fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 outstanding_extents;
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_transaction_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 113                                            u64 len, u64 orig_start,
 114                                            u64 block_start, u64 block_len,
 115                                            u64 orig_block_len, u64 ram_bytes,
 116                                            int type);
 117
 118 static int btrfs_dirty_inode(struct inode *inode);
 119
 120 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 121 void btrfs_test_inode_set_ops(struct inode *inode)
 122 {
 123         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 124 }
 125 #endif
 126
 127 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 128                                      struct inode *inode,  struct inode *dir,
 129                                      const struct qstr *qstr)
 130 {
 131         int err;
 132
 133         err = btrfs_init_acl(trans, inode, dir);
 134         if (!err)
 135                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 136         return err;
 137 }
 138
 139 /*
 140  * this does all the hard work for inserting an inline extent into
 141  * the btree.  The caller should have done a btrfs_drop_extents so that
 142  * no overlapping inline items exist in the btree
 143  */
 144 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 145                                 struct btrfs_path *path, int extent_inserted,
 146                                 struct btrfs_root *root, struct inode *inode,
 147                                 u64 start, size_t size, size_t compressed_size,
 148                                 int compress_type,
 149                                 struct page **compressed_pages)
 150 {
 151         struct extent_buffer *leaf;
 152         struct page *page = NULL;
 153         char *kaddr;
 154         unsigned long ptr;
 155         struct btrfs_file_extent_item *ei;
 156         int err = 0;
 157         int ret;
 158         size_t cur_size = size;
 159         unsigned long offset;
 160
 161         if (compressed_size && compressed_pages)
 162                 cur_size = compressed_size;
 163
 164         inode_add_bytes(inode, size);
 165
 166         if (!extent_inserted) {
 167                 struct btrfs_key key;
 168                 size_t datasize;
 169
 170                 key.objectid = btrfs_ino(inode);
 171                 key.offset = start;
 172                 key.type = BTRFS_EXTENT_DATA_KEY;
 173
 174                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 175                 path->leave_spinning = 1;
 176                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 177                                               datasize);
 178                 if (ret) {
 179                         err = ret;
 180                         goto fail;
 181                 }
 182         }
 183         leaf = path->nodes[0];
 184         ei = btrfs_item_ptr(leaf, path->slots[0],
 185                             struct btrfs_file_extent_item);
 186         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 187         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 188         btrfs_set_file_extent_encryption(leaf, ei, 0);
 189         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 190         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 191         ptr = btrfs_file_extent_inline_start(ei);
 192
 193         if (compress_type != BTRFS_COMPRESS_NONE) {
 194                 struct page *cpage;
 195                 int i = 0;
 196                 while (compressed_size > 0) {
 197                         cpage = compressed_pages[i];
 198                         cur_size = min_t(unsigned long, compressed_size,
 199                                        PAGE_SIZE);
 200
 201                         kaddr = kmap_atomic(cpage);
 202                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 203                         kunmap_atomic(kaddr);
 204
 205                         i++;
 206                         ptr += cur_size;
 207                         compressed_size -= cur_size;
 208                 }
 209                 btrfs_set_file_extent_compression(leaf, ei,
 210                                                   compress_type);
 211         } else {
 212                 page = find_get_page(inode->i_mapping,
 213                                      start >> PAGE_SHIFT);
 214                 btrfs_set_file_extent_compression(leaf, ei, 0);
 215                 kaddr = kmap_atomic(page);
 216                 offset = start & (PAGE_SIZE - 1);
 217                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 218                 kunmap_atomic(kaddr);
 219                 put_page(page);
 220         }
 221         btrfs_mark_buffer_dirty(leaf);
 222         btrfs_release_path(path);
 223
 224         /*
 225          * we're an inline extent, so nobody can
 226          * extend the file past i_size without locking
 227          * a page we already have locked.
 228          *
 229          * We must do any isize and inode updates
 230          * before we unlock the pages.  Otherwise we
 231          * could end up racing with unlink.
 232          */
 233         BTRFS_I(inode)->disk_i_size = inode->i_size;
 234         ret = btrfs_update_inode(trans, root, inode);
 235
 236         return ret;
 237 fail:
 238         return err;
 239 }
 240
 241
 242 /*
 243  * conditionally insert an inline extent into the file.  This
 244  * does the checks required to make sure the data is small enough
 245  * to fit as an inline extent.
 246  */
 247 static noinline int cow_file_range_inline(struct btrfs_root *root,
 248                                           struct inode *inode, u64 start,
 249                                           u64 end, size_t compressed_size,
 250                                           int compress_type,
 251                                           struct page **compressed_pages)
 252 {
 253         struct btrfs_trans_handle *trans;
 254         u64 isize = i_size_read(inode);
 255         u64 actual_end = min(end + 1, isize);
 256         u64 inline_len = actual_end - start;
 257         u64 aligned_end = ALIGN(end, root->sectorsize);
 258         u64 data_len = inline_len;
 259         int ret;
 260         struct btrfs_path *path;
 261         int extent_inserted = 0;
 262         u32 extent_item_size;
 263
 264         if (compressed_size)
 265                 data_len = compressed_size;
 266
 267         if (start > 0 ||
 268             actual_end > root->sectorsize ||
 269             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 270             (!compressed_size &&
 271             (actual_end & (root->sectorsize - 1)) == 0) ||
 272             end + 1 < isize ||
 273             data_len > root->fs_info->max_inline) {
 274                 return 1;
 275         }
 276
 277         path = btrfs_alloc_path();
 278         if (!path)
 279                 return -ENOMEM;
 280
 281         trans = btrfs_join_transaction(root);
 282         if (IS_ERR(trans)) {
 283                 btrfs_free_path(path);
 284                 return PTR_ERR(trans);
 285         }
 286         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 287
 288         if (compressed_size && compressed_pages)
 289                 extent_item_size = btrfs_file_extent_calc_inline_size(
 290                    compressed_size);
 291         else
 292                 extent_item_size = btrfs_file_extent_calc_inline_size(
 293                     inline_len);
 294
 295         ret = __btrfs_drop_extents(trans, root, inode, path,
 296                                    start, aligned_end, NULL,
 297                                    1, 1, extent_item_size, &extent_inserted);
 298         if (ret) {
 299                 btrfs_abort_transaction(trans, ret);
 300                 goto out;
 301         }
 302
 303         if (isize > actual_end)
 304                 inline_len = min_t(u64, isize, actual_end);
 305         ret = insert_inline_extent(trans, path, extent_inserted,
 306                                    root, inode, start,
 307                                    inline_len, compressed_size,
 308                                    compress_type, compressed_pages);
 309         if (ret && ret != -ENOSPC) {
 310                 btrfs_abort_transaction(trans, ret);
 311                 goto out;
 312         } else if (ret == -ENOSPC) {
 313                 ret = 1;
 314                 goto out;
 315         }
 316
 317         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 318         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 319         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 320 out:
 321         /*
 322          * Don't forget to free the reserved space, as for inlined extent
 323          * it won't count as data extent, free them directly here.
 324          * And at reserve time, it's always aligned to page size, so
 325          * just free one page here.
 326          */
 327         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 328         btrfs_free_path(path);
 329         btrfs_end_transaction(trans, root);
 330         return ret;
 331 }
 332
 333 struct async_extent {
 334         u64 start;
 335         u64 ram_size;
 336         u64 compressed_size;
 337         struct page **pages;
 338         unsigned long nr_pages;
 339         int compress_type;
 340         struct list_head list;
 341 };
 342
 343 struct async_cow {
 344         struct inode *inode;
 345         struct btrfs_root *root;
 346         struct page *locked_page;
 347         u64 start;
 348         u64 end;
 349         struct list_head extents;
 350         struct btrfs_work work;
 351 };
 352
 353 static noinline int add_async_extent(struct async_cow *cow,
 354                                      u64 start, u64 ram_size,
 355                                      u64 compressed_size,
 356                                      struct page **pages,
 357                                      unsigned long nr_pages,
 358                                      int compress_type)
 359 {
 360         struct async_extent *async_extent;
 361
 362         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 363         BUG_ON(!async_extent); /* -ENOMEM */
 364         async_extent->start = start;
 365         async_extent->ram_size = ram_size;
 366         async_extent->compressed_size = compressed_size;
 367         async_extent->pages = pages;
 368         async_extent->nr_pages = nr_pages;
 369         async_extent->compress_type = compress_type;
 370         list_add_tail(&async_extent->list, &cow->extents);
 371         return 0;
 372 }
 373
 374 static inline int inode_need_compress(struct inode *inode)
 375 {
 376         struct btrfs_root *root = BTRFS_I(inode)->root;
 377
 378         /* force compress */
 379         if (btrfs_test_opt(root->fs_info, FORCE_COMPRESS))
 380                 return 1;
 381         /* bad compression ratios */
 382         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 383                 return 0;
 384         if (btrfs_test_opt(root->fs_info, COMPRESS) ||
 385             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 386             BTRFS_I(inode)->force_compress)
 387                 return 1;
 388         return 0;
 389 }
 390
 391 /*
 392  * we create compressed extents in two phases.  The first
 393  * phase compresses a range of pages that have already been
 394  * locked (both pages and state bits are locked).
 395  *
 396  * This is done inside an ordered work queue, and the compression
 397  * is spread across many cpus.  The actual IO submission is step
 398  * two, and the ordered work queue takes care of making sure that
 399  * happens in the same order things were put onto the queue by
 400  * writepages and friends.
 401  *
 402  * If this code finds it can't get good compression, it puts an
 403  * entry onto the work queue to write the uncompressed bytes.  This
 404  * makes sure that both compressed inodes and uncompressed inodes
 405  * are written in the same order that the flusher thread sent them
 406  * down.
 407  */
 408 static noinline void compress_file_range(struct inode *inode,
 409                                         struct page *locked_page,
 410                                         u64 start, u64 end,
 411                                         struct async_cow *async_cow,
 412                                         int *num_added)
 413 {
 414         struct btrfs_root *root = BTRFS_I(inode)->root;
 415         u64 num_bytes;
 416         u64 blocksize = root->sectorsize;
 417         u64 actual_end;
 418         u64 isize = i_size_read(inode);
 419         int ret = 0;
 420         struct page **pages = NULL;
 421         unsigned long nr_pages;
 422         unsigned long nr_pages_ret = 0;
 423         unsigned long total_compressed = 0;
 424         unsigned long total_in = 0;
 425         unsigned long max_compressed = SZ_128K;
 426         unsigned long max_uncompressed = SZ_128K;
 427         int i;
 428         int will_compress;
 429         int compress_type = root->fs_info->compress_type;
 430         int redirty = 0;
 431
 432         /* if this is a small write inside eof, kick off a defrag */
 433         if ((end - start + 1) < SZ_16K &&
 434             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 435                 btrfs_add_inode_defrag(NULL, inode);
 436
 437         actual_end = min_t(u64, isize, end + 1);
 438 again:
 439         will_compress = 0;
 440         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 441         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 442
 443         /*
 444          * we don't want to send crud past the end of i_size through
 445          * compression, that's just a waste of CPU time.  So, if the
 446          * end of the file is before the start of our current
 447          * requested range of bytes, we bail out to the uncompressed
 448          * cleanup code that can deal with all of this.
 449          *
 450          * It isn't really the fastest way to fix things, but this is a
 451          * very uncommon corner.
 452          */
 453         if (actual_end <= start)
 454                 goto cleanup_and_bail_uncompressed;
 455
 456         total_compressed = actual_end - start;
 457
 458         /*
 459          * skip compression for a small file range(<=blocksize) that
 460          * isn't an inline extent, since it doesn't save disk space at all.
 461          */
 462         if (total_compressed <= blocksize &&
 463            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 464                 goto cleanup_and_bail_uncompressed;
 465
 466         /* we want to make sure that amount of ram required to uncompress
 467          * an extent is reasonable, so we limit the total size in ram
 468          * of a compressed extent to 128k.  This is a crucial number
 469          * because it also controls how easily we can spread reads across
 470          * cpus for decompression.
 471          *
 472          * We also want to make sure the amount of IO required to do
 473          * a random read is reasonably small, so we limit the size of
 474          * a compressed extent to 128k.
 475          */
 476         total_compressed = min(total_compressed, max_uncompressed);
 477         num_bytes = ALIGN(end - start + 1, blocksize);
 478         num_bytes = max(blocksize,  num_bytes);
 479         total_in = 0;
 480         ret = 0;
 481
 482         /*
 483          * we do compression for mount -o compress and when the
 484          * inode has not been flagged as nocompress.  This flag can
 485          * change at any time if we discover bad compression ratios.
 486          */
 487         if (inode_need_compress(inode)) {
 488                 WARN_ON(pages);
 489                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 490                 if (!pages) {
 491                         /* just bail out to the uncompressed code */
 492                         goto cont;
 493                 }
 494
 495                 if (BTRFS_I(inode)->force_compress)
 496                         compress_type = BTRFS_I(inode)->force_compress;
 497
 498                 /*
 499                  * we need to call clear_page_dirty_for_io on each
 500                  * page in the range.  Otherwise applications with the file
 501                  * mmap'd can wander in and change the page contents while
 502                  * we are compressing them.
 503                  *
 504                  * If the compression fails for any reason, we set the pages
 505                  * dirty again later on.
 506                  */
 507                 extent_range_clear_dirty_for_io(inode, start, end);
 508                 redirty = 1;
 509                 ret = btrfs_compress_pages(compress_type,
 510                                            inode->i_mapping, start,
 511                                            total_compressed, pages,
 512                                            nr_pages, &nr_pages_ret,
 513                                            &total_in,
 514                                            &total_compressed,
 515                                            max_compressed);
 516
 517                 if (!ret) {
 518                         unsigned long offset = total_compressed &
 519                                 (PAGE_SIZE - 1);
 520                         struct page *page = pages[nr_pages_ret - 1];
 521                         char *kaddr;
 522
 523                         /* zero the tail end of the last page, we might be
 524                          * sending it down to disk
 525                          */
 526                         if (offset) {
 527                                 kaddr = kmap_atomic(page);
 528                                 memset(kaddr + offset, 0,
 529                                        PAGE_SIZE - offset);
 530                                 kunmap_atomic(kaddr);
 531                         }
 532                         will_compress = 1;
 533                 }
 534         }
 535 cont:
 536         if (start == 0) {
 537                 /* lets try to make an inline extent */
 538                 if (ret || total_in < (actual_end - start)) {
 539                         /* we didn't compress the entire range, try
 540                          * to make an uncompressed inline extent.
 541                          */
 542                         ret = cow_file_range_inline(root, inode, start, end,
 543                                                     0, 0, NULL);
 544                 } else {
 545                         /* try making a compressed inline extent */
 546                         ret = cow_file_range_inline(root, inode, start, end,
 547                                                     total_compressed,
 548                                                     compress_type, pages);
 549                 }
 550                 if (ret <= 0) {
 551                         unsigned long clear_flags = EXTENT_DELALLOC |
 552                                 EXTENT_DEFRAG;
 553                         unsigned long page_error_op;
 554
 555                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 556                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 557
 558                         /*
 559                          * inline extent creation worked or returned error,
 560                          * we don't need to create any more async work items.
 561                          * Unlock and free up our temp pages.
 562                          */
 563                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 564                                                      clear_flags, PAGE_UNLOCK |
 565                                                      PAGE_CLEAR_DIRTY |
 566                                                      PAGE_SET_WRITEBACK |
 567                                                      page_error_op |
 568                                                      PAGE_END_WRITEBACK);
 569                         goto free_pages_out;
 570                 }
 571         }
 572
 573         if (will_compress) {
 574                 /*
 575                  * we aren't doing an inline extent round the compressed size
 576                  * up to a block size boundary so the allocator does sane
 577                  * things
 578                  */
 579                 total_compressed = ALIGN(total_compressed, blocksize);
 580
 581                 /*
 582                  * one last check to make sure the compression is really a
 583                  * win, compare the page count read with the blocks on disk
 584                  */
 585                 total_in = ALIGN(total_in, PAGE_SIZE);
 586                 if (total_compressed >= total_in) {
 587                         will_compress = 0;
 588                 } else {
 589                         num_bytes = total_in;
 590                         *num_added += 1;
 591
 592                         /*
 593                          * The async work queues will take care of doing actual
 594                          * allocation on disk for these compressed pages, and
 595                          * will submit them to the elevator.
 596                          */
 597                         add_async_extent(async_cow, start, num_bytes,
 598                                         total_compressed, pages, nr_pages_ret,
 599                                         compress_type);
 600
 601                         if (start + num_bytes < end) {
 602                                 start += num_bytes;
 603                                 pages = NULL;
 604                                 cond_resched();
 605                                 goto again;
 606                         }
 607                         return;
 608                 }
 609         }
 610         if (pages) {
 611                 /*
 612                  * the compression code ran but failed to make things smaller,
 613                  * free any pages it allocated and our page pointer array
 614                  */
 615                 for (i = 0; i < nr_pages_ret; i++) {
 616                         WARN_ON(pages[i]->mapping);
 617                         put_page(pages[i]);
 618                 }
 619                 kfree(pages);
 620                 pages = NULL;
 621                 total_compressed = 0;
 622                 nr_pages_ret = 0;
 623
 624                 /* flag the file so we don't compress in the future */
 625                 if (!btrfs_test_opt(root->fs_info, FORCE_COMPRESS) &&
 626                     !(BTRFS_I(inode)->force_compress)) {
 627                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 628                 }
 629         }
 630 cleanup_and_bail_uncompressed:
 631         /*
 632          * No compression, but we still need to write the pages in the file
 633          * we've been given so far.  redirty the locked page if it corresponds
 634          * to our extent and set things up for the async work queue to run
 635          * cow_file_range to do the normal delalloc dance.
 636          */
 637         if (page_offset(locked_page) >= start &&
 638             page_offset(locked_page) <= end)
 639                 __set_page_dirty_nobuffers(locked_page);
 640                 /* unlocked later on in the async handlers */
 641
 642         if (redirty)
 643                 extent_range_redirty_for_io(inode, start, end);
 644         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 645                          BTRFS_COMPRESS_NONE);
 646         *num_added += 1;
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 put_page(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 put_page(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              async_extent->start +
 716                                              async_extent->ram_size - 1,
 717                                              &page_started, &nr_written, 0,
 718                                              NULL);
 719
 720                         /* JDM XXX */
 721
 722                         /*
 723                          * if page_started, cow_file_range inserted an
 724                          * inline extent and took care of all the unlocking
 725                          * and IO for us.  Otherwise, we need to submit
 726                          * all those pages down to the drive.
 727                          */
 728                         if (!page_started && !ret)
 729                                 extent_write_locked_range(io_tree,
 730                                                   inode, async_extent->start,
 731                                                   async_extent->start +
 732                                                   async_extent->ram_size - 1,
 733                                                   btrfs_get_extent,
 734                                                   WB_SYNC_ALL);
 735                         else if (ret)
 736                                 unlock_page(async_cow->locked_page);
 737                         kfree(async_extent);
 738                         cond_resched();
 739                         continue;
 740                 }
 741
 742                 lock_extent(io_tree, async_extent->start,
 743                             async_extent->start + async_extent->ram_size - 1);
 744
 745                 ret = btrfs_reserve_extent(root,
 746                                            async_extent->compressed_size,
 747                                            async_extent->compressed_size,
 748                                            0, alloc_hint, &ins, 1, 1);
 749                 if (ret) {
 750                         free_async_extent_pages(async_extent);
 751
 752                         if (ret == -ENOSPC) {
 753                                 unlock_extent(io_tree, async_extent->start,
 754                                               async_extent->start +
 755                                               async_extent->ram_size - 1);
 756
 757                                 /*
 758                                  * we need to redirty the pages if we decide to
 759                                  * fallback to uncompressed IO, otherwise we
 760                                  * will not submit these pages down to lower
 761                                  * layers.
 762                                  */
 763                                 extent_range_redirty_for_io(inode,
 764                                                 async_extent->start,
 765                                                 async_extent->start +
 766                                                 async_extent->ram_size - 1);
 767
 768                                 goto retry;
 769                         }
 770                         goto out_free;
 771                 }
 772                 /*
 773                  * here we're doing allocation and writeback of the
 774                  * compressed pages
 775                  */
 776                 btrfs_drop_extent_cache(inode, async_extent->start,
 777                                         async_extent->start +
 778                                         async_extent->ram_size - 1, 0);
 779
 780                 em = alloc_extent_map();
 781                 if (!em) {
 782                         ret = -ENOMEM;
 783                         goto out_free_reserve;
 784                 }
 785                 em->start = async_extent->start;
 786                 em->len = async_extent->ram_size;
 787                 em->orig_start = em->start;
 788                 em->mod_start = em->start;
 789                 em->mod_len = em->len;
 790
 791                 em->block_start = ins.objectid;
 792                 em->block_len = ins.offset;
 793                 em->orig_block_len = ins.offset;
 794                 em->ram_bytes = async_extent->ram_size;
 795                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 796                 em->compress_type = async_extent->compress_type;
 797                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 798                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 799                 em->generation = -1;
 800
 801                 while (1) {
 802                         write_lock(&em_tree->lock);
 803                         ret = add_extent_mapping(em_tree, em, 1);
 804                         write_unlock(&em_tree->lock);
 805                         if (ret != -EEXIST) {
 806                                 free_extent_map(em);
 807                                 break;
 808                         }
 809                         btrfs_drop_extent_cache(inode, async_extent->start,
 810                                                 async_extent->start +
 811                                                 async_extent->ram_size - 1, 0);
 812                 }
 813
 814                 if (ret)
 815                         goto out_free_reserve;
 816
 817                 ret = btrfs_add_ordered_extent_compress(inode,
 818                                                 async_extent->start,
 819                                                 ins.objectid,
 820                                                 async_extent->ram_size,
 821                                                 ins.offset,
 822                                                 BTRFS_ORDERED_COMPRESSED,
 823                                                 async_extent->compress_type);
 824                 if (ret) {
 825                         btrfs_drop_extent_cache(inode, async_extent->start,
 826                                                 async_extent->start +
 827                                                 async_extent->ram_size - 1, 0);
 828                         goto out_free_reserve;
 829                 }
 830                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 831
 832                 /*
 833                  * clear dirty, set writeback and unlock the pages.
 834                  */
 835                 extent_clear_unlock_delalloc(inode, async_extent->start,
 836                                 async_extent->start +
 837                                 async_extent->ram_size - 1,
 838                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 839                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 840                                 PAGE_SET_WRITEBACK);
 841                 ret = btrfs_submit_compressed_write(inode,
 842                                     async_extent->start,
 843                                     async_extent->ram_size,
 844                                     ins.objectid,
 845                                     ins.offset, async_extent->pages,
 846                                     async_extent->nr_pages);
 847                 if (ret) {
 848                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 849                         struct page *p = async_extent->pages[0];
 850                         const u64 start = async_extent->start;
 851                         const u64 end = start + async_extent->ram_size - 1;
 852
 853                         p->mapping = inode->i_mapping;
 854                         tree->ops->writepage_end_io_hook(p, start, end,
 855                                                          NULL, 0);
 856                         p->mapping = NULL;
 857                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 858                                                      PAGE_END_WRITEBACK |
 859                                                      PAGE_SET_ERROR);
 860                         free_async_extent_pages(async_extent);
 861                 }
 862                 alloc_hint = ins.objectid + ins.offset;
 863                 kfree(async_extent);
 864                 cond_resched();
 865         }
 866         return;
 867 out_free_reserve:
 868         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 869         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 870 out_free:
 871         extent_clear_unlock_delalloc(inode, async_extent->start,
 872                                      async_extent->start +
 873                                      async_extent->ram_size - 1,
 874                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 875                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 876                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 877                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 878                                      PAGE_SET_ERROR);
 879         free_async_extent_pages(async_extent);
 880         kfree(async_extent);
 881         goto again;
 882 }
 883
 884 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 885                                       u64 num_bytes)
 886 {
 887         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 888         struct extent_map *em;
 889         u64 alloc_hint = 0;
 890
 891         read_lock(&em_tree->lock);
 892         em = search_extent_mapping(em_tree, start, num_bytes);
 893         if (em) {
 894                 /*
 895                  * if block start isn't an actual block number then find the
 896                  * first block in this inode and use that as a hint.  If that
 897                  * block is also bogus then just don't worry about it.
 898                  */
 899                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 900                         free_extent_map(em);
 901                         em = search_extent_mapping(em_tree, 0, 0);
 902                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 903                                 alloc_hint = em->block_start;
 904                         if (em)
 905                                 free_extent_map(em);
 906                 } else {
 907                         alloc_hint = em->block_start;
 908                         free_extent_map(em);
 909                 }
 910         }
 911         read_unlock(&em_tree->lock);
 912
 913         return alloc_hint;
 914 }
 915
 916 /*
 917  * when extent_io.c finds a delayed allocation range in the file,
 918  * the call backs end up in this code.  The basic idea is to
 919  * allocate extents on disk for the range, and create ordered data structs
 920  * in ram to track those extents.
 921  *
 922  * locked_page is the page that writepage had locked already.  We use
 923  * it to make sure we don't do extra locks or unlocks.
 924  *
 925  * *page_started is set to one if we unlock locked_page and do everything
 926  * required to start IO on it.  It may be clean and already done with
 927  * IO when we return.
 928  */
 929 static noinline int cow_file_range(struct inode *inode,
 930                                    struct page *locked_page,
 931                                    u64 start, u64 end, u64 delalloc_end,
 932                                    int *page_started, unsigned long *nr_written,
 933                                    int unlock, struct btrfs_dedupe_hash *hash)
 934 {
 935         struct btrfs_root *root = BTRFS_I(inode)->root;
 936         u64 alloc_hint = 0;
 937         u64 num_bytes;
 938         unsigned long ram_size;
 939         u64 disk_num_bytes;
 940         u64 cur_alloc_size;
 941         u64 blocksize = root->sectorsize;
 942         struct btrfs_key ins;
 943         struct extent_map *em;
 944         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 945         int ret = 0;
 946
 947         if (btrfs_is_free_space_inode(inode)) {
 948                 WARN_ON_ONCE(1);
 949                 ret = -EINVAL;
 950                 goto out_unlock;
 951         }
 952
 953         num_bytes = ALIGN(end - start + 1, blocksize);
 954         num_bytes = max(blocksize,  num_bytes);
 955         disk_num_bytes = num_bytes;
 956
 957         /* if this is a small write inside eof, kick off defrag */
 958         if (num_bytes < SZ_64K &&
 959             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 960                 btrfs_add_inode_defrag(NULL, inode);
 961
 962         if (start == 0) {
 963                 /* lets try to make an inline extent */
 964                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 965                                             NULL);
 966                 if (ret == 0) {
 967                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 968                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 969                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 970                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 971                                      PAGE_END_WRITEBACK);
 972
 973                         *nr_written = *nr_written +
 974                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 975                         *page_started = 1;
 976                         goto out;
 977                 } else if (ret < 0) {
 978                         goto out_unlock;
 979                 }
 980         }
 981
 982         BUG_ON(disk_num_bytes >
 983                btrfs_super_total_bytes(root->fs_info->super_copy));
 984
 985         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 986         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 987
 988         while (disk_num_bytes > 0) {
 989                 unsigned long op;
 990
 991                 cur_alloc_size = disk_num_bytes;
 992                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 993                                            root->sectorsize, 0, alloc_hint,
 994                                            &ins, 1, 1);
 995                 if (ret < 0)
 996                         goto out_unlock;
 997
 998                 em = alloc_extent_map();
 999                 if (!em) {
1000                         ret = -ENOMEM;
1001                         goto out_reserve;
1002                 }
1003                 em->start = start;
1004                 em->orig_start = em->start;
1005                 ram_size = ins.offset;
1006                 em->len = ins.offset;
1007                 em->mod_start = em->start;
1008                 em->mod_len = em->len;
1009
1010                 em->block_start = ins.objectid;
1011                 em->block_len = ins.offset;
1012                 em->orig_block_len = ins.offset;
1013                 em->ram_bytes = ram_size;
1014                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1015                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1016                 em->generation = -1;
1017
1018                 while (1) {
1019                         write_lock(&em_tree->lock);
1020                         ret = add_extent_mapping(em_tree, em, 1);
1021                         write_unlock(&em_tree->lock);
1022                         if (ret != -EEXIST) {
1023                                 free_extent_map(em);
1024                                 break;
1025                         }
1026                         btrfs_drop_extent_cache(inode, start,
1027                                                 start + ram_size - 1, 0);
1028                 }
1029                 if (ret)
1030                         goto out_reserve;
1031
1032                 cur_alloc_size = ins.offset;
1033                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1034                                                ram_size, cur_alloc_size, 0);
1035                 if (ret)
1036                         goto out_drop_extent_cache;
1037
1038                 if (root->root_key.objectid ==
1039                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1040                         ret = btrfs_reloc_clone_csums(inode, start,
1041                                                       cur_alloc_size);
1042                         if (ret)
1043                                 goto out_drop_extent_cache;
1044                 }
1045
1046                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1047
1048                 if (disk_num_bytes < cur_alloc_size)
1049                         break;
1050
1051                 /* we're not doing compressed IO, don't unlock the first
1052                  * page (which the caller expects to stay locked), don't
1053                  * clear any dirty bits and don't set any writeback bits
1054                  *
1055                  * Do set the Private2 bit so we know this page was properly
1056                  * setup for writepage
1057                  */
1058                 op = unlock ? PAGE_UNLOCK : 0;
1059                 op |= PAGE_SET_PRIVATE2;
1060
1061                 extent_clear_unlock_delalloc(inode, start,
1062                                              start + ram_size - 1, locked_page,
1063                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1064                                              op);
1065                 disk_num_bytes -= cur_alloc_size;
1066                 num_bytes -= cur_alloc_size;
1067                 alloc_hint = ins.objectid + ins.offset;
1068                 start += cur_alloc_size;
1069         }
1070 out:
1071         return ret;
1072
1073 out_drop_extent_cache:
1074         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1075 out_reserve:
1076         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1077         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1078 out_unlock:
1079         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1080                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1081                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1082                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1083                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1084         goto out;
1085 }
1086
1087 /*
1088  * work queue call back to started compression on a file and pages
1089  */
1090 static noinline void async_cow_start(struct btrfs_work *work)
1091 {
1092         struct async_cow *async_cow;
1093         int num_added = 0;
1094         async_cow = container_of(work, struct async_cow, work);
1095
1096         compress_file_range(async_cow->inode, async_cow->locked_page,
1097                             async_cow->start, async_cow->end, async_cow,
1098                             &num_added);
1099         if (num_added == 0) {
1100                 btrfs_add_delayed_iput(async_cow->inode);
1101                 async_cow->inode = NULL;
1102         }
1103 }
1104
1105 /*
1106  * work queue call back to submit previously compressed pages
1107  */
1108 static noinline void async_cow_submit(struct btrfs_work *work)
1109 {
1110         struct async_cow *async_cow;
1111         struct btrfs_root *root;
1112         unsigned long nr_pages;
1113
1114         async_cow = container_of(work, struct async_cow, work);
1115
1116         root = async_cow->root;
1117         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1118                 PAGE_SHIFT;
1119
1120         /*
1121          * atomic_sub_return implies a barrier for waitqueue_active
1122          */
1123         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1124             5 * SZ_1M &&
1125             waitqueue_active(&root->fs_info->async_submit_wait))
1126                 wake_up(&root->fs_info->async_submit_wait);
1127
1128         if (async_cow->inode)
1129                 submit_compressed_extents(async_cow->inode, async_cow);
1130 }
1131
1132 static noinline void async_cow_free(struct btrfs_work *work)
1133 {
1134         struct async_cow *async_cow;
1135         async_cow = container_of(work, struct async_cow, work);
1136         if (async_cow->inode)
1137                 btrfs_add_delayed_iput(async_cow->inode);
1138         kfree(async_cow);
1139 }
1140
1141 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1142                                 u64 start, u64 end, int *page_started,
1143                                 unsigned long *nr_written)
1144 {
1145         struct async_cow *async_cow;
1146         struct btrfs_root *root = BTRFS_I(inode)->root;
1147         unsigned long nr_pages;
1148         u64 cur_end;
1149         int limit = 10 * SZ_1M;
1150
1151         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1152                          1, 0, NULL, GFP_NOFS);
1153         while (start < end) {
1154                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1155                 BUG_ON(!async_cow); /* -ENOMEM */
1156                 async_cow->inode = igrab(inode);
1157                 async_cow->root = root;
1158                 async_cow->locked_page = locked_page;
1159                 async_cow->start = start;
1160
1161                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1162                     !btrfs_test_opt(root->fs_info, FORCE_COMPRESS))
1163                         cur_end = end;
1164                 else
1165                         cur_end = min(end, start + SZ_512K - 1);
1166
1167                 async_cow->end = cur_end;
1168                 INIT_LIST_HEAD(&async_cow->extents);
1169
1170                 btrfs_init_work(&async_cow->work,
1171                                 btrfs_delalloc_helper,
1172                                 async_cow_start, async_cow_submit,
1173                                 async_cow_free);
1174
1175                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1176                         PAGE_SHIFT;
1177                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1178
1179                 btrfs_queue_work(root->fs_info->delalloc_workers,
1180                                  &async_cow->work);
1181
1182                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1183                         wait_event(root->fs_info->async_submit_wait,
1184                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1185                             limit));
1186                 }
1187
1188                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1189                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1190                         wait_event(root->fs_info->async_submit_wait,
1191                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1192                            0));
1193                 }
1194
1195                 *nr_written += nr_pages;
1196                 start = cur_end + 1;
1197         }
1198         *page_started = 1;
1199         return 0;
1200 }
1201
1202 static noinline int csum_exist_in_range(struct btrfs_root *root,
1203                                         u64 bytenr, u64 num_bytes)
1204 {
1205         int ret;
1206         struct btrfs_ordered_sum *sums;
1207         LIST_HEAD(list);
1208
1209         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1210                                        bytenr + num_bytes - 1, &list, 0);
1211         if (ret == 0 && list_empty(&list))
1212                 return 0;
1213
1214         while (!list_empty(&list)) {
1215                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1216                 list_del(&sums->list);
1217                 kfree(sums);
1218         }
1219         return 1;
1220 }
1221
1222 /*
1223  * when nowcow writeback call back.  This checks for snapshots or COW copies
1224  * of the extents that exist in the file, and COWs the file as required.
1225  *
1226  * If no cow copies or snapshots exist, we write directly to the existing
1227  * blocks on disk
1228  */
1229 static noinline int run_delalloc_nocow(struct inode *inode,
1230                                        struct page *locked_page,
1231                               u64 start, u64 end, int *page_started, int force,
1232                               unsigned long *nr_written)
1233 {
1234         struct btrfs_root *root = BTRFS_I(inode)->root;
1235         struct btrfs_trans_handle *trans;
1236         struct extent_buffer *leaf;
1237         struct btrfs_path *path;
1238         struct btrfs_file_extent_item *fi;
1239         struct btrfs_key found_key;
1240         u64 cow_start;
1241         u64 cur_offset;
1242         u64 extent_end;
1243         u64 extent_offset;
1244         u64 disk_bytenr;
1245         u64 num_bytes;
1246         u64 disk_num_bytes;
1247         u64 ram_bytes;
1248         int extent_type;
1249         int ret, err;
1250         int type;
1251         int nocow;
1252         int check_prev = 1;
1253         bool nolock;
1254         u64 ino = btrfs_ino(inode);
1255
1256         path = btrfs_alloc_path();
1257         if (!path) {
1258                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1259                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1260                                              EXTENT_DO_ACCOUNTING |
1261                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1262                                              PAGE_CLEAR_DIRTY |
1263                                              PAGE_SET_WRITEBACK |
1264                                              PAGE_END_WRITEBACK);
1265                 return -ENOMEM;
1266         }
1267
1268         nolock = btrfs_is_free_space_inode(inode);
1269
1270         if (nolock)
1271                 trans = btrfs_join_transaction_nolock(root);
1272         else
1273                 trans = btrfs_join_transaction(root);
1274
1275         if (IS_ERR(trans)) {
1276                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1277                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1278                                              EXTENT_DO_ACCOUNTING |
1279                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1280                                              PAGE_CLEAR_DIRTY |
1281                                              PAGE_SET_WRITEBACK |
1282                                              PAGE_END_WRITEBACK);
1283                 btrfs_free_path(path);
1284                 return PTR_ERR(trans);
1285         }
1286
1287         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1288
1289         cow_start = (u64)-1;
1290         cur_offset = start;
1291         while (1) {
1292                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1293                                                cur_offset, 0);
1294                 if (ret < 0)
1295                         goto error;
1296                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1297                         leaf = path->nodes[0];
1298                         btrfs_item_key_to_cpu(leaf, &found_key,
1299                                               path->slots[0] - 1);
1300                         if (found_key.objectid == ino &&
1301                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1302                                 path->slots[0]--;
1303                 }
1304                 check_prev = 0;
1305 next_slot:
1306                 leaf = path->nodes[0];
1307                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1308                         ret = btrfs_next_leaf(root, path);
1309                         if (ret < 0)
1310                                 goto error;
1311                         if (ret > 0)
1312                                 break;
1313                         leaf = path->nodes[0];
1314                 }
1315
1316                 nocow = 0;
1317                 disk_bytenr = 0;
1318                 num_bytes = 0;
1319                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1320
1321                 if (found_key.objectid > ino)
1322                         break;
1323                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1324                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1325                         path->slots[0]++;
1326                         goto next_slot;
1327                 }
1328                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1329                     found_key.offset > end)
1330                         break;
1331
1332                 if (found_key.offset > cur_offset) {
1333                         extent_end = found_key.offset;
1334                         extent_type = 0;
1335                         goto out_check;
1336                 }
1337
1338                 fi = btrfs_item_ptr(leaf, path->slots[0],
1339                                     struct btrfs_file_extent_item);
1340                 extent_type = btrfs_file_extent_type(leaf, fi);
1341
1342                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1343                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1344                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1345                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1346                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1347                         extent_end = found_key.offset +
1348                                 btrfs_file_extent_num_bytes(leaf, fi);
1349                         disk_num_bytes =
1350                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1351                         if (extent_end <= start) {
1352                                 path->slots[0]++;
1353                                 goto next_slot;
1354                         }
1355                         if (disk_bytenr == 0)
1356                                 goto out_check;
1357                         if (btrfs_file_extent_compression(leaf, fi) ||
1358                             btrfs_file_extent_encryption(leaf, fi) ||
1359                             btrfs_file_extent_other_encoding(leaf, fi))
1360                                 goto out_check;
1361                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1362                                 goto out_check;
1363                         if (btrfs_extent_readonly(root, disk_bytenr))
1364                                 goto out_check;
1365                         if (btrfs_cross_ref_exist(trans, root, ino,
1366                                                   found_key.offset -
1367                                                   extent_offset, disk_bytenr))
1368                                 goto out_check;
1369                         disk_bytenr += extent_offset;
1370                         disk_bytenr += cur_offset - found_key.offset;
1371                         num_bytes = min(end + 1, extent_end) - cur_offset;
1372                         /*
1373                          * if there are pending snapshots for this root,
1374                          * we fall into common COW way.
1375                          */
1376                         if (!nolock) {
1377                                 err = btrfs_start_write_no_snapshoting(root);
1378                                 if (!err)
1379                                         goto out_check;
1380                         }
1381                         /*
1382                          * force cow if csum exists in the range.
1383                          * this ensure that csum for a given extent are
1384                          * either valid or do not exist.
