put_ordered_extent(ordered);
8745                 if (!inode_evicting) {
8746                         cached_state = NULL;
8747                         lock_extent_bits(tree, start, end,
8748                                          &cached_state);
8749                 }
8750
8751                 start = end + 1;
8752                 if (start < page_end)
8753                         goto again;
8754         }
8755
8756         /*
8757          * Qgroup reserved space handler
8758          * Page here will be either
8759          * 1) Already written to disk
8760          *    In this case, its reserved space is released from data rsv map
8761          *    and will be freed by delayed_ref handler finally.
8762          *    So even we call qgroup_free_data(), it won't decrease reserved
8763          *    space.
8764          * 2) Not written to disk
8765          *    This means the reserved space should be freed here. However,
8766          *    if a truncate invalidates the page (by clearing PageDirty)
8767          *    and the page is accounted for while allocating extent
8768          *    in btrfs_check_data_free_space() we let delayed_ref to
8769          *    free the entire extent.
8770          */
8771         if (PageDirty(page))
8772                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8773         if (!inode_evicting) {
8774                 clear_extent_bit(tree, page_start, page_end,
8775                                  EXTENT_LOCKED | EXTENT_DIRTY |
8776                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8777                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8778                                  &cached_state);
8779
8780                 __btrfs_releasepage(page, GFP_NOFS);
8781         }
8782
8783         ClearPageChecked(page);
8784         if (PagePrivate(page)) {
8785                 ClearPagePrivate(page);
8786                 set_page_private(page, 0);
8787                 put_page(page);
8788         }
8789 }
8790
8791 /*
8792  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8793  * called from a page fault handler when a page is first dirtied. Hence we must
8794  * be careful to check for EOF conditions here. We set the page up correctly
8795  * for a written page which means we get ENOSPC checking when writing into
8796  * holes and correct delalloc and unwritten extent mapping on filesystems that
8797  * support these features.
8798  *
8799  * We are not allowed to take the i_mutex here so we have to play games to
8800  * protect against truncate races as the page could now be beyond EOF.  Because
8801  * truncate_setsize() writes the inode size before removing pages, once we have
8802  * the page lock we can determine safely if the page is beyond EOF. If it is not
8803  * beyond EOF, then the page is guaranteed safe against truncation until we
8804  * unlock the page.
8805  */
8806 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8807 {
8808         struct page *page = vmf->page;
8809         struct inode *inode = file_inode(vmf->vma->vm_file);
8810         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8811         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8812         struct btrfs_ordered_extent *ordered;
8813         struct extent_state *cached_state = NULL;
8814         struct extent_changeset *data_reserved = NULL;
8815         char *kaddr;
8816         unsigned long zero_start;
8817         loff_t size;
8818         vm_fault_t ret;
8819         int ret2;
8820         int reserved = 0;
8821         u64 reserved_space;
8822         u64 page_start;
8823         u64 page_end;
8824         u64 end;
8825
8826         reserved_space = PAGE_SIZE;
8827
8828         sb_start_pagefault(inode->i_sb);
8829         page_start = page_offset(page);
8830         page_end = page_start + PAGE_SIZE - 1;
8831         end = page_end;
8832
8833         /*
8834          * Reserving delalloc space after obtaining the page lock can lead to
8835          * deadlock. For example, if a dirty page is locked by this function
8836          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8837          * dirty page write out, then the btrfs_writepage() function could
8838          * end up waiting indefinitely to get a lock on the page currently
8839          * being processed by btrfs_page_mkwrite() function.
8840          */
8841         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8842                                            reserved_space);
8843         if (!ret2) {
8844                 ret2 = file_update_time(vmf->vma->vm_file);
8845                 reserved = 1;
8846         }
8847         if (ret2) {
8848                 ret = vmf_error(ret2);
8849                 if (reserved)
8850                         goto out;
8851                 goto out_noreserve;
8852         }
8853
8854         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8855 again:
8856         lock_page(page);
8857         size = i_size_read(inode);
8858
8859         if ((page->mapping != inode->i_mapping) ||
8860             (page_start >= size)) {
8861                 /* page got truncated out from underneath us */
8862                 goto out_unlock;
8863         }
8864         wait_on_page_writeback(page);
8865
8866         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8867         set_page_extent_mapped(page);
8868
8869         /*
8870          * we can't set the delalloc bits if there are pending ordered
8871          * extents.  Drop our locks and wait for them to finish
8872          */
8873         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8874                         PAGE_SIZE);
8875         if (ordered) {
8876                 unlock_extent_cached(io_tree, page_start, page_end,
8877                                      &cached_state);
8878                 unlock_page(page);
8879                 btrfs_start_ordered_extent(inode, ordered, 1);
8880                 btrfs_put_ordered_extent(ordered);
8881                 goto again;
8882         }
8883
8884         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8885                 reserved_space = round_up(size - page_start,
8886                                           fs_info->sectorsize);
8887                 if (reserved_space < PAGE_SIZE) {
8888                         end = page_start + reserved_space - 1;
8889                         btrfs_delalloc_release_space(inode, data_reserved,
8890                                         page_start, PAGE_SIZE - reserved_space,
8891                                         true);
8892                 }
8893         }
8894
8895         /*
8896          * page_mkwrite gets called when the page is firstly dirtied after it's
8897          * faulted in, but write(2) could also dirty a page and set delalloc
8898          * bits, thus in this case for space account reason, we still need to
8899          * clear any delalloc bits within this page range since we have to
8900          * reserve data&meta space before lock_page() (see above comments).
8901          */
8902         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8903                           EXTENT_DIRTY | EXTENT_DELALLOC |
8904                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8905                           0, 0, &cached_state);
8906
8907         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8908                                         &cached_state, 0);
8909         if (ret2) {
8910                 unlock_extent_cached(io_tree, page_start, page_end,
8911                                      &cached_state);
8912                 ret = VM_FAULT_SIGBUS;
8913                 goto out_unlock;
8914         }
8915         ret2 = 0;
8916
8917         /* page is wholly or partially inside EOF */
8918         if (page_start + PAGE_SIZE > size)
8919                 zero_start = offset_in_page(size);
8920         else
8921                 zero_start = PAGE_SIZE;
8922
8923         if (zero_start != PAGE_SIZE) {
8924                 kaddr = kmap(page);
8925                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8926                 flush_dcache_page(page);
8927                 kunmap(page);
8928         }
8929         ClearPageChecked(page);
8930         set_page_dirty(page);
8931         SetPageUptodate(page);
8932
8933         BTRFS_I(inode)->last_trans = fs_info->generation;
8934         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8935         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8936
8937         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8938
8939         if (!ret2) {
8940                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
8941                 sb_end_pagefault(inode->i_sb);
8942                 extent_changeset_free(data_reserved);
8943                 return VM_FAULT_LOCKED;
8944         }
8945
8946 out_unlock:
8947         unlock_page(page);
8948 out:
8949         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
8950         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8951                                      reserved_space, (ret != 0));
8952 out_noreserve:
8953         sb_end_pagefault(inode->i_sb);
8954         extent_changeset_free(data_reserved);
8955         return ret;
8956 }
8957
8958 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8959 {
8960         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8961         struct btrfs_root *root = BTRFS_I(inode)->root;
8962         struct btrfs_block_rsv *rsv;
8963         int ret;
8964         struct btrfs_trans_handle *trans;
8965         u64 mask = fs_info->sectorsize - 1;
8966         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
8967
8968         if (!skip_writeback) {
8969                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8970                                                (u64)-1);
8971                 if (ret)
8972                         return ret;
8973         }
8974
8975         /*
8976          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8977          * things going on here:
8978          *
8979          * 1) We need to reserve space to update our inode.
8980          *
8981          * 2) We need to have something to cache all the space that is going to
8982          * be free'd up by the truncate operation, but also have some slack
8983          * space reserved in case it uses space during the truncate (thank you
8984          * very much snapshotting).
8985          *
8986          * And we need these to be separate.  The fact is we can use a lot of
8987          * space doing the truncate, and we have no earthly idea how much space
8988          * we will use, so we need the truncate reservation to be separate so it
8989          * doesn't end up using space reserved for updating the inode.  We also
8990          * need to be able to stop the transaction and start a new one, which
8991          * means we need to be able to update the inode several times, and we
8992          * have no idea of knowing how many times that will be, so we can't just
8993          * reserve 1 item for the entirety of the operation, so that has to be
8994          * done separately as well.
8995          *
8996          * So that leaves us with
8997          *
8998          * 1) rsv - for the truncate reservation, which we will steal from the
8999          * transaction reservation.
9000          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9001          * updating the inode.
9002          */
9003         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9004         if (!rsv)
9005                 return -ENOMEM;
9006         rsv->size = min_size;
9007         rsv->failfast = 1;
9008
9009         /*
9010          * 1 for the truncate slack space
9011          * 1 for updating the inode.
