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, end,
 564                                                      NULL, clear_flags,
 565                                                      PAGE_UNLOCK |
 566                                                      PAGE_CLEAR_DIRTY |
 567                                                      PAGE_SET_WRITEBACK |
 568                                                      page_error_op |
 569                                                      PAGE_END_WRITEBACK);
 570                         btrfs_free_reserved_data_space_noquota(inode, start,
 571                                                 end - start + 1);
 572                         goto free_pages_out;
 573                 }
 574         }
 575
 576         if (will_compress) {
 577                 /*
 578                  * we aren't doing an inline extent round the compressed size
 579                  * up to a block size boundary so the allocator does sane
 580                  * things
 581                  */
 582                 total_compressed = ALIGN(total_compressed, blocksize);
 583
 584                 /*
 585                  * one last check to make sure the compression is really a
 586                  * win, compare the page count read with the blocks on disk
 587                  */
 588                 total_in = ALIGN(total_in, PAGE_SIZE);
 589                 if (total_compressed >= total_in) {
 590                         will_compress = 0;
 591                 } else {
 592                         num_bytes = total_in;
 593                         *num_added += 1;
 594
 595                         /*
 596                          * The async work queues will take care of doing actual
 597                          * allocation on disk for these compressed pages, and
 598                          * will submit them to the elevator.
 599                          */
 600                         add_async_extent(async_cow, start, num_bytes,
 601                                         total_compressed, pages, nr_pages_ret,
 602                                         compress_type);
 603
 604                         if (start + num_bytes < end) {
 605                                 start += num_bytes;
 606                                 pages = NULL;
 607                                 cond_resched();
 608                                 goto again;
 609                         }
 610                         return;
 611                 }
 612         }
 613         if (pages) {
 614                 /*
 615                  * the compression code ran but failed to make things smaller,
 616                  * free any pages it allocated and our page pointer array
 617                  */
 618                 for (i = 0; i < nr_pages_ret; i++) {
 619                         WARN_ON(pages[i]->mapping);
 620                         put_page(pages[i]);
 621                 }
 622                 kfree(pages);
 623                 pages = NULL;
 624                 total_compressed = 0;
 625                 nr_pages_ret = 0;
 626
 627                 /* flag the file so we don't compress in the future */
 628                 if (!btrfs_test_opt(root->fs_info, FORCE_COMPRESS) &&
 629                     !(BTRFS_I(inode)->force_compress)) {
 630                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 631                 }
 632         }
 633 cleanup_and_bail_uncompressed:
 634         /*
 635          * No compression, but we still need to write the pages in the file
 636          * we've been given so far.  redirty the locked page if it corresponds
 637          * to our extent and set things up for the async work queue to run
 638          * cow_file_range to do the normal delalloc dance.
 639          */
 640         if (page_offset(locked_page) >= start &&
 641             page_offset(locked_page) <= end)
 642                 __set_page_dirty_nobuffers(locked_page);
 643                 /* unlocked later on in the async handlers */
 644
 645         if (redirty)
 646                 extent_range_redirty_for_io(inode, start, end);
 647         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 648                          BTRFS_COMPRESS_NONE);
 649         *num_added += 1;
 650
 651         return;
 652
 653 free_pages_out:
 654         for (i = 0; i < nr_pages_ret; i++) {
 655                 WARN_ON(pages[i]->mapping);
 656                 put_page(pages[i]);
 657         }
 658         kfree(pages);
 659 }
 660
 661 static void free_async_extent_pages(struct async_extent *async_extent)
 662 {
 663         int i;
 664
 665         if (!async_extent->pages)
 666                 return;
 667
 668         for (i = 0; i < async_extent->nr_pages; i++) {
 669                 WARN_ON(async_extent->pages[i]->mapping);
 670                 put_page(async_extent->pages[i]);
 671         }
 672         kfree(async_extent->pages);
 673         async_extent->nr_pages = 0;
 674         async_extent->pages = NULL;
 675 }
 676
 677 /*
 678  * phase two of compressed writeback.  This is the ordered portion
 679  * of the code, which only gets called in the order the work was
 680  * queued.  We walk all the async extents created by compress_file_range
 681  * and send them down to the disk.
 682  */
 683 static noinline void submit_compressed_extents(struct inode *inode,
 684                                               struct async_cow *async_cow)
 685 {
 686         struct async_extent *async_extent;
 687         u64 alloc_hint = 0;
 688         struct btrfs_key ins;
 689         struct extent_map *em;
 690         struct btrfs_root *root = BTRFS_I(inode)->root;
 691         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 692         struct extent_io_tree *io_tree;
 693         int ret = 0;
 694
 695 again:
 696         while (!list_empty(&async_cow->extents)) {
 697                 async_extent = list_entry(async_cow->extents.next,
 698                                           struct async_extent, list);
 699                 list_del(&async_extent->list);
 700
 701                 io_tree = &BTRFS_I(inode)->io_tree;
 702
 703 retry:
 704                 /* did the compression code fall back to uncompressed IO? */
 705                 if (!async_extent->pages) {
 706                         int page_started = 0;
 707                         unsigned long nr_written = 0;
 708
 709                         lock_extent(io_tree, async_extent->start,
 710                                          async_extent->start +
 711                                          async_extent->ram_size - 1);
 712
 713                         /* allocate blocks */
 714                         ret = cow_file_range(inode, async_cow->locked_page,
 715                                              async_extent->start,
 716                                              async_extent->start +
 717                                              async_extent->ram_size - 1,
 718                                              async_extent->start +
 719                                              async_extent->ram_size - 1,
 720                                              &page_started, &nr_written, 0,
 721                                              NULL);
 722
 723                         /* JDM XXX */
 724
 725                         /*
 726                          * if page_started, cow_file_range inserted an
 727                          * inline extent and took care of all the unlocking
 728                          * and IO for us.  Otherwise, we need to submit
 729                          * all those pages down to the drive.
 730                          */
 731                         if (!page_started && !ret)
 732                                 extent_write_locked_range(io_tree,
 733                                                   inode, async_extent->start,
 734                                                   async_extent->start +
 735                                                   async_extent->ram_size - 1,
 736                                                   btrfs_get_extent,
 737                                                   WB_SYNC_ALL);
 738                         else if (ret)
 739                                 unlock_page(async_cow->locked_page);
 740                         kfree(async_extent);
 741                         cond_resched();
 742                         continue;
 743                 }
 744
 745                 lock_extent(io_tree, async_extent->start,
 746                             async_extent->start + async_extent->ram_size - 1);
 747
 748                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 749                                            async_extent->compressed_size,
 750                                            async_extent->compressed_size,
 751                                            0, alloc_hint, &ins, 1, 1);
 752                 if (ret) {
 753                         free_async_extent_pages(async_extent);
 754
 755                         if (ret == -ENOSPC) {
 756                                 unlock_extent(io_tree, async_extent->start,
 757                                               async_extent->start +
 758                                               async_extent->ram_size - 1);
 759
 760                                 /*
 761                                  * we need to redirty the pages if we decide to
 762                                  * fallback to uncompressed IO, otherwise we
 763                                  * will not submit these pages down to lower
 764                                  * layers.
 765                                  */
 766                                 extent_range_redirty_for_io(inode,
 767                                                 async_extent->start,
 768                                                 async_extent->start +
 769                                                 async_extent->ram_size - 1);
 770
 771                                 goto retry;
 772                         }
 773                         goto out_free;
 774                 }
 775                 /*
 776                  * here we're doing allocation and writeback of the
 777                  * compressed pages
 778                  */
 779                 btrfs_drop_extent_cache(inode, async_extent->start,
 780                                         async_extent->start +
 781                                         async_extent->ram_size - 1, 0);
 782
 783                 em = alloc_extent_map();
 784                 if (!em) {
 785                         ret = -ENOMEM;
 786                         goto out_free_reserve;
 787                 }
 788                 em->start = async_extent->start;
 789                 em->len = async_extent->ram_size;
 790                 em->orig_start = em->start;
 791                 em->mod_start = em->start;
 792                 em->mod_len = em->len;
 793
 794                 em->block_start = ins.objectid;
 795                 em->block_len = ins.offset;
 796                 em->orig_block_len = ins.offset;
 797                 em->ram_bytes = async_extent->ram_size;
 798                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 799                 em->compress_type = async_extent->compress_type;
 800                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 801                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 802                 em->generation = -1;
 803
 804                 while (1) {
 805                         write_lock(&em_tree->lock);
 806                         ret = add_extent_mapping(em_tree, em, 1);
 807                         write_unlock(&em_tree->lock);
 808                         if (ret != -EEXIST) {
 809                                 free_extent_map(em);
 810                                 break;
 811                         }
 812                         btrfs_drop_extent_cache(inode, async_extent->start,
 813                                                 async_extent->start +
 814                                                 async_extent->ram_size - 1, 0);
 815                 }
 816
 817                 if (ret)
 818                         goto out_free_reserve;
 819
 820                 ret = btrfs_add_ordered_extent_compress(inode,
 821                                                 async_extent->start,
 822                                                 ins.objectid,
 823                                                 async_extent->ram_size,
 824                                                 ins.offset,
 825                                                 BTRFS_ORDERED_COMPRESSED,
 826                                                 async_extent->compress_type);
 827                 if (ret) {
 828                         btrfs_drop_extent_cache(inode, async_extent->start,
 829                                                 async_extent->start +
 830                                                 async_extent->ram_size - 1, 0);
 831                         goto out_free_reserve;
 832                 }
 833                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 834
 835                 /*
 836                  * clear dirty, set writeback and unlock the pages.
 837                  */
 838                 extent_clear_unlock_delalloc(inode, async_extent->start,
 839                                 async_extent->start +
 840                                 async_extent->ram_size - 1,
 841                                 async_extent->start +
 842                                 async_extent->ram_size - 1,
 843                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 844                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 845                                 PAGE_SET_WRITEBACK);
 846                 ret = btrfs_submit_compressed_write(inode,
 847                                     async_extent->start,
 848                                     async_extent->ram_size,
 849                                     ins.objectid,
 850                                     ins.offset, async_extent->pages,
 851                                     async_extent->nr_pages);
 852                 if (ret) {
 853                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 854                         struct page *p = async_extent->pages[0];
 855                         const u64 start = async_extent->start;
 856                         const u64 end = start + async_extent->ram_size - 1;
 857
 858                         p->mapping = inode->i_mapping;
 859                         tree->ops->writepage_end_io_hook(p, start, end,
 860                                                          NULL, 0);
 861                         p->mapping = NULL;
 862                         extent_clear_unlock_delalloc(inode, start, end, end,
 863                                                      NULL, 0,
 864                                                      PAGE_END_WRITEBACK |
 865                                                      PAGE_SET_ERROR);
 866                         free_async_extent_pages(async_extent);
 867                 }
 868                 alloc_hint = ins.objectid + ins.offset;
 869                 kfree(async_extent);
 870                 cond_resched();
 871         }
 872         return;
 873 out_free_reserve:
 874         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 875         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 876 out_free:
 877         extent_clear_unlock_delalloc(inode, async_extent->start,
 878                                      async_extent->start +
 879                                      async_extent->ram_size - 1,
 880                                      async_extent->start +
 881                                      async_extent->ram_size - 1,
 882                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 883                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 884                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 885                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 886                                      PAGE_SET_ERROR);
 887         free_async_extent_pages(async_extent);
 888         kfree(async_extent);
 889         goto again;
 890 }
 891
 892 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 893                                       u64 num_bytes)
 894 {
 895         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 896         struct extent_map *em;
 897         u64 alloc_hint = 0;
 898
 899         read_lock(&em_tree->lock);
 900         em = search_extent_mapping(em_tree, start, num_bytes);
 901         if (em) {
 902                 /*
 903                  * if block start isn't an actual block number then find the
 904                  * first block in this inode and use that as a hint.  If that
 905                  * block is also bogus then just don't worry about it.
 906                  */
 907                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 908                         free_extent_map(em);
 909                         em = search_extent_mapping(em_tree, 0, 0);
 910                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 911                                 alloc_hint = em->block_start;
 912                         if (em)
 913                                 free_extent_map(em);
 914                 } else {
 915                         alloc_hint = em->block_start;
 916                         free_extent_map(em);
 917                 }
 918         }
 919         read_unlock(&em_tree->lock);
 920
 921         return alloc_hint;
 922 }
 923
 924 /*
 925  * when extent_io.c finds a delayed allocation range in the file,
 926  * the call backs end up in this code.  The basic idea is to
 927  * allocate extents on disk for the range, and create ordered data structs
 928  * in ram to track those extents.
 929  *
 930  * locked_page is the page that writepage had locked already.  We use
 931  * it to make sure we don't do extra locks or unlocks.
 932  *
 933  * *page_started is set to one if we unlock locked_page and do everything
 934  * required to start IO on it.  It may be clean and already done with
 935  * IO when we return.
 936  */
 937 static noinline int cow_file_range(struct inode *inode,
 938                                    struct page *locked_page,
 939                                    u64 start, u64 end, u64 delalloc_end,
 940                                    int *page_started, unsigned long *nr_written,
 941                                    int unlock, struct btrfs_dedupe_hash *hash)
 942 {
 943         struct btrfs_root *root = BTRFS_I(inode)->root;
 944         u64 alloc_hint = 0;
 945         u64 num_bytes;
 946         unsigned long ram_size;
 947         u64 disk_num_bytes;
 948         u64 cur_alloc_size;
 949         u64 blocksize = root->sectorsize;
 950         struct btrfs_key ins;
 951         struct extent_map *em;
 952         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 953         int ret = 0;
 954
 955         if (btrfs_is_free_space_inode(inode)) {
 956                 WARN_ON_ONCE(1);
 957                 ret = -EINVAL;
 958                 goto out_unlock;
 959         }
 960
 961         num_bytes = ALIGN(end - start + 1, blocksize);
 962         num_bytes = max(blocksize,  num_bytes);
 963         disk_num_bytes = num_bytes;
 964
 965         /* if this is a small write inside eof, kick off defrag */
 966         if (num_bytes < SZ_64K &&
 967             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 968                 btrfs_add_inode_defrag(NULL, inode);
 969
 970         if (start == 0) {
 971                 /* lets try to make an inline extent */
 972                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 973                                             NULL);
 974                 if (ret == 0) {
 975                         extent_clear_unlock_delalloc(inode, start, end,
 976                                      delalloc_end, NULL,
 977                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 978                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 979                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 980                                      PAGE_END_WRITEBACK);
 981                         btrfs_free_reserved_data_space_noquota(inode, start,
 982                                                 end - start + 1);
 983                         *nr_written = *nr_written +
 984                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 985                         *page_started = 1;
 986                         goto out;
 987                 } else if (ret < 0) {
 988                         goto out_unlock;
 989                 }
 990         }
 991
 992         BUG_ON(disk_num_bytes >
 993                btrfs_super_total_bytes(root->fs_info->super_copy));
 994
 995         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 996         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 997
 998         while (disk_num_bytes > 0) {
 999                 unsigned long op;
1000
1001                 cur_alloc_size = disk_num_bytes;
1002                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1003                                            root->sectorsize, 0, alloc_hint,
1004                                            &ins, 1, 1);
1005                 if (ret < 0)
1006                         goto out_unlock;
1007
1008                 em = alloc_extent_map();
1009                 if (!em) {
1010                         ret = -ENOMEM;
1011                         goto out_reserve;
1012                 }
1013                 em->start = start;
1014                 em->orig_start = em->start;
1015                 ram_size = ins.offset;
1016                 em->len = ins.offset;
1017                 em->mod_start = em->start;
1018                 em->mod_len = em->len;
1019
1020                 em->block_start = ins.objectid;
1021                 em->block_len = ins.offset;
1022                 em->orig_block_len = ins.offset;
1023                 em->ram_bytes = ram_size;
1024                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1025                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1026                 em->generation = -1;
1027
1028                 while (1) {
1029                         write_lock(&em_tree->lock);
1030                         ret = add_extent_mapping(em_tree, em, 1);
1031                         write_unlock(&em_tree->lock);
1032                         if (ret != -EEXIST) {
1033                                 free_extent_map(em);
1034                                 break;
1035                         }
1036                         btrfs_drop_extent_cache(inode, start,
1037                                                 start + ram_size - 1, 0);
1038                 }
1039                 if (ret)
1040                         goto out_reserve;
1041
1042                 cur_alloc_size = ins.offset;
1043                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1044                                                ram_size, cur_alloc_size, 0);
1045                 if (ret)
1046                         goto out_drop_extent_cache;
1047
1048                 if (root->root_key.objectid ==
1049                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1050                         ret = btrfs_reloc_clone_csums(inode, start,
1051                                                       cur_alloc_size);
1052                         if (ret)
1053                                 goto out_drop_extent_cache;
1054                 }
1055
1056                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1057
1058                 if (disk_num_bytes < cur_alloc_size)
1059                         break;
1060
1061                 /* we're not doing compressed IO, don't unlock the first
1062                  * page (which the caller expects to stay locked), don't
1063                  * clear any dirty bits and don't set any writeback bits
1064                  *
1065                  * Do set the Private2 bit so we know this page was properly
1066                  * setup for writepage
1067                  */
1068                 op = unlock ? PAGE_UNLOCK : 0;
1069                 op |= PAGE_SET_PRIVATE2;
1070
1071                 extent_clear_unlock_delalloc(inode, start,
1072                                              start + ram_size - 1,
1073                                              delalloc_end, locked_page,
1074                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1075                                              op);
1076                 disk_num_bytes -= cur_alloc_size;
1077                 num_bytes -= cur_alloc_size;
1078                 alloc_hint = ins.objectid + ins.offset;
1079                 start += cur_alloc_size;
1080         }
1081 out:
1082         return ret;
1083
1084 out_drop_extent_cache:
1085         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1086 out_reserve:
1087         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1088         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1089 out_unlock:
1090         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1091                                      locked_page,
1092                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1093                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1094                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1095                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1096         goto out;
1097 }
1098
1099 /*
1100  * work queue call back to started compression on a file and pages
1101  */
1102 static noinline void async_cow_start(struct btrfs_work *work)
1103 {
1104         struct async_cow *async_cow;
1105         int num_added = 0;
1106         async_cow = container_of(work, struct async_cow, work);
1107
1108         compress_file_range(async_cow->inode, async_cow->locked_page,
1109                             async_cow->start, async_cow->end, async_cow,
1110                             &num_added);
1111         if (num_added == 0) {
1112                 btrfs_add_delayed_iput(async_cow->inode);
1113                 async_cow->inode = NULL;
1114         }
1115 }
1116
1117 /*
1118  * work queue call back to submit previously compressed pages
1119  */
1120 static noinline void async_cow_submit(struct btrfs_work *work)
1121 {
1122         struct async_cow *async_cow;
1123         struct btrfs_root *root;
1124         unsigned long nr_pages;
1125
1126         async_cow = container_of(work, struct async_cow, work);
1127
1128         root = async_cow->root;
1129         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1130                 PAGE_SHIFT;
1131
1132         /*
1133          * atomic_sub_return implies a barrier for waitqueue_active
1134          */
1135         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1136             5 * SZ_1M &&
1137             waitqueue_active(&root->fs_info->async_submit_wait))
1138                 wake_up(&root->fs_info->async_submit_wait);
1139
1140         if (async_cow->inode)
1141                 submit_compressed_extents(async_cow->inode, async_cow);
1142 }
1143
1144 static noinline void async_cow_free(struct btrfs_work *work)
1145 {
1146         struct async_cow *async_cow;
1147         async_cow = container_of(work, struct async_cow, work);
1148         if (async_cow->inode)
1149                 btrfs_add_delayed_iput(async_cow->inode);
1150         kfree(async_cow);
1151 }
1152
1153 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1154                                 u64 start, u64 end, int *page_started,
1155                                 unsigned long *nr_written)
1156 {
1157         struct async_cow *async_cow;
1158         struct btrfs_root *root = BTRFS_I(inode)->root;
1159         unsigned long nr_pages;
1160         u64 cur_end;
1161         int limit = 10 * SZ_1M;
1162
1163         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1164                          1, 0, NULL, GFP_NOFS);
1165         while (start < end) {
1166                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1167                 BUG_ON(!async_cow); /* -ENOMEM */
1168                 async_cow->inode = igrab(inode);
1169                 async_cow->root = root;
1170                 async_cow->locked_page = locked_page;
1171                 async_cow->start = start;
1172
1173                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1174                     !btrfs_test_opt(root->fs_info, FORCE_COMPRESS))
1175                         cur_end = end;
1176                 else
1177                         cur_end = min(end, start + SZ_512K - 1);
1178
1179                 async_cow->end = cur_end;
1180                 INIT_LIST_HEAD(&async_cow->extents);
1181
1182                 btrfs_init_work(&async_cow->work,
1183                                 btrfs_delalloc_helper,
1184                                 async_cow_start, async_cow_submit,
1185                                 async_cow_free);
1186
1187                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1188                         PAGE_SHIFT;
1189                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1190
1191                 btrfs_queue_work(root->fs_info->delalloc_workers,
1192                                  &async_cow->work);
1193
1194                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1195                         wait_event(root->fs_info->async_submit_wait,
1196                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1197                             limit));
1198                 }
1199
1200                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1201                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1202                         wait_event(root->fs_info->async_submit_wait,
1203                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1204                            0));
1205                 }
1206
1207                 *nr_written += nr_pages;
1208                 start = cur_end + 1;
1209         }
1210         *page_started = 1;
1211         return 0;
1212 }
1213
1214 static noinline int csum_exist_in_range(struct btrfs_root *root,
1215                                         u64 bytenr, u64 num_bytes)
1216 {
1217         int ret;
1218         struct btrfs_ordered_sum *sums;
1219         LIST_HEAD(list);
1220
1221         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1222                                        bytenr + num_bytes - 1, &list, 0);
1223         if (ret == 0 && list_empty(&list))
1224                 return 0;
1225
1226         while (!list_empty(&list)) {
1227                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1228                 list_del(&sums->list);
1229                 kfree(sums);
1230         }
1231         return 1;
1232 }
1233
1234 /*
1235  * when nowcow writeback call back.  This checks for snapshots or COW copies
1236  * of the extents that exist in the file, and COWs the file as required.
1237  *
1238  * If no cow copies or snapshots exist, we write directly to the existing
1239  * blocks on disk
1240  */
1241 static noinline int run_delalloc_nocow(struct inode *inode,
1242                                        struct page *locked_page,
1243                               u64 start, u64 end, int *page_started, int force,
1244                               unsigned long *nr_written)
1245 {
1246         struct btrfs_root *root = BTRFS_I(inode)->root;
1247         struct btrfs_trans_handle *trans;
1248         struct extent_buffer *leaf;
1249         struct btrfs_path *path;
1250         struct btrfs_file_extent_item *fi;
1251         struct btrfs_key found_key;
1252         u64 cow_start;
1253         u64 cur_offset;
1254         u64 extent_end;
1255         u64 extent_offset;
1256         u64 disk_bytenr;
1257         u64 num_bytes;
1258         u64 disk_num_bytes;
1259         u64 ram_bytes;
1260         int extent_type;
1261         int ret, err;
1262         int type;
1263         int nocow;
1264         int check_prev = 1;
1265         bool nolock;
1266         u64 ino = btrfs_ino(inode);
1267
1268         path = btrfs_alloc_path();
1269         if (!path) {
1270                 extent_clear_unlock_delalloc(inode, start, end, end,
1271                                              locked_page,
1272                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1273                                              EXTENT_DO_ACCOUNTING |
1274                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1275                                              PAGE_CLEAR_DIRTY |
1276                                              PAGE_SET_WRITEBACK |
1277                                              PAGE_END_WRITEBACK);
1278                 return -ENOMEM;
1279         }
1280
1281         nolock = btrfs_is_free_space_inode(inode);
1282
1283         if (nolock)
1284                 trans = btrfs_join_transaction_nolock(root);
1285         else
1286                 trans = btrfs_join_transaction(root);
1287
1288         if (IS_ERR(trans)) {
1289                 extent_clear_unlock_delalloc(inode, start, end, end,
1290                                              locked_page,
1291                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1292                                              EXTENT_DO_ACCOUNTING |
1293                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1294                                              PAGE_CLEAR_DIRTY |
1295                                              PAGE_SET_WRITEBACK |
1296                                              PAGE_END_WRITEBACK);
1297                 btrfs_free_path(path);
1298                 return PTR_ERR(trans);
1299         }
1300
1301         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1302
1303         cow_start = (u64)-1;
1304         cur_offset = start;
1305         while (1) {
1306                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1307                                                cur_offset, 0);
1308                 if (ret < 0)
1309                         goto error;
1310                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1311                         leaf = path->nodes[0];
1312                         btrfs_item_key_to_cpu(leaf, &found_key,
1313                                               path->slots[0] - 1);
1314                         if (found_key.objectid == ino &&
1315                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1316                                 path->slots[0]--;
1317                 }
1318                 check_prev = 0;
1319 next_slot:
1320                 leaf = path->nodes[0];
1321                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1322                         ret = btrfs_next_leaf(root, path);
1323                         if (ret < 0)
1324                                 goto error;
1325                         if (ret > 0)
1326                                 break;
1327                         leaf = path->nodes[0];
1328                 }
1329
1330                 nocow = 0;
1331                 disk_bytenr = 0;
1332                 num_bytes = 0;
1333                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1334
1335                 if (found_key.objectid > ino)
1336                         break;
1337                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1338                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1339                         path->slots[0]++;
1340                         goto next_slot;
1341                 }
1342                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1343                     found_key.offset > end)
1344                         break;
1345
1346                 if (found_key.offset > cur_offset) {
1347                         extent_end = found_key.offset;
1348                         extent_type = 0;
1349                         goto out_check;
1350                 }
1351
1352                 fi = btrfs_item_ptr(leaf, path->slots[0],
1353                                     struct btrfs_file_extent_item);
1354                 extent_type = btrfs_file_extent_type(leaf, fi);
1355
1356                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1357                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1358                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1359                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1360                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1361                         extent_end = found_key.offset +
1362                                 btrfs_file_extent_num_bytes(leaf, fi);
1363                         disk_num_bytes =
1364                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1365                         if (extent_end <= start) {
1366                                 path->slots[0]++;
1367                                 goto next_slot;
1368                         }
1369                         if (disk_bytenr == 0)
1370                                 goto out_check;
1371                         if (btrfs_file_extent_compression(leaf, fi) ||
1372                             btrfs_file_extent_encryption(leaf, fi) ||
1373                             btrfs_file_extent_other_encoding(leaf, fi))
1374                                 goto out_check;
1375                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1376                                 goto out_check;
1377                         if (btrfs_extent_readonly(root, disk_bytenr))
1378                                 goto out_check;
1379                         if (btrfs_cross_ref_exist(trans, root, ino,
1380                                                   found_key.offset -
1381                                                   extent_offset, disk_bytenr))
1382                                 goto out_check;
1383                         disk_bytenr += extent_offset;
1384                         disk_bytenr += cur_offset - found_key.offset;
1385                         num_bytes = min(end + 1, extent_end) - cur_offset;
1386                         /*
1387                          * if there are pending snapshots for this root,
1388                          * we fall into common COW way.
1389                          */
1390                         if (!nolock) {
1391                                 err = btrfs_start_write_no_snapshoting(root);
1392                                 if (!err)
1393                                         goto out_check;
1394                         }
1395                         /*
1396                          * force cow if csum exists in the range.
1397                          * this ensure that csum for a given extent are
1398                          * either valid or do not exist.