1385                          */
1386                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1387                                 goto out_check;
1388                         if (!btrfs_inc_nocow_writers(root->fs_info,
1389                                                      disk_bytenr))
1390                                 goto out_check;
1391                         nocow = 1;
1392                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1393                         extent_end = found_key.offset +
1394                                 btrfs_file_extent_inline_len(leaf,
1395                                                      path->slots[0], fi);
1396                         extent_end = ALIGN(extent_end, root->sectorsize);
1397                 } else {
1398                         BUG_ON(1);
1399                 }
1400 out_check:
1401                 if (extent_end <= start) {
1402                         path->slots[0]++;
1403                         if (!nolock && nocow)
1404                                 btrfs_end_write_no_snapshoting(root);
1405                         if (nocow)
1406                                 btrfs_dec_nocow_writers(root->fs_info,
1407                                                         disk_bytenr);
1408                         goto next_slot;
1409                 }
1410                 if (!nocow) {
1411                         if (cow_start == (u64)-1)
1412                                 cow_start = cur_offset;
1413                         cur_offset = extent_end;
1414                         if (cur_offset > end)
1415                                 break;
1416                         path->slots[0]++;
1417                         goto next_slot;
1418                 }
1419
1420                 btrfs_release_path(path);
1421                 if (cow_start != (u64)-1) {
1422                         ret = cow_file_range(inode, locked_page,
1423                                              cow_start, found_key.offset - 1,
1424                                              end, page_started, nr_written, 1,
1425                                              NULL);
1426                         if (ret) {
1427                                 if (!nolock && nocow)
1428                                         btrfs_end_write_no_snapshoting(root);
1429                                 if (nocow)
1430                                         btrfs_dec_nocow_writers(root->fs_info,
1431                                                                 disk_bytenr);
1432                                 goto error;
1433                         }
1434                         cow_start = (u64)-1;
1435                 }
1436
1437                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1438                         struct extent_map *em;
1439                         struct extent_map_tree *em_tree;
1440                         em_tree = &BTRFS_I(inode)->extent_tree;
1441                         em = alloc_extent_map();
1442                         BUG_ON(!em); /* -ENOMEM */
1443                         em->start = cur_offset;
1444                         em->orig_start = found_key.offset - extent_offset;
1445                         em->len = num_bytes;
1446                         em->block_len = num_bytes;
1447                         em->block_start = disk_bytenr;
1448                         em->orig_block_len = disk_num_bytes;
1449                         em->ram_bytes = ram_bytes;
1450                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1451                         em->mod_start = em->start;
1452                         em->mod_len = em->len;
1453                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1454                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1455                         em->generation = -1;
1456                         while (1) {
1457                                 write_lock(&em_tree->lock);
1458                                 ret = add_extent_mapping(em_tree, em, 1);
1459                                 write_unlock(&em_tree->lock);
1460                                 if (ret != -EEXIST) {
1461                                         free_extent_map(em);
1462                                         break;
1463                                 }
1464                                 btrfs_drop_extent_cache(inode, em->start,
1465                                                 em->start + em->len - 1, 0);
1466                         }
1467                         type = BTRFS_ORDERED_PREALLOC;
1468                 } else {
1469                         type = BTRFS_ORDERED_NOCOW;
1470                 }
1471
1472                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1473                                                num_bytes, num_bytes, type);
1474                 if (nocow)
1475                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1476                 BUG_ON(ret); /* -ENOMEM */
1477
1478                 if (root->root_key.objectid ==
1479                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1480                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1481                                                       num_bytes);
1482                         if (ret) {
1483                                 if (!nolock && nocow)
1484                                         btrfs_end_write_no_snapshoting(root);
1485                                 goto error;
1486                         }
1487                 }
1488
1489                 extent_clear_unlock_delalloc(inode, cur_offset,
1490                                              cur_offset + num_bytes - 1,
1491                                              locked_page, EXTENT_LOCKED |
1492                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1493                                              PAGE_SET_PRIVATE2);
1494                 if (!nolock && nocow)
1495                         btrfs_end_write_no_snapshoting(root);
1496                 cur_offset = extent_end;
1497                 if (cur_offset > end)
1498                         break;
1499         }
1500         btrfs_release_path(path);
1501
1502         if (cur_offset <= end && cow_start == (u64)-1) {
1503                 cow_start = cur_offset;
1504                 cur_offset = end;
1505         }
1506
1507         if (cow_start != (u64)-1) {
1508                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1509                                      page_started, nr_written, 1, NULL);
1510                 if (ret)
1511                         goto error;
1512         }
1513
1514 error:
1515         err = btrfs_end_transaction(trans, root);
1516         if (!ret)
1517                 ret = err;
1518
1519         if (ret && cur_offset < end)
1520                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1521                                              locked_page, EXTENT_LOCKED |
1522                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1523                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1524                                              PAGE_CLEAR_DIRTY |
1525                                              PAGE_SET_WRITEBACK |
1526                                              PAGE_END_WRITEBACK);
1527         btrfs_free_path(path);
1528         return ret;
1529 }
1530
1531 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1532 {
1533
1534         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1535             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1536                 return 0;
1537
1538         /*
1539          * @defrag_bytes is a hint value, no spinlock held here,
1540          * if is not zero, it means the file is defragging.
1541          * Force cow if given extent needs to be defragged.
1542          */
1543         if (BTRFS_I(inode)->defrag_bytes &&
1544             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1545                            EXTENT_DEFRAG, 0, NULL))
1546                 return 1;
1547
1548         return 0;
1549 }
1550
1551 /*
1552  * extent_io.c call back to do delayed allocation processing
1553  */
1554 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1555                               u64 start, u64 end, int *page_started,
1556                               unsigned long *nr_written)
1557 {
1558         int ret;
1559         int force_cow = need_force_cow(inode, start, end);
1560
1561         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1562                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1563                                          page_started, 1, nr_written);
1564         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1565                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1566                                          page_started, 0, nr_written);
1567         } else if (!inode_need_compress(inode)) {
1568                 ret = cow_file_range(inode, locked_page, start, end, end,
1569                                       page_started, nr_written, 1, NULL);
1570         } else {
1571                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1572                         &BTRFS_I(inode)->runtime_flags);
1573                 ret = cow_file_range_async(inode, locked_page, start, end,
1574                                            page_started, nr_written);
1575         }
1576         return ret;
1577 }
1578
1579 static void btrfs_split_extent_hook(struct inode *inode,
1580                                     struct extent_state *orig, u64 split)
1581 {
1582         u64 size;
1583
1584         /* not delalloc, ignore it */
1585         if (!(orig->state & EXTENT_DELALLOC))
1586                 return;
1587
1588         size = orig->end - orig->start + 1;
1589         if (size > BTRFS_MAX_EXTENT_SIZE) {
1590                 u64 num_extents;
1591                 u64 new_size;
1592
1593                 /*
1594                  * See the explanation in btrfs_merge_extent_hook, the same
1595                  * applies here, just in reverse.
1596                  */
1597                 new_size = orig->end - split + 1;
1598                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1599                                         BTRFS_MAX_EXTENT_SIZE);
1600                 new_size = split - orig->start;
1601                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1602                                         BTRFS_MAX_EXTENT_SIZE);
1603                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1604                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1605                         return;
1606         }
1607
1608         spin_lock(&BTRFS_I(inode)->lock);
1609         BTRFS_I(inode)->outstanding_extents++;
1610         spin_unlock(&BTRFS_I(inode)->lock);
1611 }
1612
1613 /*
1614  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1615  * extents so we can keep track of new extents that are just merged onto old
1616  * extents, such as when we are doing sequential writes, so we can properly
1617  * account for the metadata space we'll need.
1618  */
1619 static void btrfs_merge_extent_hook(struct inode *inode,
1620                                     struct extent_state *new,
1621                                     struct extent_state *other)
1622 {
1623         u64 new_size, old_size;
1624         u64 num_extents;
1625
1626         /* not delalloc, ignore it */
1627         if (!(other->state & EXTENT_DELALLOC))
1628                 return;
1629
1630         if (new->start > other->start)
1631                 new_size = new->end - other->start + 1;
1632         else
1633                 new_size = other->end - new->start + 1;
1634
1635         /* we're not bigger than the max, unreserve the space and go */
1636         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1637                 spin_lock(&BTRFS_I(inode)->lock);
1638                 BTRFS_I(inode)->outstanding_extents--;
1639                 spin_unlock(&BTRFS_I(inode)->lock);
1640                 return;
1641         }
1642
1643         /*
1644          * We have to add up either side to figure out how many extents were
1645          * accounted for before we merged into one big extent.  If the number of
1646          * extents we accounted for is <= the amount we need for the new range
1647          * then we can return, otherwise drop.  Think of it like this
1648          *
1649          * [ 4k][MAX_SIZE]
1650          *
1651          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1652          * need 2 outstanding extents, on one side we have 1 and the other side
1653          * we have 1 so they are == and we can return.  But in this case
1654          *
1655          * [MAX_SIZE+4k][MAX_SIZE+4k]
1656          *
1657          * Each range on their own accounts for 2 extents, but merged together
1658          * they are only 3 extents worth of accounting, so we need to drop in
1659          * this case.
1660          */
1661         old_size = other->end - other->start + 1;
1662         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1663                                 BTRFS_MAX_EXTENT_SIZE);
1664         old_size = new->end - new->start + 1;
1665         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1666                                  BTRFS_MAX_EXTENT_SIZE);
1667
1668         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1669                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1670                 return;
1671
1672         spin_lock(&BTRFS_I(inode)->lock);
1673         BTRFS_I(inode)->outstanding_extents--;
1674         spin_unlock(&BTRFS_I(inode)->lock);
1675 }
1676
1677 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1678                                       struct inode *inode)
1679 {
1680         spin_lock(&root->delalloc_lock);
1681         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1682                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1683                               &root->delalloc_inodes);
1684                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1685                         &BTRFS_I(inode)->runtime_flags);
1686                 root->nr_delalloc_inodes++;
1687                 if (root->nr_delalloc_inodes == 1) {
1688                         spin_lock(&root->fs_info->delalloc_root_lock);
1689                         BUG_ON(!list_empty(&root->delalloc_root));
1690                         list_add_tail(&root->delalloc_root,
1691                                       &root->fs_info->delalloc_roots);
1692                         spin_unlock(&root->fs_info->delalloc_root_lock);
1693                 }
1694         }
1695         spin_unlock(&root->delalloc_lock);
1696 }
1697
1698 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1699                                      struct inode *inode)
1700 {
1701         spin_lock(&root->delalloc_lock);
1702         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1703                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1704                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1705                           &BTRFS_I(inode)->runtime_flags);
1706                 root->nr_delalloc_inodes--;
1707                 if (!root->nr_delalloc_inodes) {
1708                         spin_lock(&root->fs_info->delalloc_root_lock);
1709                         BUG_ON(list_empty(&root->delalloc_root));
1710                         list_del_init(&root->delalloc_root);
1711                         spin_unlock(&root->fs_info->delalloc_root_lock);
1712                 }
1713         }
1714         spin_unlock(&root->delalloc_lock);
1715 }
1716
1717 /*
1718  * extent_io.c set_bit_hook, used to track delayed allocation
1719  * bytes in this file, and to maintain the list of inodes that
1720  * have pending delalloc work to be done.
1721  */
1722 static void btrfs_set_bit_hook(struct inode *inode,
1723                                struct extent_state *state, unsigned *bits)
1724 {
1725
1726         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1727                 WARN_ON(1);
1728         /*
1729          * set_bit and clear bit hooks normally require _irqsave/restore
1730          * but in this case, we are only testing for the DELALLOC
1731          * bit, which is only set or cleared with irqs on
1732          */
1733         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1734                 struct btrfs_root *root = BTRFS_I(inode)->root;
1735                 u64 len = state->end + 1 - state->start;
1736                 bool do_list = !btrfs_is_free_space_inode(inode);
1737
1738                 if (*bits & EXTENT_FIRST_DELALLOC) {
1739                         *bits &= ~EXTENT_FIRST_DELALLOC;
1740                 } else {
1741                         spin_lock(&BTRFS_I(inode)->lock);
1742                         BTRFS_I(inode)->outstanding_extents++;
1743                         spin_unlock(&BTRFS_I(inode)->lock);
1744                 }
1745
1746                 /* For sanity tests */
1747                 if (btrfs_is_testing(root->fs_info))
1748                         return;
1749
1750                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1751                                      root->fs_info->delalloc_batch);
1752                 spin_lock(&BTRFS_I(inode)->lock);
1753                 BTRFS_I(inode)->delalloc_bytes += len;
1754                 if (*bits & EXTENT_DEFRAG)
1755                         BTRFS_I(inode)->defrag_bytes += len;
1756                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1757                                          &BTRFS_I(inode)->runtime_flags))
1758                         btrfs_add_delalloc_inodes(root, inode);
1759                 spin_unlock(&BTRFS_I(inode)->lock);
1760         }
1761 }
1762
1763 /*
1764  * extent_io.c clear_bit_hook, see set_bit_hook for why
1765  */
1766 static void btrfs_clear_bit_hook(struct inode *inode,
1767                                  struct extent_state *state,
1768                                  unsigned *bits)
1769 {
1770         u64 len = state->end + 1 - state->start;
1771         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1772                                     BTRFS_MAX_EXTENT_SIZE);
1773
1774         spin_lock(&BTRFS_I(inode)->lock);
1775         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1776                 BTRFS_I(inode)->defrag_bytes -= len;
1777         spin_unlock(&BTRFS_I(inode)->lock);
1778
1779         /*
1780          * set_bit and clear bit hooks normally require _irqsave/restore
1781          * but in this case, we are only testing for the DELALLOC
1782          * bit, which is only set or cleared with irqs on
1783          */
1784         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1785                 struct btrfs_root *root = BTRFS_I(inode)->root;
1786                 bool do_list = !btrfs_is_free_space_inode(inode);
1787
1788                 if (*bits & EXTENT_FIRST_DELALLOC) {
1789                         *bits &= ~EXTENT_FIRST_DELALLOC;
1790                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1791                         spin_lock(&BTRFS_I(inode)->lock);
1792                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1793                         spin_unlock(&BTRFS_I(inode)->lock);
1794                 }
1795
1796                 /*
1797                  * We don't reserve metadata space for space cache inodes so we
1798                  * don't need to call dellalloc_release_metadata if there is an
1799                  * error.
1800                  */
1801                 if (*bits & EXTENT_DO_ACCOUNTING &&
1802                     root != root->fs_info->tree_root)
1803                         btrfs_delalloc_release_metadata(inode, len);
1804
1805                 /* For sanity tests. */
1806                 if (btrfs_is_testing(root->fs_info))
1807                         return;
1808
1809                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1810                     && do_list && !(state->state & EXTENT_NORESERVE))
1811                         btrfs_free_reserved_data_space_noquota(inode,
1812                                         state->start, len);
1813
1814                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1815                                      root->fs_info->delalloc_batch);
1816                 spin_lock(&BTRFS_I(inode)->lock);
1817                 BTRFS_I(inode)->delalloc_bytes -= len;
1818                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1819                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1820                              &BTRFS_I(inode)->runtime_flags))
1821                         btrfs_del_delalloc_inode(root, inode);
1822                 spin_unlock(&BTRFS_I(inode)->lock);
1823         }
1824 }
1825
1826 /*
1827  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1828  * we don't create bios that span stripes or chunks
1829  *
1830  * return 1 if page cannot be merged to bio
1831  * return 0 if page can be merged to bio
1832  * return error otherwise
1833  */
1834 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1835                          size_t size, struct bio *bio,
1836                          unsigned long bio_flags)
1837 {
1838         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1839         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1840         u64 length = 0;
1841         u64 map_length;
1842         int ret;
1843
1844         if (bio_flags & EXTENT_BIO_COMPRESSED)
1845                 return 0;
1846
1847         length = bio->bi_iter.bi_size;
1848         map_length = length;
1849         ret = btrfs_map_block(root->fs_info, rw, logical,
1850                               &map_length, NULL, 0);
1851         if (ret < 0)
1852                 return ret;
1853         if (map_length < length + size)
1854                 return 1;
1855         return 0;
1856 }
1857
1858 /*
1859  * in order to insert checksums into the metadata in large chunks,
1860  * we wait until bio submission time.   All the pages in the bio are
1861  * checksummed and sums are attached onto the ordered extent record.
1862  *
1863  * At IO completion time the cums attached on the ordered extent record
1864  * are inserted into the btree
1865  */
1866 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1867                                     struct bio *bio, int mirror_num,
1868                                     unsigned long bio_flags,
1869                                     u64 bio_offset)
1870 {
1871         struct btrfs_root *root = BTRFS_I(inode)->root;
1872         int ret = 0;
1873
1874         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1875         BUG_ON(ret); /* -ENOMEM */
1876         return 0;
1877 }
1878
1879 /*
1880  * in order to insert checksums into the metadata in large chunks,
1881  * we wait until bio submission time.   All the pages in the bio are
1882  * checksummed and sums are attached onto the ordered extent record.
1883  *
1884  * At IO completion time the cums attached on the ordered extent record
1885  * are inserted into the btree
1886  */
1887 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1888                           int mirror_num, unsigned long bio_flags,
1889                           u64 bio_offset)
1890 {
1891         struct btrfs_root *root = BTRFS_I(inode)->root;
1892         int ret;
1893
1894         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1895         if (ret) {
1896                 bio->bi_error = ret;
1897                 bio_endio(bio);
1898         }
1899         return ret;
1900 }
1901
1902 /*
1903  * extent_io.c submission hook. This does the right thing for csum calculation
1904  * on write, or reading the csums from the tree before a read
1905  */
1906 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1907                           int mirror_num, unsigned long bio_flags,
1908                           u64 bio_offset)
1909 {
1910         struct btrfs_root *root = BTRFS_I(inode)->root;
1911         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1912         int ret = 0;
1913         int skip_sum;
1914         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1915
1916         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1917
1918         if (btrfs_is_free_space_inode(inode))
1919                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1920
1921         if (!(rw & REQ_WRITE)) {
1922                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1923                 if (ret)
1924                         goto out;
1925
1926                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1927                         ret = btrfs_submit_compressed_read(inode, bio,
1928                                                            mirror_num,
1929                                                            bio_flags);
1930                         goto out;
1931                 } else if (!skip_sum) {
1932                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1933                         if (ret)
1934                                 goto out;
1935                 }
1936                 goto mapit;
1937         } else if (async && !skip_sum) {
1938                 /* csum items have already been cloned */
1939                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1940                         goto mapit;
1941                 /* we're doing a write, do the async checksumming */
1942                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1943                                    inode, rw, bio, mirror_num,
1944                                    bio_flags, bio_offset,
1945                                    __btrfs_submit_bio_start,
1946                                    __btrfs_submit_bio_done);
1947                 goto out;
1948         } else if (!skip_sum) {
1949                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1950                 if (ret)
1951                         goto out;
1952         }
1953
1954 mapit:
1955         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1956
1957 out:
1958         if (ret < 0) {
1959                 bio->bi_error = ret;
1960                 bio_endio(bio);
1961         }
1962         return ret;
1963 }
1964
1965 /*
1966  * given a list of ordered sums record them in the inode.  This happens
1967  * at IO completion time based on sums calculated at bio submission time.
1968  */
1969 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1970                              struct inode *inode, u64 file_offset,
1971                              struct list_head *list)
1972 {
1973         struct btrfs_ordered_sum *sum;
1974
1975         list_for_each_entry(sum, list, list) {
1976                 trans->adding_csums = 1;
1977                 btrfs_csum_file_blocks(trans,
1978                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1979                 trans->adding_csums = 0;
1980         }
1981         return 0;
1982 }
1983
1984 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1985                               struct extent_state **cached_state)
1986 {
1987         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1988         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1989                                    cached_state);
1990 }
1991
1992 /* see btrfs_writepage_start_hook for details on why this is required */
1993 struct btrfs_writepage_fixup {
1994         struct page *page;
1995         struct btrfs_work work;
1996 };
1997
1998 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1999 {
2000         struct btrfs_writepage_fixup *fixup;
2001         struct btrfs_ordered_extent *ordered;
2002         struct extent_state *cached_state = NULL;
2003         struct page *page;
2004         struct inode *inode;
2005         u64 page_start;
2006         u64 page_end;
2007         int ret;
2008
2009         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2010         page = fixup->page;
2011 again:
2012         lock_page(page);
2013         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2014                 ClearPageChecked(page);
2015                 goto out_page;
2016         }
2017
2018         inode = page->mapping->host;
2019         page_start = page_offset(page);
2020         page_end = page_offset(page) + PAGE_SIZE - 1;
2021
2022         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2023                          &cached_state);
2024
2025         /* already ordered? We're done */
2026         if (PagePrivate2(page))
2027                 goto out;
2028
2029         ordered = btrfs_lookup_ordered_range(inode, page_start,
2030                                         PAGE_SIZE);
2031         if (ordered) {
2032                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2033                                      page_end, &cached_state, GFP_NOFS);
2034                 unlock_page(page);
2035                 btrfs_start_ordered_extent(inode, ordered, 1);
2036                 btrfs_put_ordered_extent(ordered);
2037                 goto again;
2038         }
2039
2040         ret = btrfs_delalloc_reserve_space(inode, page_start,
2041                                            PAGE_SIZE);
2042         if (ret) {
2043                 mapping_set_error(page->mapping, ret);
2044                 end_extent_writepage(page, ret, page_start, page_end);
2045                 ClearPageChecked(page);
2046                 goto out;
2047          }
2048
2049         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2050         ClearPageChecked(page);
2051         set_page_dirty(page);
2052 out:
2053         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2054                              &cached_state, GFP_NOFS);
2055 out_page:
2056         unlock_page(page);
2057         put_page(page);
2058         kfree(fixup);
2059 }
2060
2061 /*
2062  * There are a few paths in the higher layers of the kernel that directly
2063  * set the page dirty bit without asking the filesystem if it is a
2064  * good idea.  This causes problems because we want to make sure COW
2065  * properly happens and the data=ordered rules are followed.
2066  *
2067  * In our case any range that doesn't have the ORDERED bit set
2068  * hasn't been properly setup for IO.  We kick off an async process
2069  * to fix it up.  The async helper will wait for ordered extents, set
2070  * the delalloc bit and make it safe to write the page.
2071  */
2072 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2073 {
2074         struct inode *inode = page->mapping->host;
2075         struct btrfs_writepage_fixup *fixup;
2076         struct btrfs_root *root = BTRFS_I(inode)->root;
2077
2078         /* this page is properly in the ordered list */
2079         if (TestClearPagePrivate2(page))
2080                 return 0;
2081
2082         if (PageChecked(page))
2083                 return -EAGAIN;
2084
2085         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2086         if (!fixup)
2087                 return -EAGAIN;
2088
2089         SetPageChecked(page);
2090         get_page(page);
2091         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2092                         btrfs_writepage_fixup_worker, NULL, NULL);
2093         fixup->page = page;
2094         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2095         return -EBUSY;
2096 }
2097
2098 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2099                                        struct inode *inode, u64 file_pos,
2100                                        u64 disk_bytenr, u64 disk_num_bytes,
2101                                        u64 num_bytes, u64 ram_bytes,
2102                                        u8 compression, u8 encryption,
2103                                        u16 other_encoding, int extent_type)
2104 {
2105         struct btrfs_root *root = BTRFS_I(inode)->root;
2106         struct btrfs_file_extent_item *fi;
2107         struct btrfs_path *path;
2108         struct extent_buffer *leaf;
2109         struct btrfs_key ins;
2110         int extent_inserted = 0;
2111         int ret;
2112
2113         path = btrfs_alloc_path();
2114         if (!path)
2115                 return -ENOMEM;
2116
2117         /*
2118          * we may be replacing one extent in the tree with another.
2119          * The new extent is pinned in the extent map, and we don't want
2120          * to drop it from the cache until it is completely in the btree.
2121          *
2122          * So, tell btrfs_drop_extents to leave this extent in the cache.
2123          * the caller is expected to unpin it and allow it to be merged
2124          * with the others.
2125          */
2126         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2127                                    file_pos + num_bytes, NULL, 0,
2128                                    1, sizeof(*fi), &extent_inserted);
2129         if (ret)
2130                 goto out;
2131
2132         if (!extent_inserted) {
2133                 ins.objectid = btrfs_ino(inode);
2134                 ins.offset = file_pos;
2135                 ins.type = BTRFS_EXTENT_DATA_KEY;
2136
2137                 path->leave_spinning = 1;
2138                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2139                                               sizeof(*fi));
2140                 if (ret)
2141                         goto out;
2142         }
2143         leaf = path->nodes[0];
2144         fi = btrfs_item_ptr(leaf, path->slots[0],
2145                             struct btrfs_file_extent_item);
2146         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2147         btrfs_set_file_extent_type(leaf, fi, extent_type);
2148         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2149         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2150         btrfs_set_file_extent_offset(leaf, fi, 0);
2151         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2152         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2153         btrfs_set_file_extent_compression(leaf, fi, compression);
2154         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2155         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2156
2157         btrfs_mark_buffer_dirty(leaf);
2158         btrfs_release_path(path);
2159
2160         inode_add_bytes(inode, num_bytes);
2161
2162         ins.objectid = disk_bytenr;
2163         ins.offset = disk_num_bytes;
2164         ins.type = BTRFS_EXTENT_ITEM_KEY;
2165         ret = btrfs_alloc_reserved_file_extent(trans, root,
2166                                         root->root_key.objectid,
2167                                         btrfs_ino(inode), file_pos,
2168                                         ram_bytes, &ins);
2169         /*
2170          * Release the reserved range from inode dirty range map, as it is
2171          * already moved into delayed_ref_head
2172          */
2173         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2174 out:
2175         btrfs_free_path(path);
2176
2177         return ret;
2178 }
2179
2180 /* snapshot-aware defrag */
2181 struct sa_defrag_extent_backref {
2182         struct rb_node node;
2183         struct old_sa_defrag_extent *old;
2184         u64 root_id;
2185         u64 inum;
2186         u64 file_pos;
2187         u64 extent_offset;
2188         u64 num_bytes;
2189         u64 generation;
2190 };
2191
2192 struct old_sa_defrag_extent {
2193         struct list_head list;
2194         struct new_sa_defrag_extent *new;
2195
2196         u64 extent_offset;
2197         u64 bytenr;
2198         u64 offset;
2199         u64 len;
2200         int count;
2201 };
2202
2203 struct new_sa_defrag_extent {
2204         struct rb_root root;
2205         struct list_head head;
2206         struct btrfs_path *path;
2207         struct inode *inode;
2208         u64 file_pos;
2209         u64 len;
2210         u64 bytenr;
2211         u64 disk_len;
2212         u8 compress_type;
2213 };
2214
2215 static int backref_comp(struct sa_defrag_extent_backref *b1,
2216                         struct sa_defrag_extent_backref *b2)
2217 {
2218         if (b1->root_id < b2->root_id)
2219                 return -1;
2220         else if (b1->root_id > b2->root_id)
2221                 return 1;
2222
2223         if (b1->inum < b2->inum)
2224                 return -1;
2225         else if (b1->inum > b2->inum)
2226                 return 1;
2227
2228         if (b1->file_pos < b2->file_pos)
2229                 return -1;
2230         else if (b1->file_pos > b2->file_pos)
2231                 return 1;
2232
2233         /*
2234          * [------------------------------] ===> (a range of space)
2235          *     |<--->|   |<---->| =============> (fs/file tree A)
2236          * |<---------------------------->| ===> (fs/file tree B)
2237          *
2238          * A range of space can refer to two file extents in one tree while
2239          * refer to only one file extent in another tree.
2240          *
2241          * So we may process a disk offset more than one time(two extents in A)
2242          * and locate at the same extent(one extent in B), then insert two same
2243          * backrefs(both refer to the extent in B).
2244          */
2245         return 0;
2246 }
2247
2248 static void backref_insert(struct rb_root *root,
2249                            struct sa_defrag_extent_backref *backref)
2250 {
2251         struct rb_node **p = &root->rb_node;
2252         struct rb_node *parent = NULL;
2253         struct sa_defrag_extent_backref *entry;
2254         int ret;
2255
2256         while (*p) {
2257                 parent = *p;
2258                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2259
2260                 ret = backref_comp(backref, entry);
2261                 if (ret < 0)
2262                         p = &(*p)->rb_left;
2263                 else
2264                         p = &(*p)->rb_right;
2265         }
2266
2267         rb_link_node(&backref->node, parent, p);
2268         rb_insert_color(&backref->node, root);
2269 }
2270
2271 /*
2272  * Note the backref might has changed, and in this case we just return 0.
2273  */
2274 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2275                                        void *ctx)
2276 {
2277         struct btrfs_file_extent_item *extent;
2278         struct btrfs_fs_info *fs_info;
2279         struct old_sa_defrag_extent *old = ctx;
2280         struct new_sa_defrag_extent *new = old->new;
2281         struct btrfs_path *path = new->path;
2282         struct btrfs_key key;
2283         struct btrfs_root *root;
2284         struct sa_defrag_extent_backref *backref;
2285         struct extent_buffer *leaf;
2286         struct inode *inode = new->inode;
2287         int slot;
2288         int ret;
2289         u64 extent_offset;
2290         u64 num_bytes;
2291
2292         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2293             inum == btrfs_ino(inode))
2294                 return 0;
2295
2296         key.objectid = root_id;
2297         key.type = BTRFS_ROOT_ITEM_KEY;
2298         key.offset = (u64)-1;
2299
2300         fs_info = BTRFS_I(inode)->root->fs_info;
2301         root = btrfs_read_fs_root_no_name(fs_info, &key);
2302         if (IS_ERR(root)) {
2303                 if (PTR_ERR(root) == -ENOENT)
2304                         return 0;
2305                 WARN_ON(1);
2306                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2307                          inum, offset, root_id);
2308                 return PTR_ERR(root);
2309         }
2310
2311         key.objectid = inum;
2312         key.type = BTRFS_EXTENT_DATA_KEY;
2313         if (offset > (u64)-1 << 32)
2314                 key.offset = 0;
2315         else
2316                 key.offset = offset;
2317
2318         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2319         if (WARN_ON(ret < 0))
2320                 return ret;
2321         ret = 0;
2322
2323         while (1) {
2324                 cond_resched();
2325
2326                 leaf = path->nodes[0];
2327                 slot = path->slots[0];
2328
2329                 if (slot >= btrfs_header_nritems(leaf)) {
2330                         ret = btrfs_next_leaf(root, path);
2331                         if (ret < 0) {
2332                                 goto out;
2333                         } else if (ret > 0) {
2334                                 ret = 0;
2335                                 goto out;
2336                         }
2337                         continue;
2338                 }
2339
2340                 path->slots[0]++;
2341
2342                 btrfs_item_key_to_cpu(leaf, &key, slot);
2343
2344                 if (key.objectid > inum)
2345                         goto out;
2346
2347                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2348                         continue;
2349
2350                 extent = btrfs_item_ptr(leaf, slot,
2351                                         struct btrfs_file_extent_item);
2352
2353                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2354                         continue;
2355
2356                 /*
2357                  * 'offset' refers to the exact key.offset,
2358                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2359                  * (key.offset - extent_offset).
2360                  */
2361                 if (key.offset != offset)
2362                         continue;
2363
2364                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2365                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2366
2367                 if (extent_offset >= old->extent_offset + old->offset +
2368                     old->len || extent_offset + num_bytes <=
2369                     old->extent_offset + old->offset)
2370                         continue;
2371                 break;
2372         }
2373
2374         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2375         if (!backref) {
2376                 ret = -ENOENT;
2377                 goto out;
2378         }
2379
2380         backref->root_id = root_id;
2381         backref->inum = inum;
2382         backref->file_pos = offset;
2383         backref->num_bytes = num_bytes;
2384         backref->extent_offset = extent_offset;
2385         backref->generation = btrfs_file_extent_generation(leaf, extent);
2386         backref->old = old;
2387         backref_insert(&new->root, backref);
2388         old->count++;
2389 out:
2390         btrfs_release_path(path);
2391         WARN_ON(ret);
2392         return ret;
2393 }
2394
2395 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2396                                    struct new_sa_defrag_extent *new)
2397 {
2398         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2399         struct old_sa_defrag_extent *old, *tmp;
2400         int ret;
2401
2402         new->path = path;
2403
2404         list_for_each_entry_safe(old, tmp, &new->head, list) {
2405                 ret = iterate_inodes_from_logical(old->bytenr +
2406                                                   old->extent_offset, fs_info,
2407                                                   path, record_one_backref,
2408                                                   old);
2409                 if (ret < 0 && ret != -ENOENT)
2410                         return false;
2411
2412                 /* no backref to be processed for this extent */
2413                 if (!old->count) {
2414                         list_del(&old->list);
2415                         kfree(old);
2416                 }
2417         }
2418
2419         if (list_empty(&new->head))
2420                 return false;
2421
2422         return true;
2423 }
2424
2425 static int relink_is_mergable(struct extent_buffer *leaf,
2426                               struct btrfs_file_extent_item *fi,
2427                               struct new_sa_defrag_extent *new)
2428 {
2429         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2430                 return 0;
2431
2432         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2433                 return 0;
2434
2435         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2436                 return 0;
2437
2438         if (btrfs_file_extent_encryption(leaf, fi) ||
2439             btrfs_file_extent_other_encoding(leaf, fi))
2440                 return 0;
2441
2442         return 1;
2443 }
2444
2445 /*
2446  * Note the backref might has changed, and in this case we just return 0.
2447  */
2448 static noinline int relink_extent_backref(struct btrfs_path *path,
2449                                  struct sa_defrag_extent_backref *prev,
2450                                  struct sa_defrag_extent_backref *backref)
2451 {
2452         struct btrfs_file_extent_item *extent;
2453         struct btrfs_file_extent_item *item;
2454         struct btrfs_ordered_extent *ordered;
2455         struct btrfs_trans_handle *trans;
2456         struct btrfs_fs_info *fs_info;
2457         struct btrfs_root *root;
2458         struct btrfs_key key;
2459         struct extent_buffer *leaf;
2460         struct old_sa_defrag_extent *old = backref->old;
2461         struct new_sa_defrag_extent *new = old->new;
2462         struct inode *src_inode = new->inode;
2463         struct inode *inode;
2464         struct extent_state *cached = NULL;
2465         int ret = 0;
2466         u64 start;
2467         u64 len;
2468         u64 lock_start;
2469         u64 lock_end;
2470         bool merge = false;
2471         int index;
2472
2473         if (prev && prev->root_id == backref->root_id &&
2474             prev->inum == backref->inum &&
2475             prev->file_pos + prev->num_bytes == backref->file_pos)
2476                 merge = true;
2477
2478         /* step 1: get root */
2479         key.objectid = backref->root_id;
2480         key.type = BTRFS_ROOT_ITEM_KEY;
2481         key.offset = (u64)-1;
2482
2483         fs_info = BTRFS_I(src_inode)->root->fs_info;
2484         index = srcu_read_lock(&fs_info->subvol_srcu);
2485
2486         root = btrfs_read_fs_root_no_name(fs_info, &key);
2487         if (IS_ERR(root)) {
2488                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2489                 if (PTR_ERR(root) == -ENOENT)
2490                         return 0;
2491                 return PTR_ERR(root);
2492         }
2493
2494         if (btrfs_root_readonly(root)) {
2495                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2496                 return 0;
2497         }
2498
2499         /* step 2: get inode */
2500         key.objectid = backref->inum;
2501         key.type = BTRFS_INODE_ITEM_KEY;
2502         key.offset = 0;
2503
2504         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2505         if (IS_ERR(inode)) {
2506                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2507                 return 0;
2508         }
2509
2510         srcu_read_unlock(&fs_info->subvol_srcu, index);
2511
2512         /* step 3: relink backref */
2513         lock_start = backref->file_pos;
2514         lock_end = backref->file_pos + backref->num_bytes - 1;
2515         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2516                          &cached);
2517
2518         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2519         if (ordered) {
2520                 btrfs_put_ordered_extent(ordered);
2521                 goto out_unlock;
2522         }
2523
2524         trans = btrfs_join_transaction(root);
2525         if (IS_ERR(trans)) {
2526                 ret = PTR_ERR(trans);
2527                 goto out_unlock;
2528         }
2529
2530         key.objectid = backref->inum;
2531         key.type = BTRFS_EXTENT_DATA_KEY;
2532         key.offset = backref->file_pos;
2533
2534         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2535         if (ret < 0) {
2536                 goto out_free_path;
2537         } else if (ret > 0) {
2538                 ret = 0;
2539                 goto out_free_path;
2540         }
2541
2542         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2543                                 struct btrfs_file_extent_item);
2544
2545         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2546             backref->generation)
2547                 goto out_free_path;
2548
2549         btrfs_release_path(path);
2550
2551         start = backref->file_pos;
2552         if (backref->extent_offset < old->extent_offset + old->offset)
2553                 start += old->extent_offset + old->offset -
2554                          backref->extent_offset;
2555
2556         len = min(backref->extent_offset + backref->num_bytes,
2557                   old->extent_offset + old->offset + old->len);
2558         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2559
2560         ret = btrfs_drop_extents(trans, root, inode, start,
2561                                  start + len, 1);
2562         if (ret)
2563                 goto out_free_path;
2564 again:
2565         key.objectid = btrfs_ino(inode);
2566         key.type = BTRFS_EXTENT_DATA_KEY;
2567         key.offset = start;
2568
2569         path->leave_spinning = 1;
2570         if (merge) {
2571                 struct btrfs_file_extent_item *fi;
2572                 u64 extent_len;
2573                 struct btrfs_key found_key;
2574
2575                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2576                 if (ret < 0)
2577                         goto out_free_path;
2578
2579                 path->slots[0]--;
2580                 leaf = path->nodes[0];
2581                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2582
2583                 fi = btrfs_item_ptr(leaf, path->slots[0],
2584                                     struct btrfs_file_extent_item);
2585                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2586
2587                 if (extent_len + found_key.offset == start &&
2588                     relink_is_mergable(leaf, fi, new)) {
2589                         btrfs_set_file_extent_num_bytes(leaf, fi,
2590                                                         extent_len + len);
2591                         btrfs_mark_buffer_dirty(leaf);
2592                         inode_add_bytes(inode, len);
2593
2594                         ret = 1;
2595                         goto out_free_path;
2596                 } else {
2597                         merge = false;
2598                         btrfs_release_path(path);
2599                         goto again;
2600                 }
2601         }
2602
2603         ret = btrfs_insert_empty_item(trans, root, path, &key,
2604                                         sizeof(*extent));
2605         if (ret) {
2606                 btrfs_abort_transaction(trans, ret);
2607                 goto out_free_path;
2608         }
2609
2610         leaf = path->nodes[0];
2611         item = btrfs_item_ptr(leaf, path->slots[0],
2612                                 struct btrfs_file_extent_item);
2613         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2614         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2615         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2616         btrfs_set_file_extent_num_bytes(leaf, item, len);
2617         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2618         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2619         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2620         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2621         btrfs_set_file_extent_encryption(leaf, item, 0);
2622         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2623
2624         btrfs_mark_buffer_dirty(leaf);
2625         inode_add_bytes(inode, len);
2626         btrfs_release_path(path);
2627
2628         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2629                         new->disk_len, 0,
2630                         backref->root_id, backref->inum,
2631                         new->file_pos); /* start - extent_offset */
2632         if (ret) {
2633                 btrfs_abort_transaction(trans, ret);
2634                 goto out_free_path;
2635         }
2636
2637         ret = 1;
2638 out_free_path:
2639         btrfs_release_path(path);
2640         path->leave_spinning = 0;
2641         btrfs_end_transaction(trans, root);
2642 out_unlock:
2643         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2644                              &cached, GFP_NOFS);
2645         iput(inode);
2646         return ret;
2647 }
2648
2649 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2650 {
2651         struct old_sa_defrag_extent *old, *tmp;
2652
2653         if (!new)
2654                 return;
2655
2656         list_for_each_entry_safe(old, tmp, &new->head, list) {
2657                 kfree(old);
2658         }
2659         kfree(new);
2660 }
2661
2662 static void relink_file_extents(struct new_sa_defrag_extent *new)
2663 {
2664         struct btrfs_path *path;
2665         struct sa_defrag_extent_backref *backref;
2666         struct sa_defrag_extent_backref *prev = NULL;
2667         struct inode *inode;
2668         struct btrfs_root *root;
2669         struct rb_node *node;
2670         int ret;
2671
2672         inode = new->inode;
2673         root = BTRFS_I(inode)->root;
2674
2675         path = btrfs_alloc_path();
2676         if (!path)
2677                 return;
2678
2679         if (!record_extent_backrefs(path, new)) {
2680                 btrfs_free_path(path);
2681                 goto out;
2682         }
2683         btrfs_release_path(path);
2684
2685         while (1) {
2686                 node = rb_first(&new->root);
2687                 if (!node)
2688                         break;
2689                 rb_erase(node, &new->root);
2690
2691                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2692
2693                 ret = relink_extent_backref(path, prev, backref);
2694                 WARN_ON(ret < 0);
2695
2696                 kfree(prev);
2697
2698                 if (ret == 1)
2699                         prev = backref;
2700                 else
2701                         prev = NULL;
2702                 cond_resched();
2703         }
2704         kfree(prev);
2705
2706         btrfs_free_path(path);
2707 out:
2708         free_sa_defrag_extent(new);
2709
2710         atomic_dec(&root->fs_info->defrag_running);
2711         wake_up(&root->fs_info->transaction_wait);
2712 }
2713
2714 static struct new_sa_defrag_extent *
2715 record_old_file_extents(struct inode *inode,
2716                         struct btrfs_ordered_extent *ordered)
2717 {
2718         struct btrfs_root *root = BTRFS_I(inode)->root;
2719         struct btrfs_path *path;
2720         struct btrfs_key key;
2721         struct old_sa_defrag_extent *old;
2722         struct new_sa_defrag_extent *new;
2723         int ret;
2724
2725         new = kmalloc(sizeof(*new), GFP_NOFS);
2726         if (!new)
2727                 return NULL;
2728
2729         new->inode = inode;
2730         new->file_pos = ordered->file_offset;
2731         new->len = ordered->len;
2732         new->bytenr = ordered->start;
2733         new->disk_len = ordered->disk_len;
2734         new->compress_type = ordered->compress_type;
2735         new->root = RB_ROOT;
2736         INIT_LIST_HEAD(&new->head);
2737
2738         path = btrfs_alloc_path();
2739         if (!path)
2740                 goto out_kfree;
2741
2742         key.objectid = btrfs_ino(inode);
2743         key.type = BTRFS_EXTENT_DATA_KEY;
2744         key.offset = new->file_pos;
2745
2746         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2747         if (ret < 0)
2748                 goto out_free_path;
2749         if (ret > 0 && path->slots[0] > 0)
2750                 path->slots[0]--;
2751
2752         /* find out all the old extents for the file range */
2753         while (1) {
2754                 struct btrfs_file_extent_item *extent;
2755                 struct extent_buffer *l;
2756                 int slot;
2757                 u64 num_bytes;
2758                 u64 offset;
2759                 u64 end;
2760                 u64 disk_bytenr;
2761                 u64 extent_offset;
2762
2763                 l = path->nodes[0];
2764                 slot = path->slots[0];
2765
2766                 if (slot >= btrfs_header_nritems(l)) {
2767                         ret = btrfs_next_leaf(root, path);
2768                         if (ret < 0)
2769                                 goto out_free_path;
2770                         else if (ret > 0)
2771                                 break;
2772                         continue;
2773                 }
2774
2775                 btrfs_item_key_to_cpu(l, &key, slot);
2776
2777                 if (key.objectid != btrfs_ino(inode))
2778                         break;
2779                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2780                         break;
2781                 if (key.offset >= new->file_pos + new->len)
2782                         break;
2783
2784                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2785
2786                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2787                 if (key.offset + num_bytes < new->file_pos)
2788                         goto next;
2789
2790                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2791                 if (!disk_bytenr)
2792                         goto next;
2793
2794                 extent_offset = btrfs_file_extent_offset(l, extent);
2795
2796                 old = kmalloc(sizeof(*old), GFP_NOFS);
2797                 if (!old)
2798                         goto out_free_path;
2799
2800                 offset = max(new->file_pos, key.offset);
2801                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2802
2803                 old->bytenr = disk_bytenr;
2804                 old->extent_offset = extent_offset;
2805                 old->offset = offset - key.offset;
2806                 old->len = end - offset;
2807                 old->new = new;
2808                 old->count = 0;
2809                 list_add_tail(&old->list, &new->head);
2810 next:
2811                 path->slots[0]++;
2812                 cond_resched();
2813         }
2814
2815         btrfs_free_path(path);
2816         atomic_inc(&root->fs_info->defrag_running);
2817
2818         return new;
2819
2820 out_free_path:
2821         btrfs_free_path(path);
2822 out_kfree:
2823         free_sa_defrag_extent(new);
2824         return NULL;
2825 }
2826
2827 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2828                                          u64 start, u64 len)
2829 {
2830         struct btrfs_block_group_cache *cache;
2831
2832         cache = btrfs_lookup_block_group(root->fs_info, start);
2833         ASSERT(cache);
2834
2835         spin_lock(&cache->lock);
2836         cache->delalloc_bytes -= len;
2837         spin_unlock(&cache->lock);
2838
2839         btrfs_put_block_group(cache);
2840 }
2841
2842 /* as ordered data IO finishes, this gets called so we can finish
2843  * an ordered extent if the range of bytes in the file it covers are
2844  * fully written.