9012          */
9013         trans = btrfs_start_transaction(root, 2);
9014         if (IS_ERR(trans)) {
9015                 ret = PTR_ERR(trans);
9016                 goto out;
9017         }
9018
9019         /* Migrate the slack space for the truncate to our reserve */
9020         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9021                                       min_size, false);
9022         BUG_ON(ret);
9023
9024         /*
9025          * So if we truncate and then write and fsync we normally would just
9026          * write the extents that changed, which is a problem if we need to
9027          * first truncate that entire inode.  So set this flag so we write out
9028          * all of the extents in the inode to the sync log so we're completely
9029          * safe.
9030          */
9031         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9032         trans->block_rsv = rsv;
9033
9034         while (1) {
9035                 ret = btrfs_truncate_inode_items(trans, root, inode,
9036                                                  inode->i_size,
9037                                                  BTRFS_EXTENT_DATA_KEY);
9038                 trans->block_rsv = &fs_info->trans_block_rsv;
9039                 if (ret != -ENOSPC && ret != -EAGAIN)
9040                         break;
9041
9042                 ret = btrfs_update_inode(trans, root, inode);
9043                 if (ret)
9044                         break;
9045
9046                 btrfs_end_transaction(trans);
9047                 btrfs_btree_balance_dirty(fs_info);
9048
9049                 trans = btrfs_start_transaction(root, 2);
9050                 if (IS_ERR(trans)) {
9051                         ret = PTR_ERR(trans);
9052                         trans = NULL;
9053                         break;
9054                 }
9055
9056                 btrfs_block_rsv_release(fs_info, rsv, -1);
9057                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9058                                               rsv, min_size, false);
9059                 BUG_ON(ret);    /* shouldn't happen */
9060                 trans->block_rsv = rsv;
9061         }
9062
9063         /*
9064          * We can't call btrfs_truncate_block inside a trans handle as we could
9065          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9066          * we've truncated everything except the last little bit, and can do
9067          * btrfs_truncate_block and then update the disk_i_size.
9068          */
9069         if (ret == NEED_TRUNCATE_BLOCK) {
9070                 btrfs_end_transaction(trans);
9071                 btrfs_btree_balance_dirty(fs_info);
9072
9073                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9074                 if (ret)
9075                         goto out;
9076                 trans = btrfs_start_transaction(root, 1);
9077                 if (IS_ERR(trans)) {
9078                         ret = PTR_ERR(trans);
9079                         goto out;
9080                 }
9081                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9082         }
9083
9084         if (trans) {
9085                 int ret2;
9086
9087                 trans->block_rsv = &fs_info->trans_block_rsv;
9088                 ret2 = btrfs_update_inode(trans, root, inode);
9089                 if (ret2 && !ret)
9090                         ret = ret2;
9091
9092                 ret2 = btrfs_end_transaction(trans);
9093                 if (ret2 && !ret)
9094                         ret = ret2;
9095                 btrfs_btree_balance_dirty(fs_info);
9096         }
9097 out:
9098         btrfs_free_block_rsv(fs_info, rsv);
9099
9100         return ret;
9101 }
9102
9103 /*
9104  * create a new subvolume directory/inode (helper for the ioctl).
9105  */
9106 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9107                              struct btrfs_root *new_root,
9108                              struct btrfs_root *parent_root,
9109                              u64 new_dirid)
9110 {
9111         struct inode *inode;
9112         int err;
9113         u64 index = 0;
9114
9115         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9116                                 new_dirid, new_dirid,
9117                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9118                                 &index);
9119         if (IS_ERR(inode))
9120                 return PTR_ERR(inode);
9121         inode->i_op = &btrfs_dir_inode_operations;
9122         inode->i_fop = &btrfs_dir_file_operations;
9123
9124         set_nlink(inode, 1);
9125         btrfs_i_size_write(BTRFS_I(inode), 0);
9126         unlock_new_inode(inode);
9127
9128         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9129         if (err)
9130                 btrfs_err(new_root->fs_info,
9131                           "error inheriting subvolume %llu properties: %d",
9132                           new_root->root_key.objectid, err);
9133
9134         err = btrfs_update_inode(trans, new_root, inode);
9135
9136         iput(inode);
9137         return err;
9138 }
9139
9140 struct inode *btrfs_alloc_inode(struct super_block *sb)
9141 {
9142         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9143         struct btrfs_inode *ei;
9144         struct inode *inode;
9145
9146         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9147         if (!ei)
9148                 return NULL;
9149
9150         ei->root = NULL;
9151         ei->generation = 0;
9152         ei->last_trans = 0;
9153         ei->last_sub_trans = 0;
9154         ei->logged_trans = 0;
9155         ei->delalloc_bytes = 0;
9156         ei->new_delalloc_bytes = 0;
9157         ei->defrag_bytes = 0;
9158         ei->disk_i_size = 0;
9159         ei->flags = 0;
9160         ei->csum_bytes = 0;
9161         ei->index_cnt = (u64)-1;
9162         ei->dir_index = 0;
9163         ei->last_unlink_trans = 0;
9164         ei->last_link_trans = 0;
9165         ei->last_log_commit = 0;
9166
9167         spin_lock_init(&ei->lock);
9168         ei->outstanding_extents = 0;
9169         if (sb->s_magic != BTRFS_TEST_MAGIC)
9170                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9171                                               BTRFS_BLOCK_RSV_DELALLOC);
9172         ei->runtime_flags = 0;
9173         ei->prop_compress = BTRFS_COMPRESS_NONE;
9174         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9175
9176         ei->delayed_node = NULL;
9177
9178         ei->i_otime.tv_sec = 0;
9179         ei->i_otime.tv_nsec = 0;
9180
9181         inode = &ei->vfs_inode;
9182         extent_map_tree_init(&ei->extent_tree);
9183         extent_io_tree_init(&ei->io_tree, inode);
9184         extent_io_tree_init(&ei->io_failure_tree, inode);
9185         ei->io_tree.track_uptodate = 1;
9186         ei->io_failure_tree.track_uptodate = 1;
9187         atomic_set(&ei->sync_writers, 0);
9188         mutex_init(&ei->log_mutex);
9189         mutex_init(&ei->delalloc_mutex);
9190         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9191         INIT_LIST_HEAD(&ei->delalloc_inodes);
9192         INIT_LIST_HEAD(&ei->delayed_iput);
9193         RB_CLEAR_NODE(&ei->rb_node);
9194         init_rwsem(&ei->dio_sem);
9195
9196         return inode;
9197 }
9198
9199 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9200 void btrfs_test_destroy_inode(struct inode *inode)
9201 {
9202         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9203         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9204 }
9205 #endif
9206
9207 static void btrfs_i_callback(struct rcu_head *head)
9208 {
9209         struct inode *inode = container_of(head, struct inode, i_rcu);
9210         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9211 }
9212
9213 void btrfs_destroy_inode(struct inode *inode)
9214 {
9215         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9216         struct btrfs_ordered_extent *ordered;
9217         struct btrfs_root *root = BTRFS_I(inode)->root;
9218
9219         WARN_ON(!hlist_empty(&inode->i_dentry));
9220         WARN_ON(inode->i_data.nrpages);
9221         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9222         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9223         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9224         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9225         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9226         WARN_ON(BTRFS_I(inode)->csum_bytes);
9227         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9228
9229         /*
9230          * This can happen where we create an inode, but somebody else also
9231          * created the same inode and we need to destroy the one we already
9232          * created.
9233          */
9234         if (!root)
9235                 goto free;
9236
9237         while (1) {
9238                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9239                 if (!ordered)
9240                         break;
9241                 else {
9242                         btrfs_err(fs_info,
9243                                   "found ordered extent %llu %llu on inode cleanup",
9244                                   ordered->file_offset, ordered->len);
9245                         btrfs_remove_ordered_extent(inode, ordered);
9246                         btrfs_put_ordered_extent(ordered);
9247                         btrfs_put_ordered_extent(ordered);
9248                 }
9249         }
9250         btrfs_qgroup_check_reserved_leak(inode);
9251         inode_tree_del(inode);
9252         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9253 free:
9254         call_rcu(&inode->i_rcu, btrfs_i_callback);
9255 }
9256
9257 int btrfs_drop_inode(struct inode *inode)
9258 {
9259         struct btrfs_root *root = BTRFS_I(inode)->root;
9260
9261         if (root == NULL)
9262                 return 1;
9263
9264         /* the snap/subvol tree is on deleting */
9265         if (btrfs_root_refs(&root->root_item) == 0)
9266                 return 1;
9267         else
9268                 return generic_drop_inode(inode);
9269 }
9270
9271 static void init_once(void *foo)
9272 {
9273         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9274
9275         inode_init_once(&ei->vfs_inode);
9276 }
9277
9278 void __cold btrfs_destroy_cachep(void)
9279 {
9280         /*
9281          * Make sure all delayed rcu free inodes are flushed before we
9282          * destroy cache.