1399                          */
1400                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1401                                 goto out_check;
1402                         if (!btrfs_inc_nocow_writers(root->fs_info,
1403                                                      disk_bytenr))
1404                                 goto out_check;
1405                         nocow = 1;
1406                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1407                         extent_end = found_key.offset +
1408                                 btrfs_file_extent_inline_len(leaf,
1409                                                      path->slots[0], fi);
1410                         extent_end = ALIGN(extent_end, root->sectorsize);
1411                 } else {
1412                         BUG_ON(1);
1413                 }
1414 out_check:
1415                 if (extent_end <= start) {
1416                         path->slots[0]++;
1417                         if (!nolock && nocow)
1418                                 btrfs_end_write_no_snapshoting(root);
1419                         if (nocow)
1420                                 btrfs_dec_nocow_writers(root->fs_info,
1421                                                         disk_bytenr);
1422                         goto next_slot;
1423                 }
1424                 if (!nocow) {
1425                         if (cow_start == (u64)-1)
1426                                 cow_start = cur_offset;
1427                         cur_offset = extent_end;
1428                         if (cur_offset > end)
1429                                 break;
1430                         path->slots[0]++;
1431                         goto next_slot;
1432                 }
1433
1434                 btrfs_release_path(path);
1435                 if (cow_start != (u64)-1) {
1436                         ret = cow_file_range(inode, locked_page,
1437                                              cow_start, found_key.offset - 1,
1438                                              end, page_started, nr_written, 1,
1439                                              NULL);
1440                         if (ret) {
1441                                 if (!nolock && nocow)
1442                                         btrfs_end_write_no_snapshoting(root);
1443                                 if (nocow)
1444                                         btrfs_dec_nocow_writers(root->fs_info,
1445                                                                 disk_bytenr);
1446                                 goto error;
1447                         }
1448                         cow_start = (u64)-1;
1449                 }
1450
1451                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1452                         struct extent_map *em;
1453                         struct extent_map_tree *em_tree;
1454                         em_tree = &BTRFS_I(inode)->extent_tree;
1455                         em = alloc_extent_map();
1456                         BUG_ON(!em); /* -ENOMEM */
1457                         em->start = cur_offset;
1458                         em->orig_start = found_key.offset - extent_offset;
1459                         em->len = num_bytes;
1460                         em->block_len = num_bytes;
1461                         em->block_start = disk_bytenr;
1462                         em->orig_block_len = disk_num_bytes;
1463                         em->ram_bytes = ram_bytes;
1464                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1465                         em->mod_start = em->start;
1466                         em->mod_len = em->len;
1467                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1468                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1469                         em->generation = -1;
1470                         while (1) {
1471                                 write_lock(&em_tree->lock);
1472                                 ret = add_extent_mapping(em_tree, em, 1);
1473                                 write_unlock(&em_tree->lock);
1474                                 if (ret != -EEXIST) {
1475                                         free_extent_map(em);
1476                                         break;
1477                                 }
1478                                 btrfs_drop_extent_cache(inode, em->start,
1479                                                 em->start + em->len - 1, 0);
1480                         }
1481                         type = BTRFS_ORDERED_PREALLOC;
1482                 } else {
1483                         type = BTRFS_ORDERED_NOCOW;
1484                 }
1485
1486                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1487                                                num_bytes, num_bytes, type);
1488                 if (nocow)
1489                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1490                 BUG_ON(ret); /* -ENOMEM */
1491
1492                 if (root->root_key.objectid ==
1493                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1494                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1495                                                       num_bytes);
1496                         if (ret) {
1497                                 if (!nolock && nocow)
1498                                         btrfs_end_write_no_snapshoting(root);
1499                                 goto error;
1500                         }
1501                 }
1502
1503                 extent_clear_unlock_delalloc(inode, cur_offset,
1504                                              cur_offset + num_bytes - 1, end,
1505                                              locked_page, EXTENT_LOCKED |
1506                                              EXTENT_DELALLOC |
1507                                              EXTENT_CLEAR_DATA_RESV,
1508                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1509
1510                 if (!nolock && nocow)
1511                         btrfs_end_write_no_snapshoting(root);
1512                 cur_offset = extent_end;
1513                 if (cur_offset > end)
1514                         break;
1515         }
1516         btrfs_release_path(path);
1517
1518         if (cur_offset <= end && cow_start == (u64)-1) {
1519                 cow_start = cur_offset;
1520                 cur_offset = end;
1521         }
1522
1523         if (cow_start != (u64)-1) {
1524                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1525                                      page_started, nr_written, 1, NULL);
1526                 if (ret)
1527                         goto error;
1528         }
1529
1530 error:
1531         err = btrfs_end_transaction(trans, root);
1532         if (!ret)
1533                 ret = err;
1534
1535         if (ret && cur_offset < end)
1536                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1537                                              locked_page, EXTENT_LOCKED |
1538                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1539                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1540                                              PAGE_CLEAR_DIRTY |
1541                                              PAGE_SET_WRITEBACK |
1542                                              PAGE_END_WRITEBACK);
1543         btrfs_free_path(path);
1544         return ret;
1545 }
1546
1547 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1548 {
1549
1550         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1551             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1552                 return 0;
1553
1554         /*
1555          * @defrag_bytes is a hint value, no spinlock held here,
1556          * if is not zero, it means the file is defragging.
1557          * Force cow if given extent needs to be defragged.
1558          */
1559         if (BTRFS_I(inode)->defrag_bytes &&
1560             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1561                            EXTENT_DEFRAG, 0, NULL))
1562                 return 1;
1563
1564         return 0;
1565 }
1566
1567 /*
1568  * extent_io.c call back to do delayed allocation processing
1569  */
1570 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1571                               u64 start, u64 end, int *page_started,
1572                               unsigned long *nr_written)
1573 {
1574         int ret;
1575         int force_cow = need_force_cow(inode, start, end);
1576
1577         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1578                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1579                                          page_started, 1, nr_written);
1580         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1581                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1582                                          page_started, 0, nr_written);
1583         } else if (!inode_need_compress(inode)) {
1584                 ret = cow_file_range(inode, locked_page, start, end, end,
1585                                       page_started, nr_written, 1, NULL);
1586         } else {
1587                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1588                         &BTRFS_I(inode)->runtime_flags);
1589                 ret = cow_file_range_async(inode, locked_page, start, end,
1590                                            page_started, nr_written);
1591         }
1592         return ret;
1593 }
1594
1595 static void btrfs_split_extent_hook(struct inode *inode,
1596                                     struct extent_state *orig, u64 split)
1597 {
1598         u64 size;
1599
1600         /* not delalloc, ignore it */
1601         if (!(orig->state & EXTENT_DELALLOC))
1602                 return;
1603
1604         size = orig->end - orig->start + 1;
1605         if (size > BTRFS_MAX_EXTENT_SIZE) {
1606                 u64 num_extents;
1607                 u64 new_size;
1608
1609                 /*
1610                  * See the explanation in btrfs_merge_extent_hook, the same
1611                  * applies here, just in reverse.
1612                  */
1613                 new_size = orig->end - split + 1;
1614                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1615                                         BTRFS_MAX_EXTENT_SIZE);
1616                 new_size = split - orig->start;
1617                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1618                                         BTRFS_MAX_EXTENT_SIZE);
1619                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1620                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1621                         return;
1622         }
1623
1624         spin_lock(&BTRFS_I(inode)->lock);
1625         BTRFS_I(inode)->outstanding_extents++;
1626         spin_unlock(&BTRFS_I(inode)->lock);
1627 }
1628
1629 /*
1630  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1631  * extents so we can keep track of new extents that are just merged onto old
1632  * extents, such as when we are doing sequential writes, so we can properly
1633  * account for the metadata space we'll need.
1634  */
1635 static void btrfs_merge_extent_hook(struct inode *inode,
1636                                     struct extent_state *new,
1637                                     struct extent_state *other)
1638 {
1639         u64 new_size, old_size;
1640         u64 num_extents;
1641
1642         /* not delalloc, ignore it */
1643         if (!(other->state & EXTENT_DELALLOC))
1644                 return;
1645
1646         if (new->start > other->start)
1647                 new_size = new->end - other->start + 1;
1648         else
1649                 new_size = other->end - new->start + 1;
1650
1651         /* we're not bigger than the max, unreserve the space and go */
1652         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1653                 spin_lock(&BTRFS_I(inode)->lock);
1654                 BTRFS_I(inode)->outstanding_extents--;
1655                 spin_unlock(&BTRFS_I(inode)->lock);
1656                 return;
1657         }
1658
1659         /*
1660          * We have to add up either side to figure out how many extents were
1661          * accounted for before we merged into one big extent.  If the number of
1662          * extents we accounted for is <= the amount we need for the new range
1663          * then we can return, otherwise drop.  Think of it like this
1664          *
1665          * [ 4k][MAX_SIZE]
1666          *
1667          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1668          * need 2 outstanding extents, on one side we have 1 and the other side
1669          * we have 1 so they are == and we can return.  But in this case
1670          *
1671          * [MAX_SIZE+4k][MAX_SIZE+4k]
1672          *
1673          * Each range on their own accounts for 2 extents, but merged together
1674          * they are only 3 extents worth of accounting, so we need to drop in
1675          * this case.
1676          */
1677         old_size = other->end - other->start + 1;
1678         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1679                                 BTRFS_MAX_EXTENT_SIZE);
1680         old_size = new->end - new->start + 1;
1681         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1682                                  BTRFS_MAX_EXTENT_SIZE);
1683
1684         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1685                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1686                 return;
1687
1688         spin_lock(&BTRFS_I(inode)->lock);
1689         BTRFS_I(inode)->outstanding_extents--;
1690         spin_unlock(&BTRFS_I(inode)->lock);
1691 }
1692
1693 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1694                                       struct inode *inode)
1695 {
1696         spin_lock(&root->delalloc_lock);
1697         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1698                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1699                               &root->delalloc_inodes);
1700                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                         &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes++;
1703                 if (root->nr_delalloc_inodes == 1) {
1704                         spin_lock(&root->fs_info->delalloc_root_lock);
1705                         BUG_ON(!list_empty(&root->delalloc_root));
1706                         list_add_tail(&root->delalloc_root,
1707                                       &root->fs_info->delalloc_roots);
1708                         spin_unlock(&root->fs_info->delalloc_root_lock);
1709                 }
1710         }
1711         spin_unlock(&root->delalloc_lock);
1712 }
1713
1714 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1715                                      struct inode *inode)
1716 {
1717         spin_lock(&root->delalloc_lock);
1718         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1719                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1720                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1721                           &BTRFS_I(inode)->runtime_flags);
1722                 root->nr_delalloc_inodes--;
1723                 if (!root->nr_delalloc_inodes) {
1724                         spin_lock(&root->fs_info->delalloc_root_lock);
1725                         BUG_ON(list_empty(&root->delalloc_root));
1726                         list_del_init(&root->delalloc_root);
1727                         spin_unlock(&root->fs_info->delalloc_root_lock);
1728                 }
1729         }
1730         spin_unlock(&root->delalloc_lock);
1731 }
1732
1733 /*
1734  * extent_io.c set_bit_hook, used to track delayed allocation
1735  * bytes in this file, and to maintain the list of inodes that
1736  * have pending delalloc work to be done.
1737  */
1738 static void btrfs_set_bit_hook(struct inode *inode,
1739                                struct extent_state *state, unsigned *bits)
1740 {
1741
1742         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1743                 WARN_ON(1);
1744         /*
1745          * set_bit and clear bit hooks normally require _irqsave/restore
1746          * but in this case, we are only testing for the DELALLOC
1747          * bit, which is only set or cleared with irqs on
1748          */
1749         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1750                 struct btrfs_root *root = BTRFS_I(inode)->root;
1751                 u64 len = state->end + 1 - state->start;
1752                 bool do_list = !btrfs_is_free_space_inode(inode);
1753
1754                 if (*bits & EXTENT_FIRST_DELALLOC) {
1755                         *bits &= ~EXTENT_FIRST_DELALLOC;
1756                 } else {
1757                         spin_lock(&BTRFS_I(inode)->lock);
1758                         BTRFS_I(inode)->outstanding_extents++;
1759                         spin_unlock(&BTRFS_I(inode)->lock);
1760                 }
1761
1762                 /* For sanity tests */
1763                 if (btrfs_is_testing(root->fs_info))
1764                         return;
1765
1766                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1767                                      root->fs_info->delalloc_batch);
1768                 spin_lock(&BTRFS_I(inode)->lock);
1769                 BTRFS_I(inode)->delalloc_bytes += len;
1770                 if (*bits & EXTENT_DEFRAG)
1771                         BTRFS_I(inode)->defrag_bytes += len;
1772                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1773                                          &BTRFS_I(inode)->runtime_flags))
1774                         btrfs_add_delalloc_inodes(root, inode);
1775                 spin_unlock(&BTRFS_I(inode)->lock);
1776         }
1777 }
1778
1779 /*
1780  * extent_io.c clear_bit_hook, see set_bit_hook for why
1781  */
1782 static void btrfs_clear_bit_hook(struct inode *inode,
1783                                  struct extent_state *state,
1784                                  unsigned *bits)
1785 {
1786         u64 len = state->end + 1 - state->start;
1787         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1788                                     BTRFS_MAX_EXTENT_SIZE);
1789
1790         spin_lock(&BTRFS_I(inode)->lock);
1791         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1792                 BTRFS_I(inode)->defrag_bytes -= len;
1793         spin_unlock(&BTRFS_I(inode)->lock);
1794
1795         /*
1796          * set_bit and clear bit hooks normally require _irqsave/restore
1797          * but in this case, we are only testing for the DELALLOC
1798          * bit, which is only set or cleared with irqs on
1799          */
1800         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1801                 struct btrfs_root *root = BTRFS_I(inode)->root;
1802                 bool do_list = !btrfs_is_free_space_inode(inode);
1803
1804                 if (*bits & EXTENT_FIRST_DELALLOC) {
1805                         *bits &= ~EXTENT_FIRST_DELALLOC;
1806                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1807                         spin_lock(&BTRFS_I(inode)->lock);
1808                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1809                         spin_unlock(&BTRFS_I(inode)->lock);
1810                 }
1811
1812                 /*
1813                  * We don't reserve metadata space for space cache inodes so we
1814                  * don't need to call dellalloc_release_metadata if there is an
1815                  * error.
1816                  */
1817                 if (*bits & EXTENT_DO_ACCOUNTING &&
1818                     root != root->fs_info->tree_root)
1819                         btrfs_delalloc_release_metadata(inode, len);
1820
1821                 /* For sanity tests. */
1822                 if (btrfs_is_testing(root->fs_info))
1823                         return;
1824
1825                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1826                     && do_list && !(state->state & EXTENT_NORESERVE)
1827                     && (*bits & (EXTENT_DO_ACCOUNTING |
1828                     EXTENT_CLEAR_DATA_RESV)))
1829                         btrfs_free_reserved_data_space_noquota(inode,
1830                                         state->start, len);
1831
1832                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1833                                      root->fs_info->delalloc_batch);
1834                 spin_lock(&BTRFS_I(inode)->lock);
1835                 BTRFS_I(inode)->delalloc_bytes -= len;
1836                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1837                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1838                              &BTRFS_I(inode)->runtime_flags))
1839                         btrfs_del_delalloc_inode(root, inode);
1840                 spin_unlock(&BTRFS_I(inode)->lock);
1841         }
1842 }
1843
1844 /*
1845  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1846  * we don't create bios that span stripes or chunks
1847  *
1848  * return 1 if page cannot be merged to bio
1849  * return 0 if page can be merged to bio
1850  * return error otherwise
1851  */
1852 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1853                          size_t size, struct bio *bio,
1854                          unsigned long bio_flags)
1855 {
1856         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1857         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1858         u64 length = 0;
1859         u64 map_length;
1860         int ret;
1861
1862         if (bio_flags & EXTENT_BIO_COMPRESSED)
1863                 return 0;
1864
1865         length = bio->bi_iter.bi_size;
1866         map_length = length;
1867         ret = btrfs_map_block(root->fs_info, bio_op(bio), logical,
1868                               &map_length, NULL, 0);
1869         if (ret < 0)
1870                 return ret;
1871         if (map_length < length + size)
1872                 return 1;
1873         return 0;
1874 }
1875
1876 /*
1877  * in order to insert checksums into the metadata in large chunks,
1878  * we wait until bio submission time.   All the pages in the bio are
1879  * checksummed and sums are attached onto the ordered extent record.
1880  *
1881  * At IO completion time the cums attached on the ordered extent record
1882  * are inserted into the btree
1883  */
1884 static int __btrfs_submit_bio_start(struct inode *inode, struct bio *bio,
1885                                     int mirror_num, unsigned long bio_flags,
1886                                     u64 bio_offset)
1887 {
1888         struct btrfs_root *root = BTRFS_I(inode)->root;
1889         int ret = 0;
1890
1891         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1892         BUG_ON(ret); /* -ENOMEM */
1893         return 0;
1894 }
1895
1896 /*
1897  * in order to insert checksums into the metadata in large chunks,
1898  * we wait until bio submission time.   All the pages in the bio are
1899  * checksummed and sums are attached onto the ordered extent record.
1900  *
1901  * At IO completion time the cums attached on the ordered extent record
1902  * are inserted into the btree
1903  */
1904 static int __btrfs_submit_bio_done(struct inode *inode, struct bio *bio,
1905                           int mirror_num, unsigned long bio_flags,
1906                           u64 bio_offset)
1907 {
1908         struct btrfs_root *root = BTRFS_I(inode)->root;
1909         int ret;
1910
1911         ret = btrfs_map_bio(root, bio, mirror_num, 1);
1912         if (ret) {
1913                 bio->bi_error = ret;
1914                 bio_endio(bio);
1915         }
1916         return ret;
1917 }
1918
1919 /*
1920  * extent_io.c submission hook. This does the right thing for csum calculation
1921  * on write, or reading the csums from the tree before a read
1922  */
1923 static int btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
1924                           int mirror_num, unsigned long bio_flags,
1925                           u64 bio_offset)
1926 {
1927         struct btrfs_root *root = BTRFS_I(inode)->root;
1928         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1929         int ret = 0;
1930         int skip_sum;
1931         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1932
1933         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1934
1935         if (btrfs_is_free_space_inode(inode))
1936                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1937
1938         if (bio_op(bio) != REQ_OP_WRITE) {
1939                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1940                 if (ret)
1941                         goto out;
1942
1943                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1944                         ret = btrfs_submit_compressed_read(inode, bio,
1945                                                            mirror_num,
1946                                                            bio_flags);
1947                         goto out;
1948                 } else if (!skip_sum) {
1949                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1950                         if (ret)
1951                                 goto out;
1952                 }
1953                 goto mapit;
1954         } else if (async && !skip_sum) {
1955                 /* csum items have already been cloned */
1956                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1957                         goto mapit;
1958                 /* we're doing a write, do the async checksumming */
1959                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1960                                    inode, bio, mirror_num,
1961                                    bio_flags, bio_offset,
1962                                    __btrfs_submit_bio_start,
1963                                    __btrfs_submit_bio_done);
1964                 goto out;
1965         } else if (!skip_sum) {
1966                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1967                 if (ret)
1968                         goto out;
1969         }
1970
1971 mapit:
1972         ret = btrfs_map_bio(root, bio, mirror_num, 0);
1973
1974 out:
1975         if (ret < 0) {
1976                 bio->bi_error = ret;
1977                 bio_endio(bio);
1978         }
1979         return ret;
1980 }
1981
1982 /*
1983  * given a list of ordered sums record them in the inode.  This happens
1984  * at IO completion time based on sums calculated at bio submission time.
1985  */
1986 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1987                              struct inode *inode, u64 file_offset,
1988                              struct list_head *list)
1989 {
1990         struct btrfs_ordered_sum *sum;
1991
1992         list_for_each_entry(sum, list, list) {
1993                 trans->adding_csums = 1;
1994                 btrfs_csum_file_blocks(trans,
1995                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1996                 trans->adding_csums = 0;
1997         }
1998         return 0;
1999 }
2000
2001 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2002                               struct extent_state **cached_state, int dedupe)
2003 {
2004         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2005         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2006                                    cached_state);
2007 }
2008
2009 /* see btrfs_writepage_start_hook for details on why this is required */
2010 struct btrfs_writepage_fixup {
2011         struct page *page;
2012         struct btrfs_work work;
2013 };
2014
2015 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2016 {
2017         struct btrfs_writepage_fixup *fixup;
2018         struct btrfs_ordered_extent *ordered;
2019         struct extent_state *cached_state = NULL;
2020         struct page *page;
2021         struct inode *inode;
2022         u64 page_start;
2023         u64 page_end;
2024         int ret;
2025
2026         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2027         page = fixup->page;
2028 again:
2029         lock_page(page);
2030         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2031                 ClearPageChecked(page);
2032                 goto out_page;
2033         }
2034
2035         inode = page->mapping->host;
2036         page_start = page_offset(page);
2037         page_end = page_offset(page) + PAGE_SIZE - 1;
2038
2039         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2040                          &cached_state);
2041
2042         /* already ordered? We're done */
2043         if (PagePrivate2(page))
2044                 goto out;
2045
2046         ordered = btrfs_lookup_ordered_range(inode, page_start,
2047                                         PAGE_SIZE);
2048         if (ordered) {
2049                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2050                                      page_end, &cached_state, GFP_NOFS);
2051                 unlock_page(page);
2052                 btrfs_start_ordered_extent(inode, ordered, 1);
2053                 btrfs_put_ordered_extent(ordered);
2054                 goto again;
2055         }
2056
2057         ret = btrfs_delalloc_reserve_space(inode, page_start,
2058                                            PAGE_SIZE);
2059         if (ret) {
2060                 mapping_set_error(page->mapping, ret);
2061                 end_extent_writepage(page, ret, page_start, page_end);
2062                 ClearPageChecked(page);
2063                 goto out;
2064          }
2065
2066         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state,
2067                                   0);
2068         ClearPageChecked(page);
2069         set_page_dirty(page);
2070 out:
2071         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2072                              &cached_state, GFP_NOFS);
2073 out_page:
2074         unlock_page(page);
2075         put_page(page);
2076         kfree(fixup);
2077 }
2078
2079 /*
2080  * There are a few paths in the higher layers of the kernel that directly
2081  * set the page dirty bit without asking the filesystem if it is a
2082  * good idea.  This causes problems because we want to make sure COW
2083  * properly happens and the data=ordered rules are followed.
2084  *
2085  * In our case any range that doesn't have the ORDERED bit set
2086  * hasn't been properly setup for IO.  We kick off an async process
2087  * to fix it up.  The async helper will wait for ordered extents, set
2088  * the delalloc bit and make it safe to write the page.
2089  */
2090 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2091 {
2092         struct inode *inode = page->mapping->host;
2093         struct btrfs_writepage_fixup *fixup;
2094         struct btrfs_root *root = BTRFS_I(inode)->root;
2095
2096         /* this page is properly in the ordered list */
2097         if (TestClearPagePrivate2(page))
2098                 return 0;
2099
2100         if (PageChecked(page))
2101                 return -EAGAIN;
2102
2103         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2104         if (!fixup)
2105                 return -EAGAIN;
2106
2107         SetPageChecked(page);
2108         get_page(page);
2109         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2110                         btrfs_writepage_fixup_worker, NULL, NULL);
2111         fixup->page = page;
2112         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2113         return -EBUSY;
2114 }
2115
2116 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2117                                        struct inode *inode, u64 file_pos,
2118                                        u64 disk_bytenr, u64 disk_num_bytes,
2119                                        u64 num_bytes, u64 ram_bytes,
2120                                        u8 compression, u8 encryption,
2121                                        u16 other_encoding, int extent_type)
2122 {
2123         struct btrfs_root *root = BTRFS_I(inode)->root;
2124         struct btrfs_file_extent_item *fi;
2125         struct btrfs_path *path;
2126         struct extent_buffer *leaf;
2127         struct btrfs_key ins;
2128         int extent_inserted = 0;
2129         int ret;
2130
2131         path = btrfs_alloc_path();
2132         if (!path)
2133                 return -ENOMEM;
2134
2135         /*
2136          * we may be replacing one extent in the tree with another.
2137          * The new extent is pinned in the extent map, and we don't want
2138          * to drop it from the cache until it is completely in the btree.
2139          *
2140          * So, tell btrfs_drop_extents to leave this extent in the cache.
2141          * the caller is expected to unpin it and allow it to be merged
2142          * with the others.
2143          */
2144         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2145                                    file_pos + num_bytes, NULL, 0,
2146                                    1, sizeof(*fi), &extent_inserted);
2147         if (ret)
2148                 goto out;
2149
2150         if (!extent_inserted) {
2151                 ins.objectid = btrfs_ino(inode);
2152                 ins.offset = file_pos;
2153                 ins.type = BTRFS_EXTENT_DATA_KEY;
2154
2155                 path->leave_spinning = 1;
2156                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2157                                               sizeof(*fi));
2158                 if (ret)
2159                         goto out;
2160         }
2161         leaf = path->nodes[0];
2162         fi = btrfs_item_ptr(leaf, path->slots[0],
2163                             struct btrfs_file_extent_item);
2164         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2165         btrfs_set_file_extent_type(leaf, fi, extent_type);
2166         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2167         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2168         btrfs_set_file_extent_offset(leaf, fi, 0);
2169         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2170         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2171         btrfs_set_file_extent_compression(leaf, fi, compression);
2172         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2173         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2174
2175         btrfs_mark_buffer_dirty(leaf);
2176         btrfs_release_path(path);
2177
2178         inode_add_bytes(inode, num_bytes);
2179
2180         ins.objectid = disk_bytenr;
2181         ins.offset = disk_num_bytes;
2182         ins.type = BTRFS_EXTENT_ITEM_KEY;
2183         ret = btrfs_alloc_reserved_file_extent(trans, root,
2184                                         root->root_key.objectid,
2185                                         btrfs_ino(inode), file_pos,
2186                                         ram_bytes, &ins);
2187         /*
2188          * Release the reserved range from inode dirty range map, as it is
2189          * already moved into delayed_ref_head
2190          */
2191         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2192 out:
2193         btrfs_free_path(path);
2194
2195         return ret;
2196 }
2197
2198 /* snapshot-aware defrag */
2199 struct sa_defrag_extent_backref {
2200         struct rb_node node;
2201         struct old_sa_defrag_extent *old;
2202         u64 root_id;
2203         u64 inum;
2204         u64 file_pos;
2205         u64 extent_offset;
2206         u64 num_bytes;
2207         u64 generation;
2208 };
2209
2210 struct old_sa_defrag_extent {
2211         struct list_head list;
2212         struct new_sa_defrag_extent *new;
2213
2214         u64 extent_offset;
2215         u64 bytenr;
2216         u64 offset;
2217         u64 len;
2218         int count;
2219 };
2220
2221 struct new_sa_defrag_extent {
2222         struct rb_root root;
2223         struct list_head head;
2224         struct btrfs_path *path;
2225         struct inode *inode;
2226         u64 file_pos;
2227         u64 len;
2228         u64 bytenr;
2229         u64 disk_len;
2230         u8 compress_type;
2231 };
2232
2233 static int backref_comp(struct sa_defrag_extent_backref *b1,
2234                         struct sa_defrag_extent_backref *b2)
2235 {
2236         if (b1->root_id < b2->root_id)
2237                 return -1;
2238         else if (b1->root_id > b2->root_id)
2239                 return 1;
2240
2241         if (b1->inum < b2->inum)
2242                 return -1;
2243         else if (b1->inum > b2->inum)
2244                 return 1;
2245
2246         if (b1->file_pos < b2->file_pos)
2247                 return -1;
2248         else if (b1->file_pos > b2->file_pos)
2249                 return 1;
2250
2251         /*
2252          * [------------------------------] ===> (a range of space)
2253          *     |<--->|   |<---->| =============> (fs/file tree A)
2254          * |<---------------------------->| ===> (fs/file tree B)
2255          *
2256          * A range of space can refer to two file extents in one tree while
2257          * refer to only one file extent in another tree.
2258          *
2259          * So we may process a disk offset more than one time(two extents in A)
2260          * and locate at the same extent(one extent in B), then insert two same
2261          * backrefs(both refer to the extent in B).
2262          */
2263         return 0;
2264 }
2265
2266 static void backref_insert(struct rb_root *root,
2267                            struct sa_defrag_extent_backref *backref)
2268 {
2269         struct rb_node **p = &root->rb_node;
2270         struct rb_node *parent = NULL;
2271         struct sa_defrag_extent_backref *entry;
2272         int ret;
2273
2274         while (*p) {
2275                 parent = *p;
2276                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2277
2278                 ret = backref_comp(backref, entry);
2279                 if (ret < 0)
2280                         p = &(*p)->rb_left;
2281                 else
2282                         p = &(*p)->rb_right;
2283         }
2284
2285         rb_link_node(&backref->node, parent, p);
2286         rb_insert_color(&backref->node, root);
2287 }
2288
2289 /*
2290  * Note the backref might has changed, and in this case we just return 0.
2291  */
2292 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2293                                        void *ctx)
2294 {
2295         struct btrfs_file_extent_item *extent;
2296         struct btrfs_fs_info *fs_info;
2297         struct old_sa_defrag_extent *old = ctx;
2298         struct new_sa_defrag_extent *new = old->new;
2299         struct btrfs_path *path = new->path;
2300         struct btrfs_key key;
2301         struct btrfs_root *root;
2302         struct sa_defrag_extent_backref *backref;
2303         struct extent_buffer *leaf;
2304         struct inode *inode = new->inode;
2305         int slot;
2306         int ret;
2307         u64 extent_offset;
2308         u64 num_bytes;
2309
2310         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2311             inum == btrfs_ino(inode))
2312                 return 0;
2313
2314         key.objectid = root_id;
2315         key.type = BTRFS_ROOT_ITEM_KEY;
2316         key.offset = (u64)-1;
2317
2318         fs_info = BTRFS_I(inode)->root->fs_info;
2319         root = btrfs_read_fs_root_no_name(fs_info, &key);
2320         if (IS_ERR(root)) {
2321                 if (PTR_ERR(root) == -ENOENT)
2322                         return 0;
2323                 WARN_ON(1);
2324                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2325                          inum, offset, root_id);
2326                 return PTR_ERR(root);
2327         }
2328
2329         key.objectid = inum;
2330         key.type = BTRFS_EXTENT_DATA_KEY;
2331         if (offset > (u64)-1 << 32)
2332                 key.offset = 0;
2333         else
2334                 key.offset = offset;
2335
2336         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2337         if (WARN_ON(ret < 0))
2338                 return ret;
2339         ret = 0;
2340
2341         while (1) {
2342                 cond_resched();
2343
2344                 leaf = path->nodes[0];
2345                 slot = path->slots[0];
2346
2347                 if (slot >= btrfs_header_nritems(leaf)) {
2348                         ret = btrfs_next_leaf(root, path);
2349                         if (ret < 0) {
2350                                 goto out;
2351                         } else if (ret > 0) {
2352                                 ret = 0;
2353                                 goto out;
2354                         }
2355                         continue;
2356                 }
2357
2358                 path->slots[0]++;
2359
2360                 btrfs_item_key_to_cpu(leaf, &key, slot);
2361
2362                 if (key.objectid > inum)
2363                         goto out;
2364
2365                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2366                         continue;
2367
2368                 extent = btrfs_item_ptr(leaf, slot,
2369                                         struct btrfs_file_extent_item);
2370
2371                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2372                         continue;
2373
2374                 /*
2375                  * 'offset' refers to the exact key.offset,
2376                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2377                  * (key.offset - extent_offset).