2845  */
2846 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2847 {
2848         struct inode *inode = ordered_extent->inode;
2849         struct btrfs_root *root = BTRFS_I(inode)->root;
2850         struct btrfs_trans_handle *trans = NULL;
2851         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2852         struct extent_state *cached_state = NULL;
2853         struct new_sa_defrag_extent *new = NULL;
2854         int compress_type = 0;
2855         int ret = 0;
2856         u64 logical_len = ordered_extent->len;
2857         bool nolock;
2858         bool truncated = false;
2859
2860         nolock = btrfs_is_free_space_inode(inode);
2861
2862         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2863                 ret = -EIO;
2864                 goto out;
2865         }
2866
2867         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2868                                      ordered_extent->file_offset +
2869                                      ordered_extent->len - 1);
2870
2871         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2872                 truncated = true;
2873                 logical_len = ordered_extent->truncated_len;
2874                 /* Truncated the entire extent, don't bother adding */
2875                 if (!logical_len)
2876                         goto out;
2877         }
2878
2879         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2880                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2881
2882                 /*
2883                  * For mwrite(mmap + memset to write) case, we still reserve
2884                  * space for NOCOW range.
2885                  * As NOCOW won't cause a new delayed ref, just free the space
2886                  */
2887                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2888                                        ordered_extent->len);
2889                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2890                 if (nolock)
2891                         trans = btrfs_join_transaction_nolock(root);
2892                 else
2893                         trans = btrfs_join_transaction(root);
2894                 if (IS_ERR(trans)) {
2895                         ret = PTR_ERR(trans);
2896                         trans = NULL;
2897                         goto out;
2898                 }
2899                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2900                 ret = btrfs_update_inode_fallback(trans, root, inode);
2901                 if (ret) /* -ENOMEM or corruption */
2902                         btrfs_abort_transaction(trans, ret);
2903                 goto out;
2904         }
2905
2906         lock_extent_bits(io_tree, ordered_extent->file_offset,
2907                          ordered_extent->file_offset + ordered_extent->len - 1,
2908                          &cached_state);
2909
2910         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2911                         ordered_extent->file_offset + ordered_extent->len - 1,
2912                         EXTENT_DEFRAG, 1, cached_state);
2913         if (ret) {
2914                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2915                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2916                         /* the inode is shared */
2917                         new = record_old_file_extents(inode, ordered_extent);
2918
2919                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2920                         ordered_extent->file_offset + ordered_extent->len - 1,
2921                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2922         }
2923
2924         if (nolock)
2925                 trans = btrfs_join_transaction_nolock(root);
2926         else
2927                 trans = btrfs_join_transaction(root);
2928         if (IS_ERR(trans)) {
2929                 ret = PTR_ERR(trans);
2930                 trans = NULL;
2931                 goto out_unlock;
2932         }
2933
2934         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2935
2936         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2937                 compress_type = ordered_extent->compress_type;
2938         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2939                 BUG_ON(compress_type);
2940                 ret = btrfs_mark_extent_written(trans, inode,
2941                                                 ordered_extent->file_offset,
2942                                                 ordered_extent->file_offset +
2943                                                 logical_len);
2944         } else {
2945                 BUG_ON(root == root->fs_info->tree_root);
2946                 ret = insert_reserved_file_extent(trans, inode,
2947                                                 ordered_extent->file_offset,
2948                                                 ordered_extent->start,
2949                                                 ordered_extent->disk_len,
2950                                                 logical_len, logical_len,
2951                                                 compress_type, 0, 0,
2952                                                 BTRFS_FILE_EXTENT_REG);
2953                 if (!ret)
2954                         btrfs_release_delalloc_bytes(root,
2955                                                      ordered_extent->start,
2956                                                      ordered_extent->disk_len);
2957         }
2958         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2959                            ordered_extent->file_offset, ordered_extent->len,
2960                            trans->transid);
2961         if (ret < 0) {
2962                 btrfs_abort_transaction(trans, ret);
2963                 goto out_unlock;
2964         }
2965
2966         add_pending_csums(trans, inode, ordered_extent->file_offset,
2967                           &ordered_extent->list);
2968
2969         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2970         ret = btrfs_update_inode_fallback(trans, root, inode);
2971         if (ret) { /* -ENOMEM or corruption */
2972                 btrfs_abort_transaction(trans, ret);
2973                 goto out_unlock;
2974         }
2975         ret = 0;
2976 out_unlock:
2977         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2978                              ordered_extent->file_offset +
2979                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2980 out:
2981         if (root != root->fs_info->tree_root)
2982                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2983         if (trans)
2984                 btrfs_end_transaction(trans, root);
2985
2986         if (ret || truncated) {
2987                 u64 start, end;
2988
2989                 if (truncated)
2990                         start = ordered_extent->file_offset + logical_len;
2991                 else
2992                         start = ordered_extent->file_offset;
2993                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2994                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2995
2996                 /* Drop the cache for the part of the extent we didn't write. */
2997                 btrfs_drop_extent_cache(inode, start, end, 0);
2998
2999                 /*
3000                  * If the ordered extent had an IOERR or something else went
3001                  * wrong we need to return the space for this ordered extent
3002                  * back to the allocator.  We only free the extent in the
3003                  * truncated case if we didn't write out the extent at all.
3004                  */
3005                 if ((ret || !logical_len) &&
3006                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3007                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3008                         btrfs_free_reserved_extent(root, ordered_extent->start,
3009                                                    ordered_extent->disk_len, 1);
3010         }
3011
3012
3013         /*
3014          * This needs to be done to make sure anybody waiting knows we are done
3015          * updating everything for this ordered extent.
3016          */
3017         btrfs_remove_ordered_extent(inode, ordered_extent);
3018
3019         /* for snapshot-aware defrag */
3020         if (new) {
3021                 if (ret) {
3022                         free_sa_defrag_extent(new);
3023                         atomic_dec(&root->fs_info->defrag_running);
3024                 } else {
3025                         relink_file_extents(new);
3026                 }
3027         }
3028
3029         /* once for us */
3030         btrfs_put_ordered_extent(ordered_extent);
3031         /* once for the tree */
3032         btrfs_put_ordered_extent(ordered_extent);
3033
3034         return ret;
3035 }
3036
3037 static void finish_ordered_fn(struct btrfs_work *work)
3038 {
3039         struct btrfs_ordered_extent *ordered_extent;
3040         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3041         btrfs_finish_ordered_io(ordered_extent);
3042 }
3043
3044 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3045                                 struct extent_state *state, int uptodate)
3046 {
3047         struct inode *inode = page->mapping->host;
3048         struct btrfs_root *root = BTRFS_I(inode)->root;
3049         struct btrfs_ordered_extent *ordered_extent = NULL;
3050         struct btrfs_workqueue *wq;
3051         btrfs_work_func_t func;
3052
3053         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3054
3055         ClearPagePrivate2(page);
3056         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3057                                             end - start + 1, uptodate))
3058                 return 0;
3059
3060         if (btrfs_is_free_space_inode(inode)) {
3061                 wq = root->fs_info->endio_freespace_worker;
3062                 func = btrfs_freespace_write_helper;
3063         } else {
3064                 wq = root->fs_info->endio_write_workers;
3065                 func = btrfs_endio_write_helper;
3066         }
3067
3068         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3069                         NULL);
3070         btrfs_queue_work(wq, &ordered_extent->work);
3071
3072         return 0;
3073 }
3074
3075 static int __readpage_endio_check(struct inode *inode,
3076                                   struct btrfs_io_bio *io_bio,
3077                                   int icsum, struct page *page,
3078                                   int pgoff, u64 start, size_t len)
3079 {
3080         char *kaddr;
3081         u32 csum_expected;
3082         u32 csum = ~(u32)0;
3083
3084         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3085
3086         kaddr = kmap_atomic(page);
3087         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3088         btrfs_csum_final(csum, (char *)&csum);
3089         if (csum != csum_expected)
3090                 goto zeroit;
3091
3092         kunmap_atomic(kaddr);
3093         return 0;
3094 zeroit:
3095         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3096                 "csum failed ino %llu off %llu csum %u expected csum %u",
3097                            btrfs_ino(inode), start, csum, csum_expected);
3098         memset(kaddr + pgoff, 1, len);
3099         flush_dcache_page(page);
3100         kunmap_atomic(kaddr);
3101         if (csum_expected == 0)
3102                 return 0;
3103         return -EIO;
3104 }
3105
3106 /*
3107  * when reads are done, we need to check csums to verify the data is correct
3108  * if there's a match, we allow the bio to finish.  If not, the code in
3109  * extent_io.c will try to find good copies for us.
3110  */
3111 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3112                                       u64 phy_offset, struct page *page,
3113                                       u64 start, u64 end, int mirror)
3114 {
3115         size_t offset = start - page_offset(page);
3116         struct inode *inode = page->mapping->host;
3117         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3118         struct btrfs_root *root = BTRFS_I(inode)->root;
3119
3120         if (PageChecked(page)) {
3121                 ClearPageChecked(page);
3122                 return 0;
3123         }
3124
3125         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3126                 return 0;
3127
3128         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3129             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3130                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3131                 return 0;
3132         }
3133
3134         phy_offset >>= inode->i_sb->s_blocksize_bits;
3135         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3136                                       start, (size_t)(end - start + 1));
3137 }
3138
3139 void btrfs_add_delayed_iput(struct inode *inode)
3140 {
3141         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3142         struct btrfs_inode *binode = BTRFS_I(inode);
3143
3144         if (atomic_add_unless(&inode->i_count, -1, 1))
3145                 return;
3146
3147         spin_lock(&fs_info->delayed_iput_lock);
3148         if (binode->delayed_iput_count == 0) {
3149                 ASSERT(list_empty(&binode->delayed_iput));
3150                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3151         } else {
3152                 binode->delayed_iput_count++;
3153         }
3154         spin_unlock(&fs_info->delayed_iput_lock);
3155 }
3156
3157 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3158 {
3159         struct btrfs_fs_info *fs_info = root->fs_info;
3160
3161         spin_lock(&fs_info->delayed_iput_lock);
3162         while (!list_empty(&fs_info->delayed_iputs)) {
3163                 struct btrfs_inode *inode;
3164
3165                 inode = list_first_entry(&fs_info->delayed_iputs,
3166                                 struct btrfs_inode, delayed_iput);
3167                 if (inode->delayed_iput_count) {
3168                         inode->delayed_iput_count--;
3169                         list_move_tail(&inode->delayed_iput,
3170                                         &fs_info->delayed_iputs);
3171                 } else {
3172                         list_del_init(&inode->delayed_iput);
3173                 }
3174                 spin_unlock(&fs_info->delayed_iput_lock);
3175                 iput(&inode->vfs_inode);
3176                 spin_lock(&fs_info->delayed_iput_lock);
3177         }
3178         spin_unlock(&fs_info->delayed_iput_lock);
3179 }
3180
3181 /*
3182  * This is called in transaction commit time. If there are no orphan
3183  * files in the subvolume, it removes orphan item and frees block_rsv
3184  * structure.
3185  */
3186 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3187                               struct btrfs_root *root)
3188 {
3189         struct btrfs_block_rsv *block_rsv;
3190         int ret;
3191
3192         if (atomic_read(&root->orphan_inodes) ||
3193             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3194                 return;
3195
3196         spin_lock(&root->orphan_lock);
3197         if (atomic_read(&root->orphan_inodes)) {
3198                 spin_unlock(&root->orphan_lock);
3199                 return;
3200         }
3201
3202         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3203                 spin_unlock(&root->orphan_lock);
3204                 return;
3205         }
3206
3207         block_rsv = root->orphan_block_rsv;
3208         root->orphan_block_rsv = NULL;
3209         spin_unlock(&root->orphan_lock);
3210
3211         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3212             btrfs_root_refs(&root->root_item) > 0) {
3213                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3214                                             root->root_key.objectid);
3215                 if (ret)
3216                         btrfs_abort_transaction(trans, ret);
3217                 else
3218                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3219                                   &root->state);
3220         }
3221
3222         if (block_rsv) {
3223                 WARN_ON(block_rsv->size > 0);
3224                 btrfs_free_block_rsv(root, block_rsv);
3225         }
3226 }
3227
3228 /*
3229  * This creates an orphan entry for the given inode in case something goes
3230  * wrong in the middle of an unlink/truncate.
3231  *
3232  * NOTE: caller of this function should reserve 5 units of metadata for
3233  *       this function.
3234  */
3235 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3236 {
3237         struct btrfs_root *root = BTRFS_I(inode)->root;
3238         struct btrfs_block_rsv *block_rsv = NULL;
3239         int reserve = 0;
3240         int insert = 0;
3241         int ret;
3242
3243         if (!root->orphan_block_rsv) {
3244                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3245                 if (!block_rsv)
3246                         return -ENOMEM;
3247         }
3248
3249         spin_lock(&root->orphan_lock);
3250         if (!root->orphan_block_rsv) {
3251                 root->orphan_block_rsv = block_rsv;
3252         } else if (block_rsv) {
3253                 btrfs_free_block_rsv(root, block_rsv);
3254                 block_rsv = NULL;
3255         }
3256
3257         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3258                               &BTRFS_I(inode)->runtime_flags)) {
3259 #if 0
3260                 /*
3261                  * For proper ENOSPC handling, we should do orphan
3262                  * cleanup when mounting. But this introduces backward
3263                  * compatibility issue.
3264                  */
3265                 if (!xchg(&root->orphan_item_inserted, 1))
3266                         insert = 2;
3267                 else
3268                         insert = 1;
3269 #endif
3270                 insert = 1;
3271                 atomic_inc(&root->orphan_inodes);
3272         }
3273
3274         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3275                               &BTRFS_I(inode)->runtime_flags))
3276                 reserve = 1;
3277         spin_unlock(&root->orphan_lock);
3278
3279         /* grab metadata reservation from transaction handle */
3280         if (reserve) {
3281                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3282                 ASSERT(!ret);
3283                 if (ret) {
3284                         atomic_dec(&root->orphan_inodes);
3285                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3286                                   &BTRFS_I(inode)->runtime_flags);
3287                         if (insert)
3288                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3289                                           &BTRFS_I(inode)->runtime_flags);
3290                         return ret;
3291                 }
3292         }
3293
3294         /* insert an orphan item to track this unlinked/truncated file */
3295         if (insert >= 1) {
3296                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3297                 if (ret) {
3298                         atomic_dec(&root->orphan_inodes);
3299                         if (reserve) {
3300                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3301                                           &BTRFS_I(inode)->runtime_flags);
3302                                 btrfs_orphan_release_metadata(inode);
3303                         }
3304                         if (ret != -EEXIST) {
3305                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3306                                           &BTRFS_I(inode)->runtime_flags);
3307                                 btrfs_abort_transaction(trans, ret);
3308                                 return ret;
3309                         }
3310                 }
3311                 ret = 0;
3312         }
3313
3314         /* insert an orphan item to track subvolume contains orphan files */
3315         if (insert >= 2) {
3316                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3317                                                root->root_key.objectid);
3318                 if (ret && ret != -EEXIST) {
3319                         btrfs_abort_transaction(trans, ret);
3320                         return ret;
3321                 }
3322         }
3323         return 0;
3324 }
3325
3326 /*
3327  * We have done the truncate/delete so we can go ahead and remove the orphan
3328  * item for this particular inode.
3329  */
3330 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3331                             struct inode *inode)
3332 {
3333         struct btrfs_root *root = BTRFS_I(inode)->root;
3334         int delete_item = 0;
3335         int release_rsv = 0;
3336         int ret = 0;
3337
3338         spin_lock(&root->orphan_lock);
3339         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3340                                &BTRFS_I(inode)->runtime_flags))
3341                 delete_item = 1;
3342
3343         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3344                                &BTRFS_I(inode)->runtime_flags))
3345                 release_rsv = 1;
3346         spin_unlock(&root->orphan_lock);
3347
3348         if (delete_item) {
3349                 atomic_dec(&root->orphan_inodes);
3350                 if (trans)
3351                         ret = btrfs_del_orphan_item(trans, root,
3352                                                     btrfs_ino(inode));
3353         }
3354
3355         if (release_rsv)
3356                 btrfs_orphan_release_metadata(inode);
3357
3358         return ret;
3359 }
3360
3361 /*
3362  * this cleans up any orphans that may be left on the list from the last use
3363  * of this root.
3364  */
3365 int btrfs_orphan_cleanup(struct btrfs_root *root)
3366 {
3367         struct btrfs_path *path;
3368         struct extent_buffer *leaf;
3369         struct btrfs_key key, found_key;
3370         struct btrfs_trans_handle *trans;
3371         struct inode *inode;
3372         u64 last_objectid = 0;
3373         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3374
3375         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3376                 return 0;
3377
3378         path = btrfs_alloc_path();
3379         if (!path) {
3380                 ret = -ENOMEM;
3381                 goto out;
3382         }
3383         path->reada = READA_BACK;
3384
3385         key.objectid = BTRFS_ORPHAN_OBJECTID;
3386         key.type = BTRFS_ORPHAN_ITEM_KEY;
3387         key.offset = (u64)-1;
3388
3389         while (1) {
3390                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3391                 if (ret < 0)
3392                         goto out;
3393
3394                 /*
3395                  * if ret == 0 means we found what we were searching for, which
3396                  * is weird, but possible, so only screw with path if we didn't
3397                  * find the key and see if we have stuff that matches
3398                  */
3399                 if (ret > 0) {
3400                         ret = 0;
3401                         if (path->slots[0] == 0)
3402                                 break;
3403                         path->slots[0]--;
3404                 }
3405
3406                 /* pull out the item */
3407                 leaf = path->nodes[0];
3408                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3409
3410                 /* make sure the item matches what we want */
3411                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3412                         break;
3413                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3414                         break;
3415
3416                 /* release the path since we're done with it */
3417                 btrfs_release_path(path);
3418
3419                 /*
3420                  * this is where we are basically btrfs_lookup, without the
3421                  * crossing root thing.  we store the inode number in the
3422                  * offset of the orphan item.
3423                  */
3424
3425                 if (found_key.offset == last_objectid) {
3426                         btrfs_err(root->fs_info,
3427                                 "Error removing orphan entry, stopping orphan cleanup");
3428                         ret = -EINVAL;
3429                         goto out;
3430                 }
3431
3432                 last_objectid = found_key.offset;
3433
3434                 found_key.objectid = found_key.offset;
3435                 found_key.type = BTRFS_INODE_ITEM_KEY;
3436                 found_key.offset = 0;
3437                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3438                 ret = PTR_ERR_OR_ZERO(inode);
3439                 if (ret && ret != -ESTALE)
3440                         goto out;
3441
3442                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3443                         struct btrfs_root *dead_root;
3444                         struct btrfs_fs_info *fs_info = root->fs_info;
3445                         int is_dead_root = 0;
3446
3447                         /*
3448                          * this is an orphan in the tree root. Currently these
3449                          * could come from 2 sources:
3450                          *  a) a snapshot deletion in progress
3451                          *  b) a free space cache inode
3452                          * We need to distinguish those two, as the snapshot
3453                          * orphan must not get deleted.
3454                          * find_dead_roots already ran before us, so if this
3455                          * is a snapshot deletion, we should find the root
3456                          * in the dead_roots list
3457                          */
3458                         spin_lock(&fs_info->trans_lock);
3459                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3460                                             root_list) {
3461                                 if (dead_root->root_key.objectid ==
3462                                     found_key.objectid) {
3463                                         is_dead_root = 1;
3464                                         break;
3465                                 }
3466                         }
3467                         spin_unlock(&fs_info->trans_lock);
3468                         if (is_dead_root) {
3469                                 /* prevent this orphan from being found again */
3470                                 key.offset = found_key.objectid - 1;
3471                                 continue;
3472                         }
3473                 }
3474                 /*
3475                  * Inode is already gone but the orphan item is still there,
3476                  * kill the orphan item.
3477                  */
3478                 if (ret == -ESTALE) {
3479                         trans = btrfs_start_transaction(root, 1);
3480                         if (IS_ERR(trans)) {
3481                                 ret = PTR_ERR(trans);
3482                                 goto out;
3483                         }
3484                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3485                                 found_key.objectid);
3486                         ret = btrfs_del_orphan_item(trans, root,
3487                                                     found_key.objectid);
3488                         btrfs_end_transaction(trans, root);
3489                         if (ret)
3490                                 goto out;
3491                         continue;
3492                 }
3493
3494                 /*
3495                  * add this inode to the orphan list so btrfs_orphan_del does
3496                  * the proper thing when we hit it
3497                  */
3498                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3499                         &BTRFS_I(inode)->runtime_flags);
3500                 atomic_inc(&root->orphan_inodes);
3501
3502                 /* if we have links, this was a truncate, lets do that */
3503                 if (inode->i_nlink) {
3504                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3505                                 iput(inode);
3506                                 continue;
3507                         }
3508                         nr_truncate++;
3509
3510                         /* 1 for the orphan item deletion. */
3511                         trans = btrfs_start_transaction(root, 1);
3512                         if (IS_ERR(trans)) {
3513                                 iput(inode);
3514                                 ret = PTR_ERR(trans);
3515                                 goto out;
3516                         }
3517                         ret = btrfs_orphan_add(trans, inode);
3518                         btrfs_end_transaction(trans, root);
3519                         if (ret) {
3520                                 iput(inode);
3521                                 goto out;
3522                         }
3523
3524                         ret = btrfs_truncate(inode);
3525                         if (ret)
3526                                 btrfs_orphan_del(NULL, inode);
3527                 } else {
3528                         nr_unlink++;
3529                 }
3530
3531                 /* this will do delete_inode and everything for us */
3532                 iput(inode);
3533                 if (ret)
3534                         goto out;
3535         }
3536         /* release the path since we're done with it */
3537         btrfs_release_path(path);
3538
3539         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3540
3541         if (root->orphan_block_rsv)
3542                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3543                                         (u64)-1);
3544
3545         if (root->orphan_block_rsv ||
3546             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3547                 trans = btrfs_join_transaction(root);
3548                 if (!IS_ERR(trans))
3549                         btrfs_end_transaction(trans, root);
3550         }
3551
3552         if (nr_unlink)
3553                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3554         if (nr_truncate)
3555                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3556
3557 out:
3558         if (ret)
3559                 btrfs_err(root->fs_info,
3560                         "could not do orphan cleanup %d", ret);
3561         btrfs_free_path(path);
3562         return ret;
3563 }
3564
3565 /*
3566  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3567  * don't find any xattrs, we know there can't be any acls.
3568  *
3569  * slot is the slot the inode is in, objectid is the objectid of the inode
3570  */
3571 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3572                                           int slot, u64 objectid,
3573                                           int *first_xattr_slot)
3574 {
3575         u32 nritems = btrfs_header_nritems(leaf);
3576         struct btrfs_key found_key;
3577         static u64 xattr_access = 0;
3578         static u64 xattr_default = 0;
3579         int scanned = 0;
3580
3581         if (!xattr_access) {
3582                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3583                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3584                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3585                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3586         }
3587
3588         slot++;
3589         *first_xattr_slot = -1;
3590         while (slot < nritems) {
3591                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3592
3593                 /* we found a different objectid, there must not be acls */
3594                 if (found_key.objectid != objectid)
3595                         return 0;
3596
3597                 /* we found an xattr, assume we've got an acl */
3598                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3599                         if (*first_xattr_slot == -1)
3600                                 *first_xattr_slot = slot;
3601                         if (found_key.offset == xattr_access ||
3602                             found_key.offset == xattr_default)
3603                                 return 1;
3604                 }
3605
3606                 /*
3607                  * we found a key greater than an xattr key, there can't
3608                  * be any acls later on
3609                  */
3610                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3611                         return 0;
3612
3613                 slot++;
3614                 scanned++;
3615
3616                 /*
3617                  * it goes inode, inode backrefs, xattrs, extents,
3618                  * so if there are a ton of hard links to an inode there can
3619                  * be a lot of backrefs.  Don't waste time searching too hard,
3620                  * this is just an optimization
3621                  */
3622                 if (scanned >= 8)
3623                         break;
3624         }
3625         /* we hit the end of the leaf before we found an xattr or
3626          * something larger than an xattr.  We have to assume the inode
3627          * has acls
3628          */
3629         if (*first_xattr_slot == -1)
3630                 *first_xattr_slot = slot;
3631         return 1;
3632 }
3633
3634 /*
3635  * read an inode from the btree into the in-memory inode
3636  */
3637 static void btrfs_read_locked_inode(struct inode *inode)
3638 {
3639         struct btrfs_path *path;
3640         struct extent_buffer *leaf;
3641         struct btrfs_inode_item *inode_item;
3642         struct btrfs_root *root = BTRFS_I(inode)->root;
3643         struct btrfs_key location;
3644         unsigned long ptr;
3645         int maybe_acls;
3646         u32 rdev;
3647         int ret;
3648         bool filled = false;
3649         int first_xattr_slot;
3650
3651         ret = btrfs_fill_inode(inode, &rdev);
3652         if (!ret)
3653                 filled = true;
3654
3655         path = btrfs_alloc_path();
3656         if (!path)
3657                 goto make_bad;
3658
3659         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3660
3661         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3662         if (ret)
3663                 goto make_bad;
3664
3665         leaf = path->nodes[0];
3666
3667         if (filled)
3668                 goto cache_index;
3669
3670         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3671                                     struct btrfs_inode_item);
3672         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3673         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3674         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3675         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3676         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3677
3678         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3679         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3680
3681         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3682         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3683
3684         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3685         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3686
3687         BTRFS_I(inode)->i_otime.tv_sec =
3688                 btrfs_timespec_sec(leaf, &inode_item->otime);
3689         BTRFS_I(inode)->i_otime.tv_nsec =
3690                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3691
3692         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3693         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3694         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3695
3696         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3697         inode->i_generation = BTRFS_I(inode)->generation;
3698         inode->i_rdev = 0;
3699         rdev = btrfs_inode_rdev(leaf, inode_item);
3700
3701         BTRFS_I(inode)->index_cnt = (u64)-1;
3702         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3703
3704 cache_index:
3705         /*
3706          * If we were modified in the current generation and evicted from memory
3707          * and then re-read we need to do a full sync since we don't have any
3708          * idea about which extents were modified before we were evicted from
3709          * cache.
3710          *
3711          * This is required for both inode re-read from disk and delayed inode
3712          * in delayed_nodes_tree.
3713          */
3714         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3715                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3716                         &BTRFS_I(inode)->runtime_flags);
3717
3718         /*
3719          * We don't persist the id of the transaction where an unlink operation
3720          * against the inode was last made. So here we assume the inode might
3721          * have been evicted, and therefore the exact value of last_unlink_trans
3722          * lost, and set it to last_trans to avoid metadata inconsistencies
3723          * between the inode and its parent if the inode is fsync'ed and the log
3724          * replayed. For example, in the scenario:
3725          *
3726          * touch mydir/foo
3727          * ln mydir/foo mydir/bar
3728          * sync
3729          * unlink mydir/bar
3730          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3731          * xfs_io -c fsync mydir/foo
3732          * <power failure>
3733          * mount fs, triggers fsync log replay
3734          *
3735          * We must make sure that when we fsync our inode foo we also log its
3736          * parent inode, otherwise after log replay the parent still has the
3737          * dentry with the "bar" name but our inode foo has a link count of 1
3738          * and doesn't have an inode ref with the name "bar" anymore.
3739          *
3740          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3741          * but it guarantees correctness at the expense of occasional full
3742          * transaction commits on fsync if our inode is a directory, or if our
3743          * inode is not a directory, logging its parent unnecessarily.
3744          */
3745         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3746
3747         path->slots[0]++;
3748         if (inode->i_nlink != 1 ||
3749             path->slots[0] >= btrfs_header_nritems(leaf))
3750                 goto cache_acl;
3751
3752         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3753         if (location.objectid != btrfs_ino(inode))
3754                 goto cache_acl;
3755
3756         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3757         if (location.type == BTRFS_INODE_REF_KEY) {
3758                 struct btrfs_inode_ref *ref;
3759
3760                 ref = (struct btrfs_inode_ref *)ptr;
3761                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3762         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3763                 struct btrfs_inode_extref *extref;
3764
3765                 extref = (struct btrfs_inode_extref *)ptr;
3766                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3767                                                                      extref);
3768         }
3769 cache_acl:
3770         /*
3771          * try to precache a NULL acl entry for files that don't have
3772          * any xattrs or acls
3773          */
3774         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3775                                            btrfs_ino(inode), &first_xattr_slot);
3776         if (first_xattr_slot != -1) {
3777                 path->slots[0] = first_xattr_slot;
3778                 ret = btrfs_load_inode_props(inode, path);
3779                 if (ret)
3780                         btrfs_err(root->fs_info,
3781                                   "error loading props for ino %llu (root %llu): %d",
3782                                   btrfs_ino(inode),
3783                                   root->root_key.objectid, ret);
3784         }
3785         btrfs_free_path(path);
3786
3787         if (!maybe_acls)
3788                 cache_no_acl(inode);
3789
3790         switch (inode->i_mode & S_IFMT) {
3791         case S_IFREG:
3792                 inode->i_mapping->a_ops = &btrfs_aops;
3793                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3794                 inode->i_fop = &btrfs_file_operations;
3795                 inode->i_op = &btrfs_file_inode_operations;
3796                 break;
3797         case S_IFDIR:
3798                 inode->i_fop = &btrfs_dir_file_operations;
3799                 if (root == root->fs_info->tree_root)
3800                         inode->i_op = &btrfs_dir_ro_inode_operations;
3801                 else
3802                         inode->i_op = &btrfs_dir_inode_operations;
3803                 break;
3804         case S_IFLNK:
3805                 inode->i_op = &btrfs_symlink_inode_operations;
3806                 inode_nohighmem(inode);
3807                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3808                 break;
3809         default:
3810                 inode->i_op = &btrfs_special_inode_operations;
3811                 init_special_inode(inode, inode->i_mode, rdev);
3812                 break;
3813         }
3814
3815         btrfs_update_iflags(inode);
3816         return;
3817
3818 make_bad:
3819         btrfs_free_path(path);
3820         make_bad_inode(inode);
3821 }
3822
3823 /*
3824  * given a leaf and an inode, copy the inode fields into the leaf
3825  */
3826 static void fill_inode_item(struct btrfs_trans_handle *trans,
3827                             struct extent_buffer *leaf,
3828                             struct btrfs_inode_item *item,
3829                             struct inode *inode)
3830 {
3831         struct btrfs_map_token token;
3832
3833         btrfs_init_map_token(&token);
3834
3835         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3836         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3837         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3838                                    &token);
3839         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3840         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3841
3842         btrfs_set_token_timespec_sec(leaf, &item->atime,
3843                                      inode->i_atime.tv_sec, &token);
3844         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3845                                       inode->i_atime.tv_nsec, &token);
3846
3847         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3848                                      inode->i_mtime.tv_sec, &token);
3849         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3850                                       inode->i_mtime.tv_nsec, &token);
3851
3852         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3853                                      inode->i_ctime.tv_sec, &token);
3854         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3855                                       inode->i_ctime.tv_nsec, &token);
3856
3857         btrfs_set_token_timespec_sec(leaf, &item->otime,
3858                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3859         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3860                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3861
3862         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3863                                      &token);
3864         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3865                                          &token);
3866         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3867         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3868         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3869         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3870         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3871 }
3872
3873 /*
3874  * copy everything in the in-memory inode into the btree.
3875  */
3876 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3877                                 struct btrfs_root *root, struct inode *inode)
3878 {
3879         struct btrfs_inode_item *inode_item;
3880         struct btrfs_path *path;
3881         struct extent_buffer *leaf;
3882         int ret;
3883
3884         path = btrfs_alloc_path();
3885         if (!path)
3886                 return -ENOMEM;
3887
3888         path->leave_spinning = 1;
3889         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3890                                  1);
3891         if (ret) {
3892                 if (ret > 0)
3893                         ret = -ENOENT;
3894                 goto failed;
3895         }
3896
3897         leaf = path->nodes[0];
3898         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3899                                     struct btrfs_inode_item);
3900
3901         fill_inode_item(trans, leaf, inode_item, inode);
3902         btrfs_mark_buffer_dirty(leaf);
3903         btrfs_set_inode_last_trans(trans, inode);
3904         ret = 0;
3905 failed:
3906         btrfs_free_path(path);
3907         return ret;
3908 }
3909
3910 /*
3911  * copy everything in the in-memory inode into the btree.
3912  */
3913 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3914                                 struct btrfs_root *root, struct inode *inode)
3915 {
3916         int ret;
3917
3918         /*
3919          * If the inode is a free space inode, we can deadlock during commit
3920          * if we put it into the delayed code.
3921          *
3922          * The data relocation inode should also be directly updated
3923          * without delay
3924          */
3925         if (!btrfs_is_free_space_inode(inode)
3926             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3927             && !root->fs_info->log_root_recovering) {
3928                 btrfs_update_root_times(trans, root);
3929
3930                 ret = btrfs_delayed_update_inode(trans, root, inode);
3931                 if (!ret)
3932                         btrfs_set_inode_last_trans(trans, inode);
3933                 return ret;
3934         }
3935
3936         return btrfs_update_inode_item(trans, root, inode);
3937 }
3938
3939 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3940                                          struct btrfs_root *root,
3941                                          struct inode *inode)
3942 {
3943         int ret;
3944
3945         ret = btrfs_update_inode(trans, root, inode);
3946         if (ret == -ENOSPC)
3947                 return btrfs_update_inode_item(trans, root, inode);
3948         return ret;
3949 }
3950
3951 /*
3952  * unlink helper that gets used here in inode.c and in the tree logging
3953  * recovery code.  It remove a link in a directory with a given name, and
3954  * also drops the back refs in the inode to the directory
3955  */
3956 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3957                                 struct btrfs_root *root,
3958                                 struct inode *dir, struct inode *inode,
3959                                 const char *name, int name_len)
3960 {
3961         struct btrfs_path *path;
3962         int ret = 0;
3963         struct extent_buffer *leaf;
3964         struct btrfs_dir_item *di;
3965         struct btrfs_key key;
3966         u64 index;
3967         u64 ino = btrfs_ino(inode);
3968         u64 dir_ino = btrfs_ino(dir);
3969
3970         path = btrfs_alloc_path();
3971         if (!path) {
3972                 ret = -ENOMEM;
3973                 goto out;
3974         }
3975
3976         path->leave_spinning = 1;
3977         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3978                                     name, name_len, -1);
3979         if (IS_ERR(di)) {
3980                 ret = PTR_ERR(di);
3981                 goto err;
3982         }
3983         if (!di) {
3984                 ret = -ENOENT;
3985                 goto err;
3986         }
3987         leaf = path->nodes[0];
3988         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3989         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3990         if (ret)
3991                 goto err;
3992         btrfs_release_path(path);
3993
3994         /*
3995          * If we don't have dir index, we have to get it by looking up
3996          * the inode ref, since we get the inode ref, remove it directly,
3997          * it is unnecessary to do delayed deletion.
3998          *
3999          * But if we have dir index, needn't search inode ref to get it.
4000          * Since the inode ref is close to the inode item, it is better
4001          * that we delay to delete it, and just do this deletion when
4002          * we update the inode item.
4003          */
4004         if (BTRFS_I(inode)->dir_index) {
4005                 ret = btrfs_delayed_delete_inode_ref(inode);
4006                 if (!ret) {
4007                         index = BTRFS_I(inode)->dir_index;
4008                         goto skip_backref;
4009                 }
4010         }
4011
4012         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4013                                   dir_ino, &index);
4014         if (ret) {
4015                 btrfs_info(root->fs_info,
4016                         "failed to delete reference to %.*s, inode %llu parent %llu",
4017                         name_len, name, ino, dir_ino);
4018                 btrfs_abort_transaction(trans, ret);
4019                 goto err;
4020         }
4021 skip_backref:
4022         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4023         if (ret) {
4024                 btrfs_abort_transaction(trans, ret);
4025                 goto err;
4026         }
4027
4028         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4029                                          inode, dir_ino);
4030         if (ret != 0 && ret != -ENOENT) {
4031                 btrfs_abort_transaction(trans, ret);
4032                 goto err;
4033         }
4034
4035         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4036                                            dir, index);
4037         if (ret == -ENOENT)
4038                 ret = 0;
4039         else if (ret)
4040                 btrfs_abort_transaction(trans, ret);
4041 err:
4042         btrfs_free_path(path);
4043         if (ret)
4044                 goto out;
4045
4046         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4047         inode_inc_iversion(inode);
4048         inode_inc_iversion(dir);
4049         inode->i_ctime = dir->i_mtime =
4050                 dir->i_ctime = current_fs_time(inode->i_sb);
4051         ret = btrfs_update_inode(trans, root, dir);
4052 out:
4053         return ret;
4054 }
4055
4056 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4057                        struct btrfs_root *root,
4058                        struct inode *dir, struct inode *inode,
4059                        const char *name, int name_len)
4060 {
4061         int ret;
4062         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4063         if (!ret) {
4064                 drop_nlink(inode);
4065                 ret = btrfs_update_inode(trans, root, inode);
4066         }
4067         return ret;
4068 }
4069
4070 /*
4071  * helper to start transaction for unlink and rmdir.
4072  *
4073  * unlink and rmdir are special in btrfs, they do not always free space, so
4074  * if we cannot make our reservations the normal way try and see if there is
4075  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4076  * allow the unlink to occur.