9283          */
9284         rcu_barrier();
9285         kmem_cache_destroy(btrfs_inode_cachep);
9286         kmem_cache_destroy(btrfs_trans_handle_cachep);
9287         kmem_cache_destroy(btrfs_path_cachep);
9288         kmem_cache_destroy(btrfs_free_space_cachep);
9289 }
9290
9291 int __init btrfs_init_cachep(void)
9292 {
9293         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9294                         sizeof(struct btrfs_inode), 0,
9295                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9296                         init_once);
9297         if (!btrfs_inode_cachep)
9298                 goto fail;
9299
9300         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9301                         sizeof(struct btrfs_trans_handle), 0,
9302                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9303         if (!btrfs_trans_handle_cachep)
9304                 goto fail;
9305
9306         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9307                         sizeof(struct btrfs_path), 0,
9308                         SLAB_MEM_SPREAD, NULL);
9309         if (!btrfs_path_cachep)
9310                 goto fail;
9311
9312         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9313                         sizeof(struct btrfs_free_space), 0,
9314                         SLAB_MEM_SPREAD, NULL);
9315         if (!btrfs_free_space_cachep)
9316                 goto fail;
9317
9318         return 0;
9319 fail:
9320         btrfs_destroy_cachep();
9321         return -ENOMEM;
9322 }
9323
9324 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9325                          u32 request_mask, unsigned int flags)
9326 {
9327         u64 delalloc_bytes;
9328         struct inode *inode = d_inode(path->dentry);
9329         u32 blocksize = inode->i_sb->s_blocksize;
9330         u32 bi_flags = BTRFS_I(inode)->flags;
9331
9332         stat->result_mask |= STATX_BTIME;
9333         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9334         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9335         if (bi_flags & BTRFS_INODE_APPEND)
9336                 stat->attributes |= STATX_ATTR_APPEND;
9337         if (bi_flags & BTRFS_INODE_COMPRESS)
9338                 stat->attributes |= STATX_ATTR_COMPRESSED;
9339         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9340                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9341         if (bi_flags & BTRFS_INODE_NODUMP)
9342                 stat->attributes |= STATX_ATTR_NODUMP;
9343
9344         stat->attributes_mask |= (STATX_ATTR_APPEND |
9345                                   STATX_ATTR_COMPRESSED |
9346                                   STATX_ATTR_IMMUTABLE |
9347                                   STATX_ATTR_NODUMP);
9348
9349         generic_fillattr(inode, stat);
9350         stat->dev = BTRFS_I(inode)->root->anon_dev;
9351
9352         spin_lock(&BTRFS_I(inode)->lock);
9353         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9354         spin_unlock(&BTRFS_I(inode)->lock);
9355         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9356                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9357         return 0;
9358 }
9359
9360 static int btrfs_rename_exchange(struct inode *old_dir,
9361                               struct dentry *old_dentry,
9362                               struct inode *new_dir,
9363                               struct dentry *new_dentry)
9364 {
9365         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9366         struct btrfs_trans_handle *trans;
9367         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9368         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9369         struct inode *new_inode = new_dentry->d_inode;
9370         struct inode *old_inode = old_dentry->d_inode;
9371         struct timespec64 ctime = current_time(old_inode);
9372         struct dentry *parent;
9373         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9374         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9375         u64 old_idx = 0;
9376         u64 new_idx = 0;
9377         u64 root_objectid;
9378         int ret;
9379         bool root_log_pinned = false;
9380         bool dest_log_pinned = false;
9381         struct btrfs_log_ctx ctx_root;
9382         struct btrfs_log_ctx ctx_dest;
9383         bool sync_log_root = false;
9384         bool sync_log_dest = false;
9385         bool commit_transaction = false;
9386
9387         /* we only allow rename subvolume link between subvolumes */
9388         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9389                 return -EXDEV;
9390
9391         btrfs_init_log_ctx(&ctx_root, old_inode);
9392         btrfs_init_log_ctx(&ctx_dest, new_inode);
9393
9394         /* close the race window with snapshot create/destroy ioctl */
9395         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9396                 down_read(&fs_info->subvol_sem);
9397         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9398                 down_read(&fs_info->subvol_sem);
9399
9400         /*
9401          * We want to reserve the absolute worst case amount of items.  So if
9402          * both inodes are subvols and we need to unlink them then that would
9403          * require 4 item modifications, but if they are both normal inodes it
9404          * would require 5 item modifications, so we'll assume their normal
9405          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9406          * should cover the worst case number of items we'll modify.
9407          */
9408         trans = btrfs_start_transaction(root, 12);
9409         if (IS_ERR(trans)) {
9410                 ret = PTR_ERR(trans);
9411                 goto out_notrans;
9412         }
9413
9414         /*
9415          * We need to find a free sequence number both in the source and
9416          * in the destination directory for the exchange.
9417          */
9418         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9419         if (ret)
9420                 goto out_fail;
9421         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9422         if (ret)
9423                 goto out_fail;
9424
9425         BTRFS_I(old_inode)->dir_index = 0ULL;
9426         BTRFS_I(new_inode)->dir_index = 0ULL;
9427
9428         /* Reference for the source. */
9429         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9430                 /* force full log commit if subvolume involved. */
9431                 btrfs_set_log_full_commit(fs_info, trans);
9432         } else {
9433                 btrfs_pin_log_trans(root);
9434                 root_log_pinned = true;
9435                 ret = btrfs_insert_inode_ref(trans, dest,
9436                                              new_dentry->d_name.name,
9437                                              new_dentry->d_name.len,
9438                                              old_ino,
9439                                              btrfs_ino(BTRFS_I(new_dir)),
9440                                              old_idx);
9441                 if (ret)
9442                         goto out_fail;
9443         }
9444
9445         /* And now for the dest. */
9446         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9447                 /* force full log commit if subvolume involved. */
9448                 btrfs_set_log_full_commit(fs_info, trans);
9449         } else {
9450                 btrfs_pin_log_trans(dest);
9451                 dest_log_pinned = true;
9452                 ret = btrfs_insert_inode_ref(trans, root,
9453                                              old_dentry->d_name.name,
9454                                              old_dentry->d_name.len,
9455                                              new_ino,
9456                                              btrfs_ino(BTRFS_I(old_dir)),
9457                                              new_idx);
9458                 if (ret)
9459                         goto out_fail;
9460         }
9461
9462         /* Update inode version and ctime/mtime. */
9463         inode_inc_iversion(old_dir);
9464         inode_inc_iversion(new_dir);
9465         inode_inc_iversion(old_inode);
9466         inode_inc_iversion(new_inode);
9467         old_dir->i_ctime = old_dir->i_mtime = ctime;
9468         new_dir->i_ctime = new_dir->i_mtime = ctime;
9469         old_inode->i_ctime = ctime;
9470         new_inode->i_ctime = ctime;
9471
9472         if (old_dentry->d_parent != new_dentry->d_parent) {
9473                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9474                                 BTRFS_I(old_inode), 1);
9475                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9476                                 BTRFS_I(new_inode), 1);
9477         }
9478
9479         /* src is a subvolume */
9480         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9481                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9482                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9483                                           old_dentry->d_name.name,
9484                                           old_dentry->d_name.len);
9485         } else { /* src is an inode */
9486                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9487                                            BTRFS_I(old_dentry->d_inode),
9488                                            old_dentry->d_name.name,
9489                                            old_dentry->d_name.len);
9490                 if (!ret)
9491                         ret = btrfs_update_inode(trans, root, old_inode);
9492         }
9493         if (ret) {
9494                 btrfs_abort_transaction(trans, ret);
9495                 goto out_fail;
9496         }
9497
9498         /* dest is a subvolume */
9499         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9500                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9501                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9502                                           new_dentry->d_name.name,
9503                                           new_dentry->d_name.len);
9504         } else { /* dest is an inode */
9505                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9506                                            BTRFS_I(new_dentry->d_inode),
9507                                            new_dentry->d_name.name,
9508                                            new_dentry->d_name.len);
9509                 if (!ret)
9510                         ret = btrfs_update_inode(trans, dest, new_inode);
9511         }
9512         if (ret) {
9513                 btrfs_abort_transaction(trans, ret);
9514                 goto out_fail;
9515         }
9516
9517         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9518                              new_dentry->d_name.name,
9519                              new_dentry->d_name.len, 0, old_idx);
9520         if (ret) {
9521                 btrfs_abort_transaction(trans, ret);
9522                 goto out_fail;
9523         }
9524
9525         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9526                              old_dentry->d_name.name,
9527                              old_dentry->d_name.len, 0, new_idx);
9528         if (ret) {
9529                 btrfs_abort_transaction(trans, ret);
9530                 goto out_fail;
9531         }
9532
9533         if (old_inode->i_nlink == 1)
9534                 BTRFS_I(old_inode)->dir_index = old_idx;
9535         if (new_inode->i_nlink == 1)
9536                 BTRFS_I(new_inode)->dir_index = new_idx;
9537
9538         if (root_log_pinned) {
9539                 parent = new_dentry->d_parent;
9540                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9541                                          BTRFS_I(old_dir), parent,
9542                                          false, &ctx_root);
9543                 if (ret == BTRFS_NEED_LOG_SYNC)
9544                         sync_log_root = true;
9545                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9546                         commit_transaction = true;
9547                 ret = 0;
9548                 btrfs_end_log_trans(root);
9549                 root_log_pinned = false;
9550         }
9551         if (dest_log_pinned) {
9552                 if (!commit_transaction) {
9553                         parent = old_dentry->d_parent;
9554                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9555                                                  BTRFS_I(new_dir), parent,
9556                                                  false, &ctx_dest);
9557                         if (ret == BTRFS_NEED_LOG_SYNC)
9558                                 sync_log_dest = true;
9559                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9560                                 commit_transaction = true;
9561                         ret = 0;
9562                 }
9563                 btrfs_end_log_trans(dest);
9564                 dest_log_pinned = false;
9565         }
9566 out_fail:
9567         /*
9568          * If we have pinned a log and an error happened, we unpin tasks
9569          * trying to sync the log and force them to fallback to a transaction
9570          * commit if the log currently contains any of the inodes involved in
9571          * this rename operation (to ensure we do not persist a log with an
9572          * inconsistent state for any of these inodes or leading to any
9573          * inconsistencies when replayed). If the transaction was aborted, the
9574          * abortion reason is propagated to userspace when attempting to commit
9575          * the transaction. If the log does not contain any of these inodes, we
9576          * allow the tasks to sync it.