2378                  */
2379                 if (key.offset != offset)
2380                         continue;
2381
2382                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2383                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2384
2385                 if (extent_offset >= old->extent_offset + old->offset +
2386                     old->len || extent_offset + num_bytes <=
2387                     old->extent_offset + old->offset)
2388                         continue;
2389                 break;
2390         }
2391
2392         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2393         if (!backref) {
2394                 ret = -ENOENT;
2395                 goto out;
2396         }
2397
2398         backref->root_id = root_id;
2399         backref->inum = inum;
2400         backref->file_pos = offset;
2401         backref->num_bytes = num_bytes;
2402         backref->extent_offset = extent_offset;
2403         backref->generation = btrfs_file_extent_generation(leaf, extent);
2404         backref->old = old;
2405         backref_insert(&new->root, backref);
2406         old->count++;
2407 out:
2408         btrfs_release_path(path);
2409         WARN_ON(ret);
2410         return ret;
2411 }
2412
2413 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2414                                    struct new_sa_defrag_extent *new)
2415 {
2416         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2417         struct old_sa_defrag_extent *old, *tmp;
2418         int ret;
2419
2420         new->path = path;
2421
2422         list_for_each_entry_safe(old, tmp, &new->head, list) {
2423                 ret = iterate_inodes_from_logical(old->bytenr +
2424                                                   old->extent_offset, fs_info,
2425                                                   path, record_one_backref,
2426                                                   old);
2427                 if (ret < 0 && ret != -ENOENT)
2428                         return false;
2429
2430                 /* no backref to be processed for this extent */
2431                 if (!old->count) {
2432                         list_del(&old->list);
2433                         kfree(old);
2434                 }
2435         }
2436
2437         if (list_empty(&new->head))
2438                 return false;
2439
2440         return true;
2441 }
2442
2443 static int relink_is_mergable(struct extent_buffer *leaf,
2444                               struct btrfs_file_extent_item *fi,
2445                               struct new_sa_defrag_extent *new)
2446 {
2447         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2448                 return 0;
2449
2450         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2451                 return 0;
2452
2453         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2454                 return 0;
2455
2456         if (btrfs_file_extent_encryption(leaf, fi) ||
2457             btrfs_file_extent_other_encoding(leaf, fi))
2458                 return 0;
2459
2460         return 1;
2461 }
2462
2463 /*
2464  * Note the backref might has changed, and in this case we just return 0.
2465  */
2466 static noinline int relink_extent_backref(struct btrfs_path *path,
2467                                  struct sa_defrag_extent_backref *prev,
2468                                  struct sa_defrag_extent_backref *backref)
2469 {
2470         struct btrfs_file_extent_item *extent;
2471         struct btrfs_file_extent_item *item;
2472         struct btrfs_ordered_extent *ordered;
2473         struct btrfs_trans_handle *trans;
2474         struct btrfs_fs_info *fs_info;
2475         struct btrfs_root *root;
2476         struct btrfs_key key;
2477         struct extent_buffer *leaf;
2478         struct old_sa_defrag_extent *old = backref->old;
2479         struct new_sa_defrag_extent *new = old->new;
2480         struct inode *src_inode = new->inode;
2481         struct inode *inode;
2482         struct extent_state *cached = NULL;
2483         int ret = 0;
2484         u64 start;
2485         u64 len;
2486         u64 lock_start;
2487         u64 lock_end;
2488         bool merge = false;
2489         int index;
2490
2491         if (prev && prev->root_id == backref->root_id &&
2492             prev->inum == backref->inum &&
2493             prev->file_pos + prev->num_bytes == backref->file_pos)
2494                 merge = true;
2495
2496         /* step 1: get root */
2497         key.objectid = backref->root_id;
2498         key.type = BTRFS_ROOT_ITEM_KEY;
2499         key.offset = (u64)-1;
2500
2501         fs_info = BTRFS_I(src_inode)->root->fs_info;
2502         index = srcu_read_lock(&fs_info->subvol_srcu);
2503
2504         root = btrfs_read_fs_root_no_name(fs_info, &key);
2505         if (IS_ERR(root)) {
2506                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2507                 if (PTR_ERR(root) == -ENOENT)
2508                         return 0;
2509                 return PTR_ERR(root);
2510         }
2511
2512         if (btrfs_root_readonly(root)) {
2513                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2514                 return 0;
2515         }
2516
2517         /* step 2: get inode */
2518         key.objectid = backref->inum;
2519         key.type = BTRFS_INODE_ITEM_KEY;
2520         key.offset = 0;
2521
2522         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2523         if (IS_ERR(inode)) {
2524                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2525                 return 0;
2526         }
2527
2528         srcu_read_unlock(&fs_info->subvol_srcu, index);
2529
2530         /* step 3: relink backref */
2531         lock_start = backref->file_pos;
2532         lock_end = backref->file_pos + backref->num_bytes - 1;
2533         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2534                          &cached);
2535
2536         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2537         if (ordered) {
2538                 btrfs_put_ordered_extent(ordered);
2539                 goto out_unlock;
2540         }
2541
2542         trans = btrfs_join_transaction(root);
2543         if (IS_ERR(trans)) {
2544                 ret = PTR_ERR(trans);
2545                 goto out_unlock;
2546         }
2547
2548         key.objectid = backref->inum;
2549         key.type = BTRFS_EXTENT_DATA_KEY;
2550         key.offset = backref->file_pos;
2551
2552         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2553         if (ret < 0) {
2554                 goto out_free_path;
2555         } else if (ret > 0) {
2556                 ret = 0;
2557                 goto out_free_path;
2558         }
2559
2560         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2561                                 struct btrfs_file_extent_item);
2562
2563         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2564             backref->generation)
2565                 goto out_free_path;
2566
2567         btrfs_release_path(path);
2568
2569         start = backref->file_pos;
2570         if (backref->extent_offset < old->extent_offset + old->offset)
2571                 start += old->extent_offset + old->offset -
2572                          backref->extent_offset;
2573
2574         len = min(backref->extent_offset + backref->num_bytes,
2575                   old->extent_offset + old->offset + old->len);
2576         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2577
2578         ret = btrfs_drop_extents(trans, root, inode, start,
2579                                  start + len, 1);
2580         if (ret)
2581                 goto out_free_path;
2582 again:
2583         key.objectid = btrfs_ino(inode);
2584         key.type = BTRFS_EXTENT_DATA_KEY;
2585         key.offset = start;
2586
2587         path->leave_spinning = 1;
2588         if (merge) {
2589                 struct btrfs_file_extent_item *fi;
2590                 u64 extent_len;
2591                 struct btrfs_key found_key;
2592
2593                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2594                 if (ret < 0)
2595                         goto out_free_path;
2596
2597                 path->slots[0]--;
2598                 leaf = path->nodes[0];
2599                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2600
2601                 fi = btrfs_item_ptr(leaf, path->slots[0],
2602                                     struct btrfs_file_extent_item);
2603                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2604
2605                 if (extent_len + found_key.offset == start &&
2606                     relink_is_mergable(leaf, fi, new)) {
2607                         btrfs_set_file_extent_num_bytes(leaf, fi,
2608                                                         extent_len + len);
2609                         btrfs_mark_buffer_dirty(leaf);
2610                         inode_add_bytes(inode, len);
2611
2612                         ret = 1;
2613                         goto out_free_path;
2614                 } else {
2615                         merge = false;
2616                         btrfs_release_path(path);
2617                         goto again;
2618                 }
2619         }
2620
2621         ret = btrfs_insert_empty_item(trans, root, path, &key,
2622                                         sizeof(*extent));
2623         if (ret) {
2624                 btrfs_abort_transaction(trans, ret);
2625                 goto out_free_path;
2626         }
2627
2628         leaf = path->nodes[0];
2629         item = btrfs_item_ptr(leaf, path->slots[0],
2630                                 struct btrfs_file_extent_item);
2631         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2632         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2633         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2634         btrfs_set_file_extent_num_bytes(leaf, item, len);
2635         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2636         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2637         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2638         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2639         btrfs_set_file_extent_encryption(leaf, item, 0);
2640         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2641
2642         btrfs_mark_buffer_dirty(leaf);
2643         inode_add_bytes(inode, len);
2644         btrfs_release_path(path);
2645
2646         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2647                         new->disk_len, 0,
2648                         backref->root_id, backref->inum,
2649                         new->file_pos); /* start - extent_offset */
2650         if (ret) {
2651                 btrfs_abort_transaction(trans, ret);
2652                 goto out_free_path;
2653         }
2654
2655         ret = 1;
2656 out_free_path:
2657         btrfs_release_path(path);
2658         path->leave_spinning = 0;
2659         btrfs_end_transaction(trans, root);
2660 out_unlock:
2661         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2662                              &cached, GFP_NOFS);
2663         iput(inode);
2664         return ret;
2665 }
2666
2667 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2668 {
2669         struct old_sa_defrag_extent *old, *tmp;
2670
2671         if (!new)
2672                 return;
2673
2674         list_for_each_entry_safe(old, tmp, &new->head, list) {
2675                 kfree(old);
2676         }
2677         kfree(new);
2678 }
2679
2680 static void relink_file_extents(struct new_sa_defrag_extent *new)
2681 {
2682         struct btrfs_path *path;
2683         struct sa_defrag_extent_backref *backref;
2684         struct sa_defrag_extent_backref *prev = NULL;
2685         struct inode *inode;
2686         struct btrfs_root *root;
2687         struct rb_node *node;
2688         int ret;
2689
2690         inode = new->inode;
2691         root = BTRFS_I(inode)->root;
2692
2693         path = btrfs_alloc_path();
2694         if (!path)
2695                 return;
2696
2697         if (!record_extent_backrefs(path, new)) {
2698                 btrfs_free_path(path);
2699                 goto out;
2700         }
2701         btrfs_release_path(path);
2702
2703         while (1) {
2704                 node = rb_first(&new->root);
2705                 if (!node)
2706                         break;
2707                 rb_erase(node, &new->root);
2708
2709                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2710
2711                 ret = relink_extent_backref(path, prev, backref);
2712                 WARN_ON(ret < 0);
2713
2714                 kfree(prev);
2715
2716                 if (ret == 1)
2717                         prev = backref;
2718                 else
2719                         prev = NULL;
2720                 cond_resched();
2721         }
2722         kfree(prev);
2723
2724         btrfs_free_path(path);
2725 out:
2726         free_sa_defrag_extent(new);
2727
2728         atomic_dec(&root->fs_info->defrag_running);
2729         wake_up(&root->fs_info->transaction_wait);
2730 }
2731
2732 static struct new_sa_defrag_extent *
2733 record_old_file_extents(struct inode *inode,
2734                         struct btrfs_ordered_extent *ordered)
2735 {
2736         struct btrfs_root *root = BTRFS_I(inode)->root;
2737         struct btrfs_path *path;
2738         struct btrfs_key key;
2739         struct old_sa_defrag_extent *old;
2740         struct new_sa_defrag_extent *new;
2741         int ret;
2742
2743         new = kmalloc(sizeof(*new), GFP_NOFS);
2744         if (!new)
2745                 return NULL;
2746
2747         new->inode = inode;
2748         new->file_pos = ordered->file_offset;
2749         new->len = ordered->len;
2750         new->bytenr = ordered->start;
2751         new->disk_len = ordered->disk_len;
2752         new->compress_type = ordered->compress_type;
2753         new->root = RB_ROOT;
2754         INIT_LIST_HEAD(&new->head);
2755
2756         path = btrfs_alloc_path();
2757         if (!path)
2758                 goto out_kfree;
2759
2760         key.objectid = btrfs_ino(inode);
2761         key.type = BTRFS_EXTENT_DATA_KEY;
2762         key.offset = new->file_pos;
2763
2764         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2765         if (ret < 0)
2766                 goto out_free_path;
2767         if (ret > 0 && path->slots[0] > 0)
2768                 path->slots[0]--;
2769
2770         /* find out all the old extents for the file range */
2771         while (1) {
2772                 struct btrfs_file_extent_item *extent;
2773                 struct extent_buffer *l;
2774                 int slot;
2775                 u64 num_bytes;
2776                 u64 offset;
2777                 u64 end;
2778                 u64 disk_bytenr;
2779                 u64 extent_offset;
2780
2781                 l = path->nodes[0];
2782                 slot = path->slots[0];
2783
2784                 if (slot >= btrfs_header_nritems(l)) {
2785                         ret = btrfs_next_leaf(root, path);
2786                         if (ret < 0)
2787                                 goto out_free_path;
2788                         else if (ret > 0)
2789                                 break;
2790                         continue;
2791                 }
2792
2793                 btrfs_item_key_to_cpu(l, &key, slot);
2794
2795                 if (key.objectid != btrfs_ino(inode))
2796                         break;
2797                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2798                         break;
2799                 if (key.offset >= new->file_pos + new->len)
2800                         break;
2801
2802                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2803
2804                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2805                 if (key.offset + num_bytes < new->file_pos)
2806                         goto next;
2807
2808                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2809                 if (!disk_bytenr)
2810                         goto next;
2811
2812                 extent_offset = btrfs_file_extent_offset(l, extent);
2813
2814                 old = kmalloc(sizeof(*old), GFP_NOFS);
2815                 if (!old)
2816                         goto out_free_path;
2817
2818                 offset = max(new->file_pos, key.offset);
2819                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2820
2821                 old->bytenr = disk_bytenr;
2822                 old->extent_offset = extent_offset;
2823                 old->offset = offset - key.offset;
2824                 old->len = end - offset;
2825                 old->new = new;
2826                 old->count = 0;
2827                 list_add_tail(&old->list, &new->head);
2828 next:
2829                 path->slots[0]++;
2830                 cond_resched();
2831         }
2832
2833         btrfs_free_path(path);
2834         atomic_inc(&root->fs_info->defrag_running);
2835
2836         return new;
2837
2838 out_free_path:
2839         btrfs_free_path(path);
2840 out_kfree:
2841         free_sa_defrag_extent(new);
2842         return NULL;
2843 }
2844
2845 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2846                                          u64 start, u64 len)
2847 {
2848         struct btrfs_block_group_cache *cache;
2849
2850         cache = btrfs_lookup_block_group(root->fs_info, start);
2851         ASSERT(cache);
2852
2853         spin_lock(&cache->lock);
2854         cache->delalloc_bytes -= len;
2855         spin_unlock(&cache->lock);
2856
2857         btrfs_put_block_group(cache);
2858 }
2859
2860 /* as ordered data IO finishes, this gets called so we can finish
2861  * an ordered extent if the range of bytes in the file it covers are
2862  * fully written.
2863  */
2864 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2865 {
2866         struct inode *inode = ordered_extent->inode;
2867         struct btrfs_root *root = BTRFS_I(inode)->root;
2868         struct btrfs_trans_handle *trans = NULL;
2869         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2870         struct extent_state *cached_state = NULL;
2871         struct new_sa_defrag_extent *new = NULL;
2872         int compress_type = 0;
2873         int ret = 0;
2874         u64 logical_len = ordered_extent->len;
2875         bool nolock;
2876         bool truncated = false;
2877
2878         nolock = btrfs_is_free_space_inode(inode);
2879
2880         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2881                 ret = -EIO;
2882                 goto out;
2883         }
2884
2885         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2886                                      ordered_extent->file_offset +
2887                                      ordered_extent->len - 1);
2888
2889         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2890                 truncated = true;
2891                 logical_len = ordered_extent->truncated_len;
2892                 /* Truncated the entire extent, don't bother adding */
2893                 if (!logical_len)
2894                         goto out;
2895         }
2896
2897         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2898                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2899
2900                 /*
2901                  * For mwrite(mmap + memset to write) case, we still reserve
2902                  * space for NOCOW range.
2903                  * As NOCOW won't cause a new delayed ref, just free the space
2904                  */
2905                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2906                                        ordered_extent->len);
2907                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2908                 if (nolock)
2909                         trans = btrfs_join_transaction_nolock(root);
2910                 else
2911                         trans = btrfs_join_transaction(root);
2912                 if (IS_ERR(trans)) {
2913                         ret = PTR_ERR(trans);
2914                         trans = NULL;
2915                         goto out;
2916                 }
2917                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2918                 ret = btrfs_update_inode_fallback(trans, root, inode);
2919                 if (ret) /* -ENOMEM or corruption */
2920                         btrfs_abort_transaction(trans, ret);
2921                 goto out;
2922         }
2923
2924         lock_extent_bits(io_tree, ordered_extent->file_offset,
2925                          ordered_extent->file_offset + ordered_extent->len - 1,
2926                          &cached_state);
2927
2928         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2929                         ordered_extent->file_offset + ordered_extent->len - 1,
2930                         EXTENT_DEFRAG, 1, cached_state);
2931         if (ret) {
2932                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2933                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2934                         /* the inode is shared */
2935                         new = record_old_file_extents(inode, ordered_extent);
2936
2937                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2938                         ordered_extent->file_offset + ordered_extent->len - 1,
2939                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2940         }
2941
2942         if (nolock)
2943                 trans = btrfs_join_transaction_nolock(root);
2944         else
2945                 trans = btrfs_join_transaction(root);
2946         if (IS_ERR(trans)) {
2947                 ret = PTR_ERR(trans);
2948                 trans = NULL;
2949                 goto out_unlock;
2950         }
2951
2952         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2953
2954         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2955                 compress_type = ordered_extent->compress_type;
2956         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2957                 BUG_ON(compress_type);
2958                 ret = btrfs_mark_extent_written(trans, inode,
2959                                                 ordered_extent->file_offset,
2960                                                 ordered_extent->file_offset +
2961                                                 logical_len);
2962         } else {
2963                 BUG_ON(root == root->fs_info->tree_root);
2964                 ret = insert_reserved_file_extent(trans, inode,
2965                                                 ordered_extent->file_offset,
2966                                                 ordered_extent->start,
2967                                                 ordered_extent->disk_len,
2968                                                 logical_len, logical_len,
2969                                                 compress_type, 0, 0,
2970                                                 BTRFS_FILE_EXTENT_REG);
2971                 if (!ret)
2972                         btrfs_release_delalloc_bytes(root,
2973                                                      ordered_extent->start,
2974                                                      ordered_extent->disk_len);
2975         }
2976         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2977                            ordered_extent->file_offset, ordered_extent->len,
2978                            trans->transid);
2979         if (ret < 0) {
2980                 btrfs_abort_transaction(trans, ret);
2981                 goto out_unlock;
2982         }
2983
2984         add_pending_csums(trans, inode, ordered_extent->file_offset,
2985                           &ordered_extent->list);
2986
2987         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2988         ret = btrfs_update_inode_fallback(trans, root, inode);
2989         if (ret) { /* -ENOMEM or corruption */
2990                 btrfs_abort_transaction(trans, ret);
2991                 goto out_unlock;
2992         }
2993         ret = 0;
2994 out_unlock:
2995         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2996                              ordered_extent->file_offset +
2997                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2998 out:
2999         if (root != root->fs_info->tree_root)
3000                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
3001         if (trans)
3002                 btrfs_end_transaction(trans, root);
3003
3004         if (ret || truncated) {
3005                 u64 start, end;
3006
3007                 if (truncated)
3008                         start = ordered_extent->file_offset + logical_len;
3009                 else
3010                         start = ordered_extent->file_offset;
3011                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3012                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3013
3014                 /* Drop the cache for the part of the extent we didn't write. */
3015                 btrfs_drop_extent_cache(inode, start, end, 0);
3016
3017                 /*
3018                  * If the ordered extent had an IOERR or something else went
3019                  * wrong we need to return the space for this ordered extent
3020                  * back to the allocator.  We only free the extent in the
3021                  * truncated case if we didn't write out the extent at all.
3022                  */
3023                 if ((ret || !logical_len) &&
3024                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3025                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3026                         btrfs_free_reserved_extent(root, ordered_extent->start,
3027                                                    ordered_extent->disk_len, 1);
3028         }
3029
3030
3031         /*
3032          * This needs to be done to make sure anybody waiting knows we are done
3033          * updating everything for this ordered extent.
3034          */
3035         btrfs_remove_ordered_extent(inode, ordered_extent);
3036
3037         /* for snapshot-aware defrag */
3038         if (new) {
3039                 if (ret) {
3040                         free_sa_defrag_extent(new);
3041                         atomic_dec(&root->fs_info->defrag_running);
3042                 } else {
3043                         relink_file_extents(new);
3044                 }
3045         }
3046
3047         /* once for us */
3048         btrfs_put_ordered_extent(ordered_extent);
3049         /* once for the tree */
3050         btrfs_put_ordered_extent(ordered_extent);
3051
3052         return ret;
3053 }
3054
3055 static void finish_ordered_fn(struct btrfs_work *work)
3056 {
3057         struct btrfs_ordered_extent *ordered_extent;
3058         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3059         btrfs_finish_ordered_io(ordered_extent);
3060 }
3061
3062 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3063                                 struct extent_state *state, int uptodate)
3064 {
3065         struct inode *inode = page->mapping->host;
3066         struct btrfs_root *root = BTRFS_I(inode)->root;
3067         struct btrfs_ordered_extent *ordered_extent = NULL;
3068         struct btrfs_workqueue *wq;
3069         btrfs_work_func_t func;
3070
3071         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3072
3073         ClearPagePrivate2(page);
3074         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3075                                             end - start + 1, uptodate))
3076                 return 0;
3077
3078         if (btrfs_is_free_space_inode(inode)) {
3079                 wq = root->fs_info->endio_freespace_worker;
3080                 func = btrfs_freespace_write_helper;
3081         } else {
3082                 wq = root->fs_info->endio_write_workers;
3083                 func = btrfs_endio_write_helper;
3084         }
3085
3086         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3087                         NULL);
3088         btrfs_queue_work(wq, &ordered_extent->work);
3089
3090         return 0;
3091 }
3092
3093 static int __readpage_endio_check(struct inode *inode,
3094                                   struct btrfs_io_bio *io_bio,
3095                                   int icsum, struct page *page,
3096                                   int pgoff, u64 start, size_t len)
3097 {
3098         char *kaddr;
3099         u32 csum_expected;
3100         u32 csum = ~(u32)0;
3101
3102         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3103
3104         kaddr = kmap_atomic(page);
3105         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3106         btrfs_csum_final(csum, (char *)&csum);
3107         if (csum != csum_expected)
3108                 goto zeroit;
3109
3110         kunmap_atomic(kaddr);
3111         return 0;
3112 zeroit:
3113         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3114                 "csum failed ino %llu off %llu csum %u expected csum %u",
3115                            btrfs_ino(inode), start, csum, csum_expected);
3116         memset(kaddr + pgoff, 1, len);
3117         flush_dcache_page(page);
3118         kunmap_atomic(kaddr);
3119         if (csum_expected == 0)
3120                 return 0;
3121         return -EIO;
3122 }
3123
3124 /*
3125  * when reads are done, we need to check csums to verify the data is correct
3126  * if there's a match, we allow the bio to finish.  If not, the code in
3127  * extent_io.c will try to find good copies for us.
3128  */
3129 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3130                                       u64 phy_offset, struct page *page,
3131                                       u64 start, u64 end, int mirror)
3132 {
3133         size_t offset = start - page_offset(page);
3134         struct inode *inode = page->mapping->host;
3135         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3136         struct btrfs_root *root = BTRFS_I(inode)->root;
3137
3138         if (PageChecked(page)) {
3139                 ClearPageChecked(page);
3140                 return 0;
3141         }
3142
3143         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3144                 return 0;
3145
3146         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3147             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3148                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3149                 return 0;
3150         }
3151
3152         phy_offset >>= inode->i_sb->s_blocksize_bits;
3153         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3154                                       start, (size_t)(end - start + 1));
3155 }
3156
3157 void btrfs_add_delayed_iput(struct inode *inode)
3158 {
3159         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3160         struct btrfs_inode *binode = BTRFS_I(inode);
3161
3162         if (atomic_add_unless(&inode->i_count, -1, 1))
3163                 return;
3164
3165         spin_lock(&fs_info->delayed_iput_lock);
3166         if (binode->delayed_iput_count == 0) {
3167                 ASSERT(list_empty(&binode->delayed_iput));
3168                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3169         } else {
3170                 binode->delayed_iput_count++;
3171         }
3172         spin_unlock(&fs_info->delayed_iput_lock);
3173 }
3174
3175 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3176 {
3177         struct btrfs_fs_info *fs_info = root->fs_info;
3178
3179         spin_lock(&fs_info->delayed_iput_lock);
3180         while (!list_empty(&fs_info->delayed_iputs)) {
3181                 struct btrfs_inode *inode;
3182
3183                 inode = list_first_entry(&fs_info->delayed_iputs,
3184                                 struct btrfs_inode, delayed_iput);
3185                 if (inode->delayed_iput_count) {
3186                         inode->delayed_iput_count--;
3187                         list_move_tail(&inode->delayed_iput,
3188                                         &fs_info->delayed_iputs);
3189                 } else {
3190                         list_del_init(&inode->delayed_iput);
3191                 }
3192                 spin_unlock(&fs_info->delayed_iput_lock);
3193                 iput(&inode->vfs_inode);
3194                 spin_lock(&fs_info->delayed_iput_lock);
3195         }
3196         spin_unlock(&fs_info->delayed_iput_lock);
3197 }
3198
3199 /*
3200  * This is called in transaction commit time. If there are no orphan
3201  * files in the subvolume, it removes orphan item and frees block_rsv
3202  * structure.
3203  */
3204 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3205                               struct btrfs_root *root)
3206 {
3207         struct btrfs_block_rsv *block_rsv;
3208         int ret;
3209
3210         if (atomic_read(&root->orphan_inodes) ||
3211             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3212                 return;
3213
3214         spin_lock(&root->orphan_lock);
3215         if (atomic_read(&root->orphan_inodes)) {
3216                 spin_unlock(&root->orphan_lock);
3217                 return;
3218         }
3219
3220         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3221                 spin_unlock(&root->orphan_lock);
3222                 return;
3223         }
3224
3225         block_rsv = root->orphan_block_rsv;
3226         root->orphan_block_rsv = NULL;
3227         spin_unlock(&root->orphan_lock);
3228
3229         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3230             btrfs_root_refs(&root->root_item) > 0) {
3231                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3232                                             root->root_key.objectid);
3233                 if (ret)
3234                         btrfs_abort_transaction(trans, ret);
3235                 else
3236                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3237                                   &root->state);
3238         }
3239
3240         if (block_rsv) {
3241                 WARN_ON(block_rsv->size > 0);
3242                 btrfs_free_block_rsv(root, block_rsv);
3243         }
3244 }
3245
3246 /*
3247  * This creates an orphan entry for the given inode in case something goes
3248  * wrong in the middle of an unlink/truncate.
3249  *
3250  * NOTE: caller of this function should reserve 5 units of metadata for
3251  *       this function.
3252  */
3253 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3254 {
3255         struct btrfs_root *root = BTRFS_I(inode)->root;
3256         struct btrfs_block_rsv *block_rsv = NULL;
3257         int reserve = 0;
3258         int insert = 0;
3259         int ret;
3260
3261         if (!root->orphan_block_rsv) {
3262                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3263                 if (!block_rsv)
3264                         return -ENOMEM;
3265         }
3266
3267         spin_lock(&root->orphan_lock);
3268         if (!root->orphan_block_rsv) {
3269                 root->orphan_block_rsv = block_rsv;
3270         } else if (block_rsv) {
3271                 btrfs_free_block_rsv(root, block_rsv);
3272                 block_rsv = NULL;
3273         }
3274
3275         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3276                               &BTRFS_I(inode)->runtime_flags)) {
3277 #if 0
3278                 /*
3279                  * For proper ENOSPC handling, we should do orphan
3280                  * cleanup when mounting. But this introduces backward
3281                  * compatibility issue.
3282                  */
3283                 if (!xchg(&root->orphan_item_inserted, 1))
3284                         insert = 2;
3285                 else
3286                         insert = 1;
3287 #endif
3288                 insert = 1;
3289                 atomic_inc(&root->orphan_inodes);
3290         }
3291
3292         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3293                               &BTRFS_I(inode)->runtime_flags))
3294                 reserve = 1;
3295         spin_unlock(&root->orphan_lock);
3296
3297         /* grab metadata reservation from transaction handle */
3298         if (reserve) {
3299                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3300                 ASSERT(!ret);
3301                 if (ret) {
3302                         atomic_dec(&root->orphan_inodes);
3303                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3304                                   &BTRFS_I(inode)->runtime_flags);
3305                         if (insert)
3306                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3307                                           &BTRFS_I(inode)->runtime_flags);
3308                         return ret;
3309                 }
3310         }
3311
3312         /* insert an orphan item to track this unlinked/truncated file */
3313         if (insert >= 1) {
3314                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3315                 if (ret) {
3316                         atomic_dec(&root->orphan_inodes);
3317                         if (reserve) {
3318                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3319                                           &BTRFS_I(inode)->runtime_flags);
3320                                 btrfs_orphan_release_metadata(inode);
3321                         }
3322                         if (ret != -EEXIST) {
3323                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3324                                           &BTRFS_I(inode)->runtime_flags);
3325                                 btrfs_abort_transaction(trans, ret);
3326                                 return ret;
3327                         }
3328                 }
3329                 ret = 0;
3330         }
3331
3332         /* insert an orphan item to track subvolume contains orphan files */
3333         if (insert >= 2) {
3334                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3335                                                root->root_key.objectid);
3336                 if (ret && ret != -EEXIST) {
3337                         btrfs_abort_transaction(trans, ret);
3338                         return ret;
3339                 }
3340         }
3341         return 0;
3342 }
3343
3344 /*
3345  * We have done the truncate/delete so we can go ahead and remove the orphan
3346  * item for this particular inode.
3347  */
3348 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3349                             struct inode *inode)
3350 {
3351         struct btrfs_root *root = BTRFS_I(inode)->root;
3352         int delete_item = 0;
3353         int release_rsv = 0;
3354         int ret = 0;
3355
3356         spin_lock(&root->orphan_lock);
3357         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3358                                &BTRFS_I(inode)->runtime_flags))
3359                 delete_item = 1;
3360
3361         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3362                                &BTRFS_I(inode)->runtime_flags))
3363                 release_rsv = 1;
3364         spin_unlock(&root->orphan_lock);
3365
3366         if (delete_item) {
3367                 atomic_dec(&root->orphan_inodes);
3368                 if (trans)
3369                         ret = btrfs_del_orphan_item(trans, root,
3370                                                     btrfs_ino(inode));
3371         }
3372
3373         if (release_rsv)
3374                 btrfs_orphan_release_metadata(inode);
3375
3376         return ret;
3377 }
3378
3379 /*
3380  * this cleans up any orphans that may be left on the list from the last use
3381  * of this root.