4077  */
4078 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4079 {
4080         struct btrfs_root *root = BTRFS_I(dir)->root;
4081
4082         /*
4083          * 1 for the possible orphan item
4084          * 1 for the dir item
4085          * 1 for the dir index
4086          * 1 for the inode ref
4087          * 1 for the inode
4088          */
4089         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4090 }
4091
4092 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4093 {
4094         struct btrfs_root *root = BTRFS_I(dir)->root;
4095         struct btrfs_trans_handle *trans;
4096         struct inode *inode = d_inode(dentry);
4097         int ret;
4098
4099         trans = __unlink_start_trans(dir);
4100         if (IS_ERR(trans))
4101                 return PTR_ERR(trans);
4102
4103         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4104
4105         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4106                                  dentry->d_name.name, dentry->d_name.len);
4107         if (ret)
4108                 goto out;
4109
4110         if (inode->i_nlink == 0) {
4111                 ret = btrfs_orphan_add(trans, inode);
4112                 if (ret)
4113                         goto out;
4114         }
4115
4116 out:
4117         btrfs_end_transaction(trans, root);
4118         btrfs_btree_balance_dirty(root);
4119         return ret;
4120 }
4121
4122 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4123                         struct btrfs_root *root,
4124                         struct inode *dir, u64 objectid,
4125                         const char *name, int name_len)
4126 {
4127         struct btrfs_path *path;
4128         struct extent_buffer *leaf;
4129         struct btrfs_dir_item *di;
4130         struct btrfs_key key;
4131         u64 index;
4132         int ret;
4133         u64 dir_ino = btrfs_ino(dir);
4134
4135         path = btrfs_alloc_path();
4136         if (!path)
4137                 return -ENOMEM;
4138
4139         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4140                                    name, name_len, -1);
4141         if (IS_ERR_OR_NULL(di)) {
4142                 if (!di)
4143                         ret = -ENOENT;
4144                 else
4145                         ret = PTR_ERR(di);
4146                 goto out;
4147         }
4148
4149         leaf = path->nodes[0];
4150         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4151         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4152         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4153         if (ret) {
4154                 btrfs_abort_transaction(trans, ret);
4155                 goto out;
4156         }
4157         btrfs_release_path(path);
4158
4159         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4160                                  objectid, root->root_key.objectid,
4161                                  dir_ino, &index, name, name_len);
4162         if (ret < 0) {
4163                 if (ret != -ENOENT) {
4164                         btrfs_abort_transaction(trans, ret);
4165                         goto out;
4166                 }
4167                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4168                                                  name, name_len);
4169                 if (IS_ERR_OR_NULL(di)) {
4170                         if (!di)
4171                                 ret = -ENOENT;
4172                         else
4173                                 ret = PTR_ERR(di);
4174                         btrfs_abort_transaction(trans, ret);
4175                         goto out;
4176                 }
4177
4178                 leaf = path->nodes[0];
4179                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4180                 btrfs_release_path(path);
4181                 index = key.offset;
4182         }
4183         btrfs_release_path(path);
4184
4185         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4186         if (ret) {
4187                 btrfs_abort_transaction(trans, ret);
4188                 goto out;
4189         }
4190
4191         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4192         inode_inc_iversion(dir);
4193         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4194         ret = btrfs_update_inode_fallback(trans, root, dir);
4195         if (ret)
4196                 btrfs_abort_transaction(trans, ret);
4197 out:
4198         btrfs_free_path(path);
4199         return ret;
4200 }
4201
4202 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4203 {
4204         struct inode *inode = d_inode(dentry);
4205         int err = 0;
4206         struct btrfs_root *root = BTRFS_I(dir)->root;
4207         struct btrfs_trans_handle *trans;
4208
4209         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4210                 return -ENOTEMPTY;
4211         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4212                 return -EPERM;
4213
4214         trans = __unlink_start_trans(dir);
4215         if (IS_ERR(trans))
4216                 return PTR_ERR(trans);
4217
4218         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4219                 err = btrfs_unlink_subvol(trans, root, dir,
4220                                           BTRFS_I(inode)->location.objectid,
4221                                           dentry->d_name.name,
4222                                           dentry->d_name.len);
4223                 goto out;
4224         }
4225
4226         err = btrfs_orphan_add(trans, inode);
4227         if (err)
4228                 goto out;
4229
4230         /* now the directory is empty */
4231         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4232                                  dentry->d_name.name, dentry->d_name.len);
4233         if (!err)
4234                 btrfs_i_size_write(inode, 0);
4235 out:
4236         btrfs_end_transaction(trans, root);
4237         btrfs_btree_balance_dirty(root);
4238
4239         return err;
4240 }
4241
4242 static int truncate_space_check(struct btrfs_trans_handle *trans,
4243                                 struct btrfs_root *root,
4244                                 u64 bytes_deleted)
4245 {
4246         int ret;
4247
4248         /*
4249          * This is only used to apply pressure to the enospc system, we don't
4250          * intend to use this reservation at all.
4251          */
4252         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4253         bytes_deleted *= root->nodesize;
4254         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4255                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4256         if (!ret) {
4257                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4258                                               trans->transid,
4259                                               bytes_deleted, 1);
4260                 trans->bytes_reserved += bytes_deleted;
4261         }
4262         return ret;
4263
4264 }
4265
4266 static int truncate_inline_extent(struct inode *inode,
4267                                   struct btrfs_path *path,
4268                                   struct btrfs_key *found_key,
4269                                   const u64 item_end,
4270                                   const u64 new_size)
4271 {
4272         struct extent_buffer *leaf = path->nodes[0];
4273         int slot = path->slots[0];
4274         struct btrfs_file_extent_item *fi;
4275         u32 size = (u32)(new_size - found_key->offset);
4276         struct btrfs_root *root = BTRFS_I(inode)->root;
4277
4278         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4279
4280         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4281                 loff_t offset = new_size;
4282                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4283
4284                 /*
4285                  * Zero out the remaining of the last page of our inline extent,
4286                  * instead of directly truncating our inline extent here - that
4287                  * would be much more complex (decompressing all the data, then
4288                  * compressing the truncated data, which might be bigger than
4289                  * the size of the inline extent, resize the extent, etc).
4290                  * We release the path because to get the page we might need to
4291                  * read the extent item from disk (data not in the page cache).
4292                  */
4293                 btrfs_release_path(path);
4294                 return btrfs_truncate_block(inode, offset, page_end - offset,
4295                                         0);
4296         }
4297
4298         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4299         size = btrfs_file_extent_calc_inline_size(size);
4300         btrfs_truncate_item(root, path, size, 1);
4301
4302         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4303                 inode_sub_bytes(inode, item_end + 1 - new_size);
4304
4305         return 0;
4306 }
4307
4308 /*
4309  * this can truncate away extent items, csum items and directory items.
4310  * It starts at a high offset and removes keys until it can't find
4311  * any higher than new_size
4312  *
4313  * csum items that cross the new i_size are truncated to the new size
4314  * as well.
4315  *
4316  * min_type is the minimum key type to truncate down to.  If set to 0, this
4317  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4318  */
4319 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4320                                struct btrfs_root *root,
4321                                struct inode *inode,
4322                                u64 new_size, u32 min_type)
4323 {
4324         struct btrfs_path *path;
4325         struct extent_buffer *leaf;
4326         struct btrfs_file_extent_item *fi;
4327         struct btrfs_key key;
4328         struct btrfs_key found_key;
4329         u64 extent_start = 0;
4330         u64 extent_num_bytes = 0;
4331         u64 extent_offset = 0;
4332         u64 item_end = 0;
4333         u64 last_size = new_size;
4334         u32 found_type = (u8)-1;
4335         int found_extent;
4336         int del_item;
4337         int pending_del_nr = 0;
4338         int pending_del_slot = 0;
4339         int extent_type = -1;
4340         int ret;
4341         int err = 0;
4342         u64 ino = btrfs_ino(inode);
4343         u64 bytes_deleted = 0;
4344         bool be_nice = 0;
4345         bool should_throttle = 0;
4346         bool should_end = 0;
4347
4348         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4349
4350         /*
4351          * for non-free space inodes and ref cows, we want to back off from
4352          * time to time
4353          */
4354         if (!btrfs_is_free_space_inode(inode) &&
4355             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4356                 be_nice = 1;
4357
4358         path = btrfs_alloc_path();
4359         if (!path)
4360                 return -ENOMEM;
4361         path->reada = READA_BACK;
4362
4363         /*
4364          * We want to drop from the next block forward in case this new size is
4365          * not block aligned since we will be keeping the last block of the
4366          * extent just the way it is.
4367          */
4368         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4369             root == root->fs_info->tree_root)
4370                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4371                                         root->sectorsize), (u64)-1, 0);
4372
4373         /*
4374          * This function is also used to drop the items in the log tree before
4375          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4376          * it is used to drop the loged items. So we shouldn't kill the delayed
4377          * items.
4378          */
4379         if (min_type == 0 && root == BTRFS_I(inode)->root)
4380                 btrfs_kill_delayed_inode_items(inode);
4381
4382         key.objectid = ino;
4383         key.offset = (u64)-1;
4384         key.type = (u8)-1;
4385
4386 search_again:
4387         /*
4388          * with a 16K leaf size and 128MB extents, you can actually queue
4389          * up a huge file in a single leaf.  Most of the time that
4390          * bytes_deleted is > 0, it will be huge by the time we get here
4391          */
4392         if (be_nice && bytes_deleted > SZ_32M) {
4393                 if (btrfs_should_end_transaction(trans, root)) {
4394                         err = -EAGAIN;
4395                         goto error;
4396                 }
4397         }
4398
4399
4400         path->leave_spinning = 1;
4401         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4402         if (ret < 0) {
4403                 err = ret;
4404                 goto out;
4405         }
4406
4407         if (ret > 0) {
4408                 /* there are no items in the tree for us to truncate, we're
4409                  * done
4410                  */
4411                 if (path->slots[0] == 0)
4412                         goto out;
4413                 path->slots[0]--;
4414         }
4415
4416         while (1) {
4417                 fi = NULL;
4418                 leaf = path->nodes[0];
4419                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4420                 found_type = found_key.type;
4421
4422                 if (found_key.objectid != ino)
4423                         break;
4424
4425                 if (found_type < min_type)
4426                         break;
4427
4428                 item_end = found_key.offset;
4429                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4430                         fi = btrfs_item_ptr(leaf, path->slots[0],
4431                                             struct btrfs_file_extent_item);
4432                         extent_type = btrfs_file_extent_type(leaf, fi);
4433                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4434                                 item_end +=
4435                                     btrfs_file_extent_num_bytes(leaf, fi);
4436                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4437                                 item_end += btrfs_file_extent_inline_len(leaf,
4438                                                          path->slots[0], fi);
4439                         }
4440                         item_end--;
4441                 }
4442                 if (found_type > min_type) {
4443                         del_item = 1;
4444                 } else {
4445                         if (item_end < new_size)
4446                                 break;
4447                         if (found_key.offset >= new_size)
4448                                 del_item = 1;
4449                         else
4450                                 del_item = 0;
4451                 }
4452                 found_extent = 0;
4453                 /* FIXME, shrink the extent if the ref count is only 1 */
4454                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4455                         goto delete;
4456
4457                 if (del_item)
4458                         last_size = found_key.offset;
4459                 else
4460                         last_size = new_size;
4461
4462                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4463                         u64 num_dec;
4464                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4465                         if (!del_item) {
4466                                 u64 orig_num_bytes =
4467                                         btrfs_file_extent_num_bytes(leaf, fi);
4468                                 extent_num_bytes = ALIGN(new_size -
4469                                                 found_key.offset,
4470                                                 root->sectorsize);
4471                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4472                                                          extent_num_bytes);
4473                                 num_dec = (orig_num_bytes -
4474                                            extent_num_bytes);
4475                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4476                                              &root->state) &&
4477                                     extent_start != 0)
4478                                         inode_sub_bytes(inode, num_dec);
4479                                 btrfs_mark_buffer_dirty(leaf);
4480                         } else {
4481                                 extent_num_bytes =
4482                                         btrfs_file_extent_disk_num_bytes(leaf,
4483                                                                          fi);
4484                                 extent_offset = found_key.offset -
4485                                         btrfs_file_extent_offset(leaf, fi);
4486
4487                                 /* FIXME blocksize != 4096 */
4488                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4489                                 if (extent_start != 0) {
4490                                         found_extent = 1;
4491                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4492                                                      &root->state))
4493                                                 inode_sub_bytes(inode, num_dec);
4494                                 }
4495                         }
4496                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4497                         /*
4498                          * we can't truncate inline items that have had
4499                          * special encodings
4500                          */
4501                         if (!del_item &&
4502                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4503                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4504
4505                                 /*
4506                                  * Need to release path in order to truncate a
4507                                  * compressed extent. So delete any accumulated
4508                                  * extent items so far.
4509                                  */
4510                                 if (btrfs_file_extent_compression(leaf, fi) !=
4511                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4512                                         err = btrfs_del_items(trans, root, path,
4513                                                               pending_del_slot,
4514                                                               pending_del_nr);
4515                                         if (err) {
4516                                                 btrfs_abort_transaction(trans,
4517                                                                         err);
4518                                                 goto error;
4519                                         }
4520                                         pending_del_nr = 0;
4521                                 }
4522
4523                                 err = truncate_inline_extent(inode, path,
4524                                                              &found_key,
4525                                                              item_end,
4526                                                              new_size);
4527                                 if (err) {
4528                                         btrfs_abort_transaction(trans, err);
4529                                         goto error;
4530                                 }
4531                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4532                                             &root->state)) {
4533                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4534                         }
4535                 }
4536 delete:
4537                 if (del_item) {
4538                         if (!pending_del_nr) {
4539                                 /* no pending yet, add ourselves */
4540                                 pending_del_slot = path->slots[0];
4541                                 pending_del_nr = 1;
4542                         } else if (pending_del_nr &&
4543                                    path->slots[0] + 1 == pending_del_slot) {
4544                                 /* hop on the pending chunk */
4545                                 pending_del_nr++;
4546                                 pending_del_slot = path->slots[0];
4547                         } else {
4548                                 BUG();
4549                         }
4550                 } else {
4551                         break;
4552                 }
4553                 should_throttle = 0;
4554
4555                 if (found_extent &&
4556                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4557                      root == root->fs_info->tree_root)) {
4558                         btrfs_set_path_blocking(path);
4559                         bytes_deleted += extent_num_bytes;
4560                         ret = btrfs_free_extent(trans, root, extent_start,
4561                                                 extent_num_bytes, 0,
4562                                                 btrfs_header_owner(leaf),
4563                                                 ino, extent_offset);
4564                         BUG_ON(ret);
4565                         if (btrfs_should_throttle_delayed_refs(trans, root))
4566                                 btrfs_async_run_delayed_refs(root,
4567                                                              trans->transid,
4568                                         trans->delayed_ref_updates * 2, 0);
4569                         if (be_nice) {
4570                                 if (truncate_space_check(trans, root,
4571                                                          extent_num_bytes)) {
4572                                         should_end = 1;
4573                                 }
4574                                 if (btrfs_should_throttle_delayed_refs(trans,
4575                                                                        root)) {
4576                                         should_throttle = 1;
4577                                 }
4578                         }
4579                 }
4580
4581                 if (found_type == BTRFS_INODE_ITEM_KEY)
4582                         break;
4583
4584                 if (path->slots[0] == 0 ||
4585                     path->slots[0] != pending_del_slot ||
4586                     should_throttle || should_end) {
4587                         if (pending_del_nr) {
4588                                 ret = btrfs_del_items(trans, root, path,
4589                                                 pending_del_slot,
4590                                                 pending_del_nr);
4591                                 if (ret) {
4592                                         btrfs_abort_transaction(trans, ret);
4593                                         goto error;
4594                                 }
4595                                 pending_del_nr = 0;
4596                         }
4597                         btrfs_release_path(path);
4598                         if (should_throttle) {
4599                                 unsigned long updates = trans->delayed_ref_updates;
4600                                 if (updates) {
4601                                         trans->delayed_ref_updates = 0;
4602                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4603                                         if (ret && !err)
4604                                                 err = ret;
4605                                 }
4606                         }
4607                         /*
4608                          * if we failed to refill our space rsv, bail out
4609                          * and let the transaction restart
4610                          */
4611                         if (should_end) {
4612                                 err = -EAGAIN;
4613                                 goto error;
4614                         }
4615                         goto search_again;
4616                 } else {
4617                         path->slots[0]--;
4618                 }
4619         }
4620 out:
4621         if (pending_del_nr) {
4622                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4623                                       pending_del_nr);
4624                 if (ret)
4625                         btrfs_abort_transaction(trans, ret);
4626         }
4627 error:
4628         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4629                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4630
4631         btrfs_free_path(path);
4632
4633         if (be_nice && bytes_deleted > SZ_32M) {
4634                 unsigned long updates = trans->delayed_ref_updates;
4635                 if (updates) {
4636                         trans->delayed_ref_updates = 0;
4637                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4638                         if (ret && !err)
4639                                 err = ret;
4640                 }
4641         }
4642         return err;
4643 }
4644
4645 /*
4646  * btrfs_truncate_block - read, zero a chunk and write a block
4647  * @inode - inode that we're zeroing
4648  * @from - the offset to start zeroing
4649  * @len - the length to zero, 0 to zero the entire range respective to the
4650  *      offset
4651  * @front - zero up to the offset instead of from the offset on
4652  *
4653  * This will find the block for the "from" offset and cow the block and zero the
4654  * part we want to zero.  This is used with truncate and hole punching.
4655  */
4656 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4657                         int front)
4658 {
4659         struct address_space *mapping = inode->i_mapping;
4660         struct btrfs_root *root = BTRFS_I(inode)->root;
4661         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4662         struct btrfs_ordered_extent *ordered;
4663         struct extent_state *cached_state = NULL;
4664         char *kaddr;
4665         u32 blocksize = root->sectorsize;
4666         pgoff_t index = from >> PAGE_SHIFT;
4667         unsigned offset = from & (blocksize - 1);
4668         struct page *page;
4669         gfp_t mask = btrfs_alloc_write_mask(mapping);
4670         int ret = 0;
4671         u64 block_start;
4672         u64 block_end;
4673
4674         if ((offset & (blocksize - 1)) == 0 &&
4675             (!len || ((len & (blocksize - 1)) == 0)))
4676                 goto out;
4677
4678         ret = btrfs_delalloc_reserve_space(inode,
4679                         round_down(from, blocksize), blocksize);
4680         if (ret)
4681                 goto out;
4682
4683 again:
4684         page = find_or_create_page(mapping, index, mask);
4685         if (!page) {
4686                 btrfs_delalloc_release_space(inode,
4687                                 round_down(from, blocksize),
4688                                 blocksize);
4689                 ret = -ENOMEM;
4690                 goto out;
4691         }
4692
4693         block_start = round_down(from, blocksize);
4694         block_end = block_start + blocksize - 1;
4695
4696         if (!PageUptodate(page)) {
4697                 ret = btrfs_readpage(NULL, page);
4698                 lock_page(page);
4699                 if (page->mapping != mapping) {
4700                         unlock_page(page);
4701                         put_page(page);
4702                         goto again;
4703                 }
4704                 if (!PageUptodate(page)) {
4705                         ret = -EIO;
4706                         goto out_unlock;
4707                 }
4708         }
4709         wait_on_page_writeback(page);
4710
4711         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4712         set_page_extent_mapped(page);
4713
4714         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4715         if (ordered) {
4716                 unlock_extent_cached(io_tree, block_start, block_end,
4717                                      &cached_state, GFP_NOFS);
4718                 unlock_page(page);
4719                 put_page(page);
4720                 btrfs_start_ordered_extent(inode, ordered, 1);
4721                 btrfs_put_ordered_extent(ordered);
4722                 goto again;
4723         }
4724
4725         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4726                           EXTENT_DIRTY | EXTENT_DELALLOC |
4727                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4728                           0, 0, &cached_state, GFP_NOFS);
4729
4730         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4731                                         &cached_state);
4732         if (ret) {
4733                 unlock_extent_cached(io_tree, block_start, block_end,
4734                                      &cached_state, GFP_NOFS);
4735                 goto out_unlock;
4736         }
4737
4738         if (offset != blocksize) {
4739                 if (!len)
4740                         len = blocksize - offset;
4741                 kaddr = kmap(page);
4742                 if (front)
4743                         memset(kaddr + (block_start - page_offset(page)),
4744                                 0, offset);
4745                 else
4746                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4747                                 0, len);
4748                 flush_dcache_page(page);
4749                 kunmap(page);
4750         }
4751         ClearPageChecked(page);
4752         set_page_dirty(page);
4753         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4754                              GFP_NOFS);
4755
4756 out_unlock:
4757         if (ret)
4758                 btrfs_delalloc_release_space(inode, block_start,
4759                                              blocksize);
4760         unlock_page(page);
4761         put_page(page);
4762 out:
4763         return ret;
4764 }
4765
4766 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4767                              u64 offset, u64 len)
4768 {
4769         struct btrfs_trans_handle *trans;
4770         int ret;
4771
4772         /*
4773          * Still need to make sure the inode looks like it's been updated so
4774          * that any holes get logged if we fsync.
4775          */
4776         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4777                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4778                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4779                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4780                 return 0;
4781         }
4782
4783         /*
4784          * 1 - for the one we're dropping
4785          * 1 - for the one we're adding
4786          * 1 - for updating the inode.
4787          */
4788         trans = btrfs_start_transaction(root, 3);
4789         if (IS_ERR(trans))
4790                 return PTR_ERR(trans);
4791
4792         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4793         if (ret) {
4794                 btrfs_abort_transaction(trans, ret);
4795                 btrfs_end_transaction(trans, root);
4796                 return ret;
4797         }
4798
4799         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4800                                        0, 0, len, 0, len, 0, 0, 0);
4801         if (ret)
4802                 btrfs_abort_transaction(trans, ret);
4803         else
4804                 btrfs_update_inode(trans, root, inode);
4805         btrfs_end_transaction(trans, root);
4806         return ret;
4807 }
4808
4809 /*
4810  * This function puts in dummy file extents for the area we're creating a hole
4811  * for.  So if we are truncating this file to a larger size we need to insert
4812  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4813  * the range between oldsize and size
4814  */
4815 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4816 {
4817         struct btrfs_root *root = BTRFS_I(inode)->root;
4818         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4819         struct extent_map *em = NULL;
4820         struct extent_state *cached_state = NULL;
4821         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4822         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4823         u64 block_end = ALIGN(size, root->sectorsize);
4824         u64 last_byte;
4825         u64 cur_offset;
4826         u64 hole_size;
4827         int err = 0;
4828
4829         /*
4830          * If our size started in the middle of a block we need to zero out the
4831          * rest of the block before we expand the i_size, otherwise we could
4832          * expose stale data.
4833          */
4834         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4835         if (err)
4836                 return err;
4837
4838         if (size <= hole_start)
4839                 return 0;
4840
4841         while (1) {
4842                 struct btrfs_ordered_extent *ordered;
4843
4844                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4845                                  &cached_state);
4846                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4847                                                      block_end - hole_start);
4848                 if (!ordered)
4849                         break;
4850                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4851                                      &cached_state, GFP_NOFS);
4852                 btrfs_start_ordered_extent(inode, ordered, 1);
4853                 btrfs_put_ordered_extent(ordered);
4854         }
4855
4856         cur_offset = hole_start;
4857         while (1) {
4858                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4859                                 block_end - cur_offset, 0);
4860                 if (IS_ERR(em)) {
4861                         err = PTR_ERR(em);
4862                         em = NULL;
4863                         break;
4864                 }
4865                 last_byte = min(extent_map_end(em), block_end);
4866                 last_byte = ALIGN(last_byte , root->sectorsize);
4867                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4868                         struct extent_map *hole_em;
4869                         hole_size = last_byte - cur_offset;
4870
4871                         err = maybe_insert_hole(root, inode, cur_offset,
4872                                                 hole_size);
4873                         if (err)
4874                                 break;
4875                         btrfs_drop_extent_cache(inode, cur_offset,
4876                                                 cur_offset + hole_size - 1, 0);
4877                         hole_em = alloc_extent_map();
4878                         if (!hole_em) {
4879                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4880                                         &BTRFS_I(inode)->runtime_flags);
4881                                 goto next;
4882                         }
4883                         hole_em->start = cur_offset;
4884                         hole_em->len = hole_size;
4885                         hole_em->orig_start = cur_offset;
4886
4887                         hole_em->block_start = EXTENT_MAP_HOLE;
4888                         hole_em->block_len = 0;
4889                         hole_em->orig_block_len = 0;
4890                         hole_em->ram_bytes = hole_size;
4891                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4892                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4893                         hole_em->generation = root->fs_info->generation;
4894
4895                         while (1) {
4896                                 write_lock(&em_tree->lock);
4897                                 err = add_extent_mapping(em_tree, hole_em, 1);
4898                                 write_unlock(&em_tree->lock);
4899                                 if (err != -EEXIST)
4900                                         break;
4901                                 btrfs_drop_extent_cache(inode, cur_offset,
4902                                                         cur_offset +
4903                                                         hole_size - 1, 0);
4904                         }
4905                         free_extent_map(hole_em);
4906                 }
4907 next:
4908                 free_extent_map(em);
4909                 em = NULL;
4910                 cur_offset = last_byte;
4911                 if (cur_offset >= block_end)
4912                         break;
4913         }
4914         free_extent_map(em);
4915         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4916                              GFP_NOFS);
4917         return err;
4918 }
4919
4920 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4921 {
4922         struct btrfs_root *root = BTRFS_I(inode)->root;
4923         struct btrfs_trans_handle *trans;
4924         loff_t oldsize = i_size_read(inode);
4925         loff_t newsize = attr->ia_size;
4926         int mask = attr->ia_valid;
4927         int ret;
4928
4929         /*
4930          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4931          * special case where we need to update the times despite not having
4932          * these flags set.  For all other operations the VFS set these flags
4933          * explicitly if it wants a timestamp update.
4934          */
4935         if (newsize != oldsize) {
4936                 inode_inc_iversion(inode);
4937                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4938                         inode->i_ctime = inode->i_mtime =
4939                                 current_fs_time(inode->i_sb);
4940         }
4941
4942         if (newsize > oldsize) {
4943                 /*
4944                  * Don't do an expanding truncate while snapshoting is ongoing.
4945                  * This is to ensure the snapshot captures a fully consistent
4946                  * state of this file - if the snapshot captures this expanding
4947                  * truncation, it must capture all writes that happened before
4948                  * this truncation.
4949                  */
4950                 btrfs_wait_for_snapshot_creation(root);
4951                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4952                 if (ret) {
4953                         btrfs_end_write_no_snapshoting(root);
4954                         return ret;
4955                 }
4956
4957                 trans = btrfs_start_transaction(root, 1);
4958                 if (IS_ERR(trans)) {
4959                         btrfs_end_write_no_snapshoting(root);
4960                         return PTR_ERR(trans);
4961                 }
4962
4963                 i_size_write(inode, newsize);
4964                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4965                 pagecache_isize_extended(inode, oldsize, newsize);
4966                 ret = btrfs_update_inode(trans, root, inode);
4967                 btrfs_end_write_no_snapshoting(root);
4968                 btrfs_end_transaction(trans, root);
4969         } else {
4970
4971                 /*
4972                  * We're truncating a file that used to have good data down to
4973                  * zero. Make sure it gets into the ordered flush list so that
4974                  * any new writes get down to disk quickly.
4975                  */
4976                 if (newsize == 0)
4977                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4978                                 &BTRFS_I(inode)->runtime_flags);
4979
4980                 /*
4981                  * 1 for the orphan item we're going to add
4982                  * 1 for the orphan item deletion.
4983                  */
4984                 trans = btrfs_start_transaction(root, 2);
4985                 if (IS_ERR(trans))
4986                         return PTR_ERR(trans);
4987
4988                 /*
4989                  * We need to do this in case we fail at _any_ point during the
4990                  * actual truncate.  Once we do the truncate_setsize we could
4991                  * invalidate pages which forces any outstanding ordered io to
4992                  * be instantly completed which will give us extents that need
4993                  * to be truncated.  If we fail to get an orphan inode down we
4994                  * could have left over extents that were never meant to live,
4995                  * so we need to guarantee from this point on that everything
4996                  * will be consistent.
4997                  */
4998                 ret = btrfs_orphan_add(trans, inode);
4999                 btrfs_end_transaction(trans, root);
5000                 if (ret)
5001                         return ret;
5002
5003                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5004                 truncate_setsize(inode, newsize);
5005
5006                 /* Disable nonlocked read DIO to avoid the end less truncate */
5007                 btrfs_inode_block_unlocked_dio(inode);
5008                 inode_dio_wait(inode);
5009                 btrfs_inode_resume_unlocked_dio(inode);
5010
5011                 ret = btrfs_truncate(inode);
5012                 if (ret && inode->i_nlink) {
5013                         int err;
5014
5015                         /*
5016                          * failed to truncate, disk_i_size is only adjusted down
5017                          * as we remove extents, so it should represent the true
5018                          * size of the inode, so reset the in memory size and
5019                          * delete our orphan entry.
5020                          */
5021                         trans = btrfs_join_transaction(root);
5022                         if (IS_ERR(trans)) {
5023                                 btrfs_orphan_del(NULL, inode);
5024                                 return ret;
5025                         }
5026                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5027                         err = btrfs_orphan_del(trans, inode);
5028                         if (err)
5029                                 btrfs_abort_transaction(trans, err);
5030                         btrfs_end_transaction(trans, root);
5031                 }
5032         }
5033
5034         return ret;
5035 }
5036
5037 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5038 {
5039         struct inode *inode = d_inode(dentry);
5040         struct btrfs_root *root = BTRFS_I(inode)->root;
5041         int err;
5042
5043         if (btrfs_root_readonly(root))
5044                 return -EROFS;
5045
5046         err = inode_change_ok(inode, attr);
5047         if (err)
5048                 return err;
5049
5050         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5051                 err = btrfs_setsize(inode, attr);
5052                 if (err)
5053                         return err;
5054         }
5055
5056         if (attr->ia_valid) {
5057                 setattr_copy(inode, attr);
5058                 inode_inc_iversion(inode);
5059                 err = btrfs_dirty_inode(inode);
5060
5061                 if (!err && attr->ia_valid & ATTR_MODE)
5062                         err = posix_acl_chmod(inode, inode->i_mode);
5063         }
5064
5065         return err;
5066 }
5067
5068 /*
5069  * While truncating the inode pages during eviction, we get the VFS calling
5070  * btrfs_invalidatepage() against each page of the inode. This is slow because
5071  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5072  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5073  * extent_state structures over and over, wasting lots of time.
5074  *
5075  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5076  * those expensive operations on a per page basis and do only the ordered io
5077  * finishing, while we release here the extent_map and extent_state structures,
5078  * without the excessive merging and splitting.
5079  */
5080 static void evict_inode_truncate_pages(struct inode *inode)
5081 {
5082         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5083         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5084         struct rb_node *node;
5085
5086         ASSERT(inode->i_state & I_FREEING);
5087         truncate_inode_pages_final(&inode->i_data);
5088
5089         write_lock(&map_tree->lock);
5090         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5091                 struct extent_map *em;
5092
5093                 node = rb_first(&map_tree->map);
5094                 em = rb_entry(node, struct extent_map, rb_node);
5095                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5096                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5097                 remove_extent_mapping(map_tree, em);
5098                 free_extent_map(em);
5099                 if (need_resched()) {
5100                         write_unlock(&map_tree->lock);
5101                         cond_resched();
5102                         write_lock(&map_tree->lock);
5103                 }
5104         }
5105         write_unlock(&map_tree->lock);
5106
5107         /*
5108          * Keep looping until we have no more ranges in the io tree.
5109          * We can have ongoing bios started by readpages (called from readahead)
5110          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5111          * still in progress (unlocked the pages in the bio but did not yet
5112          * unlocked the ranges in the io tree). Therefore this means some
5113          * ranges can still be locked and eviction started because before
5114          * submitting those bios, which are executed by a separate task (work
5115          * queue kthread), inode references (inode->i_count) were not taken
5116          * (which would be dropped in the end io callback of each bio).
5117          * Therefore here we effectively end up waiting for those bios and
5118          * anyone else holding locked ranges without having bumped the inode's
5119          * reference count - if we don't do it, when they access the inode's
5120          * io_tree to unlock a range it may be too late, leading to an
5121          * use-after-free issue.
5122          */
5123         spin_lock(&io_tree->lock);
5124         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5125                 struct extent_state *state;
5126                 struct extent_state *cached_state = NULL;
5127                 u64 start;
5128                 u64 end;
5129
5130                 node = rb_first(&io_tree->state);
5131                 state = rb_entry(node, struct extent_state, rb_node);
5132                 start = state->start;
5133                 end = state->end;
5134                 spin_unlock(&io_tree->lock);
5135
5136                 lock_extent_bits(io_tree, start, end, &cached_state);
5137
5138                 /*
5139                  * If still has DELALLOC flag, the extent didn't reach disk,
5140                  * and its reserved space won't be freed by delayed_ref.
5141                  * So we need to free its reserved space here.
5142                  * (Refer to comment in btrfs_invalidatepage, case 2)
5143                  *
5144                  * Note, end is the bytenr of last byte, so we need + 1 here.
5145                  */
5146                 if (state->state & EXTENT_DELALLOC)
5147                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5148
5149                 clear_extent_bit(io_tree, start, end,
5150                                  EXTENT_LOCKED | EXTENT_DIRTY |
5151                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5152                                  EXTENT_DEFRAG, 1, 1,
5153                                  &cached_state, GFP_NOFS);
5154
5155                 cond_resched();
5156                 spin_lock(&io_tree->lock);
5157         }
5158         spin_unlock(&io_tree->lock);
5159 }
5160
5161 void btrfs_evict_inode(struct inode *inode)
5162 {
5163         struct btrfs_trans_handle *trans;
5164         struct btrfs_root *root = BTRFS_I(inode)->root;
5165         struct btrfs_block_rsv *rsv, *global_rsv;
5166         int steal_from_global = 0;
5167         u64 min_size;
5168         int ret;
5169
5170         trace_btrfs_inode_evict(inode);
5171
5172         if (!root) {
5173                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5174                 return;
5175         }
5176
5177         min_size = btrfs_calc_trunc_metadata_size(root, 1);
5178
5179         evict_inode_truncate_pages(inode);
5180
5181         if (inode->i_nlink &&
5182             ((btrfs_root_refs(&root->root_item) != 0 &&
5183               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5184              btrfs_is_free_space_inode(inode)))
5185                 goto no_delete;
5186
5187         if (is_bad_inode(inode)) {
5188                 btrfs_orphan_del(NULL, inode);
5189                 goto no_delete;
5190         }
5191         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5192         if (!special_file(inode->i_mode))
5193                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5194
5195         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5196
5197         if (root->fs_info->log_root_recovering) {
5198                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5199                                  &BTRFS_I(inode)->runtime_flags));
5200                 goto no_delete;
5201         }
5202
5203         if (inode->i_nlink > 0) {
5204                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5205                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5206                 goto no_delete;
5207         }
5208
5209         ret = btrfs_commit_inode_delayed_inode(inode);
5210         if (ret) {
5211                 btrfs_orphan_del(NULL, inode);
5212                 goto no_delete;
5213         }
5214
5215         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5216         if (!rsv) {
5217                 btrfs_orphan_del(NULL, inode);
5218                 goto no_delete;
5219         }
5220         rsv->size = min_size;
5221         rsv->failfast = 1;
5222         global_rsv = &root->fs_info->global_block_rsv;
5223
5224         btrfs_i_size_write(inode, 0);
5225
5226         /*
5227          * This is a bit simpler than btrfs_truncate since we've already
5228          * reserved our space for our orphan item in the unlink, so we just
5229          * need to reserve some slack space in case we add bytes and update
5230          * inode item when doing the truncate.
5231          */
5232         while (1) {
5233                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5234                                              BTRFS_RESERVE_FLUSH_LIMIT);
5235
5236                 /*
5237                  * Try and steal from the global reserve since we will
5238                  * likely not use this space anyway, we want to try as
5239                  * hard as possible to get this to work.
5240                  */
5241                 if (ret)
5242                         steal_from_global++;
5243                 else
5244                         steal_from_global = 0;
5245                 ret = 0;
5246
5247                 /*
5248                  * steal_from_global == 0: we reserved stuff, hooray!
5249                  * steal_from_global == 1: we didn't reserve stuff, boo!
5250                  * steal_from_global == 2: we've committed, still not a lot of
5251                  * room but maybe we'll have room in the global reserve this
5252                  * time.
5253                  * steal_from_global == 3: abandon all hope!
5254                  */
5255                 if (steal_from_global > 2) {
5256                         btrfs_warn(root->fs_info,
5257                                 "Could not get space for a delete, will truncate on mount %d",
5258                                 ret);
5259                         btrfs_orphan_del(NULL, inode);
5260                         btrfs_free_block_rsv(root, rsv);
5261                         goto no_delete;
5262                 }
5263
5264                 trans = btrfs_join_transaction(root);
5265                 if (IS_ERR(trans)) {
5266                         btrfs_orphan_del(NULL, inode);
5267                         btrfs_free_block_rsv(root, rsv);
5268                         goto no_delete;
5269                 }
5270
5271                 /*
5272                  * We can't just steal from the global reserve, we need to make
5273                  * sure there is room to do it, if not we need to commit and try
5274                  * again.
5275                  */
5276                 if (steal_from_global) {
5277                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5278                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5279                                                               min_size, 0);
5280                         else
5281                                 ret = -ENOSPC;
5282                 }
5283
5284                 /*
5285                  * Couldn't steal from the global reserve, we have too much
5286                  * pending stuff built up, commit the transaction and try it
5287                  * again.
5288                  */
5289                 if (ret) {
5290                         ret = btrfs_commit_transaction(trans, root);
5291                         if (ret) {
5292                                 btrfs_orphan_del(NULL, inode);
5293                                 btrfs_free_block_rsv(root, rsv);
5294                                 goto no_delete;
5295                         }
5296                         continue;
5297                 } else {
5298                         steal_from_global = 0;
5299                 }
5300
5301                 trans->block_rsv = rsv;
5302
5303                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5304                 if (ret != -ENOSPC && ret != -EAGAIN)
5305                         break;
5306
5307                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5308                 btrfs_end_transaction(trans, root);
5309                 trans = NULL;
5310                 btrfs_btree_balance_dirty(root);
5311         }
5312
5313         btrfs_free_block_rsv(root, rsv);
5314
5315         /*
5316          * Errors here aren't a big deal, it just means we leave orphan items
5317          * in the tree.  They will be cleaned up on the next mount.
5318          */
5319         if (ret == 0) {
5320                 trans->block_rsv = root->orphan_block_rsv;
5321                 btrfs_orphan_del(trans, inode);
5322         } else {
5323                 btrfs_orphan_del(NULL, inode);
5324         }
5325
5326         trans->block_rsv = &root->fs_info->trans_block_rsv;
5327         if (!(root == root->fs_info->tree_root ||
5328               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5329                 btrfs_return_ino(root, btrfs_ino(inode));
5330
5331         btrfs_end_transaction(trans, root);
5332         btrfs_btree_balance_dirty(root);
5333 no_delete:
5334         btrfs_remove_delayed_node(inode);
5335         clear_inode(inode);
5336 }
5337
5338 /*
5339  * this returns the key found in the dir entry in the location pointer.
5340  * If no dir entries were found, location->objectid is 0.
5341  */
5342 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5343                                struct btrfs_key *location)
5344 {
5345         const char *name = dentry->d_name.name;
5346         int namelen = dentry->d_name.len;
5347         struct btrfs_dir_item *di;
5348         struct btrfs_path *path;
5349         struct btrfs_root *root = BTRFS_I(dir)->root;
5350         int ret = 0;
5351
5352         path = btrfs_alloc_path();
5353         if (!path)
5354                 return -ENOMEM;
5355
5356         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5357                                     namelen, 0);
5358         if (IS_ERR(di))
5359                 ret = PTR_ERR(di);
5360
5361         if (IS_ERR_OR_NULL(di))
5362                 goto out_err;
5363
5364         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5365 out:
5366         btrfs_free_path(path);
5367         return ret;
5368 out_err:
5369         location->objectid = 0;
5370         goto out;
5371 }
5372
5373 /*
5374  * when we hit a tree root in a directory, the btrfs part of the inode
5375  * needs to be changed to reflect the root directory of the tree root.  This
5376  * is kind of like crossing a mount point.