9577          */
9578         if (ret && (root_log_pinned || dest_log_pinned)) {
9579                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9580                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9581                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9582                     (new_inode &&
9583                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9584                         btrfs_set_log_full_commit(fs_info, trans);
9585
9586                 if (root_log_pinned) {
9587                         btrfs_end_log_trans(root);
9588                         root_log_pinned = false;
9589                 }
9590                 if (dest_log_pinned) {
9591                         btrfs_end_log_trans(dest);
9592                         dest_log_pinned = false;
9593                 }
9594         }
9595         if (!ret && sync_log_root && !commit_transaction) {
9596                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9597                                      &ctx_root);
9598                 if (ret)
9599                         commit_transaction = true;
9600         }
9601         if (!ret && sync_log_dest && !commit_transaction) {
9602                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9603                                      &ctx_dest);
9604                 if (ret)
9605                         commit_transaction = true;
9606         }
9607         if (commit_transaction) {
9608                 ret = btrfs_commit_transaction(trans);
9609         } else {
9610                 int ret2;
9611
9612                 ret2 = btrfs_end_transaction(trans);
9613                 ret = ret ? ret : ret2;
9614         }
9615 out_notrans:
9616         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9617                 up_read(&fs_info->subvol_sem);
9618         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9619                 up_read(&fs_info->subvol_sem);
9620
9621         return ret;
9622 }
9623
9624 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9625                                      struct btrfs_root *root,
9626                                      struct inode *dir,
9627                                      struct dentry *dentry)
9628 {
9629         int ret;
9630         struct inode *inode;
9631         u64 objectid;
9632         u64 index;
9633
9634         ret = btrfs_find_free_ino(root, &objectid);
9635         if (ret)
9636                 return ret;
9637
9638         inode = btrfs_new_inode(trans, root, dir,
9639                                 dentry->d_name.name,
9640                                 dentry->d_name.len,
9641                                 btrfs_ino(BTRFS_I(dir)),
9642                                 objectid,
9643                                 S_IFCHR | WHITEOUT_MODE,
9644                                 &index);
9645
9646         if (IS_ERR(inode)) {
9647                 ret = PTR_ERR(inode);
9648                 return ret;
9649         }
9650
9651         inode->i_op = &btrfs_special_inode_operations;
9652         init_special_inode(inode, inode->i_mode,
9653                 WHITEOUT_DEV);
9654
9655         ret = btrfs_init_inode_security(trans, inode, dir,
9656                                 &dentry->d_name);
9657         if (ret)
9658                 goto out;
9659
9660         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9661                                 BTRFS_I(inode), 0, index);
9662         if (ret)
9663                 goto out;
9664
9665         ret = btrfs_update_inode(trans, root, inode);
9666 out:
9667         unlock_new_inode(inode);
9668         if (ret)
9669                 inode_dec_link_count(inode);
9670         iput(inode);
9671
9672         return ret;
9673 }
9674
9675 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9676                            struct inode *new_dir, struct dentry *new_dentry,
9677                            unsigned int flags)
9678 {
9679         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9680         struct btrfs_trans_handle *trans;
9681         unsigned int trans_num_items;
9682         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9683         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9684         struct inode *new_inode = d_inode(new_dentry);
9685         struct inode *old_inode = d_inode(old_dentry);
9686         u64 index = 0;
9687         u64 root_objectid;
9688         int ret;
9689         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9690         bool log_pinned = false;
9691         struct btrfs_log_ctx ctx;
9692         bool sync_log = false;
9693         bool commit_transaction = false;
9694
9695         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9696                 return -EPERM;
9697
9698         /* we only allow rename subvolume link between subvolumes */
9699         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9700                 return -EXDEV;
9701
9702         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9703             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9704                 return -ENOTEMPTY;
9705
9706         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9707             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9708                 return -ENOTEMPTY;
9709
9710
9711         /* check for collisions, even if the  name isn't there */
9712         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9713                              new_dentry->d_name.name,
9714                              new_dentry->d_name.len);
9715
9716         if (ret) {
9717                 if (ret == -EEXIST) {
9718                         /* we shouldn't get
9719                          * eexist without a new_inode */
9720                         if (WARN_ON(!new_inode)) {
9721                                 return ret;
9722                         }
9723                 } else {
9724                         /* maybe -EOVERFLOW */
9725                         return ret;
9726                 }
9727         }
9728         ret = 0;
9729
9730         /*
9731          * we're using rename to replace one file with another.  Start IO on it
9732          * now so  we don't add too much work to the end of the transaction
9733          */
9734         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9735                 filemap_flush(old_inode->i_mapping);
9736
9737         /* close the racy window with snapshot create/destroy ioctl */
9738         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9739                 down_read(&fs_info->subvol_sem);
9740         /*
9741          * We want to reserve the absolute worst case amount of items.  So if
9742          * both inodes are subvols and we need to unlink them then that would
9743          * require 4 item modifications, but if they are both normal inodes it
9744          * would require 5 item modifications, so we'll assume they are normal
9745          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9746          * should cover the worst case number of items we'll modify.
9747          * If our rename has the whiteout flag, we need more 5 units for the
9748          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9749          * when selinux is enabled).