3382  */
3383 int btrfs_orphan_cleanup(struct btrfs_root *root)
3384 {
3385         struct btrfs_path *path;
3386         struct extent_buffer *leaf;
3387         struct btrfs_key key, found_key;
3388         struct btrfs_trans_handle *trans;
3389         struct inode *inode;
3390         u64 last_objectid = 0;
3391         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3392
3393         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3394                 return 0;
3395
3396         path = btrfs_alloc_path();
3397         if (!path) {
3398                 ret = -ENOMEM;
3399                 goto out;
3400         }
3401         path->reada = READA_BACK;
3402
3403         key.objectid = BTRFS_ORPHAN_OBJECTID;
3404         key.type = BTRFS_ORPHAN_ITEM_KEY;
3405         key.offset = (u64)-1;
3406
3407         while (1) {
3408                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3409                 if (ret < 0)
3410                         goto out;
3411
3412                 /*
3413                  * if ret == 0 means we found what we were searching for, which
3414                  * is weird, but possible, so only screw with path if we didn't
3415                  * find the key and see if we have stuff that matches
3416                  */
3417                 if (ret > 0) {
3418                         ret = 0;
3419                         if (path->slots[0] == 0)
3420                                 break;
3421                         path->slots[0]--;
3422                 }
3423
3424                 /* pull out the item */
3425                 leaf = path->nodes[0];
3426                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3427
3428                 /* make sure the item matches what we want */
3429                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3430                         break;
3431                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3432                         break;
3433
3434                 /* release the path since we're done with it */
3435                 btrfs_release_path(path);
3436
3437                 /*
3438                  * this is where we are basically btrfs_lookup, without the
3439                  * crossing root thing.  we store the inode number in the
3440                  * offset of the orphan item.
3441                  */
3442
3443                 if (found_key.offset == last_objectid) {
3444                         btrfs_err(root->fs_info,
3445                                 "Error removing orphan entry, stopping orphan cleanup");
3446                         ret = -EINVAL;
3447                         goto out;
3448                 }
3449
3450                 last_objectid = found_key.offset;
3451
3452                 found_key.objectid = found_key.offset;
3453                 found_key.type = BTRFS_INODE_ITEM_KEY;
3454                 found_key.offset = 0;
3455                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3456                 ret = PTR_ERR_OR_ZERO(inode);
3457                 if (ret && ret != -ENOENT)
3458                         goto out;
3459
3460                 if (ret == -ENOENT && root == root->fs_info->tree_root) {
3461                         struct btrfs_root *dead_root;
3462                         struct btrfs_fs_info *fs_info = root->fs_info;
3463                         int is_dead_root = 0;
3464
3465                         /*
3466                          * this is an orphan in the tree root. Currently these
3467                          * could come from 2 sources:
3468                          *  a) a snapshot deletion in progress
3469                          *  b) a free space cache inode
3470                          * We need to distinguish those two, as the snapshot
3471                          * orphan must not get deleted.
3472                          * find_dead_roots already ran before us, so if this
3473                          * is a snapshot deletion, we should find the root
3474                          * in the dead_roots list
3475                          */
3476                         spin_lock(&fs_info->trans_lock);
3477                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3478                                             root_list) {
3479                                 if (dead_root->root_key.objectid ==
3480                                     found_key.objectid) {
3481                                         is_dead_root = 1;
3482                                         break;
3483                                 }
3484                         }
3485                         spin_unlock(&fs_info->trans_lock);
3486                         if (is_dead_root) {
3487                                 /* prevent this orphan from being found again */
3488                                 key.offset = found_key.objectid - 1;
3489                                 continue;
3490                         }
3491                 }
3492                 /*
3493                  * Inode is already gone but the orphan item is still there,
3494                  * kill the orphan item.
3495                  */
3496                 if (ret == -ENOENT) {
3497                         trans = btrfs_start_transaction(root, 1);
3498                         if (IS_ERR(trans)) {
3499                                 ret = PTR_ERR(trans);
3500                                 goto out;
3501                         }
3502                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3503                                 found_key.objectid);
3504                         ret = btrfs_del_orphan_item(trans, root,
3505                                                     found_key.objectid);
3506                         btrfs_end_transaction(trans, root);
3507                         if (ret)
3508                                 goto out;
3509                         continue;
3510                 }
3511
3512                 /*
3513                  * add this inode to the orphan list so btrfs_orphan_del does
3514                  * the proper thing when we hit it
3515                  */
3516                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3517                         &BTRFS_I(inode)->runtime_flags);
3518                 atomic_inc(&root->orphan_inodes);
3519
3520                 /* if we have links, this was a truncate, lets do that */
3521                 if (inode->i_nlink) {
3522                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3523                                 iput(inode);
3524                                 continue;
3525                         }
3526                         nr_truncate++;
3527
3528                         /* 1 for the orphan item deletion. */
3529                         trans = btrfs_start_transaction(root, 1);
3530                         if (IS_ERR(trans)) {
3531                                 iput(inode);
3532                                 ret = PTR_ERR(trans);
3533                                 goto out;
3534                         }
3535                         ret = btrfs_orphan_add(trans, inode);
3536                         btrfs_end_transaction(trans, root);
3537                         if (ret) {
3538                                 iput(inode);
3539                                 goto out;
3540                         }
3541
3542                         ret = btrfs_truncate(inode);
3543                         if (ret)
3544                                 btrfs_orphan_del(NULL, inode);
3545                 } else {
3546                         nr_unlink++;
3547                 }
3548
3549                 /* this will do delete_inode and everything for us */
3550                 iput(inode);
3551                 if (ret)
3552                         goto out;
3553         }
3554         /* release the path since we're done with it */
3555         btrfs_release_path(path);
3556
3557         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3558
3559         if (root->orphan_block_rsv)
3560                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3561                                         (u64)-1);
3562
3563         if (root->orphan_block_rsv ||
3564             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3565                 trans = btrfs_join_transaction(root);
3566                 if (!IS_ERR(trans))
3567                         btrfs_end_transaction(trans, root);
3568         }
3569
3570         if (nr_unlink)
3571                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3572         if (nr_truncate)
3573                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3574
3575 out:
3576         if (ret)
3577                 btrfs_err(root->fs_info,
3578                         "could not do orphan cleanup %d", ret);
3579         btrfs_free_path(path);
3580         return ret;
3581 }
3582
3583 /*
3584  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3585  * don't find any xattrs, we know there can't be any acls.
3586  *
3587  * slot is the slot the inode is in, objectid is the objectid of the inode
3588  */
3589 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3590                                           int slot, u64 objectid,
3591                                           int *first_xattr_slot)
3592 {
3593         u32 nritems = btrfs_header_nritems(leaf);
3594         struct btrfs_key found_key;
3595         static u64 xattr_access = 0;
3596         static u64 xattr_default = 0;
3597         int scanned = 0;
3598
3599         if (!xattr_access) {
3600                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3601                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3602                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3603                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3604         }
3605
3606         slot++;
3607         *first_xattr_slot = -1;
3608         while (slot < nritems) {
3609                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3610
3611                 /* we found a different objectid, there must not be acls */
3612                 if (found_key.objectid != objectid)
3613                         return 0;
3614
3615                 /* we found an xattr, assume we've got an acl */
3616                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3617                         if (*first_xattr_slot == -1)
3618                                 *first_xattr_slot = slot;
3619                         if (found_key.offset == xattr_access ||
3620                             found_key.offset == xattr_default)
3621                                 return 1;
3622                 }
3623
3624                 /*
3625                  * we found a key greater than an xattr key, there can't
3626                  * be any acls later on
3627                  */
3628                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3629                         return 0;
3630
3631                 slot++;
3632                 scanned++;
3633
3634                 /*
3635                  * it goes inode, inode backrefs, xattrs, extents,
3636                  * so if there are a ton of hard links to an inode there can
3637                  * be a lot of backrefs.  Don't waste time searching too hard,
3638                  * this is just an optimization
3639                  */
3640                 if (scanned >= 8)
3641                         break;
3642         }
3643         /* we hit the end of the leaf before we found an xattr or
3644          * something larger than an xattr.  We have to assume the inode
3645          * has acls
3646          */
3647         if (*first_xattr_slot == -1)
3648                 *first_xattr_slot = slot;
3649         return 1;
3650 }
3651
3652 /*
3653  * read an inode from the btree into the in-memory inode
3654  */
3655 static int btrfs_read_locked_inode(struct inode *inode)
3656 {
3657         struct btrfs_path *path;
3658         struct extent_buffer *leaf;
3659         struct btrfs_inode_item *inode_item;
3660         struct btrfs_root *root = BTRFS_I(inode)->root;
3661         struct btrfs_key location;
3662         unsigned long ptr;
3663         int maybe_acls;
3664         u32 rdev;
3665         int ret;
3666         bool filled = false;
3667         int first_xattr_slot;
3668
3669         ret = btrfs_fill_inode(inode, &rdev);
3670         if (!ret)
3671                 filled = true;
3672
3673         path = btrfs_alloc_path();
3674         if (!path) {
3675                 ret = -ENOMEM;
3676                 goto make_bad;
3677         }
3678
3679         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3680
3681         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3682         if (ret) {
3683                 if (ret > 0)
3684                         ret = -ENOENT;
3685                 goto make_bad;
3686         }
3687
3688         leaf = path->nodes[0];
3689
3690         if (filled)
3691                 goto cache_index;
3692
3693         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3694                                     struct btrfs_inode_item);
3695         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3696         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3697         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3698         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3699         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3700
3701         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3702         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3703
3704         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3705         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3706
3707         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3708         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3709
3710         BTRFS_I(inode)->i_otime.tv_sec =
3711                 btrfs_timespec_sec(leaf, &inode_item->otime);
3712         BTRFS_I(inode)->i_otime.tv_nsec =
3713                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3714
3715         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3716         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3717         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3718
3719         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3720         inode->i_generation = BTRFS_I(inode)->generation;
3721         inode->i_rdev = 0;
3722         rdev = btrfs_inode_rdev(leaf, inode_item);
3723
3724         BTRFS_I(inode)->index_cnt = (u64)-1;
3725         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3726
3727 cache_index:
3728         /*
3729          * If we were modified in the current generation and evicted from memory
3730          * and then re-read we need to do a full sync since we don't have any
3731          * idea about which extents were modified before we were evicted from
3732          * cache.
3733          *
3734          * This is required for both inode re-read from disk and delayed inode
3735          * in delayed_nodes_tree.
3736          */
3737         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3738                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3739                         &BTRFS_I(inode)->runtime_flags);
3740
3741         /*
3742          * We don't persist the id of the transaction where an unlink operation
3743          * against the inode was last made. So here we assume the inode might
3744          * have been evicted, and therefore the exact value of last_unlink_trans
3745          * lost, and set it to last_trans to avoid metadata inconsistencies
3746          * between the inode and its parent if the inode is fsync'ed and the log
3747          * replayed. For example, in the scenario:
3748          *
3749          * touch mydir/foo
3750          * ln mydir/foo mydir/bar
3751          * sync
3752          * unlink mydir/bar
3753          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3754          * xfs_io -c fsync mydir/foo
3755          * <power failure>
3756          * mount fs, triggers fsync log replay
3757          *
3758          * We must make sure that when we fsync our inode foo we also log its
3759          * parent inode, otherwise after log replay the parent still has the
3760          * dentry with the "bar" name but our inode foo has a link count of 1
3761          * and doesn't have an inode ref with the name "bar" anymore.
3762          *
3763          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3764          * but it guarantees correctness at the expense of occasional full
3765          * transaction commits on fsync if our inode is a directory, or if our
3766          * inode is not a directory, logging its parent unnecessarily.
3767          */
3768         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3769
3770         path->slots[0]++;
3771         if (inode->i_nlink != 1 ||
3772             path->slots[0] >= btrfs_header_nritems(leaf))
3773                 goto cache_acl;
3774
3775         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3776         if (location.objectid != btrfs_ino(inode))
3777                 goto cache_acl;
3778
3779         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3780         if (location.type == BTRFS_INODE_REF_KEY) {
3781                 struct btrfs_inode_ref *ref;
3782
3783                 ref = (struct btrfs_inode_ref *)ptr;
3784                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3785         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3786                 struct btrfs_inode_extref *extref;
3787
3788                 extref = (struct btrfs_inode_extref *)ptr;
3789                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3790                                                                      extref);
3791         }
3792 cache_acl:
3793         /*
3794          * try to precache a NULL acl entry for files that don't have
3795          * any xattrs or acls
3796          */
3797         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3798                                            btrfs_ino(inode), &first_xattr_slot);
3799         if (first_xattr_slot != -1) {
3800                 path->slots[0] = first_xattr_slot;
3801                 ret = btrfs_load_inode_props(inode, path);
3802                 if (ret)
3803                         btrfs_err(root->fs_info,
3804                                   "error loading props for ino %llu (root %llu): %d",
3805                                   btrfs_ino(inode),
3806                                   root->root_key.objectid, ret);
3807         }
3808         btrfs_free_path(path);
3809
3810         if (!maybe_acls)
3811                 cache_no_acl(inode);
3812
3813         switch (inode->i_mode & S_IFMT) {
3814         case S_IFREG:
3815                 inode->i_mapping->a_ops = &btrfs_aops;
3816                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3817                 inode->i_fop = &btrfs_file_operations;
3818                 inode->i_op = &btrfs_file_inode_operations;
3819                 break;
3820         case S_IFDIR:
3821                 inode->i_fop = &btrfs_dir_file_operations;
3822                 if (root == root->fs_info->tree_root)
3823                         inode->i_op = &btrfs_dir_ro_inode_operations;
3824                 else
3825                         inode->i_op = &btrfs_dir_inode_operations;
3826                 break;
3827         case S_IFLNK:
3828                 inode->i_op = &btrfs_symlink_inode_operations;
3829                 inode_nohighmem(inode);
3830                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3831                 break;
3832         default:
3833                 inode->i_op = &btrfs_special_inode_operations;
3834                 init_special_inode(inode, inode->i_mode, rdev);
3835                 break;
3836         }
3837
3838         btrfs_update_iflags(inode);
3839         return 0;
3840
3841 make_bad:
3842         btrfs_free_path(path);
3843         make_bad_inode(inode);
3844         return ret;
3845 }
3846
3847 /*
3848  * given a leaf and an inode, copy the inode fields into the leaf
3849  */
3850 static void fill_inode_item(struct btrfs_trans_handle *trans,
3851                             struct extent_buffer *leaf,
3852                             struct btrfs_inode_item *item,
3853                             struct inode *inode)
3854 {
3855         struct btrfs_map_token token;
3856
3857         btrfs_init_map_token(&token);
3858
3859         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3860         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3861         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3862                                    &token);
3863         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3864         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3865
3866         btrfs_set_token_timespec_sec(leaf, &item->atime,
3867                                      inode->i_atime.tv_sec, &token);
3868         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3869                                       inode->i_atime.tv_nsec, &token);
3870
3871         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3872                                      inode->i_mtime.tv_sec, &token);
3873         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3874                                       inode->i_mtime.tv_nsec, &token);
3875
3876         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3877                                      inode->i_ctime.tv_sec, &token);
3878         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3879                                       inode->i_ctime.tv_nsec, &token);
3880
3881         btrfs_set_token_timespec_sec(leaf, &item->otime,
3882                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3883         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3884                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3885
3886         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3887                                      &token);
3888         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3889                                          &token);
3890         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3891         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3892         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3893         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3894         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3895 }
3896
3897 /*
3898  * copy everything in the in-memory inode into the btree.
3899  */
3900 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3901                                 struct btrfs_root *root, struct inode *inode)
3902 {
3903         struct btrfs_inode_item *inode_item;
3904         struct btrfs_path *path;
3905         struct extent_buffer *leaf;
3906         int ret;
3907
3908         path = btrfs_alloc_path();
3909         if (!path)
3910                 return -ENOMEM;
3911
3912         path->leave_spinning = 1;
3913         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3914                                  1);
3915         if (ret) {
3916                 if (ret > 0)
3917                         ret = -ENOENT;
3918                 goto failed;
3919         }
3920
3921         leaf = path->nodes[0];
3922         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3923                                     struct btrfs_inode_item);
3924
3925         fill_inode_item(trans, leaf, inode_item, inode);
3926         btrfs_mark_buffer_dirty(leaf);
3927         btrfs_set_inode_last_trans(trans, inode);
3928         ret = 0;
3929 failed:
3930         btrfs_free_path(path);
3931         return ret;
3932 }
3933
3934 /*
3935  * copy everything in the in-memory inode into the btree.
3936  */
3937 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3938                                 struct btrfs_root *root, struct inode *inode)
3939 {
3940         int ret;
3941
3942         /*
3943          * If the inode is a free space inode, we can deadlock during commit
3944          * if we put it into the delayed code.
3945          *
3946          * The data relocation inode should also be directly updated
3947          * without delay
3948          */
3949         if (!btrfs_is_free_space_inode(inode)
3950             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3951             && !test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
3952                 btrfs_update_root_times(trans, root);
3953
3954                 ret = btrfs_delayed_update_inode(trans, root, inode);
3955                 if (!ret)
3956                         btrfs_set_inode_last_trans(trans, inode);
3957                 return ret;
3958         }
3959
3960         return btrfs_update_inode_item(trans, root, inode);
3961 }
3962
3963 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3964                                          struct btrfs_root *root,
3965                                          struct inode *inode)
3966 {
3967         int ret;
3968
3969         ret = btrfs_update_inode(trans, root, inode);
3970         if (ret == -ENOSPC)
3971                 return btrfs_update_inode_item(trans, root, inode);
3972         return ret;
3973 }
3974
3975 /*
3976  * unlink helper that gets used here in inode.c and in the tree logging
3977  * recovery code.  It remove a link in a directory with a given name, and
3978  * also drops the back refs in the inode to the directory
3979  */
3980 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3981                                 struct btrfs_root *root,
3982                                 struct inode *dir, struct inode *inode,
3983                                 const char *name, int name_len)
3984 {
3985         struct btrfs_path *path;
3986         int ret = 0;
3987         struct extent_buffer *leaf;
3988         struct btrfs_dir_item *di;
3989         struct btrfs_key key;
3990         u64 index;
3991         u64 ino = btrfs_ino(inode);
3992         u64 dir_ino = btrfs_ino(dir);
3993
3994         path = btrfs_alloc_path();
3995         if (!path) {
3996                 ret = -ENOMEM;
3997                 goto out;
3998         }
3999
4000         path->leave_spinning = 1;
4001         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4002                                     name, name_len, -1);
4003         if (IS_ERR(di)) {
4004                 ret = PTR_ERR(di);
4005                 goto err;
4006         }
4007         if (!di) {
4008                 ret = -ENOENT;
4009                 goto err;
4010         }
4011         leaf = path->nodes[0];
4012         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4013         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4014         if (ret)
4015                 goto err;
4016         btrfs_release_path(path);
4017
4018         /*
4019          * If we don't have dir index, we have to get it by looking up
4020          * the inode ref, since we get the inode ref, remove it directly,
4021          * it is unnecessary to do delayed deletion.
4022          *
4023          * But if we have dir index, needn't search inode ref to get it.
4024          * Since the inode ref is close to the inode item, it is better
4025          * that we delay to delete it, and just do this deletion when
4026          * we update the inode item.
4027          */
4028         if (BTRFS_I(inode)->dir_index) {
4029                 ret = btrfs_delayed_delete_inode_ref(inode);
4030                 if (!ret) {
4031                         index = BTRFS_I(inode)->dir_index;
4032                         goto skip_backref;
4033                 }
4034         }
4035
4036         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4037                                   dir_ino, &index);
4038         if (ret) {
4039                 btrfs_info(root->fs_info,
4040                         "failed to delete reference to %.*s, inode %llu parent %llu",
4041                         name_len, name, ino, dir_ino);
4042                 btrfs_abort_transaction(trans, ret);
4043                 goto err;
4044         }
4045 skip_backref:
4046         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4047         if (ret) {
4048                 btrfs_abort_transaction(trans, ret);
4049                 goto err;
4050         }
4051
4052         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4053                                          inode, dir_ino);
4054         if (ret != 0 && ret != -ENOENT) {
4055                 btrfs_abort_transaction(trans, ret);
4056                 goto err;
4057         }
4058
4059         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4060                                            dir, index);
4061         if (ret == -ENOENT)
4062                 ret = 0;
4063         else if (ret)
4064                 btrfs_abort_transaction(trans, ret);
4065 err:
4066         btrfs_free_path(path);
4067         if (ret)
4068                 goto out;
4069
4070         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4071         inode_inc_iversion(inode);
4072         inode_inc_iversion(dir);
4073         inode->i_ctime = dir->i_mtime =
4074                 dir->i_ctime = current_time(inode);
4075         ret = btrfs_update_inode(trans, root, dir);
4076 out:
4077         return ret;
4078 }
4079
4080 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4081                        struct btrfs_root *root,
4082                        struct inode *dir, struct inode *inode,
4083                        const char *name, int name_len)
4084 {
4085         int ret;
4086         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4087         if (!ret) {
4088                 drop_nlink(inode);
4089                 ret = btrfs_update_inode(trans, root, inode);
4090         }
4091         return ret;
4092 }
4093
4094 /*
4095  * helper to start transaction for unlink and rmdir.
4096  *
4097  * unlink and rmdir are special in btrfs, they do not always free space, so
4098  * if we cannot make our reservations the normal way try and see if there is
4099  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4100  * allow the unlink to occur.
4101  */
4102 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4103 {
4104         struct btrfs_root *root = BTRFS_I(dir)->root;
4105
4106         /*
4107          * 1 for the possible orphan item
4108          * 1 for the dir item
4109          * 1 for the dir index
4110          * 1 for the inode ref
4111          * 1 for the inode
4112          */
4113         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4114 }
4115
4116 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4117 {
4118         struct btrfs_root *root = BTRFS_I(dir)->root;
4119         struct btrfs_trans_handle *trans;
4120         struct inode *inode = d_inode(dentry);
4121         int ret;
4122
4123         trans = __unlink_start_trans(dir);
4124         if (IS_ERR(trans))
4125                 return PTR_ERR(trans);
4126
4127         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4128
4129         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4130                                  dentry->d_name.name, dentry->d_name.len);
4131         if (ret)
4132                 goto out;
4133
4134         if (inode->i_nlink == 0) {
4135                 ret = btrfs_orphan_add(trans, inode);
4136                 if (ret)
4137                         goto out;
4138         }
4139
4140 out:
4141         btrfs_end_transaction(trans, root);
4142         btrfs_btree_balance_dirty(root);
4143         return ret;
4144 }
4145
4146 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4147                         struct btrfs_root *root,
4148                         struct inode *dir, u64 objectid,
4149                         const char *name, int name_len)
4150 {
4151         struct btrfs_path *path;
4152         struct extent_buffer *leaf;
4153         struct btrfs_dir_item *di;
4154         struct btrfs_key key;
4155         u64 index;
4156         int ret;
4157         u64 dir_ino = btrfs_ino(dir);
4158
4159         path = btrfs_alloc_path();
4160         if (!path)
4161                 return -ENOMEM;
4162
4163         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4164                                    name, name_len, -1);
4165         if (IS_ERR_OR_NULL(di)) {
4166                 if (!di)
4167                         ret = -ENOENT;
4168                 else
4169                         ret = PTR_ERR(di);
4170                 goto out;
4171         }
4172
4173         leaf = path->nodes[0];
4174         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4175         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4176         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4177         if (ret) {
4178                 btrfs_abort_transaction(trans, ret);
4179                 goto out;
4180         }
4181         btrfs_release_path(path);
4182
4183         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4184                                  objectid, root->root_key.objectid,
4185                                  dir_ino, &index, name, name_len);
4186         if (ret < 0) {
4187                 if (ret != -ENOENT) {
4188                         btrfs_abort_transaction(trans, ret);
4189                         goto out;
4190                 }
4191                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4192                                                  name, name_len);
4193                 if (IS_ERR_OR_NULL(di)) {
4194                         if (!di)
4195                                 ret = -ENOENT;
4196                         else
4197                                 ret = PTR_ERR(di);
4198                         btrfs_abort_transaction(trans, ret);
4199                         goto out;
4200                 }
4201
4202                 leaf = path->nodes[0];
4203                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4204                 btrfs_release_path(path);
4205                 index = key.offset;
4206         }
4207         btrfs_release_path(path);
4208
4209         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4210         if (ret) {
4211                 btrfs_abort_transaction(trans, ret);
4212                 goto out;
4213         }
4214
4215         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4216         inode_inc_iversion(dir);
4217         dir->i_mtime = dir->i_ctime = current_time(dir);
4218         ret = btrfs_update_inode_fallback(trans, root, dir);
4219         if (ret)
4220                 btrfs_abort_transaction(trans, ret);
4221 out:
4222         btrfs_free_path(path);
4223         return ret;
4224 }
4225
4226 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4227 {
4228         struct inode *inode = d_inode(dentry);
4229         int err = 0;
4230         struct btrfs_root *root = BTRFS_I(dir)->root;
4231         struct btrfs_trans_handle *trans;
4232         u64 last_unlink_trans;
4233
4234         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4235                 return -ENOTEMPTY;
4236         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4237                 return -EPERM;
4238
4239         trans = __unlink_start_trans(dir);
4240         if (IS_ERR(trans))
4241                 return PTR_ERR(trans);
4242
4243         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4244                 err = btrfs_unlink_subvol(trans, root, dir,
4245                                           BTRFS_I(inode)->location.objectid,
4246                                           dentry->d_name.name,
4247                                           dentry->d_name.len);
4248                 goto out;
4249         }
4250
4251         err = btrfs_orphan_add(trans, inode);
4252         if (err)
4253                 goto out;
4254
4255         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4256
4257         /* now the directory is empty */
4258         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4259                                  dentry->d_name.name, dentry->d_name.len);
4260         if (!err) {
4261                 btrfs_i_size_write(inode, 0);
4262                 /*
4263                  * Propagate the last_unlink_trans value of the deleted dir to
4264                  * its parent directory. This is to prevent an unrecoverable
4265                  * log tree in the case we do something like this:
4266                  * 1) create dir foo
4267                  * 2) create snapshot under dir foo
4268                  * 3) delete the snapshot
4269                  * 4) rmdir foo
4270                  * 5) mkdir foo
4271                  * 6) fsync foo or some file inside foo
4272                  */
4273                 if (last_unlink_trans >= trans->transid)
4274                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4275         }
4276 out:
4277         btrfs_end_transaction(trans, root);
4278         btrfs_btree_balance_dirty(root);
4279
4280         return err;
4281 }
4282
4283 static int truncate_space_check(struct btrfs_trans_handle *trans,
4284                                 struct btrfs_root *root,
4285                                 u64 bytes_deleted)
4286 {
4287         int ret;
4288
4289         /*
4290          * This is only used to apply pressure to the enospc system, we don't
4291          * intend to use this reservation at all.
4292          */
4293         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4294         bytes_deleted *= root->nodesize;
4295         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4296                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4297         if (!ret) {
4298                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4299                                               trans->transid,
4300                                               bytes_deleted, 1);
4301                 trans->bytes_reserved += bytes_deleted;
4302         }
4303         return ret;
4304
4305 }
4306
4307 static int truncate_inline_extent(struct inode *inode,
4308                                   struct btrfs_path *path,
4309                                   struct btrfs_key *found_key,
4310                                   const u64 item_end,
4311                                   const u64 new_size)
4312 {
4313         struct extent_buffer *leaf = path->nodes[0];
4314         int slot = path->slots[0];
4315         struct btrfs_file_extent_item *fi;
4316         u32 size = (u32)(new_size - found_key->offset);
4317         struct btrfs_root *root = BTRFS_I(inode)->root;
4318
4319         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4320
4321         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4322                 loff_t offset = new_size;
4323                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4324
4325                 /*
4326                  * Zero out the remaining of the last page of our inline extent,
4327                  * instead of directly truncating our inline extent here - that
4328                  * would be much more complex (decompressing all the data, then
4329                  * compressing the truncated data, which might be bigger than
4330                  * the size of the inline extent, resize the extent, etc).
4331                  * We release the path because to get the page we might need to
4332                  * read the extent item from disk (data not in the page cache).
4333                  */
4334                 btrfs_release_path(path);
4335                 return btrfs_truncate_block(inode, offset, page_end - offset,
4336                                         0);
4337         }
4338
4339         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4340         size = btrfs_file_extent_calc_inline_size(size);
4341         btrfs_truncate_item(root, path, size, 1);
4342
4343         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4344                 inode_sub_bytes(inode, item_end + 1 - new_size);
4345
4346         return 0;
4347 }
4348
4349 /*
4350  * this can truncate away extent items, csum items and directory items.
4351  * It starts at a high offset and removes keys until it can't find
4352  * any higher than new_size
4353  *
4354  * csum items that cross the new i_size are truncated to the new size
4355  * as well.
4356  *
4357  * min_type is the minimum key type to truncate down to.  If set to 0, this
4358  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4359  */
4360 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4361                                struct btrfs_root *root,
4362                                struct inode *inode,
4363                                u64 new_size, u32 min_type)
4364 {
4365         struct btrfs_path *path;
4366         struct extent_buffer *leaf;
4367         struct btrfs_file_extent_item *fi;
4368         struct btrfs_key key;
4369         struct btrfs_key found_key;
4370         u64 extent_start = 0;
4371         u64 extent_num_bytes = 0;
4372         u64 extent_offset = 0;
4373         u64 item_end = 0;
4374         u64 last_size = new_size;
4375         u32 found_type = (u8)-1;
4376         int found_extent;
4377         int del_item;
4378         int pending_del_nr = 0;
4379         int pending_del_slot = 0;
4380         int extent_type = -1;
4381         int ret;
4382         int err = 0;
4383         u64 ino = btrfs_ino(inode);
4384         u64 bytes_deleted = 0;
4385         bool be_nice = 0;
4386         bool should_throttle = 0;
4387         bool should_end = 0;
4388
4389         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4390
4391         /*
4392          * for non-free space inodes and ref cows, we want to back off from
4393          * time to time
4394          */
4395         if (!btrfs_is_free_space_inode(inode) &&
4396             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4397                 be_nice = 1;
4398
4399         path = btrfs_alloc_path();
4400         if (!path)
4401                 return -ENOMEM;
4402         path->reada = READA_BACK;
4403
4404         /*
4405          * We want to drop from the next block forward in case this new size is
4406          * not block aligned since we will be keeping the last block of the
4407          * extent just the way it is.