5377  */
5378 static int fixup_tree_root_location(struct btrfs_root *root,
5379                                     struct inode *dir,
5380                                     struct dentry *dentry,
5381                                     struct btrfs_key *location,
5382                                     struct btrfs_root **sub_root)
5383 {
5384         struct btrfs_path *path;
5385         struct btrfs_root *new_root;
5386         struct btrfs_root_ref *ref;
5387         struct extent_buffer *leaf;
5388         struct btrfs_key key;
5389         int ret;
5390         int err = 0;
5391
5392         path = btrfs_alloc_path();
5393         if (!path) {
5394                 err = -ENOMEM;
5395                 goto out;
5396         }
5397
5398         err = -ENOENT;
5399         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5400         key.type = BTRFS_ROOT_REF_KEY;
5401         key.offset = location->objectid;
5402
5403         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5404                                 0, 0);
5405         if (ret) {
5406                 if (ret < 0)
5407                         err = ret;
5408                 goto out;
5409         }
5410
5411         leaf = path->nodes[0];
5412         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5413         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5414             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5415                 goto out;
5416
5417         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5418                                    (unsigned long)(ref + 1),
5419                                    dentry->d_name.len);
5420         if (ret)
5421                 goto out;
5422
5423         btrfs_release_path(path);
5424
5425         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5426         if (IS_ERR(new_root)) {
5427                 err = PTR_ERR(new_root);
5428                 goto out;
5429         }
5430
5431         *sub_root = new_root;
5432         location->objectid = btrfs_root_dirid(&new_root->root_item);
5433         location->type = BTRFS_INODE_ITEM_KEY;
5434         location->offset = 0;
5435         err = 0;
5436 out:
5437         btrfs_free_path(path);
5438         return err;
5439 }
5440
5441 static void inode_tree_add(struct inode *inode)
5442 {
5443         struct btrfs_root *root = BTRFS_I(inode)->root;
5444         struct btrfs_inode *entry;
5445         struct rb_node **p;
5446         struct rb_node *parent;
5447         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5448         u64 ino = btrfs_ino(inode);
5449
5450         if (inode_unhashed(inode))
5451                 return;
5452         parent = NULL;
5453         spin_lock(&root->inode_lock);
5454         p = &root->inode_tree.rb_node;
5455         while (*p) {
5456                 parent = *p;
5457                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5458
5459                 if (ino < btrfs_ino(&entry->vfs_inode))
5460                         p = &parent->rb_left;
5461                 else if (ino > btrfs_ino(&entry->vfs_inode))
5462                         p = &parent->rb_right;
5463                 else {
5464                         WARN_ON(!(entry->vfs_inode.i_state &
5465                                   (I_WILL_FREE | I_FREEING)));
5466                         rb_replace_node(parent, new, &root->inode_tree);
5467                         RB_CLEAR_NODE(parent);
5468                         spin_unlock(&root->inode_lock);
5469                         return;
5470                 }
5471         }
5472         rb_link_node(new, parent, p);
5473         rb_insert_color(new, &root->inode_tree);
5474         spin_unlock(&root->inode_lock);
5475 }
5476
5477 static void inode_tree_del(struct inode *inode)
5478 {
5479         struct btrfs_root *root = BTRFS_I(inode)->root;
5480         int empty = 0;
5481
5482         spin_lock(&root->inode_lock);
5483         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5484                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5485                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5486                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5487         }
5488         spin_unlock(&root->inode_lock);
5489
5490         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5491                 synchronize_srcu(&root->fs_info->subvol_srcu);
5492                 spin_lock(&root->inode_lock);
5493                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5494                 spin_unlock(&root->inode_lock);
5495                 if (empty)
5496                         btrfs_add_dead_root(root);
5497         }
5498 }
5499
5500 void btrfs_invalidate_inodes(struct btrfs_root *root)
5501 {
5502         struct rb_node *node;
5503         struct rb_node *prev;
5504         struct btrfs_inode *entry;
5505         struct inode *inode;
5506         u64 objectid = 0;
5507
5508         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5509                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5510
5511         spin_lock(&root->inode_lock);
5512 again:
5513         node = root->inode_tree.rb_node;
5514         prev = NULL;
5515         while (node) {
5516                 prev = node;
5517                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5518
5519                 if (objectid < btrfs_ino(&entry->vfs_inode))
5520                         node = node->rb_left;
5521                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5522                         node = node->rb_right;
5523                 else
5524                         break;
5525         }
5526         if (!node) {
5527                 while (prev) {
5528                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5529                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5530                                 node = prev;
5531                                 break;
5532                         }
5533                         prev = rb_next(prev);
5534                 }
5535         }
5536         while (node) {
5537                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5538                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5539                 inode = igrab(&entry->vfs_inode);
5540                 if (inode) {
5541                         spin_unlock(&root->inode_lock);
5542                         if (atomic_read(&inode->i_count) > 1)
5543                                 d_prune_aliases(inode);
5544                         /*
5545                          * btrfs_drop_inode will have it removed from
5546                          * the inode cache when its usage count
5547                          * hits zero.
5548                          */
5549                         iput(inode);
5550                         cond_resched();
5551                         spin_lock(&root->inode_lock);
5552                         goto again;
5553                 }
5554
5555                 if (cond_resched_lock(&root->inode_lock))
5556                         goto again;
5557
5558                 node = rb_next(node);
5559         }
5560         spin_unlock(&root->inode_lock);
5561 }
5562
5563 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5564 {
5565         struct btrfs_iget_args *args = p;
5566         inode->i_ino = args->location->objectid;
5567         memcpy(&BTRFS_I(inode)->location, args->location,
5568                sizeof(*args->location));
5569         BTRFS_I(inode)->root = args->root;
5570         return 0;
5571 }
5572
5573 static int btrfs_find_actor(struct inode *inode, void *opaque)
5574 {
5575         struct btrfs_iget_args *args = opaque;
5576         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5577                 args->root == BTRFS_I(inode)->root;
5578 }
5579
5580 static struct inode *btrfs_iget_locked(struct super_block *s,
5581                                        struct btrfs_key *location,
5582                                        struct btrfs_root *root)
5583 {
5584         struct inode *inode;
5585         struct btrfs_iget_args args;
5586         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5587
5588         args.location = location;
5589         args.root = root;
5590
5591         inode = iget5_locked(s, hashval, btrfs_find_actor,
5592                              btrfs_init_locked_inode,
5593                              (void *)&args);
5594         return inode;
5595 }
5596
5597 /* Get an inode object given its location and corresponding root.
5598  * Returns in *is_new if the inode was read from disk
5599  */
5600 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5601                          struct btrfs_root *root, int *new)
5602 {
5603         struct inode *inode;
5604
5605         inode = btrfs_iget_locked(s, location, root);
5606         if (!inode)
5607                 return ERR_PTR(-ENOMEM);
5608
5609         if (inode->i_state & I_NEW) {
5610                 btrfs_read_locked_inode(inode);
5611                 if (!is_bad_inode(inode)) {
5612                         inode_tree_add(inode);
5613                         unlock_new_inode(inode);
5614                         if (new)
5615                                 *new = 1;
5616                 } else {
5617                         unlock_new_inode(inode);
5618                         iput(inode);
5619                         inode = ERR_PTR(-ESTALE);
5620                 }
5621         }
5622
5623         return inode;
5624 }
5625
5626 static struct inode *new_simple_dir(struct super_block *s,
5627                                     struct btrfs_key *key,
5628                                     struct btrfs_root *root)
5629 {
5630         struct inode *inode = new_inode(s);
5631
5632         if (!inode)
5633                 return ERR_PTR(-ENOMEM);
5634
5635         BTRFS_I(inode)->root = root;
5636         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5637         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5638
5639         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5640         inode->i_op = &btrfs_dir_ro_inode_operations;
5641         inode->i_fop = &simple_dir_operations;
5642         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5643         inode->i_mtime = current_fs_time(inode->i_sb);
5644         inode->i_atime = inode->i_mtime;
5645         inode->i_ctime = inode->i_mtime;
5646         BTRFS_I(inode)->i_otime = inode->i_mtime;
5647
5648         return inode;
5649 }
5650
5651 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5652 {
5653         struct inode *inode;
5654         struct btrfs_root *root = BTRFS_I(dir)->root;
5655         struct btrfs_root *sub_root = root;
5656         struct btrfs_key location;
5657         int index;
5658         int ret = 0;
5659
5660         if (dentry->d_name.len > BTRFS_NAME_LEN)
5661                 return ERR_PTR(-ENAMETOOLONG);
5662
5663         ret = btrfs_inode_by_name(dir, dentry, &location);
5664         if (ret < 0)
5665                 return ERR_PTR(ret);
5666
5667         if (location.objectid == 0)
5668                 return ERR_PTR(-ENOENT);
5669
5670         if (location.type == BTRFS_INODE_ITEM_KEY) {
5671                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5672                 return inode;
5673         }
5674
5675         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5676
5677         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5678         ret = fixup_tree_root_location(root, dir, dentry,
5679                                        &location, &sub_root);
5680         if (ret < 0) {
5681                 if (ret != -ENOENT)
5682                         inode = ERR_PTR(ret);
5683                 else
5684                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5685         } else {
5686                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5687         }
5688         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5689
5690         if (!IS_ERR(inode) && root != sub_root) {
5691                 down_read(&root->fs_info->cleanup_work_sem);
5692                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5693                         ret = btrfs_orphan_cleanup(sub_root);
5694                 up_read(&root->fs_info->cleanup_work_sem);
5695                 if (ret) {
5696                         iput(inode);
5697                         inode = ERR_PTR(ret);
5698                 }
5699         }
5700
5701         return inode;
5702 }
5703
5704 static int btrfs_dentry_delete(const struct dentry *dentry)
5705 {
5706         struct btrfs_root *root;
5707         struct inode *inode = d_inode(dentry);
5708
5709         if (!inode && !IS_ROOT(dentry))
5710                 inode = d_inode(dentry->d_parent);
5711
5712         if (inode) {
5713                 root = BTRFS_I(inode)->root;
5714                 if (btrfs_root_refs(&root->root_item) == 0)
5715                         return 1;
5716
5717                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5718                         return 1;
5719         }
5720         return 0;
5721 }
5722
5723 static void btrfs_dentry_release(struct dentry *dentry)
5724 {
5725         kfree(dentry->d_fsdata);
5726 }
5727
5728 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5729                                    unsigned int flags)
5730 {
5731         struct inode *inode;
5732
5733         inode = btrfs_lookup_dentry(dir, dentry);
5734         if (IS_ERR(inode)) {
5735                 if (PTR_ERR(inode) == -ENOENT)
5736                         inode = NULL;
5737                 else
5738                         return ERR_CAST(inode);
5739         }
5740
5741         return d_splice_alias(inode, dentry);
5742 }
5743
5744 unsigned char btrfs_filetype_table[] = {
5745         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5746 };
5747
5748 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5749 {
5750         struct inode *inode = file_inode(file);
5751         struct btrfs_root *root = BTRFS_I(inode)->root;
5752         struct btrfs_item *item;
5753         struct btrfs_dir_item *di;
5754         struct btrfs_key key;
5755         struct btrfs_key found_key;
5756         struct btrfs_path *path;
5757         struct list_head ins_list;
5758         struct list_head del_list;
5759         int ret;
5760         struct extent_buffer *leaf;
5761         int slot;
5762         unsigned char d_type;
5763         int over = 0;
5764         u32 di_cur;
5765         u32 di_total;
5766         u32 di_len;
5767         int key_type = BTRFS_DIR_INDEX_KEY;
5768         char tmp_name[32];
5769         char *name_ptr;
5770         int name_len;
5771         int is_curr = 0;        /* ctx->pos points to the current index? */
5772         bool emitted;
5773         bool put = false;
5774
5775         /* FIXME, use a real flag for deciding about the key type */
5776         if (root->fs_info->tree_root == root)
5777                 key_type = BTRFS_DIR_ITEM_KEY;
5778
5779         if (!dir_emit_dots(file, ctx))
5780                 return 0;
5781
5782         path = btrfs_alloc_path();
5783         if (!path)
5784                 return -ENOMEM;
5785
5786         path->reada = READA_FORWARD;
5787
5788         if (key_type == BTRFS_DIR_INDEX_KEY) {
5789                 INIT_LIST_HEAD(&ins_list);
5790                 INIT_LIST_HEAD(&del_list);
5791                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5792                                                       &del_list);
5793         }
5794
5795         key.type = key_type;
5796         key.offset = ctx->pos;
5797         key.objectid = btrfs_ino(inode);
5798
5799         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5800         if (ret < 0)
5801                 goto err;
5802
5803         emitted = false;
5804         while (1) {
5805                 leaf = path->nodes[0];
5806                 slot = path->slots[0];
5807                 if (slot >= btrfs_header_nritems(leaf)) {
5808                         ret = btrfs_next_leaf(root, path);
5809                         if (ret < 0)
5810                                 goto err;
5811                         else if (ret > 0)
5812                                 break;
5813                         continue;
5814                 }
5815
5816                 item = btrfs_item_nr(slot);
5817                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5818
5819                 if (found_key.objectid != key.objectid)
5820                         break;
5821                 if (found_key.type != key_type)
5822                         break;
5823                 if (found_key.offset < ctx->pos)
5824                         goto next;
5825                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5826                     btrfs_should_delete_dir_index(&del_list,
5827                                                   found_key.offset))
5828                         goto next;
5829
5830                 ctx->pos = found_key.offset;
5831                 is_curr = 1;
5832
5833                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5834                 di_cur = 0;
5835                 di_total = btrfs_item_size(leaf, item);
5836
5837                 while (di_cur < di_total) {
5838                         struct btrfs_key location;
5839
5840                         if (verify_dir_item(root, leaf, di))
5841                                 break;
5842
5843                         name_len = btrfs_dir_name_len(leaf, di);
5844                         if (name_len <= sizeof(tmp_name)) {
5845                                 name_ptr = tmp_name;
5846                         } else {
5847                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5848                                 if (!name_ptr) {
5849                                         ret = -ENOMEM;
5850                                         goto err;
5851                                 }
5852                         }
5853                         read_extent_buffer(leaf, name_ptr,
5854                                            (unsigned long)(di + 1), name_len);
5855
5856                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5857                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5858
5859
5860                         /* is this a reference to our own snapshot? If so
5861                          * skip it.
5862                          *
5863                          * In contrast to old kernels, we insert the snapshot's
5864                          * dir item and dir index after it has been created, so
5865                          * we won't find a reference to our own snapshot. We
5866                          * still keep the following code for backward
5867                          * compatibility.
5868                          */
5869                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5870                             location.objectid == root->root_key.objectid) {
5871                                 over = 0;
5872                                 goto skip;
5873                         }
5874                         over = !dir_emit(ctx, name_ptr, name_len,
5875                                        location.objectid, d_type);
5876
5877 skip:
5878                         if (name_ptr != tmp_name)
5879                                 kfree(name_ptr);
5880
5881                         if (over)
5882                                 goto nopos;
5883                         emitted = true;
5884                         di_len = btrfs_dir_name_len(leaf, di) +
5885                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5886                         di_cur += di_len;
5887                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5888                 }
5889 next:
5890                 path->slots[0]++;
5891         }
5892
5893         if (key_type == BTRFS_DIR_INDEX_KEY) {
5894                 if (is_curr)
5895                         ctx->pos++;
5896                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5897                 if (ret)
5898                         goto nopos;
5899         }
5900
5901         /*
5902          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5903          * it was was set to the termination value in previous call. We assume
5904          * that "." and ".." were emitted if we reach this point and set the
5905          * termination value as well for an empty directory.
5906          */
5907         if (ctx->pos > 2 && !emitted)
5908                 goto nopos;
5909
5910         /* Reached end of directory/root. Bump pos past the last item. */
5911         ctx->pos++;
5912
5913         /*
5914          * Stop new entries from being returned after we return the last
5915          * entry.
5916          *
5917          * New directory entries are assigned a strictly increasing
5918          * offset.  This means that new entries created during readdir
5919          * are *guaranteed* to be seen in the future by that readdir.
5920          * This has broken buggy programs which operate on names as
5921          * they're returned by readdir.  Until we re-use freed offsets
5922          * we have this hack to stop new entries from being returned
5923          * under the assumption that they'll never reach this huge
5924          * offset.
5925          *
5926          * This is being careful not to overflow 32bit loff_t unless the
5927          * last entry requires it because doing so has broken 32bit apps
5928          * in the past.
5929          */
5930         if (key_type == BTRFS_DIR_INDEX_KEY) {
5931                 if (ctx->pos >= INT_MAX)
5932                         ctx->pos = LLONG_MAX;
5933                 else
5934                         ctx->pos = INT_MAX;
5935         }
5936 nopos:
5937         ret = 0;
5938 err:
5939         if (put)
5940                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5941         btrfs_free_path(path);
5942         return ret;
5943 }
5944
5945 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5946 {
5947         struct btrfs_root *root = BTRFS_I(inode)->root;
5948         struct btrfs_trans_handle *trans;
5949         int ret = 0;
5950         bool nolock = false;
5951
5952         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5953                 return 0;
5954
5955         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5956                 nolock = true;
5957
5958         if (wbc->sync_mode == WB_SYNC_ALL) {
5959                 if (nolock)
5960                         trans = btrfs_join_transaction_nolock(root);
5961                 else
5962                         trans = btrfs_join_transaction(root);
5963                 if (IS_ERR(trans))
5964                         return PTR_ERR(trans);
5965                 ret = btrfs_commit_transaction(trans, root);
5966         }
5967         return ret;
5968 }
5969
5970 /*
5971  * This is somewhat expensive, updating the tree every time the
5972  * inode changes.  But, it is most likely to find the inode in cache.
5973  * FIXME, needs more benchmarking...there are no reasons other than performance
5974  * to keep or drop this code.
5975  */
5976 static int btrfs_dirty_inode(struct inode *inode)
5977 {
5978         struct btrfs_root *root = BTRFS_I(inode)->root;
5979         struct btrfs_trans_handle *trans;
5980         int ret;
5981
5982         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5983                 return 0;
5984
5985         trans = btrfs_join_transaction(root);
5986         if (IS_ERR(trans))
5987                 return PTR_ERR(trans);
5988
5989         ret = btrfs_update_inode(trans, root, inode);
5990         if (ret && ret == -ENOSPC) {
5991                 /* whoops, lets try again with the full transaction */
5992                 btrfs_end_transaction(trans, root);
5993                 trans = btrfs_start_transaction(root, 1);
5994                 if (IS_ERR(trans))
5995                         return PTR_ERR(trans);
5996
5997                 ret = btrfs_update_inode(trans, root, inode);
5998         }
5999         btrfs_end_transaction(trans, root);
6000         if (BTRFS_I(inode)->delayed_node)
6001                 btrfs_balance_delayed_items(root);
6002
6003         return ret;
6004 }
6005
6006 /*
6007  * This is a copy of file_update_time.  We need this so we can return error on
6008  * ENOSPC for updating the inode in the case of file write and mmap writes.
6009  */
6010 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6011                              int flags)
6012 {
6013         struct btrfs_root *root = BTRFS_I(inode)->root;
6014
6015         if (btrfs_root_readonly(root))
6016                 return -EROFS;
6017
6018         if (flags & S_VERSION)
6019                 inode_inc_iversion(inode);
6020         if (flags & S_CTIME)
6021                 inode->i_ctime = *now;
6022         if (flags & S_MTIME)
6023                 inode->i_mtime = *now;
6024         if (flags & S_ATIME)
6025                 inode->i_atime = *now;
6026         return btrfs_dirty_inode(inode);
6027 }
6028
6029 /*
6030  * find the highest existing sequence number in a directory
6031  * and then set the in-memory index_cnt variable to reflect
6032  * free sequence numbers
6033  */
6034 static int btrfs_set_inode_index_count(struct inode *inode)
6035 {
6036         struct btrfs_root *root = BTRFS_I(inode)->root;
6037         struct btrfs_key key, found_key;
6038         struct btrfs_path *path;
6039         struct extent_buffer *leaf;
6040         int ret;
6041
6042         key.objectid = btrfs_ino(inode);
6043         key.type = BTRFS_DIR_INDEX_KEY;
6044         key.offset = (u64)-1;
6045
6046         path = btrfs_alloc_path();
6047         if (!path)
6048                 return -ENOMEM;
6049
6050         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6051         if (ret < 0)
6052                 goto out;
6053         /* FIXME: we should be able to handle this */
6054         if (ret == 0)
6055                 goto out;
6056         ret = 0;
6057
6058         /*
6059          * MAGIC NUMBER EXPLANATION:
6060          * since we search a directory based on f_pos we have to start at 2
6061          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6062          * else has to start at 2
6063          */
6064         if (path->slots[0] == 0) {
6065                 BTRFS_I(inode)->index_cnt = 2;
6066                 goto out;
6067         }
6068
6069         path->slots[0]--;
6070
6071         leaf = path->nodes[0];
6072         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6073
6074         if (found_key.objectid != btrfs_ino(inode) ||
6075             found_key.type != BTRFS_DIR_INDEX_KEY) {
6076                 BTRFS_I(inode)->index_cnt = 2;
6077                 goto out;
6078         }
6079
6080         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6081 out:
6082         btrfs_free_path(path);
6083         return ret;
6084 }
6085
6086 /*
6087  * helper to find a free sequence number in a given directory.  This current
6088  * code is very simple, later versions will do smarter things in the btree
6089  */
6090 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6091 {
6092         int ret = 0;
6093
6094         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6095                 ret = btrfs_inode_delayed_dir_index_count(dir);
6096                 if (ret) {
6097                         ret = btrfs_set_inode_index_count(dir);
6098                         if (ret)
6099                                 return ret;
6100                 }
6101         }
6102
6103         *index = BTRFS_I(dir)->index_cnt;
6104         BTRFS_I(dir)->index_cnt++;
6105
6106         return ret;
6107 }
6108
6109 static int btrfs_insert_inode_locked(struct inode *inode)
6110 {
6111         struct btrfs_iget_args args;
6112         args.location = &BTRFS_I(inode)->location;
6113         args.root = BTRFS_I(inode)->root;
6114
6115         return insert_inode_locked4(inode,
6116                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6117                    btrfs_find_actor, &args);
6118 }
6119
6120 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6121                                      struct btrfs_root *root,
6122                                      struct inode *dir,
6123                                      const char *name, int name_len,
6124                                      u64 ref_objectid, u64 objectid,
6125                                      umode_t mode, u64 *index)
6126 {
6127         struct inode *inode;
6128         struct btrfs_inode_item *inode_item;
6129         struct btrfs_key *location;
6130         struct btrfs_path *path;
6131         struct btrfs_inode_ref *ref;
6132         struct btrfs_key key[2];
6133         u32 sizes[2];
6134         int nitems = name ? 2 : 1;
6135         unsigned long ptr;
6136         int ret;
6137
6138         path = btrfs_alloc_path();
6139         if (!path)
6140                 return ERR_PTR(-ENOMEM);
6141
6142         inode = new_inode(root->fs_info->sb);
6143         if (!inode) {
6144                 btrfs_free_path(path);
6145                 return ERR_PTR(-ENOMEM);
6146         }
6147
6148         /*
6149          * O_TMPFILE, set link count to 0, so that after this point,
6150          * we fill in an inode item with the correct link count.
6151          */
6152         if (!name)
6153                 set_nlink(inode, 0);
6154
6155         /*
6156          * we have to initialize this early, so we can reclaim the inode
6157          * number if we fail afterwards in this function.
6158          */
6159         inode->i_ino = objectid;
6160
6161         if (dir && name) {
6162                 trace_btrfs_inode_request(dir);
6163
6164                 ret = btrfs_set_inode_index(dir, index);
6165                 if (ret) {
6166                         btrfs_free_path(path);
6167                         iput(inode);
6168                         return ERR_PTR(ret);
6169                 }
6170         } else if (dir) {
6171                 *index = 0;
6172         }
6173         /*
6174          * index_cnt is ignored for everything but a dir,
6175          * btrfs_get_inode_index_count has an explanation for the magic
6176          * number
6177          */
6178         BTRFS_I(inode)->index_cnt = 2;
6179         BTRFS_I(inode)->dir_index = *index;
6180         BTRFS_I(inode)->root = root;
6181         BTRFS_I(inode)->generation = trans->transid;
6182         inode->i_generation = BTRFS_I(inode)->generation;
6183
6184         /*
6185          * We could have gotten an inode number from somebody who was fsynced
6186          * and then removed in this same transaction, so let's just set full
6187          * sync since it will be a full sync anyway and this will blow away the
6188          * old info in the log.
6189          */
6190         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6191
6192         key[0].objectid = objectid;
6193         key[0].type = BTRFS_INODE_ITEM_KEY;
6194         key[0].offset = 0;
6195
6196         sizes[0] = sizeof(struct btrfs_inode_item);
6197
6198         if (name) {
6199                 /*
6200                  * Start new inodes with an inode_ref. This is slightly more
6201                  * efficient for small numbers of hard links since they will
6202                  * be packed into one item. Extended refs will kick in if we
6203                  * add more hard links than can fit in the ref item.
6204                  */
6205                 key[1].objectid = objectid;
6206                 key[1].type = BTRFS_INODE_REF_KEY;
6207                 key[1].offset = ref_objectid;
6208
6209                 sizes[1] = name_len + sizeof(*ref);
6210         }
6211
6212         location = &BTRFS_I(inode)->location;
6213         location->objectid = objectid;
6214         location->offset = 0;
6215         location->type = BTRFS_INODE_ITEM_KEY;
6216
6217         ret = btrfs_insert_inode_locked(inode);
6218         if (ret < 0)
6219                 goto fail;
6220
6221         path->leave_spinning = 1;
6222         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6223         if (ret != 0)
6224                 goto fail_unlock;
6225
6226         inode_init_owner(inode, dir, mode);
6227         inode_set_bytes(inode, 0);
6228
6229         inode->i_mtime = current_fs_time(inode->i_sb);
6230         inode->i_atime = inode->i_mtime;
6231         inode->i_ctime = inode->i_mtime;
6232         BTRFS_I(inode)->i_otime = inode->i_mtime;
6233
6234         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6235                                   struct btrfs_inode_item);
6236         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6237                              sizeof(*inode_item));
6238         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6239
6240         if (name) {
6241                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6242                                      struct btrfs_inode_ref);
6243                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6244                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6245                 ptr = (unsigned long)(ref + 1);
6246                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6247         }
6248
6249         btrfs_mark_buffer_dirty(path->nodes[0]);
6250         btrfs_free_path(path);
6251
6252         btrfs_inherit_iflags(inode, dir);
6253
6254         if (S_ISREG(mode)) {
6255                 if (btrfs_test_opt(root->fs_info, NODATASUM))
6256                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6257                 if (btrfs_test_opt(root->fs_info, NODATACOW))
6258                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6259                                 BTRFS_INODE_NODATASUM;
6260         }
6261
6262         inode_tree_add(inode);
6263
6264         trace_btrfs_inode_new(inode);
6265         btrfs_set_inode_last_trans(trans, inode);
6266
6267         btrfs_update_root_times(trans, root);
6268
6269         ret = btrfs_inode_inherit_props(trans, inode, dir);
6270         if (ret)
6271                 btrfs_err(root->fs_info,
6272                           "error inheriting props for ino %llu (root %llu): %d",
6273                           btrfs_ino(inode), root->root_key.objectid, ret);
6274
6275         return inode;
6276
6277 fail_unlock:
6278         unlock_new_inode(inode);
6279 fail:
6280         if (dir && name)
6281                 BTRFS_I(dir)->index_cnt--;
6282         btrfs_free_path(path);
6283         iput(inode);
6284         return ERR_PTR(ret);
6285 }
6286
6287 static inline u8 btrfs_inode_type(struct inode *inode)
6288 {
6289         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6290 }
6291
6292 /*
6293  * utility function to add 'inode' into 'parent_inode' with
6294  * a give name and a given sequence number.
6295  * if 'add_backref' is true, also insert a backref from the
6296  * inode to the parent directory.
6297  */
6298 int btrfs_add_link(struct btrfs_trans_handle *trans,
6299                    struct inode *parent_inode, struct inode *inode,
6300                    const char *name, int name_len, int add_backref, u64 index)
6301 {
6302         int ret = 0;
6303         struct btrfs_key key;
6304         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6305         u64 ino = btrfs_ino(inode);
6306         u64 parent_ino = btrfs_ino(parent_inode);
6307
6308         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6309                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6310         } else {
6311                 key.objectid = ino;
6312                 key.type = BTRFS_INODE_ITEM_KEY;
6313                 key.offset = 0;
6314         }
6315
6316         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6317                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6318                                          key.objectid, root->root_key.objectid,
6319                                          parent_ino, index, name, name_len);
6320         } else if (add_backref) {
6321                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6322                                              parent_ino, index);
6323         }
6324
6325         /* Nothing to clean up yet */
6326         if (ret)
6327                 return ret;
6328
6329         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6330                                     parent_inode, &key,
6331                                     btrfs_inode_type(inode), index);
6332         if (ret == -EEXIST || ret == -EOVERFLOW)
6333                 goto fail_dir_item;
6334         else if (ret) {
6335                 btrfs_abort_transaction(trans, ret);
6336                 return ret;
6337         }
6338
6339         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6340                            name_len * 2);
6341         inode_inc_iversion(parent_inode);
6342         parent_inode->i_mtime = parent_inode->i_ctime =
6343                 current_fs_time(parent_inode->i_sb);
6344         ret = btrfs_update_inode(trans, root, parent_inode);
6345         if (ret)
6346                 btrfs_abort_transaction(trans, ret);
6347         return ret;
6348
6349 fail_dir_item:
6350         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6351                 u64 local_index;
6352                 int err;
6353                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6354                                  key.objectid, root->root_key.objectid,
6355                                  parent_ino, &local_index, name, name_len);
6356
6357         } else if (add_backref) {
6358                 u64 local_index;
6359                 int err;
6360
6361                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6362                                           ino, parent_ino, &local_index);
6363         }
6364         return ret;
6365 }
6366
6367 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6368                             struct inode *dir, struct dentry *dentry,
6369                             struct inode *inode, int backref, u64 index)
6370 {
6371         int err = btrfs_add_link(trans, dir, inode,
6372                                  dentry->d_name.name, dentry->d_name.len,
6373                                  backref, index);
6374         if (err > 0)
6375                 err = -EEXIST;
6376         return err;
6377 }
6378
6379 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6380                         umode_t mode, dev_t rdev)
6381 {
6382         struct btrfs_trans_handle *trans;
6383         struct btrfs_root *root = BTRFS_I(dir)->root;
6384         struct inode *inode = NULL;
6385         int err;
6386         int drop_inode = 0;
6387         u64 objectid;
6388         u64 index = 0;
6389
6390         /*
6391          * 2 for inode item and ref
6392          * 2 for dir items
6393          * 1 for xattr if selinux is on
6394          */
6395         trans = btrfs_start_transaction(root, 5);
6396         if (IS_ERR(trans))
6397                 return PTR_ERR(trans);
6398
6399         err = btrfs_find_free_ino(root, &objectid);
6400         if (err)
6401                 goto out_unlock;
6402
6403         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6404                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6405                                 mode, &index);
6406         if (IS_ERR(inode)) {
6407                 err = PTR_ERR(inode);
6408                 goto out_unlock;
6409         }
6410
6411         /*
6412         * If the active LSM wants to access the inode during
6413         * d_instantiate it needs these. Smack checks to see
6414         * if the filesystem supports xattrs by looking at the
6415         * ops vector.
6416         */
6417         inode->i_op = &btrfs_special_inode_operations;
6418         init_special_inode(inode, inode->i_mode, rdev);
6419
6420         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6421         if (err)
6422                 goto out_unlock_inode;
6423
6424         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6425         if (err) {
6426                 goto out_unlock_inode;
6427         } else {
6428                 btrfs_update_inode(trans, root, inode);
6429                 unlock_new_inode(inode);
6430                 d_instantiate(dentry, inode);
6431         }
6432
6433 out_unlock:
6434         btrfs_end_transaction(trans, root);
6435         btrfs_balance_delayed_items(root);
6436         btrfs_btree_balance_dirty(root);
6437         if (drop_inode) {
6438                 inode_dec_link_count(inode);
6439                 iput(inode);
6440         }
6441         return err;
6442
6443 out_unlock_inode:
6444         drop_inode = 1;
6445         unlock_new_inode(inode);
6446         goto out_unlock;
6447
6448 }
6449
6450 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6451                         umode_t mode, bool excl)
6452 {
6453         struct btrfs_trans_handle *trans;
6454         struct btrfs_root *root = BTRFS_I(dir)->root;
6455         struct inode *inode = NULL;
6456         int drop_inode_on_err = 0;
6457         int err;
6458         u64 objectid;
6459         u64 index = 0;
6460
6461         /*
6462          * 2 for inode item and ref
6463          * 2 for dir items
6464          * 1 for xattr if selinux is on
6465          */
6466         trans = btrfs_start_transaction(root, 5);
6467         if (IS_ERR(trans))
6468                 return PTR_ERR(trans);
6469
6470         err = btrfs_find_free_ino(root, &objectid);
6471         if (err)
6472                 goto out_unlock;
6473
6474         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6475                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6476                                 mode, &index);
6477         if (IS_ERR(inode)) {
6478                 err = PTR_ERR(inode);
6479                 goto out_unlock;
6480         }
6481         drop_inode_on_err = 1;
6482         /*
6483         * If the active LSM wants to access the inode during
6484         * d_instantiate it needs these. Smack checks to see
6485         * if the filesystem supports xattrs by looking at the
6486         * ops vector.
6487         */
6488         inode->i_fop = &btrfs_file_operations;
6489         inode->i_op = &btrfs_file_inode_operations;
6490         inode->i_mapping->a_ops = &btrfs_aops;
6491
6492         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6493         if (err)
6494                 goto out_unlock_inode;
6495
6496         err = btrfs_update_inode(trans, root, inode);
6497         if (err)
6498                 goto out_unlock_inode;
6499
6500         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6501         if (err)
6502                 goto out_unlock_inode;
6503
6504         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6505         unlock_new_inode(inode);
6506         d_instantiate(dentry, inode);
6507
6508 out_unlock:
6509         btrfs_end_transaction(trans, root);
6510         if (err && drop_inode_on_err) {
6511                 inode_dec_link_count(inode);
6512                 iput(inode);
6513         }
6514         btrfs_balance_delayed_items(root);
6515         btrfs_btree_balance_dirty(root);
6516         return err;
6517
6518 out_unlock_inode:
6519         unlock_new_inode(inode);
6520         goto out_unlock;
6521
6522 }
6523
6524 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6525                       struct dentry *dentry)
6526 {
6527         struct btrfs_trans_handle *trans = NULL;
6528         struct btrfs_root *root = BTRFS_I(dir)->root;
6529         struct inode *inode = d_inode(old_dentry);
6530         u64 index;
6531         int err;
6532         int drop_inode = 0;
6533
6534         /* do not allow sys_link's with other subvols of the same device */
6535         if (root->objectid != BTRFS_I(inode)->root->objectid)
6536                 return -EXDEV;
6537
6538         if (inode->i_nlink >= BTRFS_LINK_MAX)
6539                 return -EMLINK;
6540
6541         err = btrfs_set_inode_index(dir, &index);
6542         if (err)
6543                 goto fail;
6544
6545         /*
6546          * 2 items for inode and inode ref
6547          * 2 items for dir items
6548          * 1 item for parent inode
6549          */
6550         trans = btrfs_start_transaction(root, 5);
6551         if (IS_ERR(trans)) {
6552                 err = PTR_ERR(trans);
6553                 trans = NULL;
6554                 goto fail;
6555         }
6556
6557         /* There are several dir indexes for this inode, clear the cache. */
6558         BTRFS_I(inode)->dir_index = 0ULL;
6559         inc_nlink(inode);
6560         inode_inc_iversion(inode);
6561         inode->i_ctime = current_fs_time(inode->i_sb);
6562         ihold(inode);
6563         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6564
6565         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6566
6567         if (err) {
6568                 drop_inode = 1;
6569         } else {
6570                 struct dentry *parent = dentry->d_parent;
6571                 err = btrfs_update_inode(trans, root, inode);
6572                 if (err)
6573                         goto fail;
6574                 if (inode->i_nlink == 1) {
6575                         /*
6576                          * If new hard link count is 1, it's a file created
6577                          * with open(2) O_TMPFILE flag.
6578                          */
6579                         err = btrfs_orphan_del(trans, inode);
6580                         if (err)
6581                                 goto fail;
6582                 }
6583                 d_instantiate(dentry, inode);
6584                 btrfs_log_new_name(trans, inode, NULL, parent);
6585         }
6586
6587         btrfs_balance_delayed_items(root);
6588 fail:
6589         if (trans)
6590                 btrfs_end_transaction(trans, root);
6591         if (drop_inode) {
6592                 inode_dec_link_count(inode);
6593                 iput(inode);
6594         }
6595         btrfs_btree_balance_dirty(root);
6596         return err;
6597 }
6598
6599 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6600 {
6601         struct inode *inode = NULL;
6602         struct btrfs_trans_handle *trans;
6603         struct btrfs_root *root = BTRFS_I(dir)->root;
6604         int err = 0;
6605         int drop_on_err = 0;
6606         u64 objectid = 0;
6607         u64 index = 0;
6608
6609         /*
6610          * 2 items for inode and ref
6611          * 2 items for dir items
6612          * 1 for xattr if selinux is on
6613          */
6614         trans = btrfs_start_transaction(root, 5);
6615         if (IS_ERR(trans))
6616                 return PTR_ERR(trans);
6617
6618         err = btrfs_find_free_ino(root, &objectid);
6619         if (err)
6620                 goto out_fail;
6621
6622         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6623                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6624                                 S_IFDIR | mode, &index);
6625         if (IS_ERR(inode)) {
6626                 err = PTR_ERR(inode);
6627                 goto out_fail;
6628         }
6629
6630         drop_on_err = 1;
6631         /* these must be set before we unlock the inode */
6632         inode->i_op = &btrfs_dir_inode_operations;
6633         inode->i_fop = &btrfs_dir_file_operations;
6634
6635         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6636         if (err)
6637                 goto out_fail_inode;
6638
6639         btrfs_i_size_write(inode, 0);
6640         err = btrfs_update_inode(trans, root, inode);
6641         if (err)
6642                 goto out_fail_inode;
6643
6644         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6645                              dentry->d_name.len, 0, index);
6646         if (err)
6647                 goto out_fail_inode;
6648
6649         d_instantiate(dentry, inode);
6650         /*
6651          * mkdir is special.  We're unlocking after we call d_instantiate
6652          * to avoid a race with nfsd calling d_instantiate.
6653          */
6654         unlock_new_inode(inode);
6655         drop_on_err = 0;
6656
6657 out_fail:
6658         btrfs_end_transaction(trans, root);
6659         if (drop_on_err) {
6660                 inode_dec_link_count(inode);
6661                 iput(inode);
6662         }
6663         btrfs_balance_delayed_items(root);
6664         btrfs_btree_balance_dirty(root);
6665         return err;
6666
6667 out_fail_inode:
6668         unlock_new_inode(inode);
6669         goto out_fail;
6670 }
6671
6672 /* Find next extent map of a given extent map, caller needs to ensure locks */
6673 static struct extent_map *next_extent_map(struct extent_map *em)
6674 {
6675         struct rb_node *next;
6676
6677         next = rb_next(&em->rb_node);
6678         if (!next)
6679                 return NULL;
6680         return container_of(next, struct extent_map, rb_node);
6681 }
6682
6683 static struct extent_map *prev_extent_map(struct extent_map *em)
6684 {
6685         struct rb_node *prev;
6686
6687         prev = rb_prev(&em->rb_node);
6688         if (!prev)
6689                 return NULL;
6690         return container_of(prev, struct extent_map, rb_node);
6691 }
6692
6693 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6694  * the existing extent is the nearest extent to map_start,
6695  * and an extent that you want to insert, deal with overlap and insert
6696  * the best fitted new extent into the tree.
6697  */
6698 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6699                                 struct extent_map *existing,
6700                                 struct extent_map *em,
6701                                 u64 map_start)
6702 {
6703         struct extent_map *prev;
6704         struct extent_map *next;
6705         u64 start;
6706         u64 end;
6707         u64 start_diff;
6708
6709         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6710
6711         if (existing->start > map_start) {
6712                 next = existing;
6713                 prev = prev_extent_map(next);
6714         } else {
6715                 prev = existing;
6716                 next = next_extent_map(prev);
6717         }
6718
6719         start = prev ? extent_map_end(prev) : em->start;
6720         start = max_t(u64, start, em->start);
6721         end = next ? next->start : extent_map_end(em);
6722         end = min_t(u64, end, extent_map_end(em));
6723         start_diff = start - em->start;
6724         em->start = start;
6725         em->len = end - start;
6726         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6727             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6728                 em->block_start += start_diff;
6729                 em->block_len -= start_diff;
6730         }
6731         return add_extent_mapping(em_tree, em, 0);
6732 }
6733
6734 static noinline int uncompress_inline(struct btrfs_path *path,
6735                                       struct page *page,
6736                                       size_t pg_offset, u64 extent_offset,
6737                                       struct btrfs_file_extent_item *item)
6738 {
6739         int ret;
6740         struct extent_buffer *leaf = path->nodes[0];
6741         char *tmp;
6742         size_t max_size;
6743         unsigned long inline_size;
6744         unsigned long ptr;
6745         int compress_type;
6746
6747         WARN_ON(pg_offset != 0);
6748         compress_type = btrfs_file_extent_compression(leaf, item);
6749         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6750         inline_size = btrfs_file_extent_inline_item_len(leaf,
6751                                         btrfs_item_nr(path->slots[0]));
6752         tmp = kmalloc(inline_size, GFP_NOFS);
6753         if (!tmp)
6754                 return -ENOMEM;
6755         ptr = btrfs_file_extent_inline_start(item);
6756
6757         read_extent_buffer(leaf, tmp, ptr, inline_size);
6758
6759         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6760         ret = btrfs_decompress(compress_type, tmp, page,
6761                                extent_offset, inline_size, max_size);
6762         kfree(tmp);
6763         return ret;
6764 }
6765
6766 /*
6767  * a bit scary, this does extent mapping from logical file offset to the disk.