9750          */
9751         trans_num_items = 11;
9752         if (flags & RENAME_WHITEOUT)
9753                 trans_num_items += 5;
9754         trans = btrfs_start_transaction(root, trans_num_items);
9755         if (IS_ERR(trans)) {
9756                 ret = PTR_ERR(trans);
9757                 goto out_notrans;
9758         }
9759
9760         if (dest != root)
9761                 btrfs_record_root_in_trans(trans, dest);
9762
9763         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9764         if (ret)
9765                 goto out_fail;
9766
9767         BTRFS_I(old_inode)->dir_index = 0ULL;
9768         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9769                 /* force full log commit if subvolume involved. */
9770                 btrfs_set_log_full_commit(fs_info, trans);
9771         } else {
9772                 btrfs_pin_log_trans(root);
9773                 log_pinned = true;
9774                 ret = btrfs_insert_inode_ref(trans, dest,
9775                                              new_dentry->d_name.name,
9776                                              new_dentry->d_name.len,
9777                                              old_ino,
9778                                              btrfs_ino(BTRFS_I(new_dir)), index);
9779                 if (ret)
9780                         goto out_fail;
9781         }
9782
9783         inode_inc_iversion(old_dir);
9784         inode_inc_iversion(new_dir);
9785         inode_inc_iversion(old_inode);
9786         old_dir->i_ctime = old_dir->i_mtime =
9787         new_dir->i_ctime = new_dir->i_mtime =
9788         old_inode->i_ctime = current_time(old_dir);
9789
9790         if (old_dentry->d_parent != new_dentry->d_parent)
9791                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9792                                 BTRFS_I(old_inode), 1);
9793
9794         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9795                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9796                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9797                                         old_dentry->d_name.name,
9798                                         old_dentry->d_name.len);
9799         } else {
9800                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9801                                         BTRFS_I(d_inode(old_dentry)),
9802                                         old_dentry->d_name.name,
9803                                         old_dentry->d_name.len);
9804                 if (!ret)
9805                         ret = btrfs_update_inode(trans, root, old_inode);
9806         }
9807         if (ret) {
9808                 btrfs_abort_transaction(trans, ret);
9809                 goto out_fail;
9810         }
9811
9812         if (new_inode) {
9813                 inode_inc_iversion(new_inode);
9814                 new_inode->i_ctime = current_time(new_inode);
9815                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9816                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9817                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9818                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9819                                                 new_dentry->d_name.name,
9820                                                 new_dentry->d_name.len);
9821                         BUG_ON(new_inode->i_nlink == 0);
9822                 } else {
9823                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9824                                                  BTRFS_I(d_inode(new_dentry)),
9825                                                  new_dentry->d_name.name,
9826                                                  new_dentry->d_name.len);
9827                 }
9828                 if (!ret && new_inode->i_nlink == 0)
9829                         ret = btrfs_orphan_add(trans,
9830                                         BTRFS_I(d_inode(new_dentry)));
9831                 if (ret) {
9832                         btrfs_abort_transaction(trans, ret);
9833                         goto out_fail;
9834                 }
9835         }
9836
9837         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9838                              new_dentry->d_name.name,
9839                              new_dentry->d_name.len, 0, index);
9840         if (ret) {
9841                 btrfs_abort_transaction(trans, ret);
9842                 goto out_fail;
9843         }
9844
9845         if (old_inode->i_nlink == 1)
9846                 BTRFS_I(old_inode)->dir_index = index;
9847
9848         if (log_pinned) {
9849                 struct dentry *parent = new_dentry->d_parent;
9850
9851                 btrfs_init_log_ctx(&ctx, old_inode);
9852                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9853                                          BTRFS_I(old_dir), parent,
9854                                          false, &ctx);
9855                 if (ret == BTRFS_NEED_LOG_SYNC)
9856                         sync_log = true;
9857                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9858                         commit_transaction = true;
9859                 ret = 0;
9860                 btrfs_end_log_trans(root);
9861                 log_pinned = false;
9862         }
9863
9864         if (flags & RENAME_WHITEOUT) {
9865                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9866                                                 old_dentry);
9867
9868                 if (ret) {
9869                         btrfs_abort_transaction(trans, ret);
9870                         goto out_fail;
9871                 }
9872         }
9873 out_fail:
9874         /*
9875          * If we have pinned the log and an error happened, we unpin tasks
9876          * trying to sync the log and force them to fallback to a transaction
9877          * commit if the log currently contains any of the inodes involved in
9878          * this rename operation (to ensure we do not persist a log with an
9879          * inconsistent state for any of these inodes or leading to any
9880          * inconsistencies when replayed). If the transaction was aborted, the
9881          * abortion reason is propagated to userspace when attempting to commit
9882          * the transaction. If the log does not contain any of these inodes, we
9883          * allow the tasks to sync it.
9884          */
9885         if (ret && log_pinned) {
9886                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9887                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9888                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9889                     (new_inode &&
9890                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9891                         btrfs_set_log_full_commit(fs_info, trans);
9892
9893                 btrfs_end_log_trans(root);
9894                 log_pinned = false;
9895         }
9896         if (!ret && sync_log) {
9897                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9898                 if (ret)
9899                         commit_transaction = true;
9900         }
9901         if (commit_transaction) {
9902                 ret = btrfs_commit_transaction(trans);
9903         } else {
9904                 int ret2;
9905
9906                 ret2 = btrfs_end_transaction(trans);
9907                 ret = ret ? ret : ret2;
9908         }
9909 out_notrans:
9910         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9911                 up_read(&fs_info->subvol_sem);
9912
9913         return ret;
9914 }
9915
9916 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9917                          struct inode *new_dir, struct dentry *new_dentry,
9918                          unsigned int flags)
9919 {
9920         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9921                 return -EINVAL;
9922
9923         if (flags & RENAME_EXCHANGE)
9924                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9925                                           new_dentry);
9926
9927         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9928 }
9929
9930 struct btrfs_delalloc_work {
9931         struct inode *inode;
9932         struct completion completion;
9933         struct list_head list;
9934         struct btrfs_work work;
9935 };
9936
9937 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9938 {
9939         struct btrfs_delalloc_work *delalloc_work;
9940         struct inode *inode;
9941
9942         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9943                                      work);
9944         inode = delalloc_work->inode;
9945         filemap_flush(inode->i_mapping);
9946         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9947                                 &BTRFS_I(inode)->runtime_flags))
9948                 filemap_flush(inode->i_mapping);
9949
9950         iput(inode);
9951         complete(&delalloc_work->completion);
9952 }
9953
9954 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9955 {
9956         struct btrfs_delalloc_work *work;
9957
9958         work = kmalloc(sizeof(*work), GFP_NOFS);
9959         if (!work)
9960                 return NULL;
9961
9962         init_completion(&work->completion);
9963         INIT_LIST_HEAD(&work->list);
9964         work->inode = inode;
9965         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9966                         btrfs_run_delalloc_work, NULL, NULL);
9967
9968         return work;
9969 }
9970
9971 /*
9972  * some fairly slow code that needs optimization. This walks the list
9973  * of all the inodes with pending delalloc and forces them to disk.
9974  */
9975 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
9976 {
9977         struct btrfs_inode *binode;
9978         struct inode *inode;
9979         struct btrfs_delalloc_work *work, *next;
9980         struct list_head works;
9981         struct list_head splice;
9982         int ret = 0;
9983
9984         INIT_LIST_HEAD(&works);
9985         INIT_LIST_HEAD(&splice);
9986
9987         mutex_lock(&root->delalloc_mutex);
9988         spin_lock(&root->delalloc_lock);
9989         list_splice_init(&root->delalloc_inodes, &splice);
9990         while (!list_empty(&splice)) {
9991                 binode = list_entry(splice.next, struct btrfs_inode,
9992                                     delalloc_inodes);
9993
9994                 list_move_tail(&binode->delalloc_inodes,
9995                                &root->delalloc_inodes);
9996                 inode = igrab(&binode->vfs_inode);
9997                 if (!inode) {
9998                         cond_resched_lock(&root->delalloc_lock);
9999                         continue;
10000                 }
10001                 spin_unlock(&root->delalloc_lock);
10002
10003                 if (snapshot)
10004                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10005                                 &binode->runtime_flags);
10006                 work = btrfs_alloc_delalloc_work(inode);
10007                 if (!work) {
10008                         iput(inode);
10009                         ret = -ENOMEM;
10010                         goto out;
10011                 }
10012                 list_add_tail(&work->list, &works);
10013                 btrfs_queue_work(root->fs_info->flush_workers,
10014                                  &work->work);
10015                 ret++;
10016                 if (nr != -1 && ret >= nr)
10017                         goto out;
10018                 cond_resched();
10019                 spin_lock(&root->delalloc_lock);
10020         }
10021         spin_unlock(&root->delalloc_lock);
10022
10023 out:
10024         list_for_each_entry_safe(work, next, &works, list) {
10025                 list_del_init(&work->list);
10026                 wait_for_completion(&work->completion);
10027                 kfree(work);
10028         }
10029
10030         if (!list_empty(&splice)) {
10031                 spin_lock(&root->delalloc_lock);
10032                 list_splice_tail(&splice, &root->delalloc_inodes);
10033                 spin_unlock(&root->delalloc_lock);
10034         }
10035         mutex_unlock(&root->delalloc_mutex);
10036         return ret;
10037 }
10038
10039 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10040 {
10041         struct btrfs_fs_info *fs_info = root->fs_info;
10042         int ret;
10043
10044         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10045                 return -EROFS;
10046
10047         ret = start_delalloc_inodes(root, -1, true);
10048         if (ret > 0)
10049                 ret = 0;
10050         return ret;
10051 }
10052
10053 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10054 {
10055         struct btrfs_root *root;
10056         struct list_head splice;
10057         int ret;
10058
10059         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10060                 return -EROFS;
10061
10062         INIT_LIST_HEAD(&splice);
10063
10064         mutex_lock(&fs_info->delalloc_root_mutex);
10065         spin_lock(&fs_info->delalloc_root_lock);
10066         list_splice_init(&fs_info->delalloc_roots, &splice);
10067         while (!list_empty(&splice) && nr) {
10068                 root = list_first_entry(&splice, struct btrfs_root,
10069                                         delalloc_root);
10070                 root = btrfs_grab_fs_root(root);
10071                 BUG_ON(!root);
10072                 list_move_tail(&root->delalloc_root,
10073                                &fs_info->delalloc_roots);
10074                 spin_unlock(&fs_info->delalloc_root_lock);
10075
10076                 ret = start_delalloc_inodes(root, nr, false);
10077                 btrfs_put_fs_root(root);
10078                 if (ret < 0)
10079                         goto out;
10080
10081                 if (nr != -1) {
10082                         nr -= ret;
10083                         WARN_ON(nr < 0);
10084                 }
10085                 spin_lock(&fs_info->delalloc_root_lock);
10086         }
10087         spin_unlock(&fs_info->delalloc_root_lock);
10088
10089         ret = 0;
10090 out:
10091         if (!list_empty(&splice)) {
10092                 spin_lock(&fs_info->delalloc_root_lock);
10093                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10094                 spin_unlock(&fs_info->delalloc_root_lock);
10095         }
10096         mutex_unlock(&fs_info->delalloc_root_mutex);
10097         return ret;
10098 }
10099
10100 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10101                          const char *symname)
10102 {
10103         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10104         struct btrfs_trans_handle *trans;
10105         struct btrfs_root *root = BTRFS_I(dir)->root;
10106         struct btrfs_path *path;
10107         struct btrfs_key key;
10108         struct inode *inode = NULL;
10109         int err;
10110         u64 objectid;
10111         u64 index = 0;
10112         int name_len;
10113         int datasize;
10114         unsigned long ptr;
10115         struct btrfs_file_extent_item *ei;
10116         struct extent_buffer *leaf;
10117
10118         name_len = strlen(symname);
10119         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10120                 return -ENAMETOOLONG;
10121
10122         /*
10123          * 2 items for inode item and ref
10124          * 2 items for dir items
10125          * 1 item for updating parent inode item
10126          * 1 item for the inline extent item
10127          * 1 item for xattr if selinux is on
10128          */
10129         trans = btrfs_start_transaction(root, 7);
10130         if (IS_ERR(trans))
10131                 return PTR_ERR(trans);
10132
10133         err = btrfs_find_free_ino(root, &objectid);
10134         if (err)
10135                 goto out_unlock;
10136
10137         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10138                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10139                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10140         if (IS_ERR(inode)) {
10141                 err = PTR_ERR(inode);
10142                 inode = NULL;
10143                 goto out_unlock;
10144         }
10145
10146         /*
10147         * If the active LSM wants to access the inode during
10148         * d_instantiate it needs these. Smack checks to see
10149         * if the filesystem supports xattrs by looking at the
10150         * ops vector.