4408          */
4409         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4410             root == root->fs_info->tree_root)
4411                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4412                                         root->sectorsize), (u64)-1, 0);
4413
4414         /*
4415          * This function is also used to drop the items in the log tree before
4416          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4417          * it is used to drop the loged items. So we shouldn't kill the delayed
4418          * items.
4419          */
4420         if (min_type == 0 && root == BTRFS_I(inode)->root)
4421                 btrfs_kill_delayed_inode_items(inode);
4422
4423         key.objectid = ino;
4424         key.offset = (u64)-1;
4425         key.type = (u8)-1;
4426
4427 search_again:
4428         /*
4429          * with a 16K leaf size and 128MB extents, you can actually queue
4430          * up a huge file in a single leaf.  Most of the time that
4431          * bytes_deleted is > 0, it will be huge by the time we get here
4432          */
4433         if (be_nice && bytes_deleted > SZ_32M) {
4434                 if (btrfs_should_end_transaction(trans, root)) {
4435                         err = -EAGAIN;
4436                         goto error;
4437                 }
4438         }
4439
4440
4441         path->leave_spinning = 1;
4442         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4443         if (ret < 0) {
4444                 err = ret;
4445                 goto out;
4446         }
4447
4448         if (ret > 0) {
4449                 /* there are no items in the tree for us to truncate, we're
4450                  * done
4451                  */
4452                 if (path->slots[0] == 0)
4453                         goto out;
4454                 path->slots[0]--;
4455         }
4456
4457         while (1) {
4458                 fi = NULL;
4459                 leaf = path->nodes[0];
4460                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4461                 found_type = found_key.type;
4462
4463                 if (found_key.objectid != ino)
4464                         break;
4465
4466                 if (found_type < min_type)
4467                         break;
4468
4469                 item_end = found_key.offset;
4470                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4471                         fi = btrfs_item_ptr(leaf, path->slots[0],
4472                                             struct btrfs_file_extent_item);
4473                         extent_type = btrfs_file_extent_type(leaf, fi);
4474                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4475                                 item_end +=
4476                                     btrfs_file_extent_num_bytes(leaf, fi);
4477                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4478                                 item_end += btrfs_file_extent_inline_len(leaf,
4479                                                          path->slots[0], fi);
4480                         }
4481                         item_end--;
4482                 }
4483                 if (found_type > min_type) {
4484                         del_item = 1;
4485                 } else {
4486                         if (item_end < new_size)
4487                                 break;
4488                         if (found_key.offset >= new_size)
4489                                 del_item = 1;
4490                         else
4491                                 del_item = 0;
4492                 }
4493                 found_extent = 0;
4494                 /* FIXME, shrink the extent if the ref count is only 1 */
4495                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4496                         goto delete;
4497
4498                 if (del_item)
4499                         last_size = found_key.offset;
4500                 else
4501                         last_size = new_size;
4502
4503                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4504                         u64 num_dec;
4505                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4506                         if (!del_item) {
4507                                 u64 orig_num_bytes =
4508                                         btrfs_file_extent_num_bytes(leaf, fi);
4509                                 extent_num_bytes = ALIGN(new_size -
4510                                                 found_key.offset,
4511                                                 root->sectorsize);
4512                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4513                                                          extent_num_bytes);
4514                                 num_dec = (orig_num_bytes -
4515                                            extent_num_bytes);
4516                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4517                                              &root->state) &&
4518                                     extent_start != 0)
4519                                         inode_sub_bytes(inode, num_dec);
4520                                 btrfs_mark_buffer_dirty(leaf);
4521                         } else {
4522                                 extent_num_bytes =
4523                                         btrfs_file_extent_disk_num_bytes(leaf,
4524                                                                          fi);
4525                                 extent_offset = found_key.offset -
4526                                         btrfs_file_extent_offset(leaf, fi);
4527
4528                                 /* FIXME blocksize != 4096 */
4529                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4530                                 if (extent_start != 0) {
4531                                         found_extent = 1;
4532                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4533                                                      &root->state))
4534                                                 inode_sub_bytes(inode, num_dec);
4535                                 }
4536                         }
4537                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4538                         /*
4539                          * we can't truncate inline items that have had
4540                          * special encodings
4541                          */
4542                         if (!del_item &&
4543                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4544                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4545
4546                                 /*
4547                                  * Need to release path in order to truncate a
4548                                  * compressed extent. So delete any accumulated
4549                                  * extent items so far.
4550                                  */
4551                                 if (btrfs_file_extent_compression(leaf, fi) !=
4552                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4553                                         err = btrfs_del_items(trans, root, path,
4554                                                               pending_del_slot,
4555                                                               pending_del_nr);
4556                                         if (err) {
4557                                                 btrfs_abort_transaction(trans,
4558                                                                         err);
4559                                                 goto error;
4560                                         }
4561                                         pending_del_nr = 0;
4562                                 }
4563
4564                                 err = truncate_inline_extent(inode, path,
4565                                                              &found_key,
4566                                                              item_end,
4567                                                              new_size);
4568                                 if (err) {
4569                                         btrfs_abort_transaction(trans, err);
4570                                         goto error;
4571                                 }
4572                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4573                                             &root->state)) {
4574                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4575                         }
4576                 }
4577 delete:
4578                 if (del_item) {
4579                         if (!pending_del_nr) {
4580                                 /* no pending yet, add ourselves */
4581                                 pending_del_slot = path->slots[0];
4582                                 pending_del_nr = 1;
4583                         } else if (pending_del_nr &&
4584                                    path->slots[0] + 1 == pending_del_slot) {
4585                                 /* hop on the pending chunk */
4586                                 pending_del_nr++;
4587                                 pending_del_slot = path->slots[0];
4588                         } else {
4589                                 BUG();
4590                         }
4591                 } else {
4592                         break;
4593                 }
4594                 should_throttle = 0;
4595
4596                 if (found_extent &&
4597                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4598                      root == root->fs_info->tree_root)) {
4599                         btrfs_set_path_blocking(path);
4600                         bytes_deleted += extent_num_bytes;
4601                         ret = btrfs_free_extent(trans, root, extent_start,
4602                                                 extent_num_bytes, 0,
4603                                                 btrfs_header_owner(leaf),
4604                                                 ino, extent_offset);
4605                         BUG_ON(ret);
4606                         if (btrfs_should_throttle_delayed_refs(trans, root))
4607                                 btrfs_async_run_delayed_refs(root,
4608                                         trans->delayed_ref_updates * 2,
4609                                         trans->transid, 0);
4610                         if (be_nice) {
4611                                 if (truncate_space_check(trans, root,
4612                                                          extent_num_bytes)) {
4613                                         should_end = 1;
4614                                 }
4615                                 if (btrfs_should_throttle_delayed_refs(trans,
4616                                                                        root)) {
4617                                         should_throttle = 1;
4618                                 }
4619                         }
4620                 }
4621
4622                 if (found_type == BTRFS_INODE_ITEM_KEY)
4623                         break;
4624
4625                 if (path->slots[0] == 0 ||
4626                     path->slots[0] != pending_del_slot ||
4627                     should_throttle || should_end) {
4628                         if (pending_del_nr) {
4629                                 ret = btrfs_del_items(trans, root, path,
4630                                                 pending_del_slot,
4631                                                 pending_del_nr);
4632                                 if (ret) {
4633                                         btrfs_abort_transaction(trans, ret);
4634                                         goto error;
4635                                 }
4636                                 pending_del_nr = 0;
4637                         }
4638                         btrfs_release_path(path);
4639                         if (should_throttle) {
4640                                 unsigned long updates = trans->delayed_ref_updates;
4641                                 if (updates) {
4642                                         trans->delayed_ref_updates = 0;
4643                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4644                                         if (ret && !err)
4645                                                 err = ret;
4646                                 }
4647                         }
4648                         /*
4649                          * if we failed to refill our space rsv, bail out
4650                          * and let the transaction restart
4651                          */
4652                         if (should_end) {
4653                                 err = -EAGAIN;
4654                                 goto error;
4655                         }
4656                         goto search_again;
4657                 } else {
4658                         path->slots[0]--;
4659                 }
4660         }
4661 out:
4662         if (pending_del_nr) {
4663                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4664                                       pending_del_nr);
4665                 if (ret)
4666                         btrfs_abort_transaction(trans, ret);
4667         }
4668 error:
4669         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4670                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4671
4672         btrfs_free_path(path);
4673
4674         if (be_nice && bytes_deleted > SZ_32M) {
4675                 unsigned long updates = trans->delayed_ref_updates;
4676                 if (updates) {
4677                         trans->delayed_ref_updates = 0;
4678                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4679                         if (ret && !err)
4680                                 err = ret;
4681                 }
4682         }
4683         return err;
4684 }
4685
4686 /*
4687  * btrfs_truncate_block - read, zero a chunk and write a block
4688  * @inode - inode that we're zeroing
4689  * @from - the offset to start zeroing
4690  * @len - the length to zero, 0 to zero the entire range respective to the
4691  *      offset
4692  * @front - zero up to the offset instead of from the offset on
4693  *
4694  * This will find the block for the "from" offset and cow the block and zero the
4695  * part we want to zero.  This is used with truncate and hole punching.
4696  */
4697 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4698                         int front)
4699 {
4700         struct address_space *mapping = inode->i_mapping;
4701         struct btrfs_root *root = BTRFS_I(inode)->root;
4702         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4703         struct btrfs_ordered_extent *ordered;
4704         struct extent_state *cached_state = NULL;
4705         char *kaddr;
4706         u32 blocksize = root->sectorsize;
4707         pgoff_t index = from >> PAGE_SHIFT;
4708         unsigned offset = from & (blocksize - 1);
4709         struct page *page;
4710         gfp_t mask = btrfs_alloc_write_mask(mapping);
4711         int ret = 0;
4712         u64 block_start;
4713         u64 block_end;
4714
4715         if ((offset & (blocksize - 1)) == 0 &&
4716             (!len || ((len & (blocksize - 1)) == 0)))
4717                 goto out;
4718
4719         ret = btrfs_delalloc_reserve_space(inode,
4720                         round_down(from, blocksize), blocksize);
4721         if (ret)
4722                 goto out;
4723
4724 again:
4725         page = find_or_create_page(mapping, index, mask);
4726         if (!page) {
4727                 btrfs_delalloc_release_space(inode,
4728                                 round_down(from, blocksize),
4729                                 blocksize);
4730                 ret = -ENOMEM;
4731                 goto out;
4732         }
4733
4734         block_start = round_down(from, blocksize);
4735         block_end = block_start + blocksize - 1;
4736
4737         if (!PageUptodate(page)) {
4738                 ret = btrfs_readpage(NULL, page);
4739                 lock_page(page);
4740                 if (page->mapping != mapping) {
4741                         unlock_page(page);
4742                         put_page(page);
4743                         goto again;
4744                 }
4745                 if (!PageUptodate(page)) {
4746                         ret = -EIO;
4747                         goto out_unlock;
4748                 }
4749         }
4750         wait_on_page_writeback(page);
4751
4752         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4753         set_page_extent_mapped(page);
4754
4755         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4756         if (ordered) {
4757                 unlock_extent_cached(io_tree, block_start, block_end,
4758                                      &cached_state, GFP_NOFS);
4759                 unlock_page(page);
4760                 put_page(page);
4761                 btrfs_start_ordered_extent(inode, ordered, 1);
4762                 btrfs_put_ordered_extent(ordered);
4763                 goto again;
4764         }
4765
4766         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4767                           EXTENT_DIRTY | EXTENT_DELALLOC |
4768                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4769                           0, 0, &cached_state, GFP_NOFS);
4770
4771         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4772                                         &cached_state, 0);
4773         if (ret) {
4774                 unlock_extent_cached(io_tree, block_start, block_end,
4775                                      &cached_state, GFP_NOFS);
4776                 goto out_unlock;
4777         }
4778
4779         if (offset != blocksize) {
4780                 if (!len)
4781                         len = blocksize - offset;
4782                 kaddr = kmap(page);
4783                 if (front)
4784                         memset(kaddr + (block_start - page_offset(page)),
4785                                 0, offset);
4786                 else
4787                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4788                                 0, len);
4789                 flush_dcache_page(page);
4790                 kunmap(page);
4791         }
4792         ClearPageChecked(page);
4793         set_page_dirty(page);
4794         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4795                              GFP_NOFS);
4796
4797 out_unlock:
4798         if (ret)
4799                 btrfs_delalloc_release_space(inode, block_start,
4800                                              blocksize);
4801         unlock_page(page);
4802         put_page(page);
4803 out:
4804         return ret;
4805 }
4806
4807 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4808                              u64 offset, u64 len)
4809 {
4810         struct btrfs_trans_handle *trans;
4811         int ret;
4812
4813         /*
4814          * Still need to make sure the inode looks like it's been updated so
4815          * that any holes get logged if we fsync.
4816          */
4817         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4818                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4819                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4820                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4821                 return 0;
4822         }
4823
4824         /*
4825          * 1 - for the one we're dropping
4826          * 1 - for the one we're adding
4827          * 1 - for updating the inode.
4828          */
4829         trans = btrfs_start_transaction(root, 3);
4830         if (IS_ERR(trans))
4831                 return PTR_ERR(trans);
4832
4833         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4834         if (ret) {
4835                 btrfs_abort_transaction(trans, ret);
4836                 btrfs_end_transaction(trans, root);
4837                 return ret;
4838         }
4839
4840         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4841                                        0, 0, len, 0, len, 0, 0, 0);
4842         if (ret)
4843                 btrfs_abort_transaction(trans, ret);
4844         else
4845                 btrfs_update_inode(trans, root, inode);
4846         btrfs_end_transaction(trans, root);
4847         return ret;
4848 }
4849
4850 /*
4851  * This function puts in dummy file extents for the area we're creating a hole
4852  * for.  So if we are truncating this file to a larger size we need to insert
4853  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4854  * the range between oldsize and size
4855  */
4856 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4857 {
4858         struct btrfs_root *root = BTRFS_I(inode)->root;
4859         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4860         struct extent_map *em = NULL;
4861         struct extent_state *cached_state = NULL;
4862         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4863         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4864         u64 block_end = ALIGN(size, root->sectorsize);
4865         u64 last_byte;
4866         u64 cur_offset;
4867         u64 hole_size;
4868         int err = 0;
4869
4870         /*
4871          * If our size started in the middle of a block we need to zero out the
4872          * rest of the block before we expand the i_size, otherwise we could
4873          * expose stale data.
4874          */
4875         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4876         if (err)
4877                 return err;
4878
4879         if (size <= hole_start)
4880                 return 0;
4881
4882         while (1) {
4883                 struct btrfs_ordered_extent *ordered;
4884
4885                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4886                                  &cached_state);
4887                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4888                                                      block_end - hole_start);
4889                 if (!ordered)
4890                         break;
4891                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4892                                      &cached_state, GFP_NOFS);
4893                 btrfs_start_ordered_extent(inode, ordered, 1);
4894                 btrfs_put_ordered_extent(ordered);
4895         }
4896
4897         cur_offset = hole_start;
4898         while (1) {
4899                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4900                                 block_end - cur_offset, 0);
4901                 if (IS_ERR(em)) {
4902                         err = PTR_ERR(em);
4903                         em = NULL;
4904                         break;
4905                 }
4906                 last_byte = min(extent_map_end(em), block_end);
4907                 last_byte = ALIGN(last_byte , root->sectorsize);
4908                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4909                         struct extent_map *hole_em;
4910                         hole_size = last_byte - cur_offset;
4911
4912                         err = maybe_insert_hole(root, inode, cur_offset,
4913                                                 hole_size);
4914                         if (err)
4915                                 break;
4916                         btrfs_drop_extent_cache(inode, cur_offset,
4917                                                 cur_offset + hole_size - 1, 0);
4918                         hole_em = alloc_extent_map();
4919                         if (!hole_em) {
4920                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4921                                         &BTRFS_I(inode)->runtime_flags);
4922                                 goto next;
4923                         }
4924                         hole_em->start = cur_offset;
4925                         hole_em->len = hole_size;
4926                         hole_em->orig_start = cur_offset;
4927
4928                         hole_em->block_start = EXTENT_MAP_HOLE;
4929                         hole_em->block_len = 0;
4930                         hole_em->orig_block_len = 0;
4931                         hole_em->ram_bytes = hole_size;
4932                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4933                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4934                         hole_em->generation = root->fs_info->generation;
4935
4936                         while (1) {
4937                                 write_lock(&em_tree->lock);
4938                                 err = add_extent_mapping(em_tree, hole_em, 1);
4939                                 write_unlock(&em_tree->lock);
4940                                 if (err != -EEXIST)
4941                                         break;
4942                                 btrfs_drop_extent_cache(inode, cur_offset,
4943                                                         cur_offset +
4944                                                         hole_size - 1, 0);
4945                         }
4946                         free_extent_map(hole_em);
4947                 }
4948 next:
4949                 free_extent_map(em);
4950                 em = NULL;
4951                 cur_offset = last_byte;
4952                 if (cur_offset >= block_end)
4953                         break;
4954         }
4955         free_extent_map(em);
4956         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4957                              GFP_NOFS);
4958         return err;
4959 }
4960
4961 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4962 {
4963         struct btrfs_root *root = BTRFS_I(inode)->root;
4964         struct btrfs_trans_handle *trans;
4965         loff_t oldsize = i_size_read(inode);
4966         loff_t newsize = attr->ia_size;
4967         int mask = attr->ia_valid;
4968         int ret;
4969
4970         /*
4971          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4972          * special case where we need to update the times despite not having
4973          * these flags set.  For all other operations the VFS set these flags
4974          * explicitly if it wants a timestamp update.
4975          */
4976         if (newsize != oldsize) {
4977                 inode_inc_iversion(inode);
4978                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4979                         inode->i_ctime = inode->i_mtime =
4980                                 current_time(inode);
4981         }
4982
4983         if (newsize > oldsize) {
4984                 /*
4985                  * Don't do an expanding truncate while snapshoting is ongoing.
4986                  * This is to ensure the snapshot captures a fully consistent
4987                  * state of this file - if the snapshot captures this expanding
4988                  * truncation, it must capture all writes that happened before
4989                  * this truncation.
4990                  */
4991                 btrfs_wait_for_snapshot_creation(root);
4992                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4993                 if (ret) {
4994                         btrfs_end_write_no_snapshoting(root);
4995                         return ret;
4996                 }
4997
4998                 trans = btrfs_start_transaction(root, 1);
4999                 if (IS_ERR(trans)) {
5000                         btrfs_end_write_no_snapshoting(root);
5001                         return PTR_ERR(trans);
5002                 }
5003
5004                 i_size_write(inode, newsize);
5005                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5006                 pagecache_isize_extended(inode, oldsize, newsize);
5007                 ret = btrfs_update_inode(trans, root, inode);
5008                 btrfs_end_write_no_snapshoting(root);
5009                 btrfs_end_transaction(trans, root);
5010         } else {
5011
5012                 /*
5013                  * We're truncating a file that used to have good data down to
5014                  * zero. Make sure it gets into the ordered flush list so that
5015                  * any new writes get down to disk quickly.
5016                  */
5017                 if (newsize == 0)
5018                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5019                                 &BTRFS_I(inode)->runtime_flags);
5020
5021                 /*
5022                  * 1 for the orphan item we're going to add
5023                  * 1 for the orphan item deletion.
5024                  */
5025                 trans = btrfs_start_transaction(root, 2);
5026                 if (IS_ERR(trans))
5027                         return PTR_ERR(trans);
5028
5029                 /*
5030                  * We need to do this in case we fail at _any_ point during the
5031                  * actual truncate.  Once we do the truncate_setsize we could
5032                  * invalidate pages which forces any outstanding ordered io to
5033                  * be instantly completed which will give us extents that need
5034                  * to be truncated.  If we fail to get an orphan inode down we
5035                  * could have left over extents that were never meant to live,
5036                  * so we need to guarantee from this point on that everything
5037                  * will be consistent.
5038                  */
5039                 ret = btrfs_orphan_add(trans, inode);
5040                 btrfs_end_transaction(trans, root);
5041                 if (ret)
5042                         return ret;
5043
5044                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5045                 truncate_setsize(inode, newsize);
5046
5047                 /* Disable nonlocked read DIO to avoid the end less truncate */
5048                 btrfs_inode_block_unlocked_dio(inode);
5049                 inode_dio_wait(inode);
5050                 btrfs_inode_resume_unlocked_dio(inode);
5051
5052                 ret = btrfs_truncate(inode);
5053                 if (ret && inode->i_nlink) {
5054                         int err;
5055
5056                         /*
5057                          * failed to truncate, disk_i_size is only adjusted down
5058                          * as we remove extents, so it should represent the true
5059                          * size of the inode, so reset the in memory size and
5060                          * delete our orphan entry.
5061                          */
5062                         trans = btrfs_join_transaction(root);
5063                         if (IS_ERR(trans)) {
5064                                 btrfs_orphan_del(NULL, inode);
5065                                 return ret;
5066                         }
5067                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5068                         err = btrfs_orphan_del(trans, inode);
5069                         if (err)
5070                                 btrfs_abort_transaction(trans, err);
5071                         btrfs_end_transaction(trans, root);
5072                 }
5073         }
5074
5075         return ret;
5076 }
5077
5078 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5079 {
5080         struct inode *inode = d_inode(dentry);
5081         struct btrfs_root *root = BTRFS_I(inode)->root;
5082         int err;
5083
5084         if (btrfs_root_readonly(root))
5085                 return -EROFS;
5086
5087         err = setattr_prepare(dentry, attr);
5088         if (err)
5089                 return err;
5090
5091         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5092                 err = btrfs_setsize(inode, attr);
5093                 if (err)
5094                         return err;
5095         }
5096
5097         if (attr->ia_valid) {
5098                 setattr_copy(inode, attr);
5099                 inode_inc_iversion(inode);
5100                 err = btrfs_dirty_inode(inode);
5101
5102                 if (!err && attr->ia_valid & ATTR_MODE)
5103                         err = posix_acl_chmod(inode, inode->i_mode);
5104         }
5105
5106         return err;
5107 }
5108
5109 /*
5110  * While truncating the inode pages during eviction, we get the VFS calling
5111  * btrfs_invalidatepage() against each page of the inode. This is slow because
5112  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5113  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5114  * extent_state structures over and over, wasting lots of time.
5115  *
5116  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5117  * those expensive operations on a per page basis and do only the ordered io
5118  * finishing, while we release here the extent_map and extent_state structures,
5119  * without the excessive merging and splitting.
5120  */
5121 static void evict_inode_truncate_pages(struct inode *inode)
5122 {
5123         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5124         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5125         struct rb_node *node;
5126
5127         ASSERT(inode->i_state & I_FREEING);
5128         truncate_inode_pages_final(&inode->i_data);
5129
5130         write_lock(&map_tree->lock);
5131         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5132                 struct extent_map *em;
5133
5134                 node = rb_first(&map_tree->map);
5135                 em = rb_entry(node, struct extent_map, rb_node);
5136                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5137                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5138                 remove_extent_mapping(map_tree, em);
5139                 free_extent_map(em);
5140                 if (need_resched()) {
5141                         write_unlock(&map_tree->lock);
5142                         cond_resched();
5143                         write_lock(&map_tree->lock);
5144                 }
5145         }
5146         write_unlock(&map_tree->lock);
5147
5148         /*
5149          * Keep looping until we have no more ranges in the io tree.
5150          * We can have ongoing bios started by readpages (called from readahead)
5151          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5152          * still in progress (unlocked the pages in the bio but did not yet
5153          * unlocked the ranges in the io tree). Therefore this means some
5154          * ranges can still be locked and eviction started because before
5155          * submitting those bios, which are executed by a separate task (work
5156          * queue kthread), inode references (inode->i_count) were not taken
5157          * (which would be dropped in the end io callback of each bio).
5158          * Therefore here we effectively end up waiting for those bios and
5159          * anyone else holding locked ranges without having bumped the inode's
5160          * reference count - if we don't do it, when they access the inode's
5161          * io_tree to unlock a range it may be too late, leading to an
5162          * use-after-free issue.
5163          */
5164         spin_lock(&io_tree->lock);
5165         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5166                 struct extent_state *state;
5167                 struct extent_state *cached_state = NULL;
5168                 u64 start;
5169                 u64 end;
5170
5171                 node = rb_first(&io_tree->state);
5172                 state = rb_entry(node, struct extent_state, rb_node);
5173                 start = state->start;
5174                 end = state->end;
5175                 spin_unlock(&io_tree->lock);
5176
5177                 lock_extent_bits(io_tree, start, end, &cached_state);
5178
5179                 /*
5180                  * If still has DELALLOC flag, the extent didn't reach disk,
5181                  * and its reserved space won't be freed by delayed_ref.
5182                  * So we need to free its reserved space here.
5183                  * (Refer to comment in btrfs_invalidatepage, case 2)
5184                  *
5185                  * Note, end is the bytenr of last byte, so we need + 1 here.
5186                  */
5187                 if (state->state & EXTENT_DELALLOC)
5188                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5189
5190                 clear_extent_bit(io_tree, start, end,
5191                                  EXTENT_LOCKED | EXTENT_DIRTY |
5192                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5193                                  EXTENT_DEFRAG, 1, 1,
5194                                  &cached_state, GFP_NOFS);
5195
5196                 cond_resched();
5197                 spin_lock(&io_tree->lock);
5198         }
5199         spin_unlock(&io_tree->lock);
5200 }
5201
5202 void btrfs_evict_inode(struct inode *inode)
5203 {
5204         struct btrfs_trans_handle *trans;
5205         struct btrfs_root *root = BTRFS_I(inode)->root;
5206         struct btrfs_block_rsv *rsv, *global_rsv;
5207         int steal_from_global = 0;
5208         u64 min_size;
5209         int ret;
5210
5211         trace_btrfs_inode_evict(inode);
5212
5213         if (!root) {
5214                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5215                 return;
5216         }
5217
5218         min_size = btrfs_calc_trunc_metadata_size(root, 1);
5219
5220         evict_inode_truncate_pages(inode);
5221
5222         if (inode->i_nlink &&
5223             ((btrfs_root_refs(&root->root_item) != 0 &&
5224               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5225              btrfs_is_free_space_inode(inode)))
5226                 goto no_delete;
5227
5228         if (is_bad_inode(inode)) {
5229                 btrfs_orphan_del(NULL, inode);
5230                 goto no_delete;
5231         }
5232         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5233         if (!special_file(inode->i_mode))
5234                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5235
5236         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5237
5238         if (test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
5239                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5240                                  &BTRFS_I(inode)->runtime_flags));
5241                 goto no_delete;
5242         }
5243
5244         if (inode->i_nlink > 0) {
5245                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5246                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5247                 goto no_delete;
5248         }
5249
5250         ret = btrfs_commit_inode_delayed_inode(inode);
5251         if (ret) {
5252                 btrfs_orphan_del(NULL, inode);
5253                 goto no_delete;
5254         }
5255
5256         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5257         if (!rsv) {
5258                 btrfs_orphan_del(NULL, inode);
5259                 goto no_delete;
5260         }
5261         rsv->size = min_size;
5262         rsv->failfast = 1;
5263         global_rsv = &root->fs_info->global_block_rsv;
5264
5265         btrfs_i_size_write(inode, 0);
5266
5267         /*
5268          * This is a bit simpler than btrfs_truncate since we've already
5269          * reserved our space for our orphan item in the unlink, so we just
5270          * need to reserve some slack space in case we add bytes and update
5271          * inode item when doing the truncate.
5272          */
5273         while (1) {
5274                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5275                                              BTRFS_RESERVE_FLUSH_LIMIT);
5276
5277                 /*
5278                  * Try and steal from the global reserve since we will
5279                  * likely not use this space anyway, we want to try as
5280                  * hard as possible to get this to work.
5281                  */
5282                 if (ret)
5283                         steal_from_global++;
5284                 else
5285                         steal_from_global = 0;
5286                 ret = 0;
5287
5288                 /*
5289                  * steal_from_global == 0: we reserved stuff, hooray!
5290                  * steal_from_global == 1: we didn't reserve stuff, boo!
5291                  * steal_from_global == 2: we've committed, still not a lot of
5292                  * room but maybe we'll have room in the global reserve this
5293                  * time.
5294                  * steal_from_global == 3: abandon all hope!
5295                  */
5296                 if (steal_from_global > 2) {
5297                         btrfs_warn(root->fs_info,
5298                                 "Could not get space for a delete, will truncate on mount %d",
5299                                 ret);
5300                         btrfs_orphan_del(NULL, inode);
5301                         btrfs_free_block_rsv(root, rsv);
5302                         goto no_delete;
5303                 }
5304
5305                 trans = btrfs_join_transaction(root);
5306                 if (IS_ERR(trans)) {
5307                         btrfs_orphan_del(NULL, inode);
5308                         btrfs_free_block_rsv(root, rsv);
5309                         goto no_delete;
5310                 }
5311
5312                 /*
5313                  * We can't just steal from the global reserve, we need to make
5314                  * sure there is room to do it, if not we need to commit and try
5315                  * again.
5316                  */
5317                 if (steal_from_global) {
5318                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5319                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5320                                                               min_size, 0);
5321                         else
5322                                 ret = -ENOSPC;
5323                 }
5324
5325                 /*
5326                  * Couldn't steal from the global reserve, we have too much
5327                  * pending stuff built up, commit the transaction and try it
5328                  * again.