6768  * the ugly parts come from merging extents from the disk with the in-ram
6769  * representation.  This gets more complex because of the data=ordered code,
6770  * where the in-ram extents might be locked pending data=ordered completion.
6771  *
6772  * This also copies inline extents directly into the page.
6773  */
6774
6775 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6776                                     size_t pg_offset, u64 start, u64 len,
6777                                     int create)
6778 {
6779         int ret;
6780         int err = 0;
6781         u64 extent_start = 0;
6782         u64 extent_end = 0;
6783         u64 objectid = btrfs_ino(inode);
6784         u32 found_type;
6785         struct btrfs_path *path = NULL;
6786         struct btrfs_root *root = BTRFS_I(inode)->root;
6787         struct btrfs_file_extent_item *item;
6788         struct extent_buffer *leaf;
6789         struct btrfs_key found_key;
6790         struct extent_map *em = NULL;
6791         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6792         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6793         struct btrfs_trans_handle *trans = NULL;
6794         const bool new_inline = !page || create;
6795
6796 again:
6797         read_lock(&em_tree->lock);
6798         em = lookup_extent_mapping(em_tree, start, len);
6799         if (em)
6800                 em->bdev = root->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 = root->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         if (!path) {
6823                 path = btrfs_alloc_path();
6824                 if (!path) {
6825                         err = -ENOMEM;
6826                         goto out;
6827                 }
6828                 /*
6829                  * Chances are we'll be called again, so go ahead and do
6830                  * readahead
6831                  */
6832                 path->reada = READA_FORWARD;
6833         }
6834
6835         ret = btrfs_lookup_file_extent(trans, root, path,
6836                                        objectid, start, trans != NULL);
6837         if (ret < 0) {
6838                 err = ret;
6839                 goto out;
6840         }
6841
6842         if (ret != 0) {
6843                 if (path->slots[0] == 0)
6844                         goto not_found;
6845                 path->slots[0]--;
6846         }
6847
6848         leaf = path->nodes[0];
6849         item = btrfs_item_ptr(leaf, path->slots[0],
6850                               struct btrfs_file_extent_item);
6851         /* are we inside the extent that was found? */
6852         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6853         found_type = found_key.type;
6854         if (found_key.objectid != objectid ||
6855             found_type != BTRFS_EXTENT_DATA_KEY) {
6856                 /*
6857                  * If we backup past the first extent we want to move forward
6858                  * and see if there is an extent in front of us, otherwise we'll
6859                  * say there is a hole for our whole search range which can
6860                  * cause problems.
6861                  */
6862                 extent_end = start;
6863                 goto next;
6864         }
6865
6866         found_type = btrfs_file_extent_type(leaf, item);
6867         extent_start = found_key.offset;
6868         if (found_type == BTRFS_FILE_EXTENT_REG ||
6869             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6870                 extent_end = extent_start +
6871                        btrfs_file_extent_num_bytes(leaf, item);
6872         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6873                 size_t size;
6874                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6875                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6876         }
6877 next:
6878         if (start >= extent_end) {
6879                 path->slots[0]++;
6880                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6881                         ret = btrfs_next_leaf(root, path);
6882                         if (ret < 0) {
6883                                 err = ret;
6884                                 goto out;
6885                         }
6886                         if (ret > 0)
6887                                 goto not_found;
6888                         leaf = path->nodes[0];
6889                 }
6890                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6891                 if (found_key.objectid != objectid ||
6892                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6893                         goto not_found;
6894                 if (start + len <= found_key.offset)
6895                         goto not_found;
6896                 if (start > found_key.offset)
6897                         goto next;
6898                 em->start = start;
6899                 em->orig_start = start;
6900                 em->len = found_key.offset - start;
6901                 goto not_found_em;
6902         }
6903
6904         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6905
6906         if (found_type == BTRFS_FILE_EXTENT_REG ||
6907             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6908                 goto insert;
6909         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6910                 unsigned long ptr;
6911                 char *map;
6912                 size_t size;
6913                 size_t extent_offset;
6914                 size_t copy_size;
6915
6916                 if (new_inline)
6917                         goto out;
6918
6919                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6920                 extent_offset = page_offset(page) + pg_offset - extent_start;
6921                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6922                                   size - extent_offset);
6923                 em->start = extent_start + extent_offset;
6924                 em->len = ALIGN(copy_size, root->sectorsize);
6925                 em->orig_block_len = em->len;
6926                 em->orig_start = em->start;
6927                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6928                 if (create == 0 && !PageUptodate(page)) {
6929                         if (btrfs_file_extent_compression(leaf, item) !=
6930                             BTRFS_COMPRESS_NONE) {
6931                                 ret = uncompress_inline(path, page, pg_offset,
6932                                                         extent_offset, item);
6933                                 if (ret) {
6934                                         err = ret;
6935                                         goto out;
6936                                 }
6937                         } else {
6938                                 map = kmap(page);
6939                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6940                                                    copy_size);
6941                                 if (pg_offset + copy_size < PAGE_SIZE) {
6942                                         memset(map + pg_offset + copy_size, 0,
6943                                                PAGE_SIZE - pg_offset -
6944                                                copy_size);
6945                                 }
6946                                 kunmap(page);
6947                         }
6948                         flush_dcache_page(page);
6949                 } else if (create && PageUptodate(page)) {
6950                         BUG();
6951                         if (!trans) {
6952                                 kunmap(page);
6953                                 free_extent_map(em);
6954                                 em = NULL;
6955
6956                                 btrfs_release_path(path);
6957                                 trans = btrfs_join_transaction(root);
6958
6959                                 if (IS_ERR(trans))
6960                                         return ERR_CAST(trans);
6961                                 goto again;
6962                         }
6963                         map = kmap(page);
6964                         write_extent_buffer(leaf, map + pg_offset, ptr,
6965                                             copy_size);
6966                         kunmap(page);
6967                         btrfs_mark_buffer_dirty(leaf);
6968                 }
6969                 set_extent_uptodate(io_tree, em->start,
6970                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6971                 goto insert;
6972         }
6973 not_found:
6974         em->start = start;
6975         em->orig_start = start;
6976         em->len = len;
6977 not_found_em:
6978         em->block_start = EXTENT_MAP_HOLE;
6979         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6980 insert:
6981         btrfs_release_path(path);
6982         if (em->start > start || extent_map_end(em) <= start) {
6983                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6984                         em->start, em->len, start, len);
6985                 err = -EIO;
6986                 goto out;
6987         }
6988
6989         err = 0;
6990         write_lock(&em_tree->lock);
6991         ret = add_extent_mapping(em_tree, em, 0);
6992         /* it is possible that someone inserted the extent into the tree
6993          * while we had the lock dropped.  It is also possible that
6994          * an overlapping map exists in the tree
6995          */
6996         if (ret == -EEXIST) {
6997                 struct extent_map *existing;
6998
6999                 ret = 0;
7000
7001                 existing = search_extent_mapping(em_tree, start, len);
7002                 /*
7003                  * existing will always be non-NULL, since there must be
7004                  * extent causing the -EEXIST.
7005                  */
7006                 if (existing->start == em->start &&
7007                     extent_map_end(existing) == extent_map_end(em) &&
7008                     em->block_start == existing->block_start) {
7009                         /*
7010                          * these two extents are the same, it happens
7011                          * with inlines especially
7012                          */
7013                         free_extent_map(em);
7014                         em = existing;
7015                         err = 0;
7016
7017                 } else if (start >= extent_map_end(existing) ||
7018                     start <= existing->start) {
7019                         /*
7020                          * The existing extent map is the one nearest to
7021                          * the [start, start + len) range which overlaps
7022                          */
7023                         err = merge_extent_mapping(em_tree, existing,
7024                                                    em, start);
7025                         free_extent_map(existing);
7026                         if (err) {
7027                                 free_extent_map(em);
7028                                 em = NULL;
7029                         }
7030                 } else {
7031                         free_extent_map(em);
7032                         em = existing;
7033                         err = 0;
7034                 }
7035         }
7036         write_unlock(&em_tree->lock);
7037 out:
7038
7039         trace_btrfs_get_extent(root, em);
7040
7041         btrfs_free_path(path);
7042         if (trans) {
7043                 ret = btrfs_end_transaction(trans, root);
7044                 if (!err)
7045                         err = ret;
7046         }
7047         if (err) {
7048                 free_extent_map(em);
7049                 return ERR_PTR(err);
7050         }
7051         BUG_ON(!em); /* Error is always set */
7052         return em;
7053 }
7054
7055 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7056                                            size_t pg_offset, u64 start, u64 len,
7057                                            int create)
7058 {
7059         struct extent_map *em;
7060         struct extent_map *hole_em = NULL;
7061         u64 range_start = start;
7062         u64 end;
7063         u64 found;
7064         u64 found_end;
7065         int err = 0;
7066
7067         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7068         if (IS_ERR(em))
7069                 return em;
7070         if (em) {
7071                 /*
7072                  * if our em maps to
7073                  * -  a hole or
7074                  * -  a pre-alloc extent,
7075                  * there might actually be delalloc bytes behind it.
7076                  */
7077                 if (em->block_start != EXTENT_MAP_HOLE &&
7078                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7079                         return em;
7080                 else
7081                         hole_em = em;
7082         }
7083
7084         /* check to see if we've wrapped (len == -1 or similar) */
7085         end = start + len;
7086         if (end < start)
7087                 end = (u64)-1;
7088         else
7089                 end -= 1;
7090
7091         em = NULL;
7092
7093         /* ok, we didn't find anything, lets look for delalloc */
7094         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7095                                  end, len, EXTENT_DELALLOC, 1);
7096         found_end = range_start + found;
7097         if (found_end < range_start)
7098                 found_end = (u64)-1;
7099
7100         /*
7101          * we didn't find anything useful, return
7102          * the original results from get_extent()
7103          */
7104         if (range_start > end || found_end <= start) {
7105                 em = hole_em;
7106                 hole_em = NULL;
7107                 goto out;
7108         }
7109
7110         /* adjust the range_start to make sure it doesn't
7111          * go backwards from the start they passed in
7112          */
7113         range_start = max(start, range_start);
7114         found = found_end - range_start;
7115
7116         if (found > 0) {
7117                 u64 hole_start = start;
7118                 u64 hole_len = len;
7119
7120                 em = alloc_extent_map();
7121                 if (!em) {
7122                         err = -ENOMEM;
7123                         goto out;
7124                 }
7125                 /*
7126                  * when btrfs_get_extent can't find anything it
7127                  * returns one huge hole
7128                  *
7129                  * make sure what it found really fits our range, and
7130                  * adjust to make sure it is based on the start from
7131                  * the caller
7132                  */
7133                 if (hole_em) {
7134                         u64 calc_end = extent_map_end(hole_em);
7135
7136                         if (calc_end <= start || (hole_em->start > end)) {
7137                                 free_extent_map(hole_em);
7138                                 hole_em = NULL;
7139                         } else {
7140                                 hole_start = max(hole_em->start, start);
7141                                 hole_len = calc_end - hole_start;
7142                         }
7143                 }
7144                 em->bdev = NULL;
7145                 if (hole_em && range_start > hole_start) {
7146                         /* our hole starts before our delalloc, so we
7147                          * have to return just the parts of the hole
7148                          * that go until  the delalloc starts
7149                          */
7150                         em->len = min(hole_len,
7151                                       range_start - hole_start);
7152                         em->start = hole_start;
7153                         em->orig_start = hole_start;
7154                         /*
7155                          * don't adjust block start at all,
7156                          * it is fixed at EXTENT_MAP_HOLE
7157                          */
7158                         em->block_start = hole_em->block_start;
7159                         em->block_len = hole_len;
7160                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7161                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7162                 } else {
7163                         em->start = range_start;
7164                         em->len = found;
7165                         em->orig_start = range_start;
7166                         em->block_start = EXTENT_MAP_DELALLOC;
7167                         em->block_len = found;
7168                 }
7169         } else if (hole_em) {
7170                 return hole_em;
7171         }
7172 out:
7173
7174         free_extent_map(hole_em);
7175         if (err) {
7176                 free_extent_map(em);
7177                 return ERR_PTR(err);
7178         }
7179         return em;
7180 }
7181
7182 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7183                                                   const u64 start,
7184                                                   const u64 len,
7185                                                   const u64 orig_start,
7186                                                   const u64 block_start,
7187                                                   const u64 block_len,
7188                                                   const u64 orig_block_len,
7189                                                   const u64 ram_bytes,
7190                                                   const int type)
7191 {
7192         struct extent_map *em = NULL;
7193         int ret;
7194
7195         down_read(&BTRFS_I(inode)->dio_sem);
7196         if (type != BTRFS_ORDERED_NOCOW) {
7197                 em = create_pinned_em(inode, start, len, orig_start,
7198                                       block_start, block_len, orig_block_len,
7199                                       ram_bytes, type);
7200                 if (IS_ERR(em))
7201                         goto out;
7202         }
7203         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7204                                            len, block_len, type);
7205         if (ret) {
7206                 if (em) {
7207                         free_extent_map(em);
7208                         btrfs_drop_extent_cache(inode, start,
7209                                                 start + len - 1, 0);
7210                 }
7211                 em = ERR_PTR(ret);
7212         }
7213  out:
7214         up_read(&BTRFS_I(inode)->dio_sem);
7215
7216         return em;
7217 }
7218
7219 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7220                                                   u64 start, u64 len)
7221 {
7222         struct btrfs_root *root = BTRFS_I(inode)->root;
7223         struct extent_map *em;
7224         struct btrfs_key ins;
7225         u64 alloc_hint;
7226         int ret;
7227
7228         alloc_hint = get_extent_allocation_hint(inode, start, len);
7229         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7230                                    alloc_hint, &ins, 1, 1);
7231         if (ret)
7232                 return ERR_PTR(ret);
7233
7234         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7235                                      ins.objectid, ins.offset, ins.offset,
7236                                      ins.offset, 0);
7237         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7238         if (IS_ERR(em))
7239                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7240
7241         return em;
7242 }
7243
7244 /*
7245  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7246  * block must be cow'd
7247  */
7248 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7249                               u64 *orig_start, u64 *orig_block_len,
7250                               u64 *ram_bytes)
7251 {
7252         struct btrfs_trans_handle *trans;
7253         struct btrfs_path *path;
7254         int ret;
7255         struct extent_buffer *leaf;
7256         struct btrfs_root *root = BTRFS_I(inode)->root;
7257         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7258         struct btrfs_file_extent_item *fi;
7259         struct btrfs_key key;
7260         u64 disk_bytenr;
7261         u64 backref_offset;
7262         u64 extent_end;
7263         u64 num_bytes;
7264         int slot;
7265         int found_type;
7266         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7267
7268         path = btrfs_alloc_path();
7269         if (!path)
7270                 return -ENOMEM;
7271
7272         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7273                                        offset, 0);
7274         if (ret < 0)
7275                 goto out;
7276
7277         slot = path->slots[0];
7278         if (ret == 1) {
7279                 if (slot == 0) {
7280                         /* can't find the item, must cow */
7281                         ret = 0;
7282                         goto out;
7283                 }
7284                 slot--;
7285         }
7286         ret = 0;
7287         leaf = path->nodes[0];
7288         btrfs_item_key_to_cpu(leaf, &key, slot);
7289         if (key.objectid != btrfs_ino(inode) ||
7290             key.type != BTRFS_EXTENT_DATA_KEY) {
7291                 /* not our file or wrong item type, must cow */
7292                 goto out;
7293         }
7294
7295         if (key.offset > offset) {
7296                 /* Wrong offset, must cow */
7297                 goto out;
7298         }
7299
7300         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7301         found_type = btrfs_file_extent_type(leaf, fi);
7302         if (found_type != BTRFS_FILE_EXTENT_REG &&
7303             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7304                 /* not a regular extent, must cow */
7305                 goto out;
7306         }
7307
7308         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7309                 goto out;
7310
7311         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7312         if (extent_end <= offset)
7313                 goto out;
7314
7315         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7316         if (disk_bytenr == 0)
7317                 goto out;
7318
7319         if (btrfs_file_extent_compression(leaf, fi) ||
7320             btrfs_file_extent_encryption(leaf, fi) ||
7321             btrfs_file_extent_other_encoding(leaf, fi))
7322                 goto out;
7323
7324         backref_offset = btrfs_file_extent_offset(leaf, fi);
7325
7326         if (orig_start) {
7327                 *orig_start = key.offset - backref_offset;
7328                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7329                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7330         }
7331
7332         if (btrfs_extent_readonly(root, disk_bytenr))
7333                 goto out;
7334
7335         num_bytes = min(offset + *len, extent_end) - offset;
7336         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7337                 u64 range_end;
7338
7339                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7340                 ret = test_range_bit(io_tree, offset, range_end,
7341                                      EXTENT_DELALLOC, 0, NULL);
7342                 if (ret) {
7343                         ret = -EAGAIN;
7344                         goto out;
7345                 }
7346         }
7347
7348         btrfs_release_path(path);
7349
7350         /*
7351          * look for other files referencing this extent, if we
7352          * find any we must cow
7353          */
7354         trans = btrfs_join_transaction(root);
7355         if (IS_ERR(trans)) {
7356                 ret = 0;
7357                 goto out;
7358         }
7359
7360         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7361                                     key.offset - backref_offset, disk_bytenr);
7362         btrfs_end_transaction(trans, root);
7363         if (ret) {
7364                 ret = 0;
7365                 goto out;
7366         }
7367
7368         /*
7369          * adjust disk_bytenr and num_bytes to cover just the bytes
7370          * in this extent we are about to write.  If there
7371          * are any csums in that range we have to cow in order
7372          * to keep the csums correct
7373          */
7374         disk_bytenr += backref_offset;
7375         disk_bytenr += offset - key.offset;
7376         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7377                                 goto out;
7378         /*
7379          * all of the above have passed, it is safe to overwrite this extent
7380          * without cow
7381          */
7382         *len = num_bytes;
7383         ret = 1;
7384 out:
7385         btrfs_free_path(path);
7386         return ret;
7387 }
7388
7389 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7390 {
7391         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7392         int found = false;
7393         void **pagep = NULL;
7394         struct page *page = NULL;
7395         int start_idx;
7396         int end_idx;
7397
7398         start_idx = start >> PAGE_SHIFT;
7399
7400         /*
7401          * end is the last byte in the last page.  end == start is legal
7402          */
7403         end_idx = end >> PAGE_SHIFT;
7404
7405         rcu_read_lock();
7406
7407         /* Most of the code in this while loop is lifted from
7408          * find_get_page.  It's been modified to begin searching from a
7409          * page and return just the first page found in that range.  If the
7410          * found idx is less than or equal to the end idx then we know that
7411          * a page exists.  If no pages are found or if those pages are
7412          * outside of the range then we're fine (yay!) */
7413         while (page == NULL &&
7414                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7415                 page = radix_tree_deref_slot(pagep);
7416                 if (unlikely(!page))
7417                         break;
7418
7419                 if (radix_tree_exception(page)) {
7420                         if (radix_tree_deref_retry(page)) {
7421                                 page = NULL;
7422                                 continue;
7423                         }
7424                         /*
7425                          * Otherwise, shmem/tmpfs must be storing a swap entry
7426                          * here as an exceptional entry: so return it without
7427                          * attempting to raise page count.
7428                          */
7429                         page = NULL;
7430                         break; /* TODO: Is this relevant for this use case? */
7431                 }
7432
7433                 if (!page_cache_get_speculative(page)) {
7434                         page = NULL;
7435                         continue;
7436                 }
7437
7438                 /*
7439                  * Has the page moved?
7440                  * This is part of the lockless pagecache protocol. See
7441                  * include/linux/pagemap.h for details.
7442                  */
7443                 if (unlikely(page != *pagep)) {
7444                         put_page(page);
7445                         page = NULL;
7446                 }
7447         }
7448
7449         if (page) {
7450                 if (page->index <= end_idx)
7451                         found = true;
7452                 put_page(page);
7453         }
7454
7455         rcu_read_unlock();
7456         return found;
7457 }
7458
7459 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7460                               struct extent_state **cached_state, int writing)
7461 {
7462         struct btrfs_ordered_extent *ordered;
7463         int ret = 0;
7464
7465         while (1) {
7466                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7467                                  cached_state);
7468                 /*
7469                  * We're concerned with the entire range that we're going to be
7470                  * doing DIO to, so we need to make sure there's no ordered
7471                  * extents in this range.
7472                  */
7473                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7474                                                      lockend - lockstart + 1);
7475
7476                 /*
7477                  * We need to make sure there are no buffered pages in this
7478                  * range either, we could have raced between the invalidate in
7479                  * generic_file_direct_write and locking the extent.  The
7480                  * invalidate needs to happen so that reads after a write do not
7481                  * get stale data.
7482                  */
7483                 if (!ordered &&
7484                     (!writing ||
7485                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7486                         break;
7487
7488                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7489                                      cached_state, GFP_NOFS);
7490
7491                 if (ordered) {
7492                         /*
7493                          * If we are doing a DIO read and the ordered extent we
7494                          * found is for a buffered write, we can not wait for it
7495                          * to complete and retry, because if we do so we can
7496                          * deadlock with concurrent buffered writes on page
7497                          * locks. This happens only if our DIO read covers more
7498                          * than one extent map, if at this point has already
7499                          * created an ordered extent for a previous extent map
7500                          * and locked its range in the inode's io tree, and a
7501                          * concurrent write against that previous extent map's
7502                          * range and this range started (we unlock the ranges
7503                          * in the io tree only when the bios complete and
7504                          * buffered writes always lock pages before attempting
7505                          * to lock range in the io tree).
7506                          */
7507                         if (writing ||
7508                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7509                                 btrfs_start_ordered_extent(inode, ordered, 1);
7510                         else
7511                                 ret = -ENOTBLK;
7512                         btrfs_put_ordered_extent(ordered);
7513                 } else {
7514                         /*
7515                          * We could trigger writeback for this range (and wait
7516                          * for it to complete) and then invalidate the pages for
7517                          * this range (through invalidate_inode_pages2_range()),
7518                          * but that can lead us to a deadlock with a concurrent
7519                          * call to readpages() (a buffered read or a defrag call
7520                          * triggered a readahead) on a page lock due to an
7521                          * ordered dio extent we created before but did not have
7522                          * yet a corresponding bio submitted (whence it can not
7523                          * complete), which makes readpages() wait for that
7524                          * ordered extent to complete while holding a lock on
7525                          * that page.
7526                          */
7527                         ret = -ENOTBLK;
7528                 }
7529
7530                 if (ret)
7531                         break;
7532
7533                 cond_resched();
7534         }
7535
7536         return ret;
7537 }
7538
7539 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7540                                            u64 len, u64 orig_start,
7541                                            u64 block_start, u64 block_len,
7542                                            u64 orig_block_len, u64 ram_bytes,
7543                                            int type)
7544 {
7545         struct extent_map_tree *em_tree;
7546         struct extent_map *em;
7547         struct btrfs_root *root = BTRFS_I(inode)->root;
7548         int ret;
7549
7550         em_tree = &BTRFS_I(inode)->extent_tree;
7551         em = alloc_extent_map();
7552         if (!em)
7553                 return ERR_PTR(-ENOMEM);
7554
7555         em->start = start;
7556         em->orig_start = orig_start;
7557         em->mod_start = start;
7558         em->mod_len = len;
7559         em->len = len;
7560         em->block_len = block_len;
7561         em->block_start = block_start;
7562         em->bdev = root->fs_info->fs_devices->latest_bdev;
7563         em->orig_block_len = orig_block_len;
7564         em->ram_bytes = ram_bytes;
7565         em->generation = -1;
7566         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7567         if (type == BTRFS_ORDERED_PREALLOC)
7568                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7569
7570         do {
7571                 btrfs_drop_extent_cache(inode, em->start,
7572                                 em->start + em->len - 1, 0);
7573                 write_lock(&em_tree->lock);
7574                 ret = add_extent_mapping(em_tree, em, 1);
7575                 write_unlock(&em_tree->lock);
7576         } while (ret == -EEXIST);
7577
7578         if (ret) {
7579                 free_extent_map(em);
7580                 return ERR_PTR(ret);
7581         }
7582
7583         return em;
7584 }
7585
7586 static void adjust_dio_outstanding_extents(struct inode *inode,
7587                                            struct btrfs_dio_data *dio_data,
7588                                            const u64 len)
7589 {
7590         unsigned num_extents;
7591
7592         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7593                                            BTRFS_MAX_EXTENT_SIZE);
7594         /*
7595          * If we have an outstanding_extents count still set then we're
7596          * within our reservation, otherwise we need to adjust our inode
7597          * counter appropriately.
7598          */
7599         if (dio_data->outstanding_extents) {
7600                 dio_data->outstanding_extents -= num_extents;
7601         } else {
7602                 spin_lock(&BTRFS_I(inode)->lock);
7603                 BTRFS_I(inode)->outstanding_extents += num_extents;
7604                 spin_unlock(&BTRFS_I(inode)->lock);
7605         }
7606 }
7607
7608 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7609                                    struct buffer_head *bh_result, int create)
7610 {
7611         struct extent_map *em;
7612         struct btrfs_root *root = BTRFS_I(inode)->root;
7613         struct extent_state *cached_state = NULL;
7614         struct btrfs_dio_data *dio_data = NULL;
7615         u64 start = iblock << inode->i_blkbits;
7616         u64 lockstart, lockend;
7617         u64 len = bh_result->b_size;
7618         int unlock_bits = EXTENT_LOCKED;
7619         int ret = 0;
7620
7621         if (create)
7622                 unlock_bits |= EXTENT_DIRTY;
7623         else
7624                 len = min_t(u64, len, root->sectorsize);
7625
7626         lockstart = start;
7627         lockend = start + len - 1;
7628
7629         if (current->journal_info) {
7630                 /*
7631                  * Need to pull our outstanding extents and set journal_info to NULL so
7632                  * that anything that needs to check if there's a transaction doesn't get
7633                  * confused.
7634                  */
7635                 dio_data = current->journal_info;
7636                 current->journal_info = NULL;
7637         }
7638
7639         /*
7640          * If this errors out it's because we couldn't invalidate pagecache for
7641          * this range and we need to fallback to buffered.
7642          */
7643         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7644                                create)) {
7645                 ret = -ENOTBLK;
7646                 goto err;
7647         }
7648
7649         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7650         if (IS_ERR(em)) {
7651                 ret = PTR_ERR(em);
7652                 goto unlock_err;
7653         }
7654
7655         /*
7656          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7657          * io.  INLINE is special, and we could probably kludge it in here, but
7658          * it's still buffered so for safety lets just fall back to the generic
7659          * buffered path.
7660          *
7661          * For COMPRESSED we _have_ to read the entire extent in so we can
7662          * decompress it, so there will be buffering required no matter what we
7663          * do, so go ahead and fallback to buffered.
7664          *
7665          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7666          * to buffered IO.  Don't blame me, this is the price we pay for using
7667          * the generic code.
7668          */
7669         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7670             em->block_start == EXTENT_MAP_INLINE) {
7671                 free_extent_map(em);
7672                 ret = -ENOTBLK;
7673                 goto unlock_err;
7674         }
7675
7676         /* Just a good old fashioned hole, return */
7677         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7678                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7679                 free_extent_map(em);
7680                 goto unlock_err;
7681         }
7682
7683         /*
7684          * We don't allocate a new extent in the following cases
7685          *
7686          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7687          * existing extent.
7688          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7689          * just use the extent.
7690          *
7691          */
7692         if (!create) {
7693                 len = min(len, em->len - (start - em->start));
7694                 lockstart = start + len;
7695                 goto unlock;
7696         }
7697
7698         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7699             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7700              em->block_start != EXTENT_MAP_HOLE)) {
7701                 int type;
7702                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7703
7704                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7705                         type = BTRFS_ORDERED_PREALLOC;
7706                 else
7707                         type = BTRFS_ORDERED_NOCOW;
7708                 len = min(len, em->len - (start - em->start));
7709                 block_start = em->block_start + (start - em->start);
7710
7711                 if (can_nocow_extent(inode, start, &len, &orig_start,
7712                                      &orig_block_len, &ram_bytes) == 1 &&
7713                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7714                         struct extent_map *em2;
7715
7716                         em2 = btrfs_create_dio_extent(inode, start, len,
7717                                                       orig_start, block_start,
7718                                                       len, orig_block_len,
7719                                                       ram_bytes, type);
7720                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7721                         if (type == BTRFS_ORDERED_PREALLOC) {
7722                                 free_extent_map(em);
7723                                 em = em2;
7724                         }
7725                         if (em2 && IS_ERR(em2)) {
7726                                 ret = PTR_ERR(em2);
7727                                 goto unlock_err;
7728                         }
7729                         goto unlock;
7730                 }
7731         }
7732
7733         /*
7734          * this will cow the extent, reset the len in case we changed
7735          * it above
7736          */
7737         len = bh_result->b_size;
7738         free_extent_map(em);
7739         em = btrfs_new_extent_direct(inode, start, len);
7740         if (IS_ERR(em)) {
7741                 ret = PTR_ERR(em);
7742                 goto unlock_err;
7743         }
7744         len = min(len, em->len - (start - em->start));
7745 unlock:
7746         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7747                 inode->i_blkbits;
7748         bh_result->b_size = len;
7749         bh_result->b_bdev = em->bdev;
7750         set_buffer_mapped(bh_result);
7751         if (create) {
7752                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7753                         set_buffer_new(bh_result);
7754
7755                 /*
7756                  * Need to update the i_size under the extent lock so buffered
7757                  * readers will get the updated i_size when we unlock.
7758                  */
7759                 if (start + len > i_size_read(inode))
7760                         i_size_write(inode, start + len);
7761
7762                 adjust_dio_outstanding_extents(inode, dio_data, len);
7763                 btrfs_free_reserved_data_space(inode, start, len);
7764                 WARN_ON(dio_data->reserve < len);
7765                 dio_data->reserve -= len;
7766                 dio_data->unsubmitted_oe_range_end = start + len;
7767                 current->journal_info = dio_data;
7768         }
7769
7770         /*
7771          * In the case of write we need to clear and unlock the entire range,
7772          * in the case of read we need to unlock only the end area that we
7773          * aren't using if there is any left over space.
7774          */
7775         if (lockstart < lockend) {
7776                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7777                                  lockend, unlock_bits, 1, 0,
7778                                  &cached_state, GFP_NOFS);
7779         } else {
7780                 free_extent_state(cached_state);
7781         }
7782
7783         free_extent_map(em);
7784
7785         return 0;
7786
7787 unlock_err:
7788         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7789                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7790 err:
7791         if (dio_data)
7792                 current->journal_info = dio_data;
7793         /*
7794          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7795          * write less data then expected, so that we don't underflow our inode's
7796          * outstanding extents counter.
7797          */
7798         if (create && dio_data)
7799                 adjust_dio_outstanding_extents(inode, dio_data, len);
7800
7801         return ret;
7802 }
7803
7804 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7805                                         int rw, int mirror_num)
7806 {
7807         struct btrfs_root *root = BTRFS_I(inode)->root;
7808         int ret;
7809
7810         BUG_ON(rw & REQ_WRITE);
7811
7812         bio_get(bio);
7813
7814         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7815                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7816         if (ret)
7817                 goto err;
7818
7819         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7820 err:
7821         bio_put(bio);
7822         return ret;
7823 }
7824
7825 static int btrfs_check_dio_repairable(struct inode *inode,
7826                                       struct bio *failed_bio,
7827                                       struct io_failure_record *failrec,
7828                                       int failed_mirror)
7829 {
7830         int num_copies;
7831
7832         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7833                                       failrec->logical, failrec->len);
7834         if (num_copies == 1) {
7835                 /*
7836                  * we only have a single copy of the data, so don't bother with
7837                  * all the retry and error correction code that follows. no
7838                  * matter what the error is, it is very likely to persist.
7839                  */
7840                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7841                          num_copies, failrec->this_mirror, failed_mirror);
7842                 return 0;
7843         }
7844
7845         failrec->failed_mirror = failed_mirror;
7846         failrec->this_mirror++;
7847         if (failrec->this_mirror == failed_mirror)
7848                 failrec->this_mirror++;
7849
7850         if (failrec->this_mirror > num_copies) {
7851                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7852                          num_copies, failrec->this_mirror, failed_mirror);
7853                 return 0;
7854         }
7855
7856         return 1;
7857 }
7858
7859 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7860                         struct page *page, unsigned int pgoff,
7861                         u64 start, u64 end, int failed_mirror,
7862                         bio_end_io_t *repair_endio, void *repair_arg)
7863 {
7864         struct io_failure_record *failrec;
7865         struct bio *bio;
7866         int isector;
7867         int read_mode;
7868         int ret;
7869
7870         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7871
7872         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7873         if (ret)
7874                 return ret;
7875
7876         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7877                                          failed_mirror);
7878         if (!ret) {
7879                 free_io_failure(inode, failrec);
7880                 return -EIO;
7881         }
7882
7883         if ((failed_bio->bi_vcnt > 1)
7884                 || (failed_bio->bi_io_vec->bv_len
7885                         > BTRFS_I(inode)->root->sectorsize))
7886                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7887         else
7888                 read_mode = READ_SYNC;
7889
7890         isector = start - btrfs_io_bio(failed_bio)->logical;
7891         isector >>= inode->i_sb->s_blocksize_bits;
7892         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7893                                 pgoff, isector, repair_endio, repair_arg);
7894         if (!bio) {
7895                 free_io_failure(inode, failrec);
7896                 return -EIO;
7897         }
7898
7899         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7900                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7901                     read_mode, failrec->this_mirror, failrec->in_validation);
7902
7903         ret = submit_dio_repair_bio(inode, bio, read_mode,
7904                                     failrec->this_mirror);
7905         if (ret) {
7906                 free_io_failure(inode, failrec);
7907                 bio_put(bio);
7908         }
7909
7910         return ret;
7911 }
7912
7913 struct btrfs_retry_complete {
7914         struct completion done;
7915         struct inode *inode;
7916         u64 start;
7917         int uptodate;
7918 };
7919
7920 static void btrfs_retry_endio_nocsum(struct bio *bio)
7921 {
7922         struct btrfs_retry_complete *done = bio->bi_private;
7923         struct inode *inode;
7924         struct bio_vec *bvec;
7925         int i;
7926
7927         if (bio->bi_error)
7928                 goto end;
7929
7930         ASSERT(bio->bi_vcnt == 1);
7931         inode = bio->bi_io_vec->bv_page->mapping->host;
7932         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7933
7934         done->uptodate = 1;
7935         bio_for_each_segment_all(bvec, bio, i)
7936                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7937 end:
7938         complete(&done->done);
7939         bio_put(bio);
7940 }
7941
7942 static int __btrfs_correct_data_nocsum(struct inode *inode,
7943                                        struct btrfs_io_bio *io_bio)
7944 {
7945         struct btrfs_fs_info *fs_info;
7946         struct bio_vec *bvec;
7947         struct btrfs_retry_complete done;
7948         u64 start;
7949         unsigned int pgoff;
7950         u32 sectorsize;
7951         int nr_sectors;
7952         int i;
7953         int ret;
7954
7955         fs_info = BTRFS_I(inode)->root->fs_info;
7956         sectorsize = BTRFS_I(inode)->root->sectorsize;
7957
7958         start = io_bio->logical;
7959         done.inode = inode;
7960
7961         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7962                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7963                 pgoff = bvec->bv_offset;
7964
7965 next_block_or_try_again:
7966                 done.uptodate = 0;
7967                 done.start = start;
7968                 init_completion(&done.done);
7969
7970                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7971                                 pgoff, start, start + sectorsize - 1,
7972                                 io_bio->mirror_num,
7973                                 btrfs_retry_endio_nocsum, &done);
7974                 if (ret)
7975                         return ret;
7976
7977                 wait_for_completion(&done.done);
7978
7979                 if (!done.uptodate) {
7980                         /* We might have another mirror, so try again */
7981                         goto next_block_or_try_again;
7982                 }
7983
7984                 start += sectorsize;
7985
7986                 if (nr_sectors--) {
7987                         pgoff += sectorsize;
7988                         goto next_block_or_try_again;
7989                 }
7990         }
7991
7992         return 0;
7993 }
7994
7995 static void btrfs_retry_endio(struct bio *bio)
7996 {
7997         struct btrfs_retry_complete *done = bio->bi_private;
7998         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7999         struct inode *inode;
8000         struct bio_vec *bvec;
8001         u64 start;
8002         int uptodate;
8003         int ret;
8004         int i;
8005
8006         if (bio->bi_error)
8007                 goto end;
8008
8009         uptodate = 1;
8010
8011         start = done->start;
8012
8013         ASSERT(bio->bi_vcnt == 1);
8014         inode = bio->bi_io_vec->bv_page->mapping->host;
8015         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8016
8017         bio_for_each_segment_all(bvec, bio, i) {
8018                 ret = __readpage_endio_check(done->inode, io_bio, i,
8019                                         bvec->bv_page, bvec->bv_offset,
8020                                         done->start, bvec->bv_len);
8021                 if (!ret)
8022                         clean_io_failure(done->inode, done->start,
8023                                         bvec->bv_page, bvec->bv_offset);
8024                 else
8025                         uptodate = 0;
8026         }
8027
8028         done->uptodate = uptodate;
8029 end:
8030         complete(&done->done);
8031         bio_put(bio);
8032 }
8033
8034 static int __btrfs_subio_endio_read(struct inode *inode,
8035                                     struct btrfs_io_bio *io_bio, int err)
8036 {
8037         struct btrfs_fs_info *fs_info;
8038         struct bio_vec *bvec;
8039         struct btrfs_retry_complete done;
8040         u64 start;
8041         u64 offset = 0;
8042         u32 sectorsize;
8043         int nr_sectors;
8044         unsigned int pgoff;
8045         int csum_pos;
8046         int i;
8047         int ret;
8048
8049         fs_info = BTRFS_I(inode)->root->fs_info;
8050         sectorsize = BTRFS_I(inode)->root->sectorsize;
8051
8052         err = 0;
8053         start = io_bio->logical;
8054         done.inode = inode;
8055
8056         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8057                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8058
8059                 pgoff = bvec->bv_offset;
8060 next_block:
8061                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8062                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8063                                         bvec->bv_page, pgoff, start,
8064                                         sectorsize);
8065                 if (likely(!ret))
8066                         goto next;
8067 try_again:
8068                 done.uptodate = 0;
8069                 done.start = start;
8070                 init_completion(&done.done);
8071
8072                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8073                                 pgoff, start, start + sectorsize - 1,
8074                                 io_bio->mirror_num,
8075                                 btrfs_retry_endio, &done);
8076                 if (ret) {
8077                         err = ret;
8078                         goto next;
8079                 }
8080
8081                 wait_for_completion(&done.done);
8082
8083                 if (!done.uptodate) {
8084                         /* We might have another mirror, so try again */
8085                         goto try_again;
8086                 }
8087 next:
8088                 offset += sectorsize;
8089                 start += sectorsize;
8090
8091                 ASSERT(nr_sectors);
8092
8093                 if (--nr_sectors) {
8094                         pgoff += sectorsize;
8095                         goto next_block;
8096                 }
8097         }
8098
8099         return err;
8100 }
8101
8102 static int btrfs_subio_endio_read(struct inode *inode,
8103                                   struct btrfs_io_bio *io_bio, int err)
8104 {
8105         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8106
8107         if (skip_csum) {
8108                 if (unlikely(err))
8109                         return __btrfs_correct_data_nocsum(inode, io_bio);
8110                 else
8111                         return 0;
8112         } else {
8113                 return __btrfs_subio_endio_read(inode, io_bio, err);
8114         }
8115 }
8116
8117 static void btrfs_endio_direct_read(struct bio *bio)
8118 {
8119         struct btrfs_dio_private *dip = bio->bi_private;
8120         struct inode *inode = dip->inode;
8121         struct bio *dio_bio;
8122         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8123         int err = bio->bi_error;
8124
8125         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8126                 err = btrfs_subio_endio_read(inode, io_bio, err);
8127
8128         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8129                       dip->logical_offset + dip->bytes - 1);
8130         dio_bio = dip->dio_bio;
8131
8132         kfree(dip);
8133
8134         dio_bio->bi_error = bio->bi_error;
8135         dio_end_io(dio_bio, bio->bi_error);
8136
8137         if (io_bio->end_io)
8138                 io_bio->end_io(io_bio, err);
8139         bio_put(bio);
8140 }
8141
8142 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8143                                                     const u64 offset,
8144                                                     const u64 bytes,
8145                                                     const int uptodate)
8146 {
8147         struct btrfs_root *root = BTRFS_I(inode)->root;
8148         struct btrfs_ordered_extent *ordered = NULL;
8149         u64 ordered_offset = offset;
8150         u64 ordered_bytes = bytes;
8151         int ret;
8152
8153 again:
8154         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8155                                                    &ordered_offset,
8156                                                    ordered_bytes,
8157                                                    uptodate);
8158         if (!ret)
8159                 goto out_test;
8160
8161         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8162                         finish_ordered_fn, NULL, NULL);
8163         btrfs_queue_work(root->fs_info->endio_write_workers,
8164                          &ordered->work);
8165 out_test:
8166         /*
8167          * our bio might span multiple ordered extents.  If we haven't
8168          * completed the accounting for the whole dio, go back and try again
8169          */
8170         if (ordered_offset < offset + bytes) {
8171                 ordered_bytes = offset + bytes - ordered_offset;
8172                 ordered = NULL;
8173                 goto again;
8174         }
8175 }
8176
8177 static void btrfs_endio_direct_write(struct bio *bio)
8178 {
8179         struct btrfs_dio_private *dip = bio->bi_private;
8180         struct bio *dio_bio = dip->dio_bio;
8181
8182         btrfs_endio_direct_write_update_ordered(dip->inode,
8183                                                 dip->logical_offset,
8184                                                 dip->bytes,
8185                                                 !bio->bi_error);
8186
8187         kfree(dip);
8188
8189         dio_bio->bi_error = bio->bi_error;
8190         dio_end_io(dio_bio, bio->bi_error);
8191         bio_put(bio);
8192 }
8193
8194 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8195                                     struct bio *bio, int mirror_num,
8196                                     unsigned long bio_flags, u64 offset)
8197 {
8198         int ret;
8199         struct btrfs_root *root = BTRFS_I(inode)->root;
8200         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8201         BUG_ON(ret); /* -ENOMEM */
8202         return 0;
8203 }
8204
8205 static void btrfs_end_dio_bio(struct bio *bio)
8206 {
8207         struct btrfs_dio_private *dip = bio->bi_private;
8208         int err = bio->bi_error;
8209
8210         if (err)
8211                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8212                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8213                            btrfs_ino(dip->inode), bio->bi_rw,
8214                            (unsigned long long)bio->bi_iter.bi_sector,
8215                            bio->bi_iter.bi_size, err);
8216
8217         if (dip->subio_endio)
8218                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8219
8220         if (err) {
8221                 dip->errors = 1;
8222
8223                 /*
8224                  * before atomic variable goto zero, we must make sure
8225                  * dip->errors is perceived to be set.