10151         */
10152         inode->i_fop = &btrfs_file_operations;
10153         inode->i_op = &btrfs_file_inode_operations;
10154         inode->i_mapping->a_ops = &btrfs_aops;
10155         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10156
10157         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10158         if (err)
10159                 goto out_unlock;
10160
10161         path = btrfs_alloc_path();
10162         if (!path) {
10163                 err = -ENOMEM;
10164                 goto out_unlock;
10165         }
10166         key.objectid = btrfs_ino(BTRFS_I(inode));
10167         key.offset = 0;
10168         key.type = BTRFS_EXTENT_DATA_KEY;
10169         datasize = btrfs_file_extent_calc_inline_size(name_len);
10170         err = btrfs_insert_empty_item(trans, root, path, &key,
10171                                       datasize);
10172         if (err) {
10173                 btrfs_free_path(path);
10174                 goto out_unlock;
10175         }
10176         leaf = path->nodes[0];
10177         ei = btrfs_item_ptr(leaf, path->slots[0],
10178                             struct btrfs_file_extent_item);
10179         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10180         btrfs_set_file_extent_type(leaf, ei,
10181                                    BTRFS_FILE_EXTENT_INLINE);
10182         btrfs_set_file_extent_encryption(leaf, ei, 0);
10183         btrfs_set_file_extent_compression(leaf, ei, 0);
10184         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10185         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10186
10187         ptr = btrfs_file_extent_inline_start(ei);
10188         write_extent_buffer(leaf, symname, ptr, name_len);
10189         btrfs_mark_buffer_dirty(leaf);
10190         btrfs_free_path(path);
10191
10192         inode->i_op = &btrfs_symlink_inode_operations;
10193         inode_nohighmem(inode);
10194         inode->i_mapping->a_ops = &btrfs_aops;
10195         inode_set_bytes(inode, name_len);
10196         btrfs_i_size_write(BTRFS_I(inode), name_len);
10197         err = btrfs_update_inode(trans, root, inode);
10198         /*
10199          * Last step, add directory indexes for our symlink inode. This is the
10200          * last step to avoid extra cleanup of these indexes if an error happens
10201          * elsewhere above.
10202          */
10203         if (!err)
10204                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10205                                 BTRFS_I(inode), 0, index);
10206         if (err)
10207                 goto out_unlock;
10208
10209         d_instantiate_new(dentry, inode);
10210
10211 out_unlock:
10212         btrfs_end_transaction(trans);
10213         if (err && inode) {
10214                 inode_dec_link_count(inode);
10215                 discard_new_inode(inode);
10216         }
10217         btrfs_btree_balance_dirty(fs_info);
10218         return err;
10219 }
10220
10221 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10222                                        u64 start, u64 num_bytes, u64 min_size,
10223                                        loff_t actual_len, u64 *alloc_hint,
10224                                        struct btrfs_trans_handle *trans)
10225 {
10226         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10227         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10228         struct extent_map *em;
10229         struct btrfs_root *root = BTRFS_I(inode)->root;
10230         struct btrfs_key ins;
10231         u64 cur_offset = start;
10232         u64 i_size;
10233         u64 cur_bytes;
10234         u64 last_alloc = (u64)-1;
10235         int ret = 0;
10236         bool own_trans = true;
10237         u64 end = start + num_bytes - 1;
10238
10239         if (trans)
10240                 own_trans = false;
10241         while (num_bytes > 0) {
10242                 if (own_trans) {
10243                         trans = btrfs_start_transaction(root, 3);
10244                         if (IS_ERR(trans)) {
10245                                 ret = PTR_ERR(trans);
10246                                 break;
10247                         }
10248                 }
10249
10250                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10251                 cur_bytes = max(cur_bytes, min_size);
10252                 /*
10253                  * If we are severely fragmented we could end up with really
10254                  * small allocations, so if the allocator is returning small
10255                  * chunks lets make its job easier by only searching for those
10256                  * sized chunks.
10257                  */
10258                 cur_bytes = min(cur_bytes, last_alloc);
10259                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10260                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10261                 if (ret) {
10262                         if (own_trans)
10263                                 btrfs_end_transaction(trans);
10264                         break;
10265                 }
10266                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10267
10268                 last_alloc = ins.offset;
10269                 ret = insert_reserved_file_extent(trans, inode,
10270                                                   cur_offset, ins.objectid,
10271                                                   ins.offset, ins.offset,
10272                                                   ins.offset, 0, 0, 0,
10273                                                   BTRFS_FILE_EXTENT_PREALLOC);
10274                 if (ret) {
10275                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10276                                                    ins.offset, 0);
10277                         btrfs_abort_transaction(trans, ret);
10278                         if (own_trans)
10279                                 btrfs_end_transaction(trans);
10280                         break;
10281                 }
10282
10283                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10284                                         cur_offset + ins.offset -1, 0);
10285
10286                 em = alloc_extent_map();
10287                 if (!em) {
10288                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10289                                 &BTRFS_I(inode)->runtime_flags);
10290                         goto next;
10291                 }
10292
10293                 em->start = cur_offset;
10294                 em->orig_start = cur_offset;
10295                 em->len = ins.offset;
10296                 em->block_start = ins.objectid;
10297                 em->block_len = ins.offset;
10298                 em->orig_block_len = ins.offset;
10299                 em->ram_bytes = ins.offset;
10300                 em->bdev = fs_info->fs_devices->latest_bdev;
10301                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10302                 em->generation = trans->transid;
10303
10304                 while (1) {
10305                         write_lock(&em_tree->lock);
10306                         ret = add_extent_mapping(em_tree, em, 1);
10307                         write_unlock(&em_tree->lock);
10308                         if (ret != -EEXIST)
10309                                 break;
10310                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10311                                                 cur_offset + ins.offset - 1,
10312                                                 0);
10313                 }
10314                 free_extent_map(em);
10315 next:
10316                 num_bytes -= ins.offset;
10317                 cur_offset += ins.offset;
10318                 *alloc_hint = ins.objectid + ins.offset;
10319
10320                 inode_inc_iversion(inode);
10321                 inode->i_ctime = current_time(inode);
10322                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10323                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10324                     (actual_len > inode->i_size) &&
10325                     (cur_offset > inode->i_size)) {
10326                         if (cur_offset > actual_len)
10327                                 i_size = actual_len;
10328                         else
10329                                 i_size = cur_offset;
10330                         i_size_write(inode, i_size);
10331                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10332                 }
10333
10334                 ret = btrfs_update_inode(trans, root, inode);
10335
10336                 if (ret) {
10337                         btrfs_abort_transaction(trans, ret);
10338                         if (own_trans)
10339                                 btrfs_end_transaction(trans);
10340                         break;
10341                 }
10342
10343                 if (own_trans)
10344                         btrfs_end_transaction(trans);
10345         }
10346         if (cur_offset < end)
10347                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10348                         end - cur_offset + 1);
10349         return ret;
10350 }
10351
10352 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10353                               u64 start, u64 num_bytes, u64 min_size,
10354                               loff_t actual_len, u64 *alloc_hint)
10355 {
10356         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10357                                            min_size, actual_len, alloc_hint,
10358                                            NULL);
10359 }
10360
10361 int btrfs_prealloc_file_range_trans(struct inode *inode,
10362                                     struct btrfs_trans_handle *trans, int mode,
10363                                     u64 start, u64 num_bytes, u64 min_size,
10364                                     loff_t actual_len, u64 *alloc_hint)
10365 {
10366         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10367                                            min_size, actual_len, alloc_hint, trans);
10368 }
10369
10370 static int btrfs_set_page_dirty(struct page *page)
10371 {
10372         return __set_page_dirty_nobuffers(page);
10373 }
10374
10375 static int btrfs_permission(struct inode *inode, int mask)
10376 {
10377         struct btrfs_root *root = BTRFS_I(inode)->root;
10378         umode_t mode = inode->i_mode;
10379
10380         if (mask & MAY_WRITE &&
10381             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10382                 if (btrfs_root_readonly(root))