5329                  */
5330                 if (ret) {
5331                         ret = btrfs_commit_transaction(trans, root);
5332                         if (ret) {
5333                                 btrfs_orphan_del(NULL, inode);
5334                                 btrfs_free_block_rsv(root, rsv);
5335                                 goto no_delete;
5336                         }
5337                         continue;
5338                 } else {
5339                         steal_from_global = 0;
5340                 }
5341
5342                 trans->block_rsv = rsv;
5343
5344                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5345                 if (ret != -ENOSPC && ret != -EAGAIN)
5346                         break;
5347
5348                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5349                 btrfs_end_transaction(trans, root);
5350                 trans = NULL;
5351                 btrfs_btree_balance_dirty(root);
5352         }
5353
5354         btrfs_free_block_rsv(root, rsv);
5355
5356         /*
5357          * Errors here aren't a big deal, it just means we leave orphan items
5358          * in the tree.  They will be cleaned up on the next mount.
5359          */
5360         if (ret == 0) {
5361                 trans->block_rsv = root->orphan_block_rsv;
5362                 btrfs_orphan_del(trans, inode);
5363         } else {
5364                 btrfs_orphan_del(NULL, inode);
5365         }
5366
5367         trans->block_rsv = &root->fs_info->trans_block_rsv;
5368         if (!(root == root->fs_info->tree_root ||
5369               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5370                 btrfs_return_ino(root, btrfs_ino(inode));
5371
5372         btrfs_end_transaction(trans, root);
5373         btrfs_btree_balance_dirty(root);
5374 no_delete:
5375         btrfs_remove_delayed_node(inode);
5376         clear_inode(inode);
5377 }
5378
5379 /*
5380  * this returns the key found in the dir entry in the location pointer.
5381  * If no dir entries were found, location->objectid is 0.
5382  */
5383 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5384                                struct btrfs_key *location)
5385 {
5386         const char *name = dentry->d_name.name;
5387         int namelen = dentry->d_name.len;
5388         struct btrfs_dir_item *di;
5389         struct btrfs_path *path;
5390         struct btrfs_root *root = BTRFS_I(dir)->root;
5391         int ret = 0;
5392
5393         path = btrfs_alloc_path();
5394         if (!path)
5395                 return -ENOMEM;
5396
5397         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5398                                     namelen, 0);
5399         if (IS_ERR(di))
5400                 ret = PTR_ERR(di);
5401
5402         if (IS_ERR_OR_NULL(di))
5403                 goto out_err;
5404
5405         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5406 out:
5407         btrfs_free_path(path);
5408         return ret;
5409 out_err:
5410         location->objectid = 0;
5411         goto out;
5412 }
5413
5414 /*
5415  * when we hit a tree root in a directory, the btrfs part of the inode
5416  * needs to be changed to reflect the root directory of the tree root.  This
5417  * is kind of like crossing a mount point.
5418  */
5419 static int fixup_tree_root_location(struct btrfs_root *root,
5420                                     struct inode *dir,
5421                                     struct dentry *dentry,
5422                                     struct btrfs_key *location,
5423                                     struct btrfs_root **sub_root)
5424 {
5425         struct btrfs_path *path;
5426         struct btrfs_root *new_root;
5427         struct btrfs_root_ref *ref;
5428         struct extent_buffer *leaf;
5429         struct btrfs_key key;
5430         int ret;
5431         int err = 0;
5432
5433         path = btrfs_alloc_path();
5434         if (!path) {
5435                 err = -ENOMEM;
5436                 goto out;
5437         }
5438
5439         err = -ENOENT;
5440         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5441         key.type = BTRFS_ROOT_REF_KEY;
5442         key.offset = location->objectid;
5443
5444         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5445                                 0, 0);
5446         if (ret) {
5447                 if (ret < 0)
5448                         err = ret;
5449                 goto out;
5450         }
5451
5452         leaf = path->nodes[0];
5453         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5454         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5455             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5456                 goto out;
5457
5458         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5459                                    (unsigned long)(ref + 1),
5460                                    dentry->d_name.len);
5461         if (ret)
5462                 goto out;
5463
5464         btrfs_release_path(path);
5465
5466         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5467         if (IS_ERR(new_root)) {
5468                 err = PTR_ERR(new_root);
5469                 goto out;
5470         }
5471
5472         *sub_root = new_root;
5473         location->objectid = btrfs_root_dirid(&new_root->root_item);
5474         location->type = BTRFS_INODE_ITEM_KEY;
5475         location->offset = 0;
5476         err = 0;
5477 out:
5478         btrfs_free_path(path);
5479         return err;
5480 }
5481
5482 static void inode_tree_add(struct inode *inode)
5483 {
5484         struct btrfs_root *root = BTRFS_I(inode)->root;
5485         struct btrfs_inode *entry;
5486         struct rb_node **p;
5487         struct rb_node *parent;
5488         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5489         u64 ino = btrfs_ino(inode);
5490
5491         if (inode_unhashed(inode))
5492                 return;
5493         parent = NULL;
5494         spin_lock(&root->inode_lock);
5495         p = &root->inode_tree.rb_node;
5496         while (*p) {
5497                 parent = *p;
5498                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5499
5500                 if (ino < btrfs_ino(&entry->vfs_inode))
5501                         p = &parent->rb_left;
5502                 else if (ino > btrfs_ino(&entry->vfs_inode))
5503                         p = &parent->rb_right;
5504                 else {
5505                         WARN_ON(!(entry->vfs_inode.i_state &
5506                                   (I_WILL_FREE | I_FREEING)));
5507                         rb_replace_node(parent, new, &root->inode_tree);
5508                         RB_CLEAR_NODE(parent);
5509                         spin_unlock(&root->inode_lock);
5510                         return;
5511                 }
5512         }
5513         rb_link_node(new, parent, p);
5514         rb_insert_color(new, &root->inode_tree);
5515         spin_unlock(&root->inode_lock);
5516 }
5517
5518 static void inode_tree_del(struct inode *inode)
5519 {
5520         struct btrfs_root *root = BTRFS_I(inode)->root;
5521         int empty = 0;
5522
5523         spin_lock(&root->inode_lock);
5524         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5525                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5526                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5527                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5528         }
5529         spin_unlock(&root->inode_lock);
5530
5531         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5532                 synchronize_srcu(&root->fs_info->subvol_srcu);
5533                 spin_lock(&root->inode_lock);
5534                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5535                 spin_unlock(&root->inode_lock);
5536                 if (empty)
5537                         btrfs_add_dead_root(root);
5538         }
5539 }
5540
5541 void btrfs_invalidate_inodes(struct btrfs_root *root)
5542 {
5543         struct rb_node *node;
5544         struct rb_node *prev;
5545         struct btrfs_inode *entry;
5546         struct inode *inode;
5547         u64 objectid = 0;
5548
5549         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5550                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5551
5552         spin_lock(&root->inode_lock);
5553 again:
5554         node = root->inode_tree.rb_node;
5555         prev = NULL;
5556         while (node) {
5557                 prev = node;
5558                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5559
5560                 if (objectid < btrfs_ino(&entry->vfs_inode))
5561                         node = node->rb_left;
5562                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5563                         node = node->rb_right;
5564                 else
5565                         break;
5566         }
5567         if (!node) {
5568                 while (prev) {
5569                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5570                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5571                                 node = prev;
5572                                 break;
5573                         }
5574                         prev = rb_next(prev);
5575                 }
5576         }
5577         while (node) {
5578                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5579                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5580                 inode = igrab(&entry->vfs_inode);
5581                 if (inode) {
5582                         spin_unlock(&root->inode_lock);
5583                         if (atomic_read(&inode->i_count) > 1)
5584                                 d_prune_aliases(inode);
5585                         /*
5586                          * btrfs_drop_inode will have it removed from
5587                          * the inode cache when its usage count
5588                          * hits zero.
5589                          */
5590                         iput(inode);
5591                         cond_resched();
5592                         spin_lock(&root->inode_lock);
5593                         goto again;
5594                 }
5595
5596                 if (cond_resched_lock(&root->inode_lock))
5597                         goto again;
5598
5599                 node = rb_next(node);
5600         }
5601         spin_unlock(&root->inode_lock);
5602 }
5603
5604 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5605 {
5606         struct btrfs_iget_args *args = p;
5607         inode->i_ino = args->location->objectid;
5608         memcpy(&BTRFS_I(inode)->location, args->location,
5609                sizeof(*args->location));
5610         BTRFS_I(inode)->root = args->root;
5611         return 0;
5612 }
5613
5614 static int btrfs_find_actor(struct inode *inode, void *opaque)
5615 {
5616         struct btrfs_iget_args *args = opaque;
5617         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5618                 args->root == BTRFS_I(inode)->root;
5619 }
5620
5621 static struct inode *btrfs_iget_locked(struct super_block *s,
5622                                        struct btrfs_key *location,
5623                                        struct btrfs_root *root)
5624 {
5625         struct inode *inode;
5626         struct btrfs_iget_args args;
5627         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5628
5629         args.location = location;
5630         args.root = root;
5631
5632         inode = iget5_locked(s, hashval, btrfs_find_actor,
5633                              btrfs_init_locked_inode,
5634                              (void *)&args);
5635         return inode;
5636 }
5637
5638 /* Get an inode object given its location and corresponding root.
5639  * Returns in *is_new if the inode was read from disk
5640  */
5641 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5642                          struct btrfs_root *root, int *new)
5643 {
5644         struct inode *inode;
5645
5646         inode = btrfs_iget_locked(s, location, root);
5647         if (!inode)
5648                 return ERR_PTR(-ENOMEM);
5649
5650         if (inode->i_state & I_NEW) {
5651                 int ret;
5652
5653                 ret = btrfs_read_locked_inode(inode);
5654                 if (!is_bad_inode(inode)) {
5655                         inode_tree_add(inode);
5656                         unlock_new_inode(inode);
5657                         if (new)
5658                                 *new = 1;
5659                 } else {
5660                         unlock_new_inode(inode);
5661                         iput(inode);
5662                         ASSERT(ret < 0);
5663                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5664                 }
5665         }
5666
5667         return inode;
5668 }
5669
5670 static struct inode *new_simple_dir(struct super_block *s,
5671                                     struct btrfs_key *key,
5672                                     struct btrfs_root *root)
5673 {
5674         struct inode *inode = new_inode(s);
5675
5676         if (!inode)
5677                 return ERR_PTR(-ENOMEM);
5678
5679         BTRFS_I(inode)->root = root;
5680         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5681         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5682
5683         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5684         inode->i_op = &btrfs_dir_ro_inode_operations;
5685         inode->i_fop = &simple_dir_operations;
5686         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5687         inode->i_mtime = current_time(inode);
5688         inode->i_atime = inode->i_mtime;
5689         inode->i_ctime = inode->i_mtime;
5690         BTRFS_I(inode)->i_otime = inode->i_mtime;
5691
5692         return inode;
5693 }
5694
5695 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5696 {
5697         struct inode *inode;
5698         struct btrfs_root *root = BTRFS_I(dir)->root;
5699         struct btrfs_root *sub_root = root;
5700         struct btrfs_key location;
5701         int index;
5702         int ret = 0;
5703
5704         if (dentry->d_name.len > BTRFS_NAME_LEN)
5705                 return ERR_PTR(-ENAMETOOLONG);
5706
5707         ret = btrfs_inode_by_name(dir, dentry, &location);
5708         if (ret < 0)
5709                 return ERR_PTR(ret);
5710
5711         if (location.objectid == 0)
5712                 return ERR_PTR(-ENOENT);
5713
5714         if (location.type == BTRFS_INODE_ITEM_KEY) {
5715                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5716                 return inode;
5717         }
5718
5719         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5720
5721         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5722         ret = fixup_tree_root_location(root, dir, dentry,
5723                                        &location, &sub_root);
5724         if (ret < 0) {
5725                 if (ret != -ENOENT)
5726                         inode = ERR_PTR(ret);
5727                 else
5728                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5729         } else {
5730                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5731         }
5732         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5733
5734         if (!IS_ERR(inode) && root != sub_root) {
5735                 down_read(&root->fs_info->cleanup_work_sem);
5736                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5737                         ret = btrfs_orphan_cleanup(sub_root);
5738                 up_read(&root->fs_info->cleanup_work_sem);
5739                 if (ret) {
5740                         iput(inode);
5741                         inode = ERR_PTR(ret);
5742                 }
5743         }
5744
5745         return inode;
5746 }
5747
5748 static int btrfs_dentry_delete(const struct dentry *dentry)
5749 {
5750         struct btrfs_root *root;
5751         struct inode *inode = d_inode(dentry);
5752
5753         if (!inode && !IS_ROOT(dentry))
5754                 inode = d_inode(dentry->d_parent);
5755
5756         if (inode) {
5757                 root = BTRFS_I(inode)->root;
5758                 if (btrfs_root_refs(&root->root_item) == 0)
5759                         return 1;
5760
5761                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5762                         return 1;
5763         }
5764         return 0;
5765 }
5766
5767 static void btrfs_dentry_release(struct dentry *dentry)
5768 {
5769         kfree(dentry->d_fsdata);
5770 }
5771
5772 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5773                                    unsigned int flags)
5774 {
5775         struct inode *inode;
5776
5777         inode = btrfs_lookup_dentry(dir, dentry);
5778         if (IS_ERR(inode)) {
5779                 if (PTR_ERR(inode) == -ENOENT)
5780                         inode = NULL;
5781                 else
5782                         return ERR_CAST(inode);
5783         }
5784
5785         return d_splice_alias(inode, dentry);
5786 }
5787
5788 unsigned char btrfs_filetype_table[] = {
5789         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5790 };
5791
5792 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5793 {
5794         struct inode *inode = file_inode(file);
5795         struct btrfs_root *root = BTRFS_I(inode)->root;
5796         struct btrfs_item *item;
5797         struct btrfs_dir_item *di;
5798         struct btrfs_key key;
5799         struct btrfs_key found_key;
5800         struct btrfs_path *path;
5801         struct list_head ins_list;
5802         struct list_head del_list;
5803         int ret;
5804         struct extent_buffer *leaf;
5805         int slot;
5806         unsigned char d_type;
5807         int over = 0;
5808         u32 di_cur;
5809         u32 di_total;
5810         u32 di_len;
5811         int key_type = BTRFS_DIR_INDEX_KEY;
5812         char tmp_name[32];
5813         char *name_ptr;
5814         int name_len;
5815         int is_curr = 0;        /* ctx->pos points to the current index? */
5816         bool emitted;
5817         bool put = false;
5818
5819         /* FIXME, use a real flag for deciding about the key type */
5820         if (root->fs_info->tree_root == root)
5821                 key_type = BTRFS_DIR_ITEM_KEY;
5822
5823         if (!dir_emit_dots(file, ctx))
5824                 return 0;
5825
5826         path = btrfs_alloc_path();
5827         if (!path)
5828                 return -ENOMEM;
5829
5830         path->reada = READA_FORWARD;
5831
5832         if (key_type == BTRFS_DIR_INDEX_KEY) {
5833                 INIT_LIST_HEAD(&ins_list);
5834                 INIT_LIST_HEAD(&del_list);
5835                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5836                                                       &del_list);
5837         }
5838
5839         key.type = key_type;
5840         key.offset = ctx->pos;
5841         key.objectid = btrfs_ino(inode);
5842
5843         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5844         if (ret < 0)
5845                 goto err;
5846
5847         emitted = false;
5848         while (1) {
5849                 leaf = path->nodes[0];
5850                 slot = path->slots[0];
5851                 if (slot >= btrfs_header_nritems(leaf)) {
5852                         ret = btrfs_next_leaf(root, path);
5853                         if (ret < 0)
5854                                 goto err;
5855                         else if (ret > 0)
5856                                 break;
5857                         continue;
5858                 }
5859
5860                 item = btrfs_item_nr(slot);
5861                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5862
5863                 if (found_key.objectid != key.objectid)
5864                         break;
5865                 if (found_key.type != key_type)
5866                         break;
5867                 if (found_key.offset < ctx->pos)
5868                         goto next;
5869                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5870                     btrfs_should_delete_dir_index(&del_list,
5871                                                   found_key.offset))
5872                         goto next;
5873
5874                 ctx->pos = found_key.offset;
5875                 is_curr = 1;
5876
5877                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5878                 di_cur = 0;
5879                 di_total = btrfs_item_size(leaf, item);
5880
5881                 while (di_cur < di_total) {
5882                         struct btrfs_key location;
5883
5884                         if (verify_dir_item(root, leaf, di))
5885                                 break;
5886
5887                         name_len = btrfs_dir_name_len(leaf, di);
5888                         if (name_len <= sizeof(tmp_name)) {
5889                                 name_ptr = tmp_name;
5890                         } else {
5891                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5892                                 if (!name_ptr) {
5893                                         ret = -ENOMEM;
5894                                         goto err;
5895                                 }
5896                         }
5897                         read_extent_buffer(leaf, name_ptr,
5898                                            (unsigned long)(di + 1), name_len);
5899
5900                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5901                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5902
5903
5904                         /* is this a reference to our own snapshot? If so
5905                          * skip it.
5906                          *
5907                          * In contrast to old kernels, we insert the snapshot's
5908                          * dir item and dir index after it has been created, so
5909                          * we won't find a reference to our own snapshot. We
5910                          * still keep the following code for backward
5911                          * compatibility.
5912                          */
5913                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5914                             location.objectid == root->root_key.objectid) {
5915                                 over = 0;
5916                                 goto skip;
5917                         }
5918                         over = !dir_emit(ctx, name_ptr, name_len,
5919                                        location.objectid, d_type);
5920
5921 skip:
5922                         if (name_ptr != tmp_name)
5923                                 kfree(name_ptr);
5924
5925                         if (over)
5926                                 goto nopos;
5927                         emitted = true;
5928                         di_len = btrfs_dir_name_len(leaf, di) +
5929                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5930                         di_cur += di_len;
5931                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5932                 }
5933 next:
5934                 path->slots[0]++;
5935         }
5936
5937         if (key_type == BTRFS_DIR_INDEX_KEY) {
5938                 if (is_curr)
5939                         ctx->pos++;
5940                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5941                 if (ret)
5942                         goto nopos;
5943         }
5944
5945         /*
5946          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5947          * it was was set to the termination value in previous call. We assume
5948          * that "." and ".." were emitted if we reach this point and set the
5949          * termination value as well for an empty directory.
5950          */
5951         if (ctx->pos > 2 && !emitted)
5952                 goto nopos;
5953
5954         /* Reached end of directory/root. Bump pos past the last item. */
5955         ctx->pos++;
5956
5957         /*
5958          * Stop new entries from being returned after we return the last
5959          * entry.
5960          *
5961          * New directory entries are assigned a strictly increasing
5962          * offset.  This means that new entries created during readdir
5963          * are *guaranteed* to be seen in the future by that readdir.
5964          * This has broken buggy programs which operate on names as
5965          * they're returned by readdir.  Until we re-use freed offsets
5966          * we have this hack to stop new entries from being returned
5967          * under the assumption that they'll never reach this huge
5968          * offset.
5969          *
5970          * This is being careful not to overflow 32bit loff_t unless the
5971          * last entry requires it because doing so has broken 32bit apps
5972          * in the past.
5973          */
5974         if (key_type == BTRFS_DIR_INDEX_KEY) {
5975                 if (ctx->pos >= INT_MAX)
5976                         ctx->pos = LLONG_MAX;
5977                 else
5978                         ctx->pos = INT_MAX;
5979         }
5980 nopos:
5981         ret = 0;
5982 err:
5983         if (put)
5984                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5985         btrfs_free_path(path);
5986         return ret;
5987 }
5988
5989 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5990 {
5991         struct btrfs_root *root = BTRFS_I(inode)->root;
5992         struct btrfs_trans_handle *trans;
5993         int ret = 0;
5994         bool nolock = false;
5995
5996         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5997                 return 0;
5998
5999         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
6000                 nolock = true;
6001
6002         if (wbc->sync_mode == WB_SYNC_ALL) {
6003                 if (nolock)
6004                         trans = btrfs_join_transaction_nolock(root);
6005                 else
6006                         trans = btrfs_join_transaction(root);
6007                 if (IS_ERR(trans))
6008                         return PTR_ERR(trans);
6009                 ret = btrfs_commit_transaction(trans, root);
6010         }
6011         return ret;
6012 }
6013
6014 /*
6015  * This is somewhat expensive, updating the tree every time the
6016  * inode changes.  But, it is most likely to find the inode in cache.
6017  * FIXME, needs more benchmarking...there are no reasons other than performance
6018  * to keep or drop this code.
6019  */
6020 static int btrfs_dirty_inode(struct inode *inode)
6021 {
6022         struct btrfs_root *root = BTRFS_I(inode)->root;
6023         struct btrfs_trans_handle *trans;
6024         int ret;
6025
6026         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6027                 return 0;
6028
6029         trans = btrfs_join_transaction(root);
6030         if (IS_ERR(trans))
6031                 return PTR_ERR(trans);
6032
6033         ret = btrfs_update_inode(trans, root, inode);
6034         if (ret && ret == -ENOSPC) {
6035                 /* whoops, lets try again with the full transaction */
6036                 btrfs_end_transaction(trans, root);
6037                 trans = btrfs_start_transaction(root, 1);
6038                 if (IS_ERR(trans))
6039                         return PTR_ERR(trans);
6040
6041                 ret = btrfs_update_inode(trans, root, inode);
6042         }
6043         btrfs_end_transaction(trans, root);
6044         if (BTRFS_I(inode)->delayed_node)
6045                 btrfs_balance_delayed_items(root);
6046
6047         return ret;
6048 }
6049
6050 /*
6051  * This is a copy of file_update_time.  We need this so we can return error on
6052  * ENOSPC for updating the inode in the case of file write and mmap writes.
6053  */
6054 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6055                              int flags)
6056 {
6057         struct btrfs_root *root = BTRFS_I(inode)->root;
6058
6059         if (btrfs_root_readonly(root))
6060                 return -EROFS;
6061
6062         if (flags & S_VERSION)
6063                 inode_inc_iversion(inode);
6064         if (flags & S_CTIME)
6065                 inode->i_ctime = *now;
6066         if (flags & S_MTIME)
6067                 inode->i_mtime = *now;
6068         if (flags & S_ATIME)
6069                 inode->i_atime = *now;
6070         return btrfs_dirty_inode(inode);
6071 }
6072
6073 /*
6074  * find the highest existing sequence number in a directory
6075  * and then set the in-memory index_cnt variable to reflect
6076  * free sequence numbers
6077  */
6078 static int btrfs_set_inode_index_count(struct inode *inode)
6079 {
6080         struct btrfs_root *root = BTRFS_I(inode)->root;
6081         struct btrfs_key key, found_key;
6082         struct btrfs_path *path;
6083         struct extent_buffer *leaf;
6084         int ret;
6085
6086         key.objectid = btrfs_ino(inode);
6087         key.type = BTRFS_DIR_INDEX_KEY;
6088         key.offset = (u64)-1;
6089
6090         path = btrfs_alloc_path();
6091         if (!path)
6092                 return -ENOMEM;
6093
6094         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6095         if (ret < 0)
6096                 goto out;
6097         /* FIXME: we should be able to handle this */
6098         if (ret == 0)
6099                 goto out;
6100         ret = 0;
6101
6102         /*
6103          * MAGIC NUMBER EXPLANATION:
6104          * since we search a directory based on f_pos we have to start at 2
6105          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6106          * else has to start at 2
6107          */
6108         if (path->slots[0] == 0) {
6109                 BTRFS_I(inode)->index_cnt = 2;
6110                 goto out;
6111         }
6112
6113         path->slots[0]--;
6114
6115         leaf = path->nodes[0];
6116         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6117
6118         if (found_key.objectid != btrfs_ino(inode) ||
6119             found_key.type != BTRFS_DIR_INDEX_KEY) {
6120                 BTRFS_I(inode)->index_cnt = 2;
6121                 goto out;
6122         }
6123
6124         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6125 out:
6126         btrfs_free_path(path);
6127         return ret;
6128 }
6129
6130 /*
6131  * helper to find a free sequence number in a given directory.  This current
6132  * code is very simple, later versions will do smarter things in the btree
6133  */
6134 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6135 {
6136         int ret = 0;
6137
6138         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6139                 ret = btrfs_inode_delayed_dir_index_count(dir);
6140                 if (ret) {
6141                         ret = btrfs_set_inode_index_count(dir);
6142                         if (ret)
6143                                 return ret;
6144                 }
6145         }
6146
6147         *index = BTRFS_I(dir)->index_cnt;
6148         BTRFS_I(dir)->index_cnt++;
6149
6150         return ret;
6151 }
6152
6153 static int btrfs_insert_inode_locked(struct inode *inode)
6154 {
6155         struct btrfs_iget_args args;
6156         args.location = &BTRFS_I(inode)->location;
6157         args.root = BTRFS_I(inode)->root;
6158
6159         return insert_inode_locked4(inode,
6160                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6161                    btrfs_find_actor, &args);
6162 }
6163
6164 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6165                                      struct btrfs_root *root,
6166                                      struct inode *dir,
6167                                      const char *name, int name_len,
6168                                      u64 ref_objectid, u64 objectid,
6169                                      umode_t mode, u64 *index)
6170 {
6171         struct inode *inode;
6172         struct btrfs_inode_item *inode_item;
6173         struct btrfs_key *location;
6174         struct btrfs_path *path;
6175         struct btrfs_inode_ref *ref;
6176         struct btrfs_key key[2];
6177         u32 sizes[2];
6178         int nitems = name ? 2 : 1;
6179         unsigned long ptr;
6180         int ret;
6181
6182         path = btrfs_alloc_path();
6183         if (!path)
6184                 return ERR_PTR(-ENOMEM);
6185
6186         inode = new_inode(root->fs_info->sb);
6187         if (!inode) {
6188                 btrfs_free_path(path);
6189                 return ERR_PTR(-ENOMEM);
6190         }
6191
6192         /*
6193          * O_TMPFILE, set link count to 0, so that after this point,
6194          * we fill in an inode item with the correct link count.
6195          */
6196         if (!name)
6197                 set_nlink(inode, 0);
6198
6199         /*
6200          * we have to initialize this early, so we can reclaim the inode
6201          * number if we fail afterwards in this function.
6202          */
6203         inode->i_ino = objectid;
6204
6205         if (dir && name) {
6206                 trace_btrfs_inode_request(dir);
6207
6208                 ret = btrfs_set_inode_index(dir, index);
6209                 if (ret) {
6210                         btrfs_free_path(path);
6211                         iput(inode);
6212                         return ERR_PTR(ret);
6213                 }
6214         } else if (dir) {
6215                 *index = 0;
6216         }
6217         /*
6218          * index_cnt is ignored for everything but a dir,
6219          * btrfs_get_inode_index_count has an explanation for the magic
6220          * number
6221          */
6222         BTRFS_I(inode)->index_cnt = 2;
6223         BTRFS_I(inode)->dir_index = *index;
6224         BTRFS_I(inode)->root = root;
6225         BTRFS_I(inode)->generation = trans->transid;
6226         inode->i_generation = BTRFS_I(inode)->generation;
6227
6228         /*
6229          * We could have gotten an inode number from somebody who was fsynced
6230          * and then removed in this same transaction, so let's just set full
6231          * sync since it will be a full sync anyway and this will blow away the
6232          * old info in the log.
6233          */
6234         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6235
6236         key[0].objectid = objectid;
6237         key[0].type = BTRFS_INODE_ITEM_KEY;
6238         key[0].offset = 0;
6239
6240         sizes[0] = sizeof(struct btrfs_inode_item);
6241
6242         if (name) {
6243                 /*
6244                  * Start new inodes with an inode_ref. This is slightly more
6245                  * efficient for small numbers of hard links since they will
6246                  * be packed into one item. Extended refs will kick in if we
6247                  * add more hard links than can fit in the ref item.
6248                  */
6249                 key[1].objectid = objectid;
6250                 key[1].type = BTRFS_INODE_REF_KEY;
6251                 key[1].offset = ref_objectid;
6252
6253                 sizes[1] = name_len + sizeof(*ref);
6254         }
6255
6256         location = &BTRFS_I(inode)->location;
6257         location->objectid = objectid;
6258         location->offset = 0;
6259         location->type = BTRFS_INODE_ITEM_KEY;
6260
6261         ret = btrfs_insert_inode_locked(inode);
6262         if (ret < 0)
6263                 goto fail;
6264
6265         path->leave_spinning = 1;
6266         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6267         if (ret != 0)
6268                 goto fail_unlock;
6269
6270         inode_init_owner(inode, dir, mode);
6271         inode_set_bytes(inode, 0);
6272
6273         inode->i_mtime = current_time(inode);
6274         inode->i_atime = inode->i_mtime;
6275         inode->i_ctime = inode->i_mtime;
6276         BTRFS_I(inode)->i_otime = inode->i_mtime;
6277
6278         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6279                                   struct btrfs_inode_item);
6280         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6281                              sizeof(*inode_item));
6282         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6283
6284         if (name) {
6285                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6286                                      struct btrfs_inode_ref);
6287                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6288                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6289                 ptr = (unsigned long)(ref + 1);
6290                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6291         }
6292
6293         btrfs_mark_buffer_dirty(path->nodes[0]);
6294         btrfs_free_path(path);
6295
6296         btrfs_inherit_iflags(inode, dir);
6297
6298         if (S_ISREG(mode)) {
6299                 if (btrfs_test_opt(root->fs_info, NODATASUM))
6300                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6301                 if (btrfs_test_opt(root->fs_info, NODATACOW))
6302                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6303                                 BTRFS_INODE_NODATASUM;
6304         }
6305
6306         inode_tree_add(inode);
6307
6308         trace_btrfs_inode_new(inode);
6309         btrfs_set_inode_last_trans(trans, inode);
6310
6311         btrfs_update_root_times(trans, root);
6312
6313         ret = btrfs_inode_inherit_props(trans, inode, dir);
6314         if (ret)
6315                 btrfs_err(root->fs_info,
6316                           "error inheriting props for ino %llu (root %llu): %d",
6317                           btrfs_ino(inode), root->root_key.objectid, ret);
6318
6319         return inode;
6320
6321 fail_unlock:
6322         unlock_new_inode(inode);
6323 fail:
6324         if (dir && name)
6325                 BTRFS_I(dir)->index_cnt--;
6326         btrfs_free_path(path);
6327         iput(inode);
6328         return ERR_PTR(ret);
6329 }
6330
6331 static inline u8 btrfs_inode_type(struct inode *inode)
6332 {
6333         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6334 }
6335
6336 /*
6337  * utility function to add 'inode' into 'parent_inode' with
6338  * a give name and a given sequence number.