8226                  */
8227                 smp_mb__before_atomic();
8228         }
8229
8230         /* if there are more bios still pending for this dio, just exit */
8231         if (!atomic_dec_and_test(&dip->pending_bios))
8232                 goto out;
8233
8234         if (dip->errors) {
8235                 bio_io_error(dip->orig_bio);
8236         } else {
8237                 dip->dio_bio->bi_error = 0;
8238                 bio_endio(dip->orig_bio);
8239         }
8240 out:
8241         bio_put(bio);
8242 }
8243
8244 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8245                                        u64 first_sector, gfp_t gfp_flags)
8246 {
8247         struct bio *bio;
8248         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8249         if (bio)
8250                 bio_associate_current(bio);
8251         return bio;
8252 }
8253
8254 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8255                                                  struct inode *inode,
8256                                                  struct btrfs_dio_private *dip,
8257                                                  struct bio *bio,
8258                                                  u64 file_offset)
8259 {
8260         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8261         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8262         int ret;
8263
8264         /*
8265          * We load all the csum data we need when we submit
8266          * the first bio to reduce the csum tree search and
8267          * contention.
8268          */
8269         if (dip->logical_offset == file_offset) {
8270                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8271                                                 file_offset);
8272                 if (ret)
8273                         return ret;
8274         }
8275
8276         if (bio == dip->orig_bio)
8277                 return 0;
8278
8279         file_offset -= dip->logical_offset;
8280         file_offset >>= inode->i_sb->s_blocksize_bits;
8281         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8282
8283         return 0;
8284 }
8285
8286 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8287                                          int rw, u64 file_offset, int skip_sum,
8288                                          int async_submit)
8289 {
8290         struct btrfs_dio_private *dip = bio->bi_private;
8291         int write = rw & REQ_WRITE;
8292         struct btrfs_root *root = BTRFS_I(inode)->root;
8293         int ret;
8294
8295         if (async_submit)
8296                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8297
8298         bio_get(bio);
8299
8300         if (!write) {
8301                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8302                                 BTRFS_WQ_ENDIO_DATA);
8303                 if (ret)
8304                         goto err;
8305         }
8306
8307         if (skip_sum)
8308                 goto map;
8309
8310         if (write && async_submit) {
8311                 ret = btrfs_wq_submit_bio(root->fs_info,
8312                                    inode, rw, bio, 0, 0,
8313                                    file_offset,
8314                                    __btrfs_submit_bio_start_direct_io,
8315                                    __btrfs_submit_bio_done);
8316                 goto err;
8317         } else if (write) {
8318                 /*
8319                  * If we aren't doing async submit, calculate the csum of the
8320                  * bio now.
8321                  */
8322                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8323                 if (ret)
8324                         goto err;
8325         } else {
8326                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8327                                                      file_offset);
8328                 if (ret)
8329                         goto err;
8330         }
8331 map:
8332         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8333 err:
8334         bio_put(bio);
8335         return ret;
8336 }
8337
8338 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8339                                     int skip_sum)
8340 {
8341         struct inode *inode = dip->inode;
8342         struct btrfs_root *root = BTRFS_I(inode)->root;
8343         struct bio *bio;
8344         struct bio *orig_bio = dip->orig_bio;
8345         struct bio_vec *bvec = orig_bio->bi_io_vec;
8346         u64 start_sector = orig_bio->bi_iter.bi_sector;
8347         u64 file_offset = dip->logical_offset;
8348         u64 submit_len = 0;
8349         u64 map_length;
8350         u32 blocksize = root->sectorsize;
8351         int async_submit = 0;
8352         int nr_sectors;
8353         int ret;
8354         int i;
8355
8356         map_length = orig_bio->bi_iter.bi_size;
8357         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8358                               &map_length, NULL, 0);
8359         if (ret)
8360                 return -EIO;
8361
8362         if (map_length >= orig_bio->bi_iter.bi_size) {
8363                 bio = orig_bio;
8364                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8365                 goto submit;
8366         }
8367
8368         /* async crcs make it difficult to collect full stripe writes. */
8369         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8370                 async_submit = 0;
8371         else
8372                 async_submit = 1;
8373
8374         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8375         if (!bio)
8376                 return -ENOMEM;
8377
8378         bio->bi_private = dip;
8379         bio->bi_end_io = btrfs_end_dio_bio;
8380         btrfs_io_bio(bio)->logical = file_offset;
8381         atomic_inc(&dip->pending_bios);
8382
8383         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8384                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8385                 i = 0;
8386 next_block:
8387                 if (unlikely(map_length < submit_len + blocksize ||
8388                     bio_add_page(bio, bvec->bv_page, blocksize,
8389                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8390                         /*
8391                          * inc the count before we submit the bio so
8392                          * we know the end IO handler won't happen before
8393                          * we inc the count. Otherwise, the dip might get freed
8394                          * before we're done setting it up
8395                          */
8396                         atomic_inc(&dip->pending_bios);
8397                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8398                                                      file_offset, skip_sum,
8399                                                      async_submit);
8400                         if (ret) {
8401                                 bio_put(bio);
8402                                 atomic_dec(&dip->pending_bios);
8403                                 goto out_err;
8404                         }
8405
8406                         start_sector += submit_len >> 9;
8407                         file_offset += submit_len;
8408
8409                         submit_len = 0;
8410
8411                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8412                                                   start_sector, GFP_NOFS);
8413                         if (!bio)
8414                                 goto out_err;
8415                         bio->bi_private = dip;
8416                         bio->bi_end_io = btrfs_end_dio_bio;
8417                         btrfs_io_bio(bio)->logical = file_offset;
8418
8419                         map_length = orig_bio->bi_iter.bi_size;
8420                         ret = btrfs_map_block(root->fs_info, rw,
8421                                               start_sector << 9,
8422                                               &map_length, NULL, 0);
8423                         if (ret) {
8424                                 bio_put(bio);
8425                                 goto out_err;
8426                         }
8427
8428                         goto next_block;
8429                 } else {
8430                         submit_len += blocksize;
8431                         if (--nr_sectors) {
8432                                 i++;
8433                                 goto next_block;
8434                         }
8435                         bvec++;
8436                 }
8437         }
8438
8439 submit:
8440         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8441                                      async_submit);
8442         if (!ret)
8443                 return 0;
8444
8445         bio_put(bio);
8446 out_err:
8447         dip->errors = 1;
8448         /*
8449          * before atomic variable goto zero, we must
8450          * make sure dip->errors is perceived to be set.
8451          */
8452         smp_mb__before_atomic();
8453         if (atomic_dec_and_test(&dip->pending_bios))
8454                 bio_io_error(dip->orig_bio);
8455
8456         /* bio_end_io() will handle error, so we needn't return it */
8457         return 0;
8458 }
8459
8460 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8461                                 struct inode *inode, loff_t file_offset)
8462 {
8463         struct btrfs_dio_private *dip = NULL;
8464         struct bio *io_bio = NULL;
8465         struct btrfs_io_bio *btrfs_bio;
8466         int skip_sum;
8467         int write = rw & REQ_WRITE;
8468         int ret = 0;
8469
8470         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8471
8472         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8473         if (!io_bio) {
8474                 ret = -ENOMEM;
8475                 goto free_ordered;
8476         }
8477
8478         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8479         if (!dip) {
8480                 ret = -ENOMEM;
8481                 goto free_ordered;
8482         }
8483
8484         dip->private = dio_bio->bi_private;
8485         dip->inode = inode;
8486         dip->logical_offset = file_offset;
8487         dip->bytes = dio_bio->bi_iter.bi_size;
8488         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8489         io_bio->bi_private = dip;
8490         dip->orig_bio = io_bio;
8491         dip->dio_bio = dio_bio;
8492         atomic_set(&dip->pending_bios, 0);
8493         btrfs_bio = btrfs_io_bio(io_bio);
8494         btrfs_bio->logical = file_offset;
8495
8496         if (write) {
8497                 io_bio->bi_end_io = btrfs_endio_direct_write;
8498         } else {
8499                 io_bio->bi_end_io = btrfs_endio_direct_read;
8500                 dip->subio_endio = btrfs_subio_endio_read;
8501         }
8502
8503         /*
8504          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8505          * even if we fail to submit a bio, because in such case we do the
8506          * corresponding error handling below and it must not be done a second
8507          * time by btrfs_direct_IO().
8508          */
8509         if (write) {
8510                 struct btrfs_dio_data *dio_data = current->journal_info;
8511
8512                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8513                         dip->bytes;
8514                 dio_data->unsubmitted_oe_range_start =
8515                         dio_data->unsubmitted_oe_range_end;
8516         }
8517
8518         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8519         if (!ret)
8520                 return;
8521
8522         if (btrfs_bio->end_io)
8523                 btrfs_bio->end_io(btrfs_bio, ret);
8524
8525 free_ordered:
8526         /*
8527          * If we arrived here it means either we failed to submit the dip
8528          * or we either failed to clone the dio_bio or failed to allocate the
8529          * dip. If we cloned the dio_bio and allocated the dip, we can just
8530          * call bio_endio against our io_bio so that we get proper resource
8531          * cleanup if we fail to submit the dip, otherwise, we must do the
8532          * same as btrfs_endio_direct_[write|read] because we can't call these
8533          * callbacks - they require an allocated dip and a clone of dio_bio.
8534          */
8535         if (io_bio && dip) {
8536                 io_bio->bi_error = -EIO;
8537                 bio_endio(io_bio);
8538                 /*
8539                  * The end io callbacks free our dip, do the final put on io_bio
8540                  * and all the cleanup and final put for dio_bio (through
8541                  * dio_end_io()).
8542                  */
8543                 dip = NULL;
8544                 io_bio = NULL;
8545         } else {
8546                 if (write)
8547                         btrfs_endio_direct_write_update_ordered(inode,
8548                                                 file_offset,
8549                                                 dio_bio->bi_iter.bi_size,
8550                                                 0);
8551                 else
8552                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8553                               file_offset + dio_bio->bi_iter.bi_size - 1);
8554
8555                 dio_bio->bi_error = -EIO;
8556                 /*
8557                  * Releases and cleans up our dio_bio, no need to bio_put()
8558                  * nor bio_endio()/bio_io_error() against dio_bio.
8559                  */
8560                 dio_end_io(dio_bio, ret);
8561         }
8562         if (io_bio)
8563                 bio_put(io_bio);
8564         kfree(dip);
8565 }
8566
8567 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8568                         const struct iov_iter *iter, loff_t offset)
8569 {
8570         int seg;
8571         int i;
8572         unsigned blocksize_mask = root->sectorsize - 1;
8573         ssize_t retval = -EINVAL;
8574
8575         if (offset & blocksize_mask)
8576                 goto out;
8577
8578         if (iov_iter_alignment(iter) & blocksize_mask)
8579                 goto out;
8580
8581         /* If this is a write we don't need to check anymore */
8582         if (iov_iter_rw(iter) == WRITE)
8583                 return 0;
8584         /*
8585          * Check to make sure we don't have duplicate iov_base's in this
8586          * iovec, if so return EINVAL, otherwise we'll get csum errors
8587          * when reading back.
8588          */
8589         for (seg = 0; seg < iter->nr_segs; seg++) {
8590                 for (i = seg + 1; i < iter->nr_segs; i++) {
8591                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8592                                 goto out;
8593                 }
8594         }
8595         retval = 0;
8596 out:
8597         return retval;
8598 }
8599
8600 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8601 {
8602         struct file *file = iocb->ki_filp;
8603         struct inode *inode = file->f_mapping->host;
8604         struct btrfs_root *root = BTRFS_I(inode)->root;
8605         struct btrfs_dio_data dio_data = { 0 };
8606         loff_t offset = iocb->ki_pos;
8607         size_t count = 0;
8608         int flags = 0;
8609         bool wakeup = true;
8610         bool relock = false;
8611         ssize_t ret;
8612
8613         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8614                 return 0;
8615
8616         inode_dio_begin(inode);
8617         smp_mb__after_atomic();
8618
8619         /*
8620          * The generic stuff only does filemap_write_and_wait_range, which
8621          * isn't enough if we've written compressed pages to this area, so
8622          * we need to flush the dirty pages again to make absolutely sure
8623          * that any outstanding dirty pages are on disk.
8624          */
8625         count = iov_iter_count(iter);
8626         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8627                      &BTRFS_I(inode)->runtime_flags))
8628                 filemap_fdatawrite_range(inode->i_mapping, offset,
8629                                          offset + count - 1);
8630
8631         if (iov_iter_rw(iter) == WRITE) {
8632                 /*
8633                  * If the write DIO is beyond the EOF, we need update
8634                  * the isize, but it is protected by i_mutex. So we can
8635                  * not unlock the i_mutex at this case.
8636                  */
8637                 if (offset + count <= inode->i_size) {
8638                         inode_unlock(inode);
8639                         relock = true;
8640                 }
8641                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8642                 if (ret)
8643                         goto out;
8644                 dio_data.outstanding_extents = div64_u64(count +
8645                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8646                                                 BTRFS_MAX_EXTENT_SIZE);
8647
8648                 /*
8649                  * We need to know how many extents we reserved so that we can
8650                  * do the accounting properly if we go over the number we
8651                  * originally calculated.  Abuse current->journal_info for this.
8652                  */
8653                 dio_data.reserve = round_up(count, root->sectorsize);
8654                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8655                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8656                 current->journal_info = &dio_data;
8657         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8658                                      &BTRFS_I(inode)->runtime_flags)) {
8659                 inode_dio_end(inode);
8660                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8661                 wakeup = false;
8662         }
8663
8664         ret = __blockdev_direct_IO(iocb, inode,
8665                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8666                                    iter, btrfs_get_blocks_direct, NULL,
8667                                    btrfs_submit_direct, flags);
8668         if (iov_iter_rw(iter) == WRITE) {
8669                 current->journal_info = NULL;
8670                 if (ret < 0 && ret != -EIOCBQUEUED) {
8671                         if (dio_data.reserve)
8672                                 btrfs_delalloc_release_space(inode, offset,
8673                                                              dio_data.reserve);
8674                         /*
8675                          * On error we might have left some ordered extents
8676                          * without submitting corresponding bios for them, so
8677                          * cleanup them up to avoid other tasks getting them
8678                          * and waiting for them to complete forever.
8679                          */
8680                         if (dio_data.unsubmitted_oe_range_start <
8681                             dio_data.unsubmitted_oe_range_end)
8682                                 btrfs_endio_direct_write_update_ordered(inode,
8683                                         dio_data.unsubmitted_oe_range_start,
8684                                         dio_data.unsubmitted_oe_range_end -
8685                                         dio_data.unsubmitted_oe_range_start,
8686                                         0);
8687                 } else if (ret >= 0 && (size_t)ret < count)
8688                         btrfs_delalloc_release_space(inode, offset,
8689                                                      count - (size_t)ret);
8690         }
8691 out:
8692         if (wakeup)
8693                 inode_dio_end(inode);
8694         if (relock)
8695                 inode_lock(inode);
8696
8697         return ret;
8698 }
8699
8700 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8701
8702 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8703                 __u64 start, __u64 len)
8704 {
8705         int     ret;
8706
8707         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8708         if (ret)
8709                 return ret;
8710
8711         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8712 }
8713
8714 int btrfs_readpage(struct file *file, struct page *page)
8715 {
8716         struct extent_io_tree *tree;
8717         tree = &BTRFS_I(page->mapping->host)->io_tree;
8718         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8719 }
8720
8721 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8722 {
8723         struct extent_io_tree *tree;
8724         struct inode *inode = page->mapping->host;
8725         int ret;
8726
8727         if (current->flags & PF_MEMALLOC) {
8728                 redirty_page_for_writepage(wbc, page);
8729                 unlock_page(page);
8730                 return 0;
8731         }
8732
8733         /*
8734          * If we are under memory pressure we will call this directly from the
8735          * VM, we need to make sure we have the inode referenced for the ordered
8736          * extent.  If not just return like we didn't do anything.
8737          */
8738         if (!igrab(inode)) {
8739                 redirty_page_for_writepage(wbc, page);
8740                 return AOP_WRITEPAGE_ACTIVATE;
8741         }
8742         tree = &BTRFS_I(page->mapping->host)->io_tree;
8743         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8744         btrfs_add_delayed_iput(inode);
8745         return ret;
8746 }
8747
8748 static int btrfs_writepages(struct address_space *mapping,
8749                             struct writeback_control *wbc)
8750 {
8751         struct extent_io_tree *tree;
8752
8753         tree = &BTRFS_I(mapping->host)->io_tree;
8754         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8755 }
8756
8757 static int
8758 btrfs_readpages(struct file *file, struct address_space *mapping,
8759                 struct list_head *pages, unsigned nr_pages)
8760 {
8761         struct extent_io_tree *tree;
8762         tree = &BTRFS_I(mapping->host)->io_tree;
8763         return extent_readpages(tree, mapping, pages, nr_pages,
8764                                 btrfs_get_extent);
8765 }
8766 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8767 {
8768         struct extent_io_tree *tree;
8769         struct extent_map_tree *map;
8770         int ret;
8771
8772         tree = &BTRFS_I(page->mapping->host)->io_tree;
8773         map = &BTRFS_I(page->mapping->host)->extent_tree;
8774         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8775         if (ret == 1) {
8776                 ClearPagePrivate(page);
8777                 set_page_private(page, 0);
8778                 put_page(page);
8779         }
8780         return ret;
8781 }
8782
8783 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8784 {
8785         if (PageWriteback(page) || PageDirty(page))
8786                 return 0;
8787         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8788 }
8789
8790 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8791                                  unsigned int length)
8792 {
8793         struct inode *inode = page->mapping->host;
8794         struct extent_io_tree *tree;
8795         struct btrfs_ordered_extent *ordered;
8796         struct extent_state *cached_state = NULL;
8797         u64 page_start = page_offset(page);
8798         u64 page_end = page_start + PAGE_SIZE - 1;
8799         u64 start;
8800         u64 end;
8801         int inode_evicting = inode->i_state & I_FREEING;
8802
8803         /*
8804          * we have the page locked, so new writeback can't start,
8805          * and the dirty bit won't be cleared while we are here.
8806          *
8807          * Wait for IO on this page so that we can safely clear
8808          * the PagePrivate2 bit and do ordered accounting
8809          */
8810         wait_on_page_writeback(page);
8811
8812         tree = &BTRFS_I(inode)->io_tree;
8813         if (offset) {
8814                 btrfs_releasepage(page, GFP_NOFS);
8815                 return;
8816         }
8817
8818         if (!inode_evicting)
8819                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8820 again:
8821         start = page_start;
8822         ordered = btrfs_lookup_ordered_range(inode, start,
8823                                         page_end - start + 1);
8824         if (ordered) {
8825                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8826                 /*
8827                  * IO on this page will never be started, so we need
8828                  * to account for any ordered extents now
8829                  */
8830                 if (!inode_evicting)
8831                         clear_extent_bit(tree, start, end,
8832                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8833                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8834                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8835                                          GFP_NOFS);
8836                 /*
8837                  * whoever cleared the private bit is responsible
8838                  * for the finish_ordered_io
8839                  */
8840                 if (TestClearPagePrivate2(page)) {
8841                         struct btrfs_ordered_inode_tree *tree;
8842                         u64 new_len;
8843
8844                         tree = &BTRFS_I(inode)->ordered_tree;
8845
8846                         spin_lock_irq(&tree->lock);
8847                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8848                         new_len = start - ordered->file_offset;
8849                         if (new_len < ordered->truncated_len)
8850                                 ordered->truncated_len = new_len;
8851                         spin_unlock_irq(&tree->lock);
8852
8853                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8854                                                            start,
8855                                                            end - start + 1, 1))
8856                                 btrfs_finish_ordered_io(ordered);
8857                 }
8858                 btrfs_put_ordered_extent(ordered);
8859                 if (!inode_evicting) {
8860                         cached_state = NULL;
8861                         lock_extent_bits(tree, start, end,
8862                                          &cached_state);
8863                 }
8864
8865                 start = end + 1;
8866                 if (start < page_end)
8867                         goto again;
8868         }
8869
8870         /*
8871          * Qgroup reserved space handler
8872          * Page here will be either
8873          * 1) Already written to disk
8874          *    In this case, its reserved space is released from data rsv map
8875          *    and will be freed by delayed_ref handler finally.
8876          *    So even we call qgroup_free_data(), it won't decrease reserved
8877          *    space.
8878          * 2) Not written to disk
8879          *    This means the reserved space should be freed here.
8880          */
8881         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8882         if (!inode_evicting) {
8883                 clear_extent_bit(tree, page_start, page_end,
8884                                  EXTENT_LOCKED | EXTENT_DIRTY |
8885                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8886                                  EXTENT_DEFRAG, 1, 1,
8887                                  &cached_state, GFP_NOFS);
8888
8889                 __btrfs_releasepage(page, GFP_NOFS);
8890         }
8891
8892         ClearPageChecked(page);
8893         if (PagePrivate(page)) {
8894                 ClearPagePrivate(page);
8895                 set_page_private(page, 0);
8896                 put_page(page);
8897         }
8898 }
8899
8900 /*
8901  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8902  * called from a page fault handler when a page is first dirtied. Hence we must
8903  * be careful to check for EOF conditions here. We set the page up correctly
8904  * for a written page which means we get ENOSPC checking when writing into
8905  * holes and correct delalloc and unwritten extent mapping on filesystems that
8906  * support these features.
8907  *
8908  * We are not allowed to take the i_mutex here so we have to play games to
8909  * protect against truncate races as the page could now be beyond EOF.  Because
8910  * vmtruncate() writes the inode size before removing pages, once we have the
8911  * page lock we can determine safely if the page is beyond EOF. If it is not
8912  * beyond EOF, then the page is guaranteed safe against truncation until we
8913  * unlock the page.
8914  */
8915 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8916 {
8917         struct page *page = vmf->page;
8918         struct inode *inode = file_inode(vma->vm_file);
8919         struct btrfs_root *root = BTRFS_I(inode)->root;
8920         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8921         struct btrfs_ordered_extent *ordered;
8922         struct extent_state *cached_state = NULL;
8923         char *kaddr;
8924         unsigned long zero_start;
8925         loff_t size;
8926         int ret;
8927         int reserved = 0;
8928         u64 reserved_space;
8929         u64 page_start;
8930         u64 page_end;
8931         u64 end;
8932
8933         reserved_space = PAGE_SIZE;
8934
8935         sb_start_pagefault(inode->i_sb);
8936         page_start = page_offset(page);
8937         page_end = page_start + PAGE_SIZE - 1;
8938         end = page_end;
8939
8940         /*
8941          * Reserving delalloc space after obtaining the page lock can lead to
8942          * deadlock. For example, if a dirty page is locked by this function
8943          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8944          * dirty page write out, then the btrfs_writepage() function could
8945          * end up waiting indefinitely to get a lock on the page currently
8946          * being processed by btrfs_page_mkwrite() function.
8947          */
8948         ret = btrfs_delalloc_reserve_space(inode, page_start,
8949                                            reserved_space);
8950         if (!ret) {
8951                 ret = file_update_time(vma->vm_file);
8952                 reserved = 1;
8953         }
8954         if (ret) {
8955                 if (ret == -ENOMEM)
8956                         ret = VM_FAULT_OOM;
8957                 else /* -ENOSPC, -EIO, etc */
8958                         ret = VM_FAULT_SIGBUS;
8959                 if (reserved)
8960                         goto out;
8961                 goto out_noreserve;
8962         }
8963
8964         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8965 again:
8966         lock_page(page);
8967         size = i_size_read(inode);
8968
8969         if ((page->mapping != inode->i_mapping) ||
8970             (page_start >= size)) {
8971                 /* page got truncated out from underneath us */
8972                 goto out_unlock;
8973         }
8974         wait_on_page_writeback(page);
8975
8976         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8977         set_page_extent_mapped(page);
8978
8979         /*
8980          * we can't set the delalloc bits if there are pending ordered
8981          * extents.  Drop our locks and wait for them to finish
8982          */
8983         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8984         if (ordered) {
8985                 unlock_extent_cached(io_tree, page_start, page_end,
8986                                      &cached_state, GFP_NOFS);
8987                 unlock_page(page);
8988                 btrfs_start_ordered_extent(inode, ordered, 1);
8989                 btrfs_put_ordered_extent(ordered);
8990                 goto again;
8991         }
8992
8993         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8994                 reserved_space = round_up(size - page_start, root->sectorsize);
8995                 if (reserved_space < PAGE_SIZE) {
8996                         end = page_start + reserved_space - 1;
8997                         spin_lock(&BTRFS_I(inode)->lock);
8998                         BTRFS_I(inode)->outstanding_extents++;
8999                         spin_unlock(&BTRFS_I(inode)->lock);
9000                         btrfs_delalloc_release_space(inode, page_start,
9001                                                 PAGE_SIZE - reserved_space);
9002                 }
9003         }
9004
9005         /*
9006          * XXX - page_mkwrite gets called every time the page is dirtied, even
9007          * if it was already dirty, so for space accounting reasons we need to
9008          * clear any delalloc bits for the range we are fixing to save.  There
9009          * is probably a better way to do this, but for now keep consistent with
9010          * prepare_pages in the normal write path.
9011          */
9012         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9013                           EXTENT_DIRTY | EXTENT_DELALLOC |
9014                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9015                           0, 0, &cached_state, GFP_NOFS);
9016
9017         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9018                                         &cached_state);
9019         if (ret) {
9020                 unlock_extent_cached(io_tree, page_start, page_end,
9021                                      &cached_state, GFP_NOFS);
9022                 ret = VM_FAULT_SIGBUS;
9023                 goto out_unlock;
9024         }
9025         ret = 0;
9026
9027         /* page is wholly or partially inside EOF */
9028         if (page_start + PAGE_SIZE > size)
9029                 zero_start = size & ~PAGE_MASK;
9030         else
9031                 zero_start = PAGE_SIZE;
9032
9033         if (zero_start != PAGE_SIZE) {
9034                 kaddr = kmap(page);
9035                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9036                 flush_dcache_page(page);
9037                 kunmap(page);
9038         }
9039         ClearPageChecked(page);
9040         set_page_dirty(page);
9041         SetPageUptodate(page);
9042
9043         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9044         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9045         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9046
9047         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9048
9049 out_unlock:
9050         if (!ret) {
9051                 sb_end_pagefault(inode->i_sb);
9052                 return VM_FAULT_LOCKED;
9053         }
9054         unlock_page(page);
9055 out:
9056         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9057 out_noreserve:
9058         sb_end_pagefault(inode->i_sb);
9059         return ret;
9060 }
9061
9062 static int btrfs_truncate(struct inode *inode)
9063 {
9064         struct btrfs_root *root = BTRFS_I(inode)->root;
9065         struct btrfs_block_rsv *rsv;
9066         int ret = 0;
9067         int err = 0;
9068         struct btrfs_trans_handle *trans;
9069         u64 mask = root->sectorsize - 1;
9070         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9071
9072         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9073                                        (u64)-1);
9074         if (ret)
9075                 return ret;
9076
9077         /*
9078          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9079          * 3 things going on here
9080          *
9081          * 1) We need to reserve space for our orphan item and the space to
9082          * delete our orphan item.  Lord knows we don't want to have a dangling
9083          * orphan item because we didn't reserve space to remove it.
9084          *
9085          * 2) We need to reserve space to update our inode.
9086          *
9087          * 3) We need to have something to cache all the space that is going to
9088          * be free'd up by the truncate operation, but also have some slack
9089          * space reserved in case it uses space during the truncate (thank you
9090          * very much snapshotting).
9091          *
9092          * And we need these to all be separate.  The fact is we can use a lot of
9093          * space doing the truncate, and we have no earthly idea how much space
9094          * we will use, so we need the truncate reservation to be separate so it
9095          * doesn't end up using space reserved for updating the inode or
9096          * removing the orphan item.  We also need to be able to stop the
9097          * transaction and start a new one, which means we need to be able to
9098          * update the inode several times, and we have no idea of knowing how
9099          * many times that will be, so we can't just reserve 1 item for the
9100          * entirety of the operation, so that has to be done separately as well.
9101          * Then there is the orphan item, which does indeed need to be held on
9102          * to for the whole operation, and we need nobody to touch this reserved
9103          * space except the orphan code.
9104          *
9105          * So that leaves us with
9106          *
9107          * 1) root->orphan_block_rsv - for the orphan deletion.
9108          * 2) rsv - for the truncate reservation, which we will steal from the
9109          * transaction reservation.
9110          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9111          * updating the inode.
9112          */
9113         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9114         if (!rsv)
9115                 return -ENOMEM;
9116         rsv->size = min_size;
9117         rsv->failfast = 1;
9118
9119         /*
9120          * 1 for the truncate slack space
9121          * 1 for updating the inode.
9122          */
9123         trans = btrfs_start_transaction(root, 2);
9124         if (IS_ERR(trans)) {
9125                 err = PTR_ERR(trans);
9126                 goto out;
9127         }
9128
9129         /* Migrate the slack space for the truncate to our reserve */
9130         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9131                                       min_size, 0);
9132         BUG_ON(ret);
9133
9134         /*
9135          * So if we truncate and then write and fsync we normally would just
9136          * write the extents that changed, which is a problem if we need to
9137          * first truncate that entire inode.  So set this flag so we write out
9138          * all of the extents in the inode to the sync log so we're completely
9139          * safe.
9140          */
9141         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9142         trans->block_rsv = rsv;
9143
9144         while (1) {
9145                 ret = btrfs_truncate_inode_items(trans, root, inode,
9146                                                  inode->i_size,
9147                                                  BTRFS_EXTENT_DATA_KEY);
9148                 if (ret != -ENOSPC && ret != -EAGAIN) {
9149                         err = ret;
9150                         break;
9151                 }
9152
9153                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9154                 ret = btrfs_update_inode(trans, root, inode);
9155                 if (ret) {
9156                         err = ret;
9157                         break;
9158                 }
9159
9160                 btrfs_end_transaction(trans, root);
9161                 btrfs_btree_balance_dirty(root);
9162
9163                 trans = btrfs_start_transaction(root, 2);
9164                 if (IS_ERR(trans)) {
9165                         ret = err = PTR_ERR(trans);
9166                         trans = NULL;
9167                         break;
9168                 }
9169
9170                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9171                                               rsv, min_size, 0);
9172                 BUG_ON(ret);    /* shouldn't happen */
9173                 trans->block_rsv = rsv;
9174         }
9175
9176         if (ret == 0 && inode->i_nlink > 0) {
9177                 trans->block_rsv = root->orphan_block_rsv;
9178                 ret = btrfs_orphan_del(trans, inode);
9179                 if (ret)
9180                         err = ret;
9181         }
9182
9183         if (trans) {
9184                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9185                 ret = btrfs_update_inode(trans, root, inode);
9186                 if (ret && !err)
9187                         err = ret;
9188
9189                 ret = btrfs_end_transaction(trans, root);
9190                 btrfs_btree_balance_dirty(root);
9191         }
9192 out:
9193         btrfs_free_block_rsv(root, rsv);
9194
9195         if (ret && !err)
9196                 err = ret;
9197
9198         return err;
9199 }
9200
9201 /*
9202  * create a new subvolume directory/inode (helper for the ioctl).
9203  */
9204 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9205                              struct btrfs_root *new_root,
9206                              struct btrfs_root *parent_root,
9207                              u64 new_dirid)
9208 {
9209         struct inode *inode;
9210         int err;
9211         u64 index = 0;
9212
9213         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9214                                 new_dirid, new_dirid,
9215                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9216                                 &index);
9217         if (IS_ERR(inode))
9218                 return PTR_ERR(inode);
9219         inode->i_op = &btrfs_dir_inode_operations;
9220         inode->i_fop = &btrfs_dir_file_operations;
9221
9222         set_nlink(inode, 1);
9223         btrfs_i_size_write(inode, 0);
9224         unlock_new_inode(inode);
9225
9226         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9227         if (err)
9228                 btrfs_err(new_root->fs_info,
9229                           "error inheriting subvolume %llu properties: %d",
9230                           new_root->root_key.objectid, err);
9231
9232         err = btrfs_update_inode(trans, new_root, inode);
9233
9234         iput(inode);
9235         return err;
9236 }
9237
9238 struct inode *btrfs_alloc_inode(struct super_block *sb)
9239 {
9240         struct btrfs_inode *ei;
9241         struct inode *inode;
9242
9243         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9244         if (!ei)
9245                 return NULL;
9246
9247         ei->root = NULL;
9248         ei->generation = 0;
9249         ei->last_trans = 0;
9250         ei->last_sub_trans = 0;
9251         ei->logged_trans = 0;
9252         ei->delalloc_bytes = 0;
9253         ei->defrag_bytes = 0;
9254         ei->disk_i_size = 0;
9255         ei->flags = 0;
9256         ei->csum_bytes = 0;
9257         ei->index_cnt = (u64)-1;
9258         ei->dir_index = 0;
9259         ei->last_unlink_trans = 0;
9260         ei->last_log_commit = 0;
9261         ei->delayed_iput_count = 0;
9262
9263         spin_lock_init(&ei->lock);
9264         ei->outstanding_extents = 0;
9265         ei->reserved_extents = 0;
9266
9267         ei->runtime_flags = 0;
9268         ei->force_compress = BTRFS_COMPRESS_NONE;
9269
9270         ei->delayed_node = NULL;
9271
9272         ei->i_otime.tv_sec = 0;
9273         ei->i_otime.tv_nsec = 0;
9274
9275         inode = &ei->vfs_inode;
9276         extent_map_tree_init(&ei->extent_tree);
9277         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9278         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9279         ei->io_tree.track_uptodate = 1;
9280         ei->io_failure_tree.track_uptodate = 1;
9281         atomic_set(&ei->sync_writers, 0);
9282         mutex_init(&ei->log_mutex);
9283         mutex_init(&ei->delalloc_mutex);
9284         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9285         INIT_LIST_HEAD(&ei->delalloc_inodes);
9286         INIT_LIST_HEAD(&ei->delayed_iput);
9287         RB_CLEAR_NODE(&ei->rb_node);
9288         init_rwsem(&ei->dio_sem);
9289
9290         return inode;
9291 }
9292
9293 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9294 void btrfs_test_destroy_inode(struct inode *inode)
9295 {
9296         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9297         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9298 }
9299 #endif
9300
9301 static void btrfs_i_callback(struct rcu_head *head)
9302 {
9303         struct inode *inode = container_of(head, struct inode, i_rcu);
9304         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9305 }
9306
9307 void btrfs_destroy_inode(struct inode *inode)
9308 {
9309         struct btrfs_ordered_extent *ordered;
9310         struct btrfs_root *root = BTRFS_I(inode)->root;
9311
9312         WARN_ON(!hlist_empty(&inode->i_dentry));
9313         WARN_ON(inode->i_data.nrpages);
9314         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9315         WARN_ON(BTRFS_I(inode)->reserved_extents);
9316         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9317         WARN_ON(BTRFS_I(inode)->csum_bytes);
9318         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9319
9320         /*
9321          * This can happen where we create an inode, but somebody else also
9322          * created the same inode and we need to destroy the one we already
9323          * created.
9324          */
9325         if (!root)
9326                 goto free;
9327
9328         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9329                      &BTRFS_I(inode)->runtime_flags)) {
9330                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9331                         btrfs_ino(inode));
9332                 atomic_dec(&root->orphan_inodes);
9333         }
9334
9335         while (1) {
9336                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9337                 if (!ordered)
9338                         break;
9339                 else {
9340                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9341                                 ordered->file_offset, ordered->len);
9342                         btrfs_remove_ordered_extent(inode, ordered);
9343                         btrfs_put_ordered_extent(ordered);
9344                         btrfs_put_ordered_extent(ordered);
9345                 }
9346         }
9347         btrfs_qgroup_check_reserved_leak(inode);
9348         inode_tree_del(inode);
9349         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9350 free:
9351         call_rcu(&inode->i_rcu, btrfs_i_callback);
9352 }
9353
9354 int btrfs_drop_inode(struct inode *inode)
9355 {
9356         struct btrfs_root *root = BTRFS_I(inode)->root;
9357
9358         if (root == NULL)
9359                 return 1;
9360
9361         /* the snap/subvol tree is on deleting */
9362         if (btrfs_root_refs(&root->root_item) == 0)
9363                 return 1;
9364         else
9365                 return generic_drop_inode(inode);
9366 }
9367
9368 static void init_once(void *foo)
9369 {
9370         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9371
9372         inode_init_once(&ei->vfs_inode);
9373 }
9374
9375 void btrfs_destroy_cachep(void)
9376 {
9377         /*
9378          * Make sure all delayed rcu free inodes are flushed before we
9379          * destroy cache.