10383                         return -EROFS;
10384                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10385                         return -EACCES;
10386         }
10387         return generic_permission(inode, mask);
10388 }
10389
10390 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10391 {
10392         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10393         struct btrfs_trans_handle *trans;
10394         struct btrfs_root *root = BTRFS_I(dir)->root;
10395         struct inode *inode = NULL;
10396         u64 objectid;
10397         u64 index;
10398         int ret = 0;
10399
10400         /*
10401          * 5 units required for adding orphan entry
10402          */
10403         trans = btrfs_start_transaction(root, 5);
10404         if (IS_ERR(trans))
10405                 return PTR_ERR(trans);
10406
10407         ret = btrfs_find_free_ino(root, &objectid);
10408         if (ret)
10409                 goto out;
10410
10411         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10412                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10413         if (IS_ERR(inode)) {
10414                 ret = PTR_ERR(inode);
10415                 inode = NULL;
10416                 goto out;
10417         }
10418
10419         inode->i_fop = &btrfs_file_operations;
10420         inode->i_op = &btrfs_file_inode_operations;
10421
10422         inode->i_mapping->a_ops = &btrfs_aops;
10423         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10424
10425         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10426         if (ret)
10427                 goto out;
10428
10429         ret = btrfs_update_inode(trans, root, inode);
10430         if (ret)
10431                 goto out;
10432         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10433         if (ret)
10434                 goto out;
10435
10436         /*
10437          * We set number of links to 0 in btrfs_new_inode(), and here we set
10438          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10439          * through:
10440          *
10441          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10442          */
10443         set_nlink(inode, 1);
10444         d_tmpfile(dentry, inode);
10445         unlock_new_inode(inode);
10446         mark_inode_dirty(inode);
10447 out:
10448         btrfs_end_transaction(trans);
10449         if (ret && inode)
10450                 discard_new_inode(inode);
10451         btrfs_btree_balance_dirty(fs_info);
10452         return ret;
10453 }
10454
10455 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10456 {
10457         struct inode *inode = tree->private_data;
10458         unsigned long index = start >> PAGE_SHIFT;
10459         unsigned long end_index = end >> PAGE_SHIFT;
10460         struct page *page;
10461
10462         while (index <= end_index) {
10463                 page = find_get_page(inode->i_mapping, index);
10464                 ASSERT(page); /* Pages should be in the extent_io_tree */
10465                 set_page_writeback(page);
10466                 put_page(page);
10467                 index++;
10468         }
10469 }
10470
10471 #ifdef CONFIG_SWAP
10472 /*
10473  * Add an entry indicating a block group or device which is pinned by a
10474  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10475  * negative errno on failure.
10476  */
10477 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10478                                   bool is_block_group)
10479 {
10480         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10481         struct btrfs_swapfile_pin *sp, *entry;
10482         struct rb_node **p;
10483         struct rb_node *parent = NULL;
10484
10485         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10486         if (!sp)
10487                 return -ENOMEM;
10488         sp->ptr = ptr;
10489         sp->inode = inode;
10490         sp->is_block_group = is_block_group;
10491
10492         spin_lock(&fs_info->swapfile_pins_lock);
10493         p = &fs_info->swapfile_pins.rb_node;
10494         while (*p) {
10495                 parent = *p;
10496                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10497                 if (sp->ptr < entry->ptr ||
10498                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10499                         p = &(*p)->rb_left;
10500                 } else if (sp->ptr > entry->ptr ||
10501                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10502                         p = &(*p)->rb_right;
10503                 } else {
10504                         spin_unlock(&fs_info->swapfile_pins_lock);
10505                         kfree(sp);
10506                         return 1;
10507                 }
10508         }
10509         rb_link_node(&sp->node, parent, p);
10510         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10511         spin_unlock(&fs_info->swapfile_pins_lock);
10512         return 0;
10513 }
10514
10515 /* Free all of the entries pinned by this swapfile. */
10516 static void btrfs_free_swapfile_pins(struct inode *inode)
10517 {
10518         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10519         struct btrfs_swapfile_pin *sp;
10520         struct rb_node *node, *next;
10521
10522         spin_lock(&fs_info->swapfile_pins_lock);
10523         node = rb_first(&fs_info->swapfile_pins);
10524         while (node) {
10525                 next = rb_next(node);
10526                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10527                 if (sp->inode == inode) {
10528                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10529                         if (sp->is_block_group)
10530                                 btrfs_put_block_group(sp->ptr);
10531                         kfree(sp);
10532                 }
10533                 node = next;
10534         }
10535         spin_unlock(&fs_info->swapfile_pins_lock);
10536 }
10537
10538 struct btrfs_swap_info {
10539         u64 start;
10540         u64 block_start;
10541         u64 block_len;
10542         u64 lowest_ppage;
10543         u64 highest_ppage;
10544         unsigned long nr_pages;
10545         int nr_extents;
10546 };
10547
10548 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10549                                  struct btrfs_swap_info *bsi)
10550 {
10551         unsigned long nr_pages;
10552         u64 first_ppage, first_ppage_reported, next_ppage;
10553         int ret;
10554
10555         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10556         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10557                                 PAGE_SIZE) >> PAGE_SHIFT;
10558
10559         if (first_ppage >= next_ppage)
10560                 return 0;
10561         nr_pages = next_ppage - first_ppage;
10562
10563         first_ppage_reported = first_ppage;
10564         if (bsi->start == 0)
10565                 first_ppage_reported++;
10566         if (bsi->lowest_ppage > first_ppage_reported)
10567                 bsi->lowest_ppage = first_ppage_reported;
10568         if (bsi->highest_ppage < (next_ppage - 1))
10569                 bsi->highest_ppage = next_ppage - 1;
10570
10571         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10572         if (ret < 0)
10573                 return ret;
10574         bsi->nr_extents += ret;
10575         bsi->nr_pages += nr_pages;
10576         return 0;
10577 }
10578
10579 static void btrfs_swap_deactivate(struct file *file)
10580 {
10581         struct inode *inode = file_inode(file);
10582
10583         btrfs_free_swapfile_pins(inode);
10584         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10585 }
10586
10587 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10588                                sector_t *span)
10589 {
10590         struct inode *inode = file_inode(file);
10591         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10592         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10593         struct extent_state *cached_state = NULL;
10594         struct extent_map *em = NULL;
10595         struct btrfs_device *device = NULL;
10596         struct btrfs_swap_info bsi = {
10597                 .lowest_ppage = (sector_t)-1ULL,
10598         };
10599         int ret = 0;
10600         u64 isize;
10601         u64 start;
10602
10603         /*
10604          * If the swap file was just created, make sure delalloc is done. If the
10605          * file changes again after this, the user is doing something stupid and
10606          * we don't really care.
10607          */
10608         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10609         if (ret)
10610                 return ret;
10611
10612         /*
10613          * The inode is locked, so these flags won't change after we check them.
10614          */
10615         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10616                 btrfs_warn(fs_info, "swapfile must not be compressed");
10617                 return -EINVAL;
10618         }
10619         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10620                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10621                 return -EINVAL;
10622         }
10623         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10624                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10625                 return -EINVAL;
10626         }
10627
10628         /*
10629          * Balance or device remove/replace/resize can move stuff around from
10630          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10631          * concurrently while we are mapping the swap extents, and
10632          * fs_info->swapfile_pins prevents them from running while the swap file
10633          * is active and moving the extents. Note that this also prevents a
10634          * concurrent device add which isn't actually necessary, but it's not
10635          * really worth the trouble to allow it.
10636          */
10637         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10638                 btrfs_warn(fs_info,
10639            "cannot activate swapfile while exclusive operation is running");
10640                 return -EBUSY;
10641         }
10642         /*
10643          * Snapshots can create extents which require COW even if NODATACOW is
10644          * set. We use this counter to prevent snapshots. We must increment it
10645          * before walking the extents because we don't want a concurrent
10646          * snapshot to run after we've already checked the extents.