6339  * if 'add_backref' is true, also insert a backref from the
6340  * inode to the parent directory.
6341  */
6342 int btrfs_add_link(struct btrfs_trans_handle *trans,
6343                    struct inode *parent_inode, struct inode *inode,
6344                    const char *name, int name_len, int add_backref, u64 index)
6345 {
6346         int ret = 0;
6347         struct btrfs_key key;
6348         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6349         u64 ino = btrfs_ino(inode);
6350         u64 parent_ino = btrfs_ino(parent_inode);
6351
6352         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6353                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6354         } else {
6355                 key.objectid = ino;
6356                 key.type = BTRFS_INODE_ITEM_KEY;
6357                 key.offset = 0;
6358         }
6359
6360         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6361                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6362                                          key.objectid, root->root_key.objectid,
6363                                          parent_ino, index, name, name_len);
6364         } else if (add_backref) {
6365                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6366                                              parent_ino, index);
6367         }
6368
6369         /* Nothing to clean up yet */
6370         if (ret)
6371                 return ret;
6372
6373         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6374                                     parent_inode, &key,
6375                                     btrfs_inode_type(inode), index);
6376         if (ret == -EEXIST || ret == -EOVERFLOW)
6377                 goto fail_dir_item;
6378         else if (ret) {
6379                 btrfs_abort_transaction(trans, ret);
6380                 return ret;
6381         }
6382
6383         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6384                            name_len * 2);
6385         inode_inc_iversion(parent_inode);
6386         parent_inode->i_mtime = parent_inode->i_ctime =
6387                 current_time(parent_inode);
6388         ret = btrfs_update_inode(trans, root, parent_inode);
6389         if (ret)
6390                 btrfs_abort_transaction(trans, ret);
6391         return ret;
6392
6393 fail_dir_item:
6394         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6395                 u64 local_index;
6396                 int err;
6397                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6398                                  key.objectid, root->root_key.objectid,
6399                                  parent_ino, &local_index, name, name_len);
6400
6401         } else if (add_backref) {
6402                 u64 local_index;
6403                 int err;
6404
6405                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6406                                           ino, parent_ino, &local_index);
6407         }
6408         return ret;
6409 }
6410
6411 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6412                             struct inode *dir, struct dentry *dentry,
6413                             struct inode *inode, int backref, u64 index)
6414 {
6415         int err = btrfs_add_link(trans, dir, inode,
6416                                  dentry->d_name.name, dentry->d_name.len,
6417                                  backref, index);
6418         if (err > 0)
6419                 err = -EEXIST;
6420         return err;
6421 }
6422
6423 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6424                         umode_t mode, dev_t rdev)
6425 {
6426         struct btrfs_trans_handle *trans;
6427         struct btrfs_root *root = BTRFS_I(dir)->root;
6428         struct inode *inode = NULL;
6429         int err;
6430         int drop_inode = 0;
6431         u64 objectid;
6432         u64 index = 0;
6433
6434         /*
6435          * 2 for inode item and ref
6436          * 2 for dir items
6437          * 1 for xattr if selinux is on
6438          */
6439         trans = btrfs_start_transaction(root, 5);
6440         if (IS_ERR(trans))
6441                 return PTR_ERR(trans);
6442
6443         err = btrfs_find_free_ino(root, &objectid);
6444         if (err)
6445                 goto out_unlock;
6446
6447         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6448                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6449                                 mode, &index);
6450         if (IS_ERR(inode)) {
6451                 err = PTR_ERR(inode);
6452                 goto out_unlock;
6453         }
6454
6455         /*
6456         * If the active LSM wants to access the inode during
6457         * d_instantiate it needs these. Smack checks to see
6458         * if the filesystem supports xattrs by looking at the
6459         * ops vector.
6460         */
6461         inode->i_op = &btrfs_special_inode_operations;
6462         init_special_inode(inode, inode->i_mode, rdev);
6463
6464         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6465         if (err)
6466                 goto out_unlock_inode;
6467
6468         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6469         if (err) {
6470                 goto out_unlock_inode;
6471         } else {
6472                 btrfs_update_inode(trans, root, inode);
6473                 unlock_new_inode(inode);
6474                 d_instantiate(dentry, inode);
6475         }
6476
6477 out_unlock:
6478         btrfs_end_transaction(trans, root);
6479         btrfs_balance_delayed_items(root);
6480         btrfs_btree_balance_dirty(root);
6481         if (drop_inode) {
6482                 inode_dec_link_count(inode);
6483                 iput(inode);
6484         }
6485         return err;
6486
6487 out_unlock_inode:
6488         drop_inode = 1;
6489         unlock_new_inode(inode);
6490         goto out_unlock;
6491
6492 }
6493
6494 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6495                         umode_t mode, bool excl)
6496 {
6497         struct btrfs_trans_handle *trans;
6498         struct btrfs_root *root = BTRFS_I(dir)->root;
6499         struct inode *inode = NULL;
6500         int drop_inode_on_err = 0;
6501         int err;
6502         u64 objectid;
6503         u64 index = 0;
6504
6505         /*
6506          * 2 for inode item and ref
6507          * 2 for dir items
6508          * 1 for xattr if selinux is on
6509          */
6510         trans = btrfs_start_transaction(root, 5);
6511         if (IS_ERR(trans))
6512                 return PTR_ERR(trans);
6513
6514         err = btrfs_find_free_ino(root, &objectid);
6515         if (err)
6516                 goto out_unlock;
6517
6518         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6519                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6520                                 mode, &index);
6521         if (IS_ERR(inode)) {
6522                 err = PTR_ERR(inode);
6523                 goto out_unlock;
6524         }
6525         drop_inode_on_err = 1;
6526         /*
6527         * If the active LSM wants to access the inode during
6528         * d_instantiate it needs these. Smack checks to see
6529         * if the filesystem supports xattrs by looking at the
6530         * ops vector.
6531         */
6532         inode->i_fop = &btrfs_file_operations;
6533         inode->i_op = &btrfs_file_inode_operations;
6534         inode->i_mapping->a_ops = &btrfs_aops;
6535
6536         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6537         if (err)
6538                 goto out_unlock_inode;
6539
6540         err = btrfs_update_inode(trans, root, inode);
6541         if (err)
6542                 goto out_unlock_inode;
6543
6544         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6545         if (err)
6546                 goto out_unlock_inode;
6547
6548         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6549         unlock_new_inode(inode);
6550         d_instantiate(dentry, inode);
6551
6552 out_unlock:
6553         btrfs_end_transaction(trans, root);
6554         if (err && drop_inode_on_err) {
6555                 inode_dec_link_count(inode);
6556                 iput(inode);
6557         }
6558         btrfs_balance_delayed_items(root);
6559         btrfs_btree_balance_dirty(root);
6560         return err;
6561
6562 out_unlock_inode:
6563         unlock_new_inode(inode);
6564         goto out_unlock;
6565
6566 }
6567
6568 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6569                       struct dentry *dentry)
6570 {
6571         struct btrfs_trans_handle *trans = NULL;
6572         struct btrfs_root *root = BTRFS_I(dir)->root;
6573         struct inode *inode = d_inode(old_dentry);
6574         u64 index;
6575         int err;
6576         int drop_inode = 0;
6577
6578         /* do not allow sys_link's with other subvols of the same device */
6579         if (root->objectid != BTRFS_I(inode)->root->objectid)
6580                 return -EXDEV;
6581
6582         if (inode->i_nlink >= BTRFS_LINK_MAX)
6583                 return -EMLINK;
6584
6585         err = btrfs_set_inode_index(dir, &index);
6586         if (err)
6587                 goto fail;
6588
6589         /*
6590          * 2 items for inode and inode ref
6591          * 2 items for dir items
6592          * 1 item for parent inode
6593          */
6594         trans = btrfs_start_transaction(root, 5);
6595         if (IS_ERR(trans)) {
6596                 err = PTR_ERR(trans);
6597                 trans = NULL;
6598                 goto fail;
6599         }
6600
6601         /* There are several dir indexes for this inode, clear the cache. */
6602         BTRFS_I(inode)->dir_index = 0ULL;
6603         inc_nlink(inode);
6604         inode_inc_iversion(inode);
6605         inode->i_ctime = current_time(inode);
6606         ihold(inode);
6607         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6608
6609         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6610
6611         if (err) {
6612                 drop_inode = 1;
6613         } else {
6614                 struct dentry *parent = dentry->d_parent;
6615                 err = btrfs_update_inode(trans, root, inode);
6616                 if (err)
6617                         goto fail;
6618                 if (inode->i_nlink == 1) {
6619                         /*
6620                          * If new hard link count is 1, it's a file created
6621                          * with open(2) O_TMPFILE flag.
6622                          */
6623                         err = btrfs_orphan_del(trans, inode);
6624                         if (err)
6625                                 goto fail;
6626                 }
6627                 d_instantiate(dentry, inode);
6628                 btrfs_log_new_name(trans, inode, NULL, parent);
6629         }
6630
6631         btrfs_balance_delayed_items(root);
6632 fail:
6633         if (trans)
6634                 btrfs_end_transaction(trans, root);
6635         if (drop_inode) {
6636                 inode_dec_link_count(inode);
6637                 iput(inode);
6638         }
6639         btrfs_btree_balance_dirty(root);
6640         return err;
6641 }
6642
6643 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6644 {
6645         struct inode *inode = NULL;
6646         struct btrfs_trans_handle *trans;
6647         struct btrfs_root *root = BTRFS_I(dir)->root;
6648         int err = 0;
6649         int drop_on_err = 0;
6650         u64 objectid = 0;
6651         u64 index = 0;
6652
6653         /*
6654          * 2 items for inode and ref
6655          * 2 items for dir items
6656          * 1 for xattr if selinux is on
6657          */
6658         trans = btrfs_start_transaction(root, 5);
6659         if (IS_ERR(trans))
6660                 return PTR_ERR(trans);
6661
6662         err = btrfs_find_free_ino(root, &objectid);
6663         if (err)
6664                 goto out_fail;
6665
6666         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6667                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6668                                 S_IFDIR | mode, &index);
6669         if (IS_ERR(inode)) {
6670                 err = PTR_ERR(inode);
6671                 goto out_fail;
6672         }
6673
6674         drop_on_err = 1;
6675         /* these must be set before we unlock the inode */
6676         inode->i_op = &btrfs_dir_inode_operations;
6677         inode->i_fop = &btrfs_dir_file_operations;
6678
6679         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6680         if (err)
6681                 goto out_fail_inode;
6682
6683         btrfs_i_size_write(inode, 0);
6684         err = btrfs_update_inode(trans, root, inode);
6685         if (err)
6686                 goto out_fail_inode;
6687
6688         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6689                              dentry->d_name.len, 0, index);
6690         if (err)
6691                 goto out_fail_inode;
6692
6693         d_instantiate(dentry, inode);
6694         /*
6695          * mkdir is special.  We're unlocking after we call d_instantiate
6696          * to avoid a race with nfsd calling d_instantiate.
6697          */
6698         unlock_new_inode(inode);
6699         drop_on_err = 0;
6700
6701 out_fail:
6702         btrfs_end_transaction(trans, root);
6703         if (drop_on_err) {
6704                 inode_dec_link_count(inode);
6705                 iput(inode);
6706         }
6707         btrfs_balance_delayed_items(root);
6708         btrfs_btree_balance_dirty(root);
6709         return err;
6710
6711 out_fail_inode:
6712         unlock_new_inode(inode);
6713         goto out_fail;
6714 }
6715
6716 /* Find next extent map of a given extent map, caller needs to ensure locks */
6717 static struct extent_map *next_extent_map(struct extent_map *em)
6718 {
6719         struct rb_node *next;
6720
6721         next = rb_next(&em->rb_node);
6722         if (!next)
6723                 return NULL;
6724         return container_of(next, struct extent_map, rb_node);
6725 }
6726
6727 static struct extent_map *prev_extent_map(struct extent_map *em)
6728 {
6729         struct rb_node *prev;
6730
6731         prev = rb_prev(&em->rb_node);
6732         if (!prev)
6733                 return NULL;
6734         return container_of(prev, struct extent_map, rb_node);
6735 }
6736
6737 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6738  * the existing extent is the nearest extent to map_start,
6739  * and an extent that you want to insert, deal with overlap and insert
6740  * the best fitted new extent into the tree.
6741  */
6742 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6743                                 struct extent_map *existing,
6744                                 struct extent_map *em,
6745                                 u64 map_start)
6746 {
6747         struct extent_map *prev;
6748         struct extent_map *next;
6749         u64 start;
6750         u64 end;
6751         u64 start_diff;
6752
6753         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6754
6755         if (existing->start > map_start) {
6756                 next = existing;
6757                 prev = prev_extent_map(next);
6758         } else {
6759                 prev = existing;
6760                 next = next_extent_map(prev);
6761         }
6762
6763         start = prev ? extent_map_end(prev) : em->start;
6764         start = max_t(u64, start, em->start);
6765         end = next ? next->start : extent_map_end(em);
6766         end = min_t(u64, end, extent_map_end(em));
6767         start_diff = start - em->start;
6768         em->start = start;
6769         em->len = end - start;
6770         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6771             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6772                 em->block_start += start_diff;
6773                 em->block_len -= start_diff;
6774         }
6775         return add_extent_mapping(em_tree, em, 0);
6776 }
6777
6778 static noinline int uncompress_inline(struct btrfs_path *path,
6779                                       struct page *page,
6780                                       size_t pg_offset, u64 extent_offset,
6781                                       struct btrfs_file_extent_item *item)
6782 {
6783         int ret;
6784         struct extent_buffer *leaf = path->nodes[0];
6785         char *tmp;
6786         size_t max_size;
6787         unsigned long inline_size;
6788         unsigned long ptr;
6789         int compress_type;
6790
6791         WARN_ON(pg_offset != 0);
6792         compress_type = btrfs_file_extent_compression(leaf, item);
6793         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6794         inline_size = btrfs_file_extent_inline_item_len(leaf,
6795                                         btrfs_item_nr(path->slots[0]));
6796         tmp = kmalloc(inline_size, GFP_NOFS);
6797         if (!tmp)
6798                 return -ENOMEM;
6799         ptr = btrfs_file_extent_inline_start(item);
6800
6801         read_extent_buffer(leaf, tmp, ptr, inline_size);
6802
6803         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6804         ret = btrfs_decompress(compress_type, tmp, page,
6805                                extent_offset, inline_size, max_size);
6806         kfree(tmp);
6807         return ret;
6808 }
6809
6810 /*
6811  * a bit scary, this does extent mapping from logical file offset to the disk.
6812  * the ugly parts come from merging extents from the disk with the in-ram
6813  * representation.  This gets more complex because of the data=ordered code,
6814  * where the in-ram extents might be locked pending data=ordered completion.
6815  *
6816  * This also copies inline extents directly into the page.
6817  */
6818
6819 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6820                                     size_t pg_offset, u64 start, u64 len,
6821                                     int create)
6822 {
6823         int ret;
6824         int err = 0;
6825         u64 extent_start = 0;
6826         u64 extent_end = 0;
6827         u64 objectid = btrfs_ino(inode);
6828         u32 found_type;
6829         struct btrfs_path *path = NULL;
6830         struct btrfs_root *root = BTRFS_I(inode)->root;
6831         struct btrfs_file_extent_item *item;
6832         struct extent_buffer *leaf;
6833         struct btrfs_key found_key;
6834         struct extent_map *em = NULL;
6835         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6836         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6837         struct btrfs_trans_handle *trans = NULL;
6838         const bool new_inline = !page || create;
6839
6840 again:
6841         read_lock(&em_tree->lock);
6842         em = lookup_extent_mapping(em_tree, start, len);
6843         if (em)
6844                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6845         read_unlock(&em_tree->lock);
6846
6847         if (em) {
6848                 if (em->start > start || em->start + em->len <= start)
6849                         free_extent_map(em);
6850                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6851                         free_extent_map(em);
6852                 else
6853                         goto out;
6854         }
6855         em = alloc_extent_map();
6856         if (!em) {
6857                 err = -ENOMEM;
6858                 goto out;
6859         }
6860         em->bdev = root->fs_info->fs_devices->latest_bdev;
6861         em->start = EXTENT_MAP_HOLE;
6862         em->orig_start = EXTENT_MAP_HOLE;
6863         em->len = (u64)-1;
6864         em->block_len = (u64)-1;
6865
6866         if (!path) {
6867                 path = btrfs_alloc_path();
6868                 if (!path) {
6869                         err = -ENOMEM;
6870                         goto out;
6871                 }
6872                 /*
6873                  * Chances are we'll be called again, so go ahead and do
6874                  * readahead
6875                  */
6876                 path->reada = READA_FORWARD;
6877         }
6878
6879         ret = btrfs_lookup_file_extent(trans, root, path,
6880                                        objectid, start, trans != NULL);
6881         if (ret < 0) {
6882                 err = ret;
6883                 goto out;
6884         }
6885
6886         if (ret != 0) {
6887                 if (path->slots[0] == 0)
6888                         goto not_found;
6889                 path->slots[0]--;
6890         }
6891
6892         leaf = path->nodes[0];
6893         item = btrfs_item_ptr(leaf, path->slots[0],
6894                               struct btrfs_file_extent_item);
6895         /* are we inside the extent that was found? */
6896         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6897         found_type = found_key.type;
6898         if (found_key.objectid != objectid ||
6899             found_type != BTRFS_EXTENT_DATA_KEY) {
6900                 /*
6901                  * If we backup past the first extent we want to move forward
6902                  * and see if there is an extent in front of us, otherwise we'll
6903                  * say there is a hole for our whole search range which can
6904                  * cause problems.
6905                  */
6906                 extent_end = start;
6907                 goto next;
6908         }
6909
6910         found_type = btrfs_file_extent_type(leaf, item);
6911         extent_start = found_key.offset;
6912         if (found_type == BTRFS_FILE_EXTENT_REG ||
6913             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6914                 extent_end = extent_start +
6915                        btrfs_file_extent_num_bytes(leaf, item);
6916         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6917                 size_t size;
6918                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6919                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6920         }
6921 next:
6922         if (start >= extent_end) {
6923                 path->slots[0]++;
6924                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6925                         ret = btrfs_next_leaf(root, path);
6926                         if (ret < 0) {
6927                                 err = ret;
6928                                 goto out;
6929                         }
6930                         if (ret > 0)
6931                                 goto not_found;
6932                         leaf = path->nodes[0];
6933                 }
6934                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6935                 if (found_key.objectid != objectid ||
6936                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6937                         goto not_found;
6938                 if (start + len <= found_key.offset)
6939                         goto not_found;
6940                 if (start > found_key.offset)
6941                         goto next;
6942                 em->start = start;
6943                 em->orig_start = start;
6944                 em->len = found_key.offset - start;
6945                 goto not_found_em;
6946         }
6947
6948         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6949
6950         if (found_type == BTRFS_FILE_EXTENT_REG ||
6951             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6952                 goto insert;
6953         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6954                 unsigned long ptr;
6955                 char *map;
6956                 size_t size;
6957                 size_t extent_offset;
6958                 size_t copy_size;
6959
6960                 if (new_inline)
6961                         goto out;
6962
6963                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6964                 extent_offset = page_offset(page) + pg_offset - extent_start;
6965                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6966                                   size - extent_offset);
6967                 em->start = extent_start + extent_offset;
6968                 em->len = ALIGN(copy_size, root->sectorsize);
6969                 em->orig_block_len = em->len;
6970                 em->orig_start = em->start;
6971                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6972                 if (create == 0 && !PageUptodate(page)) {
6973                         if (btrfs_file_extent_compression(leaf, item) !=
6974                             BTRFS_COMPRESS_NONE) {
6975                                 ret = uncompress_inline(path, page, pg_offset,
6976                                                         extent_offset, item);
6977                                 if (ret) {
6978                                         err = ret;
6979                                         goto out;
6980                                 }
6981                         } else {
6982                                 map = kmap(page);
6983                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6984                                                    copy_size);
6985                                 if (pg_offset + copy_size < PAGE_SIZE) {
6986                                         memset(map + pg_offset + copy_size, 0,
6987                                                PAGE_SIZE - pg_offset -
6988                                                copy_size);
6989                                 }
6990                                 kunmap(page);
6991                         }
6992                         flush_dcache_page(page);
6993                 } else if (create && PageUptodate(page)) {
6994                         BUG();
6995                         if (!trans) {
6996                                 kunmap(page);
6997                                 free_extent_map(em);
6998                                 em = NULL;
6999
7000                                 btrfs_release_path(path);
7001                                 trans = btrfs_join_transaction(root);
7002
7003                                 if (IS_ERR(trans))
7004                                         return ERR_CAST(trans);
7005                                 goto again;
7006                         }
7007                         map = kmap(page);
7008                         write_extent_buffer(leaf, map + pg_offset, ptr,
7009                                             copy_size);
7010                         kunmap(page);
7011                         btrfs_mark_buffer_dirty(leaf);
7012                 }
7013                 set_extent_uptodate(io_tree, em->start,
7014                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7015                 goto insert;
7016         }
7017 not_found:
7018         em->start = start;
7019         em->orig_start = start;
7020         em->len = len;
7021 not_found_em:
7022         em->block_start = EXTENT_MAP_HOLE;
7023         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7024 insert:
7025         btrfs_release_path(path);
7026         if (em->start > start || extent_map_end(em) <= start) {
7027                 btrfs_err(root->fs_info,
7028                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7029                           em->start, em->len, start, len);
7030                 err = -EIO;
7031                 goto out;
7032         }
7033
7034         err = 0;
7035         write_lock(&em_tree->lock);
7036         ret = add_extent_mapping(em_tree, em, 0);
7037         /* it is possible that someone inserted the extent into the tree
7038          * while we had the lock dropped.  It is also possible that
7039          * an overlapping map exists in the tree
7040          */
7041         if (ret == -EEXIST) {
7042                 struct extent_map *existing;
7043
7044                 ret = 0;
7045
7046                 existing = search_extent_mapping(em_tree, start, len);
7047                 /*
7048                  * existing will always be non-NULL, since there must be
7049                  * extent causing the -EEXIST.
7050                  */
7051                 if (existing->start == em->start &&
7052                     extent_map_end(existing) == extent_map_end(em) &&
7053                     em->block_start == existing->block_start) {
7054                         /*
7055                          * these two extents are the same, it happens
7056                          * with inlines especially
7057                          */
7058                         free_extent_map(em);
7059                         em = existing;
7060                         err = 0;
7061
7062                 } else if (start >= extent_map_end(existing) ||
7063                     start <= existing->start) {
7064                         /*
7065                          * The existing extent map is the one nearest to
7066                          * the [start, start + len) range which overlaps
7067                          */
7068                         err = merge_extent_mapping(em_tree, existing,
7069                                                    em, start);
7070                         free_extent_map(existing);
7071                         if (err) {
7072                                 free_extent_map(em);
7073                                 em = NULL;
7074                         }
7075                 } else {
7076                         free_extent_map(em);
7077                         em = existing;
7078                         err = 0;
7079                 }
7080         }
7081         write_unlock(&em_tree->lock);
7082 out:
7083
7084         trace_btrfs_get_extent(root, em);
7085
7086         btrfs_free_path(path);
7087         if (trans) {
7088                 ret = btrfs_end_transaction(trans, root);
7089                 if (!err)
7090                         err = ret;
7091         }
7092         if (err) {
7093                 free_extent_map(em);
7094                 return ERR_PTR(err);
7095         }
7096         BUG_ON(!em); /* Error is always set */
7097         return em;
7098 }
7099
7100 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7101                                            size_t pg_offset, u64 start, u64 len,
7102                                            int create)
7103 {
7104         struct extent_map *em;
7105         struct extent_map *hole_em = NULL;
7106         u64 range_start = start;
7107         u64 end;
7108         u64 found;
7109         u64 found_end;
7110         int err = 0;
7111
7112         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7113         if (IS_ERR(em))
7114                 return em;
7115         if (em) {
7116                 /*
7117                  * if our em maps to
7118                  * -  a hole or
7119                  * -  a pre-alloc extent,
7120                  * there might actually be delalloc bytes behind it.