9380          */
9381         rcu_barrier();
9382         kmem_cache_destroy(btrfs_inode_cachep);
9383         kmem_cache_destroy(btrfs_trans_handle_cachep);
9384         kmem_cache_destroy(btrfs_transaction_cachep);
9385         kmem_cache_destroy(btrfs_path_cachep);
9386         kmem_cache_destroy(btrfs_free_space_cachep);
9387 }
9388
9389 int btrfs_init_cachep(void)
9390 {
9391         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9392                         sizeof(struct btrfs_inode), 0,
9393                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9394                         init_once);
9395         if (!btrfs_inode_cachep)
9396                 goto fail;
9397
9398         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9399                         sizeof(struct btrfs_trans_handle), 0,
9400                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9401         if (!btrfs_trans_handle_cachep)
9402                 goto fail;
9403
9404         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9405                         sizeof(struct btrfs_transaction), 0,
9406                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9407         if (!btrfs_transaction_cachep)
9408                 goto fail;
9409
9410         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9411                         sizeof(struct btrfs_path), 0,
9412                         SLAB_MEM_SPREAD, NULL);
9413         if (!btrfs_path_cachep)
9414                 goto fail;
9415
9416         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9417                         sizeof(struct btrfs_free_space), 0,
9418                         SLAB_MEM_SPREAD, NULL);
9419         if (!btrfs_free_space_cachep)
9420                 goto fail;
9421
9422         return 0;
9423 fail:
9424         btrfs_destroy_cachep();
9425         return -ENOMEM;
9426 }
9427
9428 static int btrfs_getattr(struct vfsmount *mnt,
9429                          struct dentry *dentry, struct kstat *stat)
9430 {
9431         u64 delalloc_bytes;
9432         struct inode *inode = d_inode(dentry);
9433         u32 blocksize = inode->i_sb->s_blocksize;
9434
9435         generic_fillattr(inode, stat);
9436         stat->dev = BTRFS_I(inode)->root->anon_dev;
9437
9438         spin_lock(&BTRFS_I(inode)->lock);
9439         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9440         spin_unlock(&BTRFS_I(inode)->lock);
9441         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9442                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9443         return 0;
9444 }
9445
9446 static int btrfs_rename_exchange(struct inode *old_dir,
9447                               struct dentry *old_dentry,
9448                               struct inode *new_dir,
9449                               struct dentry *new_dentry)
9450 {
9451         struct btrfs_trans_handle *trans;
9452         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9453         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9454         struct inode *new_inode = new_dentry->d_inode;
9455         struct inode *old_inode = old_dentry->d_inode;
9456         struct timespec ctime = CURRENT_TIME;
9457         struct dentry *parent;
9458         u64 old_ino = btrfs_ino(old_inode);
9459         u64 new_ino = btrfs_ino(new_inode);
9460         u64 old_idx = 0;
9461         u64 new_idx = 0;
9462         u64 root_objectid;
9463         int ret;
9464         bool root_log_pinned = false;
9465         bool dest_log_pinned = false;
9466
9467         /* we only allow rename subvolume link between subvolumes */
9468         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9469                 return -EXDEV;
9470
9471         /* close the race window with snapshot create/destroy ioctl */
9472         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9473                 down_read(&root->fs_info->subvol_sem);
9474         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9475                 down_read(&dest->fs_info->subvol_sem);
9476
9477         /*
9478          * We want to reserve the absolute worst case amount of items.  So if
9479          * both inodes are subvols and we need to unlink them then that would
9480          * require 4 item modifications, but if they are both normal inodes it
9481          * would require 5 item modifications, so we'll assume their normal
9482          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9483          * should cover the worst case number of items we'll modify.
9484          */
9485         trans = btrfs_start_transaction(root, 12);
9486         if (IS_ERR(trans)) {
9487                 ret = PTR_ERR(trans);
9488                 goto out_notrans;
9489         }
9490
9491         /*
9492          * We need to find a free sequence number both in the source and
9493          * in the destination directory for the exchange.
9494          */
9495         ret = btrfs_set_inode_index(new_dir, &old_idx);
9496         if (ret)
9497                 goto out_fail;
9498         ret = btrfs_set_inode_index(old_dir, &new_idx);
9499         if (ret)
9500                 goto out_fail;
9501
9502         BTRFS_I(old_inode)->dir_index = 0ULL;
9503         BTRFS_I(new_inode)->dir_index = 0ULL;
9504
9505         /* Reference for the source. */
9506         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9507                 /* force full log commit if subvolume involved. */
9508                 btrfs_set_log_full_commit(root->fs_info, trans);
9509         } else {
9510                 btrfs_pin_log_trans(root);
9511                 root_log_pinned = true;
9512                 ret = btrfs_insert_inode_ref(trans, dest,
9513                                              new_dentry->d_name.name,
9514                                              new_dentry->d_name.len,
9515                                              old_ino,
9516                                              btrfs_ino(new_dir), old_idx);
9517                 if (ret)
9518                         goto out_fail;
9519         }
9520
9521         /* And now for the dest. */
9522         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9523                 /* force full log commit if subvolume involved. */
9524                 btrfs_set_log_full_commit(dest->fs_info, trans);
9525         } else {
9526                 btrfs_pin_log_trans(dest);
9527                 dest_log_pinned = true;
9528                 ret = btrfs_insert_inode_ref(trans, root,
9529                                              old_dentry->d_name.name,
9530                                              old_dentry->d_name.len,
9531                                              new_ino,
9532                                              btrfs_ino(old_dir), new_idx);
9533                 if (ret)
9534                         goto out_fail;
9535         }
9536
9537         /* Update inode version and ctime/mtime. */
9538         inode_inc_iversion(old_dir);
9539         inode_inc_iversion(new_dir);
9540         inode_inc_iversion(old_inode);
9541         inode_inc_iversion(new_inode);
9542         old_dir->i_ctime = old_dir->i_mtime = ctime;
9543         new_dir->i_ctime = new_dir->i_mtime = ctime;
9544         old_inode->i_ctime = ctime;
9545         new_inode->i_ctime = ctime;
9546
9547         if (old_dentry->d_parent != new_dentry->d_parent) {
9548                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9549                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9550         }
9551
9552         /* src is a subvolume */
9553         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9554                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9555                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9556                                           root_objectid,
9557                                           old_dentry->d_name.name,
9558                                           old_dentry->d_name.len);
9559         } else { /* src is an inode */
9560                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9561                                            old_dentry->d_inode,
9562                                            old_dentry->d_name.name,
9563                                            old_dentry->d_name.len);
9564                 if (!ret)
9565                         ret = btrfs_update_inode(trans, root, old_inode);
9566         }
9567         if (ret) {
9568                 btrfs_abort_transaction(trans, ret);
9569                 goto out_fail;
9570         }
9571
9572         /* dest is a subvolume */
9573         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9574                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9575                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9576                                           root_objectid,
9577                                           new_dentry->d_name.name,
9578                                           new_dentry->d_name.len);
9579         } else { /* dest is an inode */
9580                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9581                                            new_dentry->d_inode,
9582                                            new_dentry->d_name.name,
9583                                            new_dentry->d_name.len);
9584                 if (!ret)
9585                         ret = btrfs_update_inode(trans, dest, new_inode);
9586         }
9587         if (ret) {
9588                 btrfs_abort_transaction(trans, ret);
9589                 goto out_fail;
9590         }
9591
9592         ret = btrfs_add_link(trans, new_dir, old_inode,
9593                              new_dentry->d_name.name,
9594                              new_dentry->d_name.len, 0, old_idx);
9595         if (ret) {
9596                 btrfs_abort_transaction(trans, ret);
9597                 goto out_fail;
9598         }
9599
9600         ret = btrfs_add_link(trans, old_dir, new_inode,
9601                              old_dentry->d_name.name,
9602                              old_dentry->d_name.len, 0, new_idx);
9603         if (ret) {
9604                 btrfs_abort_transaction(trans, ret);
9605                 goto out_fail;
9606         }
9607
9608         if (old_inode->i_nlink == 1)
9609                 BTRFS_I(old_inode)->dir_index = old_idx;
9610         if (new_inode->i_nlink == 1)
9611                 BTRFS_I(new_inode)->dir_index = new_idx;
9612
9613         if (root_log_pinned) {
9614                 parent = new_dentry->d_parent;
9615                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9616                 btrfs_end_log_trans(root);
9617                 root_log_pinned = false;
9618         }
9619         if (dest_log_pinned) {
9620                 parent = old_dentry->d_parent;
9621                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9622                 btrfs_end_log_trans(dest);
9623                 dest_log_pinned = false;
9624         }
9625 out_fail:
9626         /*
9627          * If we have pinned a log and an error happened, we unpin tasks
9628          * trying to sync the log and force them to fallback to a transaction
9629          * commit if the log currently contains any of the inodes involved in
9630          * this rename operation (to ensure we do not persist a log with an
9631          * inconsistent state for any of these inodes or leading to any
9632          * inconsistencies when replayed). If the transaction was aborted, the
9633          * abortion reason is propagated to userspace when attempting to commit
9634          * the transaction. If the log does not contain any of these inodes, we
9635          * allow the tasks to sync it.
9636          */
9637         if (ret && (root_log_pinned || dest_log_pinned)) {
9638                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9639                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9640                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9641                     (new_inode &&
9642                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9643                     btrfs_set_log_full_commit(root->fs_info, trans);
9644
9645                 if (root_log_pinned) {
9646                         btrfs_end_log_trans(root);
9647                         root_log_pinned = false;
9648                 }
9649                 if (dest_log_pinned) {
9650                         btrfs_end_log_trans(dest);
9651                         dest_log_pinned = false;
9652                 }
9653         }
9654         ret = btrfs_end_transaction(trans, root);
9655 out_notrans:
9656         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9657                 up_read(&dest->fs_info->subvol_sem);
9658         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9659                 up_read(&root->fs_info->subvol_sem);
9660
9661         return ret;
9662 }
9663
9664 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9665                                      struct btrfs_root *root,
9666                                      struct inode *dir,
9667                                      struct dentry *dentry)
9668 {
9669         int ret;
9670         struct inode *inode;
9671         u64 objectid;
9672         u64 index;
9673
9674         ret = btrfs_find_free_ino(root, &objectid);
9675         if (ret)
9676                 return ret;
9677
9678         inode = btrfs_new_inode(trans, root, dir,
9679                                 dentry->d_name.name,
9680                                 dentry->d_name.len,
9681                                 btrfs_ino(dir),
9682                                 objectid,
9683                                 S_IFCHR | WHITEOUT_MODE,
9684                                 &index);
9685
9686         if (IS_ERR(inode)) {
9687                 ret = PTR_ERR(inode);
9688                 return ret;
9689         }
9690
9691         inode->i_op = &btrfs_special_inode_operations;
9692         init_special_inode(inode, inode->i_mode,
9693                 WHITEOUT_DEV);
9694
9695         ret = btrfs_init_inode_security(trans, inode, dir,
9696                                 &dentry->d_name);
9697         if (ret)
9698                 goto out;
9699
9700         ret = btrfs_add_nondir(trans, dir, dentry,
9701                                 inode, 0, index);
9702         if (ret)
9703                 goto out;
9704
9705         ret = btrfs_update_inode(trans, root, inode);
9706 out:
9707         unlock_new_inode(inode);
9708         if (ret)
9709                 inode_dec_link_count(inode);
9710         iput(inode);
9711
9712         return ret;
9713 }
9714
9715 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9716                            struct inode *new_dir, struct dentry *new_dentry,
9717                            unsigned int flags)
9718 {
9719         struct btrfs_trans_handle *trans;
9720         unsigned int trans_num_items;
9721         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9722         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9723         struct inode *new_inode = d_inode(new_dentry);
9724         struct inode *old_inode = d_inode(old_dentry);
9725         u64 index = 0;
9726         u64 root_objectid;
9727         int ret;
9728         u64 old_ino = btrfs_ino(old_inode);
9729         bool log_pinned = false;
9730
9731         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9732                 return -EPERM;
9733
9734         /* we only allow rename subvolume link between subvolumes */
9735         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9736                 return -EXDEV;
9737
9738         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9739             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9740                 return -ENOTEMPTY;
9741
9742         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9743             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9744                 return -ENOTEMPTY;
9745
9746
9747         /* check for collisions, even if the  name isn't there */
9748         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9749                              new_dentry->d_name.name,
9750                              new_dentry->d_name.len);
9751
9752         if (ret) {
9753                 if (ret == -EEXIST) {
9754                         /* we shouldn't get
9755                          * eexist without a new_inode */
9756                         if (WARN_ON(!new_inode)) {
9757                                 return ret;
9758                         }
9759                 } else {
9760                         /* maybe -EOVERFLOW */
9761                         return ret;
9762                 }
9763         }
9764         ret = 0;
9765
9766         /*
9767          * we're using rename to replace one file with another.  Start IO on it
9768          * now so  we don't add too much work to the end of the transaction
9769          */
9770         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9771                 filemap_flush(old_inode->i_mapping);
9772
9773         /* close the racy window with snapshot create/destroy ioctl */
9774         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9775                 down_read(&root->fs_info->subvol_sem);
9776         /*
9777          * We want to reserve the absolute worst case amount of items.  So if
9778          * both inodes are subvols and we need to unlink them then that would
9779          * require 4 item modifications, but if they are both normal inodes it
9780          * would require 5 item modifications, so we'll assume they are normal
9781          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9782          * should cover the worst case number of items we'll modify.
9783          * If our rename has the whiteout flag, we need more 5 units for the
9784          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9785          * when selinux is enabled).
9786          */
9787         trans_num_items = 11;
9788         if (flags & RENAME_WHITEOUT)
9789                 trans_num_items += 5;
9790         trans = btrfs_start_transaction(root, trans_num_items);
9791         if (IS_ERR(trans)) {
9792                 ret = PTR_ERR(trans);
9793                 goto out_notrans;
9794         }
9795
9796         if (dest != root)
9797                 btrfs_record_root_in_trans(trans, dest);
9798
9799         ret = btrfs_set_inode_index(new_dir, &index);
9800         if (ret)
9801                 goto out_fail;
9802
9803         BTRFS_I(old_inode)->dir_index = 0ULL;
9804         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9805                 /* force full log commit if subvolume involved. */
9806                 btrfs_set_log_full_commit(root->fs_info, trans);
9807         } else {
9808                 btrfs_pin_log_trans(root);
9809                 log_pinned = true;
9810                 ret = btrfs_insert_inode_ref(trans, dest,
9811                                              new_dentry->d_name.name,
9812                                              new_dentry->d_name.len,
9813                                              old_ino,
9814                                              btrfs_ino(new_dir), index);
9815                 if (ret)
9816                         goto out_fail;
9817         }
9818
9819         inode_inc_iversion(old_dir);
9820         inode_inc_iversion(new_dir);
9821         inode_inc_iversion(old_inode);
9822         old_dir->i_ctime = old_dir->i_mtime =
9823         new_dir->i_ctime = new_dir->i_mtime =
9824         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9825
9826         if (old_dentry->d_parent != new_dentry->d_parent)
9827                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9828
9829         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9830                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9831                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9832                                         old_dentry->d_name.name,
9833                                         old_dentry->d_name.len);
9834         } else {
9835                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9836                                         d_inode(old_dentry),
9837                                         old_dentry->d_name.name,
9838                                         old_dentry->d_name.len);
9839                 if (!ret)
9840                         ret = btrfs_update_inode(trans, root, old_inode);
9841         }
9842         if (ret) {
9843                 btrfs_abort_transaction(trans, ret);
9844                 goto out_fail;
9845         }
9846
9847         if (new_inode) {
9848                 inode_inc_iversion(new_inode);
9849                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9850                 if (unlikely(btrfs_ino(new_inode) ==
9851                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9852                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9853                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9854                                                 root_objectid,
9855                                                 new_dentry->d_name.name,
9856                                                 new_dentry->d_name.len);
9857                         BUG_ON(new_inode->i_nlink == 0);
9858                 } else {
9859                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9860                                                  d_inode(new_dentry),
9861                                                  new_dentry->d_name.name,
9862                                                  new_dentry->d_name.len);
9863                 }
9864                 if (!ret && new_inode->i_nlink == 0)
9865                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9866                 if (ret) {
9867                         btrfs_abort_transaction(trans, ret);
9868                         goto out_fail;
9869                 }
9870         }
9871
9872         ret = btrfs_add_link(trans, new_dir, old_inode,
9873                              new_dentry->d_name.name,
9874                              new_dentry->d_name.len, 0, index);
9875         if (ret) {
9876                 btrfs_abort_transaction(trans, ret);
9877                 goto out_fail;
9878         }
9879
9880         if (old_inode->i_nlink == 1)
9881                 BTRFS_I(old_inode)->dir_index = index;
9882
9883         if (log_pinned) {
9884                 struct dentry *parent = new_dentry->d_parent;
9885
9886                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9887                 btrfs_end_log_trans(root);
9888                 log_pinned = false;
9889         }
9890
9891         if (flags & RENAME_WHITEOUT) {
9892                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9893                                                 old_dentry);
9894
9895                 if (ret) {
9896                         btrfs_abort_transaction(trans, ret);
9897                         goto out_fail;
9898                 }
9899         }
9900 out_fail:
9901         /*
9902          * If we have pinned the log and an error happened, we unpin tasks
9903          * trying to sync the log and force them to fallback to a transaction
9904          * commit if the log currently contains any of the inodes involved in
9905          * this rename operation (to ensure we do not persist a log with an
9906          * inconsistent state for any of these inodes or leading to any
9907          * inconsistencies when replayed). If the transaction was aborted, the
9908          * abortion reason is propagated to userspace when attempting to commit
9909          * the transaction. If the log does not contain any of these inodes, we
9910          * allow the tasks to sync it.
9911          */
9912         if (ret && log_pinned) {
9913                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9914                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9915                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9916                     (new_inode &&
9917                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9918                     btrfs_set_log_full_commit(root->fs_info, trans);
9919
9920                 btrfs_end_log_trans(root);
9921                 log_pinned = false;
9922         }
9923         btrfs_end_transaction(trans, root);
9924 out_notrans:
9925         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9926                 up_read(&root->fs_info->subvol_sem);
9927
9928         return ret;
9929 }
9930
9931 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9932                          struct inode *new_dir, struct dentry *new_dentry,
9933                          unsigned int flags)
9934 {
9935         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9936                 return -EINVAL;
9937
9938         if (flags & RENAME_EXCHANGE)
9939                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9940                                           new_dentry);
9941
9942         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9943 }
9944
9945 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9946 {
9947         struct btrfs_delalloc_work *delalloc_work;
9948         struct inode *inode;
9949
9950         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9951                                      work);
9952         inode = delalloc_work->inode;
9953         filemap_flush(inode->i_mapping);
9954         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9955                                 &BTRFS_I(inode)->runtime_flags))
9956                 filemap_flush(inode->i_mapping);
9957
9958         if (delalloc_work->delay_iput)
9959                 btrfs_add_delayed_iput(inode);
9960         else
9961                 iput(inode);
9962         complete(&delalloc_work->completion);
9963 }
9964
9965 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9966                                                     int delay_iput)
9967 {
9968         struct btrfs_delalloc_work *work;
9969
9970         work = kmalloc(sizeof(*work), GFP_NOFS);
9971         if (!work)
9972                 return NULL;
9973
9974         init_completion(&work->completion);
9975         INIT_LIST_HEAD(&work->list);
9976         work->inode = inode;
9977         work->delay_iput = delay_iput;
9978         WARN_ON_ONCE(!inode);
9979         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9980                         btrfs_run_delalloc_work, NULL, NULL);
9981
9982         return work;
9983 }
9984
9985 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9986 {
9987         wait_for_completion(&work->completion);
9988         kfree(work);
9989 }
9990
9991 /*
9992  * some fairly slow code that needs optimization. This walks the list
9993  * of all the inodes with pending delalloc and forces them to disk.
9994  */
9995 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9996                                    int nr)
9997 {
9998         struct btrfs_inode *binode;
9999         struct inode *inode;
10000         struct btrfs_delalloc_work *work, *next;
10001         struct list_head works;
10002         struct list_head splice;
10003         int ret = 0;
10004
10005         INIT_LIST_HEAD(&works);
10006         INIT_LIST_HEAD(&splice);
10007
10008         mutex_lock(&root->delalloc_mutex);
10009         spin_lock(&root->delalloc_lock);
10010         list_splice_init(&root->delalloc_inodes, &splice);
10011         while (!list_empty(&splice)) {
10012                 binode = list_entry(splice.next, struct btrfs_inode,
10013                                     delalloc_inodes);
10014
10015                 list_move_tail(&binode->delalloc_inodes,
10016                                &root->delalloc_inodes);
10017                 inode = igrab(&binode->vfs_inode);
10018                 if (!inode) {
10019                         cond_resched_lock(&root->delalloc_lock);
10020                         continue;
10021                 }
10022                 spin_unlock(&root->delalloc_lock);
10023
10024                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10025                 if (!work) {
10026                         if (delay_iput)
10027                                 btrfs_add_delayed_iput(inode);
10028                         else
10029                                 iput(inode);
10030                         ret = -ENOMEM;
10031                         goto out;
10032                 }
10033                 list_add_tail(&work->list, &works);
10034                 btrfs_queue_work(root->fs_info->flush_workers,
10035                                  &work->work);
10036                 ret++;
10037                 if (nr != -1 && ret >= nr)
10038                         goto out;
10039                 cond_resched();
10040                 spin_lock(&root->delalloc_lock);
10041         }
10042         spin_unlock(&root->delalloc_lock);
10043
10044 out:
10045         list_for_each_entry_safe(work, next, &works, list) {
10046                 list_del_init(&work->list);
10047                 btrfs_wait_and_free_delalloc_work(work);
10048         }
10049
10050         if (!list_empty_careful(&splice)) {
10051                 spin_lock(&root->delalloc_lock);
10052                 list_splice_tail(&splice, &root->delalloc_inodes);
10053                 spin_unlock(&root->delalloc_lock);
10054         }
10055         mutex_unlock(&root->delalloc_mutex);
10056         return ret;
10057 }
10058
10059 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10060 {
10061         int ret;
10062
10063         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10064                 return -EROFS;
10065
10066         ret = __start_delalloc_inodes(root, delay_iput, -1);
10067         if (ret > 0)
10068                 ret = 0;
10069         /*
10070          * the filemap_flush will queue IO into the worker threads, but
10071          * we have to make sure the IO is actually started and that
10072          * ordered extents get created before we return
10073          */
10074         atomic_inc(&root->fs_info->async_submit_draining);
10075         while (atomic_read(&root->fs_info->nr_async_submits) ||
10076               atomic_read(&root->fs_info->async_delalloc_pages)) {
10077                 wait_event(root->fs_info->async_submit_wait,
10078                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10079                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10080         }
10081         atomic_dec(&root->fs_info->async_submit_draining);
10082         return ret;
10083 }
10084
10085 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10086                                int nr)
10087 {
10088         struct btrfs_root *root;
10089         struct list_head splice;
10090         int ret;
10091
10092         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10093                 return -EROFS;
10094
10095         INIT_LIST_HEAD(&splice);
10096
10097         mutex_lock(&fs_info->delalloc_root_mutex);
10098         spin_lock(&fs_info->delalloc_root_lock);
10099         list_splice_init(&fs_info->delalloc_roots, &splice);
10100         while (!list_empty(&splice) && nr) {
10101                 root = list_first_entry(&splice, struct btrfs_root,
10102                                         delalloc_root);
10103                 root = btrfs_grab_fs_root(root);
10104                 BUG_ON(!root);
10105                 list_move_tail(&root->delalloc_root,
10106                                &fs_info->delalloc_roots);
10107                 spin_unlock(&fs_info->delalloc_root_lock);
10108
10109                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10110                 btrfs_put_fs_root(root);
10111                 if (ret < 0)
10112                         goto out;
10113
10114                 if (nr != -1) {
10115                         nr -= ret;
10116                         WARN_ON(nr < 0);
10117                 }
10118                 spin_lock(&fs_info->delalloc_root_lock);
10119         }
10120         spin_unlock(&fs_info->delalloc_root_lock);
10121
10122         ret = 0;
10123         atomic_inc(&fs_info->async_submit_draining);
10124         while (atomic_read(&fs_info->nr_async_submits) ||
10125               atomic_read(&fs_info->async_delalloc_pages)) {
10126                 wait_event(fs_info->async_submit_wait,
10127                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10128                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10129         }
10130         atomic_dec(&fs_info->async_submit_draining);
10131 out:
10132         if (!list_empty_careful(&splice)) {
10133                 spin_lock(&fs_info->delalloc_root_lock);
10134                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10135                 spin_unlock(&fs_info->delalloc_root_lock);
10136         }
10137         mutex_unlock(&fs_info->delalloc_root_mutex);
10138         return ret;
10139 }
10140
10141 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10142                          const char *symname)
10143 {
10144         struct btrfs_trans_handle *trans;
10145         struct btrfs_root *root = BTRFS_I(dir)->root;
10146         struct btrfs_path *path;
10147         struct btrfs_key key;
10148         struct inode *inode = NULL;
10149         int err;
10150         int drop_inode = 0;
10151         u64 objectid;
10152         u64 index = 0;
10153         int name_len;
10154         int datasize;
10155         unsigned long ptr;
10156         struct btrfs_file_extent_item *ei;
10157         struct extent_buffer *leaf;
10158
10159         name_len = strlen(symname);
10160         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10161                 return -ENAMETOOLONG;
10162
10163         /*
10164          * 2 items for inode item and ref
10165          * 2 items for dir items
10166          * 1 item for updating parent inode item
10167          * 1 item for the inline extent item
10168          * 1 item for xattr if selinux is on
10169          */
10170         trans = btrfs_start_transaction(root, 7);
10171         if (IS_ERR(trans))
10172                 return PTR_ERR(trans);
10173
10174         err = btrfs_find_free_ino(root, &objectid);
10175         if (err)
10176                 goto out_unlock;
10177
10178         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10179                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10180                                 S_IFLNK|S_IRWXUGO, &index);
10181         if (IS_ERR(inode)) {
10182                 err = PTR_ERR(inode);
10183                 goto out_unlock;
10184         }
10185
10186         /*
10187         * If the active LSM wants to access the inode during
10188         * d_instantiate it needs these. Smack checks to see
10189         * if the filesystem supports xattrs by looking at the
10190         * ops vector.
10191         */
10192         inode->i_fop = &btrfs_file_operations;
10193         inode->i_op = &btrfs_file_inode_operations;
10194         inode->i_mapping->a_ops = &btrfs_aops;
10195         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10196
10197         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10198         if (err)
10199                 goto out_unlock_inode;
10200
10201         path = btrfs_alloc_path();
10202         if (!path) {
10203                 err = -ENOMEM;
10204                 goto out_unlock_inode;
10205         }
10206         key.objectid = btrfs_ino(inode);
10207         key.offset = 0;
10208         key.type = BTRFS_EXTENT_DATA_KEY;
10209         datasize = btrfs_file_extent_calc_inline_size(name_len);
10210         err = btrfs_insert_empty_item(trans, root, path, &key,
10211                                       datasize);
10212         if (err) {
10213                 btrfs_free_path(path);
10214                 goto out_unlock_inode;
10215         }
10216         leaf = path->nodes[0];
10217         ei = btrfs_item_ptr(leaf, path->slots[0],
10218                             struct btrfs_file_extent_item);
10219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10220         btrfs_set_file_extent_type(leaf, ei,
10221                                    BTRFS_FILE_EXTENT_INLINE);
10222         btrfs_set_file_extent_encryption(leaf, ei, 0);
10223         btrfs_set_file_extent_compression(leaf, ei, 0);
10224         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10225         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10226
10227         ptr = btrfs_file_extent_inline_start(ei);
10228         write_extent_buffer(leaf, symname, ptr, name_len);
10229         btrfs_mark_buffer_dirty(leaf);
10230         btrfs_free_path(path);
10231
10232         inode->i_op = &btrfs_symlink_inode_operations;
10233         inode_nohighmem(inode);
10234         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10235         inode_set_bytes(inode, name_len);
10236         btrfs_i_size_write(inode, name_len);
10237         err = btrfs_update_inode(trans, root, inode);
10238         /*
10239          * Last step, add directory indexes for our symlink inode. This is the
10240          * last step to avoid extra cleanup of these indexes if an error happens
10241          * elsewhere above.
10242          */
10243         if (!err)
10244                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10245         if (err) {
10246                 drop_inode = 1;
10247                 goto out_unlock_inode;
10248         }
10249
10250         unlock_new_inode(inode);
10251         d_instantiate(dentry, inode);
10252
10253 out_unlock:
10254         btrfs_end_transaction(trans, root);
10255         if (drop_inode) {
10256                 inode_dec_link_count(inode);
10257                 iput(inode);
10258         }
10259         btrfs_btree_balance_dirty(root);
10260         return err;
10261
10262 out_unlock_inode:
10263         drop_inode = 1;
10264         unlock_new_inode(inode);
10265         goto out_unlock;
10266 }
10267
10268 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10269                                        u64 start, u64 num_bytes, u64 min_size,
10270                                        loff_t actual_len, u64 *alloc_hint,
10271                                        struct btrfs_trans_handle *trans)
10272 {
10273         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10274         struct extent_map *em;
10275         struct btrfs_root *root = BTRFS_I(inode)->root;
10276         struct btrfs_key ins;
10277         u64 cur_offset = start;
10278         u64 i_size;
10279         u64 cur_bytes;
10280         u64 last_alloc = (u64)-1;
10281         int ret = 0;
10282         bool own_trans = true;
10283
10284         if (trans)
10285                 own_trans = false;
10286         while (num_bytes > 0) {
10287                 if (own_trans) {
10288                         trans = btrfs_start_transaction(root, 3);
10289                         if (IS_ERR(trans)) {
10290                                 ret = PTR_ERR(trans);
10291                                 break;
10292                         }
10293                 }
10294
10295                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10296                 cur_bytes = max(cur_bytes, min_size);
10297                 /*
10298                  * If we are severely fragmented we could end up with really
10299                  * small allocations, so if the allocator is returning small
10300                  * chunks lets make its job easier by only searching for those
10301                  * sized chunks.
10302                  */
10303                 cur_bytes = min(cur_bytes, last_alloc);
10304                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10305                                            *alloc_hint, &ins, 1, 0);
10306                 if (ret) {
10307                         if (own_trans)
10308                                 btrfs_end_transaction(trans, root);
10309                         break;
10310                 }
10311                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10312
10313                 last_alloc = ins.offset;
10314                 ret = insert_reserved_file_extent(trans, inode,
10315                                                   cur_offset, ins.objectid,
10316                                                   ins.offset, ins.offset,
10317                                                   ins.offset, 0, 0, 0,
10318                                                   BTRFS_FILE_EXTENT_PREALLOC);
10319                 if (ret) {
10320                         btrfs_free_reserved_extent(root, ins.objectid,
10321                                                    ins.offset, 0);
10322                         btrfs_abort_transaction(trans, ret);
10323                         if (own_trans)
10324                                 btrfs_end_transaction(trans, root);
10325                         break;
10326                 }
10327
10328                 btrfs_drop_extent_cache(inode, cur_offset,
10329                                         cur_offset + ins.offset -1, 0);
10330
10331                 em = alloc_extent_map();
10332                 if (!em) {
10333                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10334                                 &BTRFS_I(inode)->runtime_flags);
10335                         goto next;
10336                 }
10337
10338                 em->start = cur_offset;
10339                 em->orig_start = cur_offset;
10340                 em->len = ins.offset;
10341                 em->block_start = ins.objectid;
10342                 em->block_len = ins.offset;
10343                 em->orig_block_len = ins.offset;
10344                 em->ram_bytes = ins.offset;
10345                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10346                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10347                 em->generation = trans->transid;
10348
10349                 while (1) {
10350                         write_lock(&em_tree->lock);
10351                         ret = add_extent_mapping(em_tree, em, 1);
10352                         write_unlock(&em_tree->lock);
10353                         if (ret != -EEXIST)
10354                                 break;
10355                         btrfs_drop_extent_cache(inode, cur_offset,
10356                                                 cur_offset + ins.offset - 1,
10357                                                 0);
10358                 }
10359                 free_extent_map(em);
10360 next:
10361                 num_bytes -= ins.offset;
10362                 cur_offset += ins.offset;
10363                 *alloc_hint = ins.objectid + ins.offset;
10364
10365                 inode_inc_iversion(inode);
10366                 inode->i_ctime = current_fs_time(inode->i_sb);
10367                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10368                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10369                     (actual_len > inode->i_size) &&
10370                     (cur_offset > inode->i_size)) {
10371                         if (cur_offset > actual_len)
10372                                 i_size = actual_len;
10373                         else
10374                                 i_size = cur_offset;
10375                         i_size_write(inode, i_size);
10376                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10377                 }
10378
10379                 ret = btrfs_update_inode(trans, root, inode);
10380
10381                 if (ret) {
10382                         btrfs_abort_transaction(trans, ret);
10383                         if (own_trans)
10384                                 btrfs_end_transaction(trans, root);
10385                         break;
10386                 }
10387
10388                 if (own_trans)
10389                         btrfs_end_transaction(trans, root);
10390         }
10391         return ret;
10392 }
10393
10394 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10395                               u64 start, u64 num_bytes, u64 min_size,
10396                               loff_t actual_len, u64 *alloc_hint)
10397 {
10398         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10399                                            min_size, actual_len, alloc_hint,
10400                                            NULL);
10401 }
10402
10403 int btrfs_prealloc_file_range_trans(struct inode *inode,
10404                                     struct btrfs_trans_handle *trans, int mode,
10405                                     u64 start, u64 num_bytes, u64 min_size,
10406                                     loff_t actual_len, u64 *alloc_hint)
10407 {
10408         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10409                                            min_size, actual_len, alloc_hint, trans);
10410 }
10411
10412 static int btrfs_set_page_dirty(struct page *page)
10413 {
10414         return __set_page_dirty_nobuffers(page);
10415 }
10416
10417 static int btrfs_permission(struct inode *inode, int mask)
10418 {
10419         struct btrfs_root *root = BTRFS_I(inode)->root;
10420         umode_t mode = inode->i_mode;
10421
10422         if (mask & MAY_WRITE &&
10423             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10424                 if (btrfs_root_readonly(root))
10425                         return -EROFS;
10426                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10427                         return -EACCES;
10428         }
10429         return generic_permission(inode, mask);
10430 }
10431
10432 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10433 {
10434         struct btrfs_trans_handle *trans;
10435         struct btrfs_root *root = BTRFS_I(dir)->root;
10436         struct inode *inode = NULL;
10437         u64 objectid;
10438         u64 index;
10439         int ret = 0;
10440
10441         /*
10442          * 5 units required for adding orphan entry
10443          */
10444         trans = btrfs_start_transaction(root, 5);
10445         if (IS_ERR(trans))
10446                 return PTR_ERR(trans);
10447
10448         ret = btrfs_find_free_ino(root, &objectid);
10449         if (ret)
10450                 goto out;
10451
10452         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10453                                 btrfs_ino(dir), objectid, mode, &index);
10454         if (IS_ERR(inode)) {
10455                 ret = PTR_ERR(inode);
10456                 inode = NULL;
10457                 goto out;
10458         }
10459
10460         inode->i_fop = &btrfs_file_operations;
10461         inode->i_op = &btrfs_file_inode_operations;
10462
10463         inode->i_mapping->a_ops = &btrfs_aops;
10464         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10465
10466         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10467         if (ret)
10468                 goto out_inode;
10469
10470         ret = btrfs_update_inode(trans, root, inode);
10471         if (ret)
10472                 goto out_inode;
10473         ret = btrfs_orphan_add(trans, inode);
10474         if (ret)
10475                 goto out_inode;
10476
10477         /*
10478          * We set number of links to 0 in btrfs_new_inode(), and here we set
10479          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10480          * through:
10481          *
10482          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10483          */
10484         set_nlink(inode, 1);
10485         unlock_new_inode(inode);
10486         d_tmpfile(dentry, inode);
10487         mark_inode_dirty(inode);
10488
10489 out:
10490         btrfs_end_transaction(trans, root);
10491         if (ret)
10492                 iput(inode);
10493         btrfs_balance_delayed_items(root);
10494         btrfs_btree_balance_dirty(root);
10495         return ret;
10496
10497 out_inode:
10498         unlock_new_inode(inode);
10499         goto out;
10500
10501 }
10502
10503 /* Inspired by filemap_check_errors() */
10504 int btrfs_inode_check_errors(struct inode *inode)
10505 {
10506         int ret = 0;
10507
10508         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10509             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10510                 ret = -ENOSPC;
10511         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10512             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10513                 ret = -EIO;
10514
10515         return ret;
10516 }
10517
10518 static const struct inode_operations btrfs_dir_inode_operations = {
10519         .getattr        = btrfs_getattr,
10520         .lookup         = btrfs_lookup,
10521         .create         = btrfs_create,
10522         .unlink         = btrfs_unlink,
10523         .link           = btrfs_link,
10524         .mkdir          = btrfs_mkdir,
10525         .rmdir          = btrfs_rmdir,
10526         .rename2        = btrfs_rename2,
10527         .symlink        = btrfs_symlink,
10528         .setattr        = btrfs_setattr,
10529         .mknod          = btrfs_mknod,
10530         .setxattr       = generic_setxattr,
10531         .getxattr       = generic_getxattr,
10532         .listxattr      = btrfs_listxattr,
10533         .removexattr    = generic_removexattr,
10534         .permission     = btrfs_permission,
10535         .get_acl        = btrfs_get_acl,
10536         .set_acl        = btrfs_set_acl,
10537         .update_time    = btrfs_update_time,
10538         .tmpfile        = btrfs_tmpfile,
10539 };
10540 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10541         .lookup         = btrfs_lookup,
10542         .permission     = btrfs_permission,
10543         .get_acl        = btrfs_get_acl,
10544         .set_acl        = btrfs_set_acl,
10545         .update_time    = btrfs_update_time,
10546 };
10547
10548 static const struct file_operations btrfs_dir_file_operations = {
10549         .llseek         = generic_file_llseek,
10550         .read           = generic_read_dir,
10551         .iterate_shared = btrfs_real_readdir,
10552         .unlocked_ioctl = btrfs_ioctl,
10553 #ifdef CONFIG_COMPAT
10554         .compat_ioctl   = btrfs_compat_ioctl,
10555 #endif
10556         .release        = btrfs_release_file,
10557         .fsync          = btrfs_sync_file,
10558 };
10559
10560 static const struct extent_io_ops btrfs_extent_io_ops = {
10561         .fill_delalloc = run_delalloc_range,
10562         .submit_bio_hook = btrfs_submit_bio_hook,
10563         .merge_bio_hook = btrfs_merge_bio_hook,
10564         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10565         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10566         .writepage_start_hook = btrfs_writepage_start_hook,
10567         .set_bit_hook = btrfs_set_bit_hook,
10568         .clear_bit_hook = btrfs_clear_bit_hook,
10569         .merge_extent_hook = btrfs_merge_extent_hook,
10570         .split_extent_hook = btrfs_split_extent_hook,
10571 };
10572
10573 /*
10574  * btrfs doesn't support the bmap operation because swapfiles
10575  * use bmap to make a mapping of extents in the file.  They assume
10576  * these extents won't change over the life of the file and they
10577  * use the bmap result to do IO directly to the drive.
10578  *
10579  * the btrfs bmap call would return logical addresses that aren't
10580  * suitable for IO and they also will change frequently as COW
10581  * operations happen.  So, swapfile + btrfs == corruption.
10582  *
10583  * For now we're avoiding this by dropping bmap.
10584  */
10585 static const struct address_space_operations btrfs_aops = {
10586         .readpage       = btrfs_readpage,
10587         .writepage      = btrfs_writepage,
10588         .writepages     = btrfs_writepages,
10589         .readpages      = btrfs_readpages,
10590         .direct_IO      = btrfs_direct_IO,
10591         .invalidatepage = btrfs_invalidatepage,
10592         .releasepage    = btrfs_releasepage,
10593         .set_page_dirty = btrfs_set_page_dirty,
10594         .error_remove_page = generic_error_remove_page,
10595 };
10596
10597 static const struct address_space_operations btrfs_symlink_aops = {
10598         .readpage       = btrfs_readpage,
10599         .writepage      = btrfs_writepage,
10600         .invalidatepage = btrfs_invalidatepage,
10601         .releasepage    = btrfs_releasepage,
10602 };
10603
10604 static const struct inode_operations btrfs_file_inode_operations = {
10605         .getattr        = btrfs_getattr,
10606         .setattr        = btrfs_setattr,
10607         .setxattr       = generic_setxattr,
10608         .getxattr       = generic_getxattr,
10609         .listxattr      = btrfs_listxattr,
10610         .removexattr    = generic_removexattr,
10611         .permission     = btrfs_permission,
10612         .fiemap         = btrfs_fiemap,
10613         .get_acl        = btrfs_get_acl,
10614         .set_acl        = btrfs_set_acl,
10615         .update_time    = btrfs_update_time,
10616 };
10617 static const struct inode_operations btrfs_special_inode_operations = {
10618         .getattr        = btrfs_getattr,
10619         .setattr        = btrfs_setattr,
10620         .permission     = btrfs_permission,
10621         .setxattr       = generic_setxattr,
10622         .getxattr       = generic_getxattr,
10623         .listxattr      = btrfs_listxattr,
10624         .removexattr    = generic_removexattr,
10625         .get_acl        = btrfs_get_acl,
10626         .set_acl        = btrfs_set_acl,
10627         .update_time    = btrfs_update_time,
10628 };
10629 static const struct inode_operations btrfs_symlink_inode_operations = {
10630         .readlink       = generic_readlink,
10631         .get_link       = page_get_link,
10632         .getattr        = btrfs_getattr,
10633         .setattr        = btrfs_setattr,
10634         .permission     = btrfs_permission,
10635         .setxattr       = generic_setxattr,
10636         .getxattr       = generic_getxattr,
10637         .listxattr      = btrfs_listxattr,
10638         .removexattr    = generic_removexattr,
10639         .update_time    = btrfs_update_time,
10640 };
10641
10642 const struct dentry_operations btrfs_dentry_operations = {
10643         .d_delete       = btrfs_dentry_delete,
10644         .d_release      = btrfs_dentry_release,
10645 };