10647          */
10648         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10649
10650         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10651
10652         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10653         start = 0;
10654         while (start < isize) {
10655                 u64 logical_block_start, physical_block_start;
10656                 struct btrfs_block_group_cache *bg;
10657                 u64 len = isize - start;
10658
10659                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10660                 if (IS_ERR(em)) {
10661                         ret = PTR_ERR(em);
10662                         goto out;
10663                 }
10664
10665                 if (em->block_start == EXTENT_MAP_HOLE) {
10666                         btrfs_warn(fs_info, "swapfile must not have holes");
10667                         ret = -EINVAL;
10668                         goto out;
10669                 }
10670                 if (em->block_start == EXTENT_MAP_INLINE) {
10671                         /*
10672                          * It's unlikely we'll ever actually find ourselves
10673                          * here, as a file small enough to fit inline won't be
10674                          * big enough to store more than the swap header, but in
10675                          * case something changes in the future, let's catch it
10676                          * here rather than later.
10677                          */
10678                         btrfs_warn(fs_info, "swapfile must not be inline");
10679                         ret = -EINVAL;
10680                         goto out;
10681                 }
10682                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10683                         btrfs_warn(fs_info, "swapfile must not be compressed");
10684                         ret = -EINVAL;
10685                         goto out;
10686                 }
10687
10688                 logical_block_start = em->block_start + (start - em->start);
10689                 len = min(len, em->len - (start - em->start));
10690                 free_extent_map(em);
10691                 em = NULL;
10692
10693                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10694                 if (ret < 0) {
10695                         goto out;
10696                 } else if (ret) {
10697                         ret = 0;
10698                 } else {
10699                         btrfs_warn(fs_info,
10700                                    "swapfile must not be copy-on-write");
10701                         ret = -EINVAL;
10702                         goto out;
10703                 }
10704
10705                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10706                 if (IS_ERR(em)) {
10707                         ret = PTR_ERR(em);
10708                         goto out;
10709                 }
10710
10711                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10712                         btrfs_warn(fs_info,
10713                                    "swapfile must have single data profile");
10714                         ret = -EINVAL;
10715                         goto out;
10716                 }
10717
10718                 if (device == NULL) {
10719                         device = em->map_lookup->stripes[0].dev;
10720                         ret = btrfs_add_swapfile_pin(inode, device, false);
10721                         if (ret == 1)
10722                                 ret = 0;
10723                         else if (ret)
10724                                 goto out;
10725                 } else if (device != em->map_lookup->stripes[0].dev) {
10726                         btrfs_warn(fs_info, "swapfile must be on one device");
10727                         ret = -EINVAL;
10728                         goto out;
10729                 }
10730
10731                 physical_block_start = (em->map_lookup->stripes[0].physical +
10732                                         (logical_block_start - em->start));
10733                 len = min(len, em->len - (logical_block_start - em->start));
10734                 free_extent_map(em);
10735                 em = NULL;
10736
10737                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10738                 if (!bg) {
10739                         btrfs_warn(fs_info,
10740                            "could not find block group containing swapfile");
10741                         ret = -EINVAL;
10742                         goto out;
10743                 }
10744
10745                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10746                 if (ret) {
10747                         btrfs_put_block_group(bg);
10748                         if (ret == 1)
10749                                 ret = 0;
10750                         else
10751                                 goto out;
10752                 }
10753
10754                 if (bsi.block_len &&
10755                     bsi.block_start + bsi.block_len == physical_block_start) {
10756                         bsi.block_len += len;
10757                 } else {
10758                         if (bsi.block_len) {
10759                                 ret = btrfs_add_swap_extent(sis, &bsi);
10760                                 if (ret)
10761                                         goto out;
10762                         }
10763                         bsi.start = start;
10764                         bsi.block_start = physical_block_start;
10765                         bsi.block_len = len;
10766                 }
10767
10768                 start += len;
10769         }
10770
10771         if (bsi.block_len)
10772                 ret = btrfs_add_swap_extent(sis, &bsi);
10773
10774 out:
10775         if (!IS_ERR_OR_NULL(em))
10776                 free_extent_map(em);
10777
10778         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10779
10780         if (ret)
10781                 btrfs_swap_deactivate(file);
10782
10783         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10784
10785         if (ret)
10786                 return ret;
10787
10788         if (device)
10789                 sis->bdev = device->bdev;
10790         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10791         sis->max = bsi.nr_pages;
10792         sis->pages = bsi.nr_pages - 1;
10793         sis->highest_bit = bsi.nr_pages - 1;
10794         return bsi.nr_extents;
10795 }
10796 #else
10797 static void btrfs_swap_deactivate(struct file *file)
10798 {
10799 }
10800
10801 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10802                                sector_t *span)
10803 {
10804         return -EOPNOTSUPP;
10805 }
10806 #endif
10807
10808 static const struct inode_operations btrfs_dir_inode_operations = {
10809         .getattr        = btrfs_getattr,
10810         .lookup         = btrfs_lookup,
10811         .create         = btrfs_create,
10812         .unlink         = btrfs_unlink,
10813         .link           = btrfs_link,
10814         .mkdir          = btrfs_mkdir,
10815         .rmdir          = btrfs_rmdir,
10816         .rename         = btrfs_rename2,
10817         .symlink        = btrfs_symlink,
10818         .setattr        = btrfs_setattr,
10819         .mknod          = btrfs_mknod,
10820         .listxattr      = btrfs_listxattr,
10821         .permission     = btrfs_permission,
10822         .get_acl        = btrfs_get_acl,
10823         .set_acl        = btrfs_set_acl,
10824         .update_time    = btrfs_update_time,
10825         .tmpfile        = btrfs_tmpfile,
10826 };
10827 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10828         .lookup         = btrfs_lookup,
10829         .permission     = btrfs_permission,
10830         .update_time    = btrfs_update_time,
10831 };
10832
10833 static const struct file_operations btrfs_dir_file_operations = {
10834         .llseek         = generic_file_llseek,
10835         .read           = generic_read_dir,
10836         .iterate_shared = btrfs_real_readdir,
10837         .open           = btrfs_opendir,
10838         .unlocked_ioctl = btrfs_ioctl,
10839 #ifdef CONFIG_COMPAT
10840         .compat_ioctl   = btrfs_compat_ioctl,
10841 #endif
10842         .release        = btrfs_release_file,
10843         .fsync          = btrfs_sync_file,
10844 };
10845
10846 static const struct extent_io_ops btrfs_extent_io_ops = {
10847         /* mandatory callbacks */
10848         .submit_bio_hook = btrfs_submit_bio_hook,
10849         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10850 };
10851
10852 /*
10853  * btrfs doesn't support the bmap operation because swapfiles
10854  * use bmap to make a mapping of extents in the file.  They assume
10855  * these extents won't change over the life of the file and they
10856  * use the bmap result to do IO directly to the drive.
10857  *
10858  * the btrfs bmap call would return logical addresses that aren't
10859  * suitable for IO and they also will change frequently as COW
10860  * operations happen.  So, swapfile + btrfs == corruption.
10861  *
10862  * For now we're avoiding this by dropping bmap.
10863  */
10864 static const struct address_space_operations btrfs_aops = {
10865         .readpage       = btrfs_readpage,
10866         .writepage      = btrfs_writepage,
10867         .writepages     = btrfs_writepages,
10868         .readpages      = btrfs_readpages,
10869         .direct_IO      = btrfs_direct_IO,
10870         .invalidatepage = btrfs_invalidatepage,
10871         .releasepage    = btrfs_releasepage,
10872         .set_page_dirty = btrfs_set_page_dirty,
10873         .error_remove_page = generic_error_remove_page,
10874         .swap_activate  = btrfs_swap_activate,
10875         .swap_deactivate = btrfs_swap_deactivate,
10876 };
10877
10878 static const struct inode_operations btrfs_file_inode_operations = {
10879         .getattr        = btrfs_getattr,
10880         .setattr        = btrfs_setattr,
10881         .listxattr      = btrfs_listxattr,
10882         .permission     = btrfs_permission,
10883         .fiemap         = btrfs_fiemap,
10884         .get_acl        = btrfs_get_acl,
10885         .set_acl        = btrfs_set_acl,
10886         .update_time    = btrfs_update_time,
10887 };
10888 static const struct inode_operations btrfs_special_inode_operations = {
10889         .getattr        = btrfs_getattr,
10890         .setattr        = btrfs_setattr,
10891         .permission     = btrfs_permission,
10892         .listxattr      = btrfs_listxattr,
10893         .get_acl        = btrfs_get_acl,
10894         .set_acl        = btrfs_set_acl,
10895         .update_time    = btrfs_update_time,
10896 };
10897 static const struct inode_operations btrfs_symlink_inode_operations = {
10898         .get_link       = page_get_link,
10899         .getattr        = btrfs_getattr,
10900         .setattr        = btrfs_setattr,
10901         .permission     = btrfs_permission,
10902         .listxattr      = btrfs_listxattr,
10903         .update_time    = btrfs_update_time,
10904 };
10905
10906 const struct dentry_operations btrfs_dentry_operations = {
10907         .d_delete       = btrfs_dentry_delete,
10908 };