7121                  */
7122                 if (em->block_start != EXTENT_MAP_HOLE &&
7123                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7124                         return em;
7125                 else
7126                         hole_em = em;
7127         }
7128
7129         /* check to see if we've wrapped (len == -1 or similar) */
7130         end = start + len;
7131         if (end < start)
7132                 end = (u64)-1;
7133         else
7134                 end -= 1;
7135
7136         em = NULL;
7137
7138         /* ok, we didn't find anything, lets look for delalloc */
7139         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7140                                  end, len, EXTENT_DELALLOC, 1);
7141         found_end = range_start + found;
7142         if (found_end < range_start)
7143                 found_end = (u64)-1;
7144
7145         /*
7146          * we didn't find anything useful, return
7147          * the original results from get_extent()
7148          */
7149         if (range_start > end || found_end <= start) {
7150                 em = hole_em;
7151                 hole_em = NULL;
7152                 goto out;
7153         }
7154
7155         /* adjust the range_start to make sure it doesn't
7156          * go backwards from the start they passed in
7157          */
7158         range_start = max(start, range_start);
7159         found = found_end - range_start;
7160
7161         if (found > 0) {
7162                 u64 hole_start = start;
7163                 u64 hole_len = len;
7164
7165                 em = alloc_extent_map();
7166                 if (!em) {
7167                         err = -ENOMEM;
7168                         goto out;
7169                 }
7170                 /*
7171                  * when btrfs_get_extent can't find anything it
7172                  * returns one huge hole
7173                  *
7174                  * make sure what it found really fits our range, and
7175                  * adjust to make sure it is based on the start from
7176                  * the caller
7177                  */
7178                 if (hole_em) {
7179                         u64 calc_end = extent_map_end(hole_em);
7180
7181                         if (calc_end <= start || (hole_em->start > end)) {
7182                                 free_extent_map(hole_em);
7183                                 hole_em = NULL;
7184                         } else {
7185                                 hole_start = max(hole_em->start, start);
7186                                 hole_len = calc_end - hole_start;
7187                         }
7188                 }
7189                 em->bdev = NULL;
7190                 if (hole_em && range_start > hole_start) {
7191                         /* our hole starts before our delalloc, so we
7192                          * have to return just the parts of the hole
7193                          * that go until  the delalloc starts
7194                          */
7195                         em->len = min(hole_len,
7196                                       range_start - hole_start);
7197                         em->start = hole_start;
7198                         em->orig_start = hole_start;
7199                         /*
7200                          * don't adjust block start at all,
7201                          * it is fixed at EXTENT_MAP_HOLE
7202                          */
7203                         em->block_start = hole_em->block_start;
7204                         em->block_len = hole_len;
7205                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7206                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7207                 } else {
7208                         em->start = range_start;
7209                         em->len = found;
7210                         em->orig_start = range_start;
7211                         em->block_start = EXTENT_MAP_DELALLOC;
7212                         em->block_len = found;
7213                 }
7214         } else if (hole_em) {
7215                 return hole_em;
7216         }
7217 out:
7218
7219         free_extent_map(hole_em);
7220         if (err) {
7221                 free_extent_map(em);
7222                 return ERR_PTR(err);
7223         }
7224         return em;
7225 }
7226
7227 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7228                                                   const u64 start,
7229                                                   const u64 len,
7230                                                   const u64 orig_start,
7231                                                   const u64 block_start,
7232                                                   const u64 block_len,
7233                                                   const u64 orig_block_len,
7234                                                   const u64 ram_bytes,
7235                                                   const int type)
7236 {
7237         struct extent_map *em = NULL;
7238         int ret;
7239
7240         down_read(&BTRFS_I(inode)->dio_sem);
7241         if (type != BTRFS_ORDERED_NOCOW) {
7242                 em = create_pinned_em(inode, start, len, orig_start,
7243                                       block_start, block_len, orig_block_len,
7244                                       ram_bytes, type);
7245                 if (IS_ERR(em))
7246                         goto out;
7247         }
7248         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7249                                            len, block_len, type);
7250         if (ret) {
7251                 if (em) {
7252                         free_extent_map(em);
7253                         btrfs_drop_extent_cache(inode, start,
7254                                                 start + len - 1, 0);
7255                 }
7256                 em = ERR_PTR(ret);
7257         }
7258  out:
7259         up_read(&BTRFS_I(inode)->dio_sem);
7260
7261         return em;
7262 }
7263
7264 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7265                                                   u64 start, u64 len)
7266 {
7267         struct btrfs_root *root = BTRFS_I(inode)->root;
7268         struct extent_map *em;
7269         struct btrfs_key ins;
7270         u64 alloc_hint;
7271         int ret;
7272
7273         alloc_hint = get_extent_allocation_hint(inode, start, len);
7274         ret = btrfs_reserve_extent(root, len, len, root->sectorsize, 0,
7275                                    alloc_hint, &ins, 1, 1);
7276         if (ret)
7277                 return ERR_PTR(ret);
7278
7279         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7280                                      ins.objectid, ins.offset, ins.offset,
7281                                      ins.offset, 0);
7282         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7283         if (IS_ERR(em))
7284                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7285
7286         return em;
7287 }
7288
7289 /*
7290  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7291  * block must be cow'd
7292  */
7293 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7294                               u64 *orig_start, u64 *orig_block_len,
7295                               u64 *ram_bytes)
7296 {
7297         struct btrfs_trans_handle *trans;
7298         struct btrfs_path *path;
7299         int ret;
7300         struct extent_buffer *leaf;
7301         struct btrfs_root *root = BTRFS_I(inode)->root;
7302         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7303         struct btrfs_file_extent_item *fi;
7304         struct btrfs_key key;
7305         u64 disk_bytenr;
7306         u64 backref_offset;
7307         u64 extent_end;
7308         u64 num_bytes;
7309         int slot;
7310         int found_type;
7311         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7312
7313         path = btrfs_alloc_path();
7314         if (!path)
7315                 return -ENOMEM;
7316
7317         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7318                                        offset, 0);
7319         if (ret < 0)
7320                 goto out;
7321
7322         slot = path->slots[0];
7323         if (ret == 1) {
7324                 if (slot == 0) {
7325                         /* can't find the item, must cow */
7326                         ret = 0;
7327                         goto out;
7328                 }
7329                 slot--;
7330         }
7331         ret = 0;
7332         leaf = path->nodes[0];
7333         btrfs_item_key_to_cpu(leaf, &key, slot);
7334         if (key.objectid != btrfs_ino(inode) ||
7335             key.type != BTRFS_EXTENT_DATA_KEY) {
7336                 /* not our file or wrong item type, must cow */
7337                 goto out;
7338         }
7339
7340         if (key.offset > offset) {
7341                 /* Wrong offset, must cow */
7342                 goto out;
7343         }
7344
7345         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7346         found_type = btrfs_file_extent_type(leaf, fi);
7347         if (found_type != BTRFS_FILE_EXTENT_REG &&
7348             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7349                 /* not a regular extent, must cow */
7350                 goto out;
7351         }
7352
7353         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7354                 goto out;
7355
7356         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7357         if (extent_end <= offset)
7358                 goto out;
7359
7360         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7361         if (disk_bytenr == 0)
7362                 goto out;
7363
7364         if (btrfs_file_extent_compression(leaf, fi) ||
7365             btrfs_file_extent_encryption(leaf, fi) ||
7366             btrfs_file_extent_other_encoding(leaf, fi))
7367                 goto out;
7368
7369         backref_offset = btrfs_file_extent_offset(leaf, fi);
7370
7371         if (orig_start) {
7372                 *orig_start = key.offset - backref_offset;
7373                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7374                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7375         }
7376
7377         if (btrfs_extent_readonly(root, disk_bytenr))
7378                 goto out;
7379
7380         num_bytes = min(offset + *len, extent_end) - offset;
7381         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7382                 u64 range_end;
7383
7384                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7385                 ret = test_range_bit(io_tree, offset, range_end,
7386                                      EXTENT_DELALLOC, 0, NULL);
7387                 if (ret) {
7388                         ret = -EAGAIN;
7389                         goto out;
7390                 }
7391         }
7392
7393         btrfs_release_path(path);
7394
7395         /*
7396          * look for other files referencing this extent, if we
7397          * find any we must cow
7398          */
7399         trans = btrfs_join_transaction(root);
7400         if (IS_ERR(trans)) {
7401                 ret = 0;
7402                 goto out;
7403         }
7404
7405         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7406                                     key.offset - backref_offset, disk_bytenr);
7407         btrfs_end_transaction(trans, root);
7408         if (ret) {
7409                 ret = 0;
7410                 goto out;
7411         }
7412
7413         /*
7414          * adjust disk_bytenr and num_bytes to cover just the bytes
7415          * in this extent we are about to write.  If there
7416          * are any csums in that range we have to cow in order
7417          * to keep the csums correct
7418          */
7419         disk_bytenr += backref_offset;
7420         disk_bytenr += offset - key.offset;
7421         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7422                                 goto out;
7423         /*
7424          * all of the above have passed, it is safe to overwrite this extent
7425          * without cow
7426          */
7427         *len = num_bytes;
7428         ret = 1;
7429 out:
7430         btrfs_free_path(path);
7431         return ret;
7432 }
7433
7434 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7435 {
7436         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7437         int found = false;
7438         void **pagep = NULL;
7439         struct page *page = NULL;
7440         int start_idx;
7441         int end_idx;
7442
7443         start_idx = start >> PAGE_SHIFT;
7444
7445         /*
7446          * end is the last byte in the last page.  end == start is legal
7447          */
7448         end_idx = end >> PAGE_SHIFT;
7449
7450         rcu_read_lock();
7451
7452         /* Most of the code in this while loop is lifted from
7453          * find_get_page.  It's been modified to begin searching from a
7454          * page and return just the first page found in that range.  If the
7455          * found idx is less than or equal to the end idx then we know that
7456          * a page exists.  If no pages are found or if those pages are
7457          * outside of the range then we're fine (yay!) */
7458         while (page == NULL &&
7459                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7460                 page = radix_tree_deref_slot(pagep);
7461                 if (unlikely(!page))
7462                         break;
7463
7464                 if (radix_tree_exception(page)) {
7465                         if (radix_tree_deref_retry(page)) {
7466                                 page = NULL;
7467                                 continue;
7468                         }
7469                         /*
7470                          * Otherwise, shmem/tmpfs must be storing a swap entry
7471                          * here as an exceptional entry: so return it without
7472                          * attempting to raise page count.
7473                          */
7474                         page = NULL;
7475                         break; /* TODO: Is this relevant for this use case? */
7476                 }
7477
7478                 if (!page_cache_get_speculative(page)) {
7479                         page = NULL;
7480                         continue;
7481                 }
7482
7483                 /*
7484                  * Has the page moved?
7485                  * This is part of the lockless pagecache protocol. See
7486                  * include/linux/pagemap.h for details.
7487                  */
7488                 if (unlikely(page != *pagep)) {
7489                         put_page(page);
7490                         page = NULL;
7491                 }
7492         }
7493
7494         if (page) {
7495                 if (page->index <= end_idx)
7496                         found = true;
7497                 put_page(page);
7498         }
7499
7500         rcu_read_unlock();
7501         return found;
7502 }
7503
7504 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7505                               struct extent_state **cached_state, int writing)
7506 {
7507         struct btrfs_ordered_extent *ordered;
7508         int ret = 0;
7509
7510         while (1) {
7511                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7512                                  cached_state);
7513                 /*
7514                  * We're concerned with the entire range that we're going to be
7515                  * doing DIO to, so we need to make sure there's no ordered
7516                  * extents in this range.
7517                  */
7518                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7519                                                      lockend - lockstart + 1);
7520
7521                 /*
7522                  * We need to make sure there are no buffered pages in this
7523                  * range either, we could have raced between the invalidate in
7524                  * generic_file_direct_write and locking the extent.  The
7525                  * invalidate needs to happen so that reads after a write do not
7526                  * get stale data.
7527                  */
7528                 if (!ordered &&
7529                     (!writing ||
7530                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7531                         break;
7532
7533                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7534                                      cached_state, GFP_NOFS);
7535
7536                 if (ordered) {
7537                         /*
7538                          * If we are doing a DIO read and the ordered extent we
7539                          * found is for a buffered write, we can not wait for it
7540                          * to complete and retry, because if we do so we can
7541                          * deadlock with concurrent buffered writes on page
7542                          * locks. This happens only if our DIO read covers more
7543                          * than one extent map, if at this point has already
7544                          * created an ordered extent for a previous extent map
7545                          * and locked its range in the inode's io tree, and a
7546                          * concurrent write against that previous extent map's
7547                          * range and this range started (we unlock the ranges
7548                          * in the io tree only when the bios complete and
7549                          * buffered writes always lock pages before attempting
7550                          * to lock range in the io tree).
7551                          */
7552                         if (writing ||
7553                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7554                                 btrfs_start_ordered_extent(inode, ordered, 1);
7555                         else
7556                                 ret = -ENOTBLK;
7557                         btrfs_put_ordered_extent(ordered);
7558                 } else {
7559                         /*
7560                          * We could trigger writeback for this range (and wait
7561                          * for it to complete) and then invalidate the pages for
7562                          * this range (through invalidate_inode_pages2_range()),
7563                          * but that can lead us to a deadlock with a concurrent
7564                          * call to readpages() (a buffered read or a defrag call
7565                          * triggered a readahead) on a page lock due to an
7566                          * ordered dio extent we created before but did not have
7567                          * yet a corresponding bio submitted (whence it can not
7568                          * complete), which makes readpages() wait for that
7569                          * ordered extent to complete while holding a lock on
7570                          * that page.
7571                          */
7572                         ret = -ENOTBLK;
7573                 }
7574
7575                 if (ret)
7576                         break;
7577
7578                 cond_resched();
7579         }
7580
7581         return ret;
7582 }
7583
7584 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7585                                            u64 len, u64 orig_start,
7586                                            u64 block_start, u64 block_len,
7587                                            u64 orig_block_len, u64 ram_bytes,
7588                                            int type)
7589 {
7590         struct extent_map_tree *em_tree;
7591         struct extent_map *em;
7592         struct btrfs_root *root = BTRFS_I(inode)->root;
7593         int ret;
7594
7595         em_tree = &BTRFS_I(inode)->extent_tree;
7596         em = alloc_extent_map();
7597         if (!em)
7598                 return ERR_PTR(-ENOMEM);
7599
7600         em->start = start;
7601         em->orig_start = orig_start;
7602         em->mod_start = start;
7603         em->mod_len = len;
7604         em->len = len;
7605         em->block_len = block_len;
7606         em->block_start = block_start;
7607         em->bdev = root->fs_info->fs_devices->latest_bdev;
7608         em->orig_block_len = orig_block_len;
7609         em->ram_bytes = ram_bytes;
7610         em->generation = -1;
7611         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7612         if (type == BTRFS_ORDERED_PREALLOC)
7613                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7614
7615         do {
7616                 btrfs_drop_extent_cache(inode, em->start,
7617                                 em->start + em->len - 1, 0);
7618                 write_lock(&em_tree->lock);
7619                 ret = add_extent_mapping(em_tree, em, 1);
7620                 write_unlock(&em_tree->lock);
7621         } while (ret == -EEXIST);
7622
7623         if (ret) {
7624                 free_extent_map(em);
7625                 return ERR_PTR(ret);
7626         }
7627
7628         return em;
7629 }
7630
7631 static void adjust_dio_outstanding_extents(struct inode *inode,
7632                                            struct btrfs_dio_data *dio_data,
7633                                            const u64 len)
7634 {
7635         unsigned num_extents;
7636
7637         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7638                                            BTRFS_MAX_EXTENT_SIZE);
7639         /*
7640          * If we have an outstanding_extents count still set then we're
7641          * within our reservation, otherwise we need to adjust our inode
7642          * counter appropriately.
7643          */
7644         if (dio_data->outstanding_extents) {
7645                 dio_data->outstanding_extents -= num_extents;
7646         } else {
7647                 spin_lock(&BTRFS_I(inode)->lock);
7648                 BTRFS_I(inode)->outstanding_extents += num_extents;
7649                 spin_unlock(&BTRFS_I(inode)->lock);
7650         }
7651 }
7652
7653 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7654                                    struct buffer_head *bh_result, int create)
7655 {
7656         struct extent_map *em;
7657         struct btrfs_root *root = BTRFS_I(inode)->root;
7658         struct extent_state *cached_state = NULL;
7659         struct btrfs_dio_data *dio_data = NULL;
7660         u64 start = iblock << inode->i_blkbits;
7661         u64 lockstart, lockend;
7662         u64 len = bh_result->b_size;
7663         int unlock_bits = EXTENT_LOCKED;
7664         int ret = 0;
7665
7666         if (create)
7667                 unlock_bits |= EXTENT_DIRTY;
7668         else
7669                 len = min_t(u64, len, root->sectorsize);
7670
7671         lockstart = start;
7672         lockend = start + len - 1;
7673
7674         if (current->journal_info) {
7675                 /*
7676                  * Need to pull our outstanding extents and set journal_info to NULL so
7677                  * that anything that needs to check if there's a transaction doesn't get
7678                  * confused.
7679                  */
7680                 dio_data = current->journal_info;
7681                 current->journal_info = NULL;
7682         }
7683
7684         /*
7685          * If this errors out it's because we couldn't invalidate pagecache for
7686          * this range and we need to fallback to buffered.
7687          */
7688         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7689                                create)) {
7690                 ret = -ENOTBLK;
7691                 goto err;
7692         }
7693
7694         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7695         if (IS_ERR(em)) {
7696                 ret = PTR_ERR(em);
7697                 goto unlock_err;
7698         }
7699
7700         /*
7701          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7702          * io.  INLINE is special, and we could probably kludge it in here, but
7703          * it's still buffered so for safety lets just fall back to the generic
7704          * buffered path.
7705          *
7706          * For COMPRESSED we _have_ to read the entire extent in so we can
7707          * decompress it, so there will be buffering required no matter what we
7708          * do, so go ahead and fallback to buffered.
7709          *
7710          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7711          * to buffered IO.  Don't blame me, this is the price we pay for using
7712          * the generic code.
7713          */
7714         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7715             em->block_start == EXTENT_MAP_INLINE) {
7716                 free_extent_map(em);
7717                 ret = -ENOTBLK;
7718                 goto unlock_err;
7719         }
7720
7721         /* Just a good old fashioned hole, return */
7722         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7723                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7724                 free_extent_map(em);
7725                 goto unlock_err;
7726         }
7727
7728         /*
7729          * We don't allocate a new extent in the following cases
7730          *
7731          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7732          * existing extent.
7733          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7734          * just use the extent.
7735          *
7736          */
7737         if (!create) {
7738                 len = min(len, em->len - (start - em->start));
7739                 lockstart = start + len;
7740                 goto unlock;
7741         }
7742
7743         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7744             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7745              em->block_start != EXTENT_MAP_HOLE)) {
7746                 int type;
7747                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7748
7749                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7750                         type = BTRFS_ORDERED_PREALLOC;
7751                 else
7752                         type = BTRFS_ORDERED_NOCOW;
7753                 len = min(len, em->len - (start - em->start));
7754                 block_start = em->block_start + (start - em->start);
7755
7756                 if (can_nocow_extent(inode, start, &len, &orig_start,
7757                                      &orig_block_len, &ram_bytes) == 1 &&
7758                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7759                         struct extent_map *em2;
7760
7761                         em2 = btrfs_create_dio_extent(inode, start, len,
7762                                                       orig_start, block_start,
7763                                                       len, orig_block_len,
7764                                                       ram_bytes, type);
7765                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7766                         if (type == BTRFS_ORDERED_PREALLOC) {
7767                                 free_extent_map(em);
7768                                 em = em2;
7769                         }
7770                         if (em2 && IS_ERR(em2)) {
7771                                 ret = PTR_ERR(em2);
7772                                 goto unlock_err;
7773                         }
7774                         /*
7775                          * For inode marked NODATACOW or extent marked PREALLOC,
7776                          * use the existing or preallocated extent, so does not
7777                          * need to adjust btrfs_space_info's bytes_may_use.
7778                          */
7779                         btrfs_free_reserved_data_space_noquota(inode,
7780                                         start, len);
7781                         goto unlock;
7782                 }
7783         }
7784
7785         /*
7786          * this will cow the extent, reset the len in case we changed
7787          * it above
7788          */
7789         len = bh_result->b_size;
7790         free_extent_map(em);
7791         em = btrfs_new_extent_direct(inode, start, len);
7792         if (IS_ERR(em)) {
7793                 ret = PTR_ERR(em);
7794                 goto unlock_err;
7795         }
7796         len = min(len, em->len - (start - em->start));
7797 unlock:
7798         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7799                 inode->i_blkbits;
7800         bh_result->b_size = len;
7801         bh_result->b_bdev = em->bdev;
7802         set_buffer_mapped(bh_result);
7803         if (create) {
7804                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7805                         set_buffer_new(bh_result);
7806
7807                 /*
7808                  * Need to update the i_size under the extent lock so buffered
7809                  * readers will get the updated i_size when we unlock.
7810                  */
7811                 if (start + len > i_size_read(inode))
7812                         i_size_write(inode, start + len);
7813
7814                 adjust_dio_outstanding_extents(inode, dio_data, len);
7815                 WARN_ON(dio_data->reserve < len);
7816                 dio_data->reserve -= len;
7817                 dio_data->unsubmitted_oe_range_end = start + len;
7818                 current->journal_info = dio_data;
7819         }
7820
7821         /*
7822          * In the case of write we need to clear and unlock the entire range,
7823          * in the case of read we need to unlock only the end area that we
7824          * aren't using if there is any left over space.
7825          */
7826         if (lockstart < lockend) {
7827                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7828                                  lockend, unlock_bits, 1, 0,
7829                                  &cached_state, GFP_NOFS);
7830         } else {
7831                 free_extent_state(cached_state);
7832         }
7833
7834         free_extent_map(em);
7835
7836         return 0;
7837
7838 unlock_err:
7839         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7840                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7841 err:
7842         if (dio_data)
7843                 current->journal_info = dio_data;
7844         /*
7845          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7846          * write less data then expected, so that we don't underflow our inode's
7847          * outstanding extents counter.
7848          */
7849         if (create && dio_data)
7850                 adjust_dio_outstanding_extents(inode, dio_data, len);
7851
7852         return ret;
7853 }
7854
7855 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7856                                         int mirror_num)
7857 {
7858         struct btrfs_root *root = BTRFS_I(inode)->root;
7859         int ret;
7860
7861         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7862
7863         bio_get(bio);
7864
7865         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7866                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7867         if (ret)
7868                 goto err;
7869
7870         ret = btrfs_map_bio(root, bio, mirror_num, 0);
7871 err:
7872         bio_put(bio);
7873         return ret;
7874 }
7875
7876 static int btrfs_check_dio_repairable(struct inode *inode,
7877                                       struct bio *failed_bio,
7878                                       struct io_failure_record *failrec,
7879                                       int failed_mirror)
7880 {
7881         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7882         int num_copies;
7883
7884         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7885         if (num_copies == 1) {
7886                 /*
7887                  * we only have a single copy of the data, so don't bother with
7888                  * all the retry and error correction code that follows. no
7889                  * matter what the error is, it is very likely to persist.
7890                  */
7891                 btrfs_debug(fs_info,
7892                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7893                         num_copies, failrec->this_mirror, failed_mirror);
7894                 return 0;
7895         }
7896
7897         failrec->failed_mirror = failed_mirror;
7898         failrec->this_mirror++;
7899         if (failrec->this_mirror == failed_mirror)
7900                 failrec->this_mirror++;
7901
7902         if (failrec->this_mirror > num_copies) {
7903                 btrfs_debug(fs_info,
7904                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7905                         num_copies, failrec->this_mirror, failed_mirror);
7906                 return 0;
7907         }
7908
7909         return 1;
7910 }
7911
7912 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7913                         struct page *page, unsigned int pgoff,
7914                         u64 start, u64 end, int failed_mirror,
7915                         bio_end_io_t *repair_endio, void *repair_arg)
7916 {
7917         struct io_failure_record *failrec;
7918         struct bio *bio;
7919         int isector;
7920         int read_mode;
7921         int ret;
7922
7923         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7924
7925         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7926         if (ret)
7927                 return ret;
7928
7929         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7930                                          failed_mirror);
7931         if (!ret) {
7932                 free_io_failure(inode, failrec);
7933                 return -EIO;
7934         }
7935
7936         if ((failed_bio->bi_vcnt > 1)
7937                 || (failed_bio->bi_io_vec->bv_len
7938                         > BTRFS_I(inode)->root->sectorsize))
7939                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7940         else
7941                 read_mode = READ_SYNC;
7942
7943         isector = start - btrfs_io_bio(failed_bio)->logical;
7944         isector >>= inode->i_sb->s_blocksize_bits;
7945         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7946                                 pgoff, isector, repair_endio, repair_arg);
7947         if (!bio) {
7948                 free_io_failure(inode, failrec);
7949                 return -EIO;
7950         }
7951         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7952
7953         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7954                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7955                     read_mode, failrec->this_mirror, failrec->in_validation);
7956
7957         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7958         if (ret) {
7959                 free_io_failure(inode, failrec);
7960                 bio_put(bio);
7961         }
7962
7963         return ret;
7964 }
7965
7966 struct btrfs_retry_complete {
7967         struct completion done;
7968         struct inode *inode;
7969         u64 start;
7970         int uptodate;
7971 };
7972
7973 static void btrfs_retry_endio_nocsum(struct bio *bio)
7974 {
7975         struct btrfs_retry_complete *done = bio->bi_private;
7976         struct inode *inode;
7977         struct bio_vec *bvec;
7978         int i;
7979
7980         if (bio->bi_error)
7981                 goto end;
7982
7983         ASSERT(bio->bi_vcnt == 1);
7984         inode = bio->bi_io_vec->bv_page->mapping->host;
7985         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7986
7987         done->uptodate = 1;
7988         bio_for_each_segment_all(bvec, bio, i)
7989                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7990 end:
7991         complete(&done->done);
7992         bio_put(bio);
7993 }
7994
7995 static int __btrfs_correct_data_nocsum(struct inode *inode,
7996                                        struct btrfs_io_bio *io_bio)
7997 {
7998         struct btrfs_fs_info *fs_info;
7999         struct bio_vec *bvec;
8000         struct btrfs_retry_complete done;
8001         u64 start;
8002         unsigned int pgoff;
8003         u32 sectorsize;
8004         int nr_sectors;
8005         int i;
8006         int ret;
8007
8008         fs_info = BTRFS_I(inode)->root->fs_info;
8009         sectorsize = BTRFS_I(inode)->root->sectorsize;
8010
8011         start = io_bio->logical;
8012         done.inode = inode;
8013
8014         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8015                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8016                 pgoff = bvec->bv_offset;
8017
8018 next_block_or_try_again:
8019                 done.uptodate = 0;
8020                 done.start = start;
8021                 init_completion(&done.done);
8022
8023                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8024                                 pgoff, start, start + sectorsize - 1,
8025                                 io_bio->mirror_num,
8026                                 btrfs_retry_endio_nocsum, &done);
8027                 if (ret)
8028                         return ret;
8029
8030                 wait_for_completion(&done.done);
8031
8032                 if (!done.uptodate) {
8033                         /* We might have another mirror, so try again */
8034                         goto next_block_or_try_again;
8035                 }
8036
8037                 start += sectorsize;
8038
8039                 if (nr_sectors--) {
8040                         pgoff += sectorsize;
8041                         goto next_block_or_try_again;
8042                 }
8043         }
8044
8045         return 0;
8046 }
8047
8048 static void btrfs_retry_endio(struct bio *bio)
8049 {
8050         struct btrfs_retry_complete *done = bio->bi_private;
8051         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8052         struct inode *inode;
8053         struct bio_vec *bvec;
8054         u64 start;
8055         int uptodate;
8056         int ret;
8057         int i;
8058
8059         if (bio->bi_error)
8060                 goto end;
8061
8062         uptodate = 1;
8063
8064         start = done->start;
8065
8066         ASSERT(bio->bi_vcnt == 1);
8067         inode = bio->bi_io_vec->bv_page->mapping->host;
8068         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8069
8070         bio_for_each_segment_all(bvec, bio, i) {
8071                 ret = __readpage_endio_check(done->inode, io_bio, i,
8072                                         bvec->bv_page, bvec->bv_offset,
8073                                         done->start, bvec->bv_len);
8074                 if (!ret)
8075                         clean_io_failure(done->inode, done->start,
8076                                         bvec->bv_page, bvec->bv_offset);
8077                 else
8078                         uptodate = 0;
8079         }
8080
8081         done->uptodate = uptodate;
8082 end:
8083         complete(&done->done);
8084         bio_put(bio);
8085 }
8086
8087 static int __btrfs_subio_endio_read(struct inode *inode,
8088                                     struct btrfs_io_bio *io_bio, int err)
8089 {
8090         struct btrfs_fs_info *fs_info;
8091         struct bio_vec *bvec;
8092         struct btrfs_retry_complete done;
8093         u64 start;
8094         u64 offset = 0;
8095         u32 sectorsize;
8096         int nr_sectors;
8097         unsigned int pgoff;
8098         int csum_pos;
8099         int i;
8100         int ret;
8101
8102         fs_info = BTRFS_I(inode)->root->fs_info;
8103         sectorsize = BTRFS_I(inode)->root->sectorsize;
8104
8105         err = 0;
8106         start = io_bio->logical;
8107         done.inode = inode;
8108
8109         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8110                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8111
8112                 pgoff = bvec->bv_offset;
8113 next_block:
8114                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8115                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8116                                         bvec->bv_page, pgoff, start,
8117                                         sectorsize);
8118                 if (likely(!ret))
8119                         goto next;
8120 try_again:
8121                 done.uptodate = 0;
8122                 done.start = start;
8123                 init_completion(&done.done);
8124
8125                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8126                                 pgoff, start, start + sectorsize - 1,
8127                                 io_bio->mirror_num,
8128                                 btrfs_retry_endio, &done);
8129                 if (ret) {
8130                         err = ret;
8131                         goto next;
8132                 }
8133
8134                 wait_for_completion(&done.done);
8135
8136                 if (!done.uptodate) {
8137                         /* We might have another mirror, so try again */
8138                         goto try_again;
8139                 }
8140 next:
8141                 offset += sectorsize;
8142                 start += sectorsize;
8143
8144                 ASSERT(nr_sectors);
8145
8146                 if (--nr_sectors) {
8147                         pgoff += sectorsize;
8148                         goto next_block;
8149                 }
8150         }
8151
8152         return err;
8153 }
8154
8155 static int btrfs_subio_endio_read(struct inode *inode,
8156                                   struct btrfs_io_bio *io_bio, int err)
8157 {
8158         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8159
8160         if (skip_csum) {
8161                 if (unlikely(err))
8162                         return __btrfs_correct_data_nocsum(inode, io_bio);
8163                 else
8164                         return 0;
8165         } else {
8166                 return __btrfs_subio_endio_read(inode, io_bio, err);
8167         }
8168 }
8169
8170 static void btrfs_endio_direct_read(struct bio *bio)
8171 {
8172         struct btrfs_dio_private *dip = bio->bi_private;
8173         struct inode *inode = dip->inode;
8174         struct bio *dio_bio;
8175         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8176         int err = bio->bi_error;
8177
8178         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8179                 err = btrfs_subio_endio_read(inode, io_bio, err);
8180
8181         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8182                       dip->logical_offset + dip->bytes - 1);
8183         dio_bio = dip->dio_bio;
8184
8185         kfree(dip);
8186
8187         dio_bio->bi_error = bio->bi_error;
8188         dio_end_io(dio_bio, bio->bi_error);
8189
8190         if (io_bio->end_io)
8191                 io_bio->end_io(io_bio, err);
8192         bio_put(bio);
8193 }
8194
8195 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8196                                                     const u64 offset,
8197                                                     const u64 bytes,
8198                                                     const int uptodate)
8199 {
8200         struct btrfs_root *root = BTRFS_I(inode)->root;
8201         struct btrfs_ordered_extent *ordered = NULL;
8202         u64 ordered_offset = offset;
8203         u64 ordered_bytes = bytes;
8204         int ret;
8205
8206 again:
8207         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8208                                                    &ordered_offset,
8209                                                    ordered_bytes,
8210                                                    uptodate);
8211         if (!ret)
8212                 goto out_test;
8213
8214         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8215                         finish_ordered_fn, NULL, NULL);
8216         btrfs_queue_work(root->fs_info->endio_write_workers,
8217                          &ordered->work);
8218 out_test:
8219         /*
8220          * our bio might span multiple ordered extents.  If we haven't
8221          * completed the accounting for the whole dio, go back and try again
8222          */
8223         if (ordered_offset < offset + bytes) {
8224                 ordered_bytes = offset + bytes - ordered_offset;
8225                 ordered = NULL;
8226                 goto again;
8227         }
8228 }
8229
8230 static void btrfs_endio_direct_write(struct bio *bio)
8231 {
8232         struct btrfs_dio_private *dip = bio->bi_private;
8233         struct bio *dio_bio = dip->dio_bio;
8234
8235         btrfs_endio_direct_write_update_ordered(dip->inode,
8236                                                 dip->logical_offset,
8237                                                 dip->bytes,
8238                                                 !bio->bi_error);
8239
8240         kfree(dip);
8241
8242         dio_bio->bi_error = bio->bi_error;
8243         dio_end_io(dio_bio, bio->bi_error);
8244         bio_put(bio);
8245 }
8246
8247 static int __btrfs_submit_bio_start_direct_io(struct inode *inode,
8248                                     struct bio *bio, int mirror_num,
8249                                     unsigned long bio_flags, u64 offset)
8250 {
8251         int ret;
8252         struct btrfs_root *root = BTRFS_I(inode)->root;
8253         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8254         BUG_ON(ret); /* -ENOMEM */
8255         return 0;
8256 }
8257
8258 static void btrfs_end_dio_bio(struct bio *bio)
8259 {
8260         struct btrfs_dio_private *dip = bio->bi_private;
8261         int err = bio->bi_error;
8262
8263         if (err)
8264                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8265                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8266                            btrfs_ino(dip->inode), bio_op(bio), bio->bi_opf,
8267                            (unsigned long long)bio->bi_iter.bi_sector,
8268                            bio->bi_iter.bi_size, err);
8269
8270         if (dip->subio_endio)
8271                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8272
8273         if (err) {
8274                 dip->errors = 1;
8275
8276                 /*
8277                  * before atomic variable goto zero, we must make sure
8278                  * dip->errors is perceived to be set.
8279                  */
8280                 smp_mb__before_atomic();
8281         }
8282
8283         /* if there are more bios still pending for this dio, just exit */
8284         if (!atomic_dec_and_test(&dip->pending_bios))
8285                 goto out;
8286
8287         if (dip->errors) {
8288                 bio_io_error(dip->orig_bio);
8289         } else {
8290                 dip->dio_bio->bi_error = 0;
8291                 bio_endio(dip->orig_bio);
8292         }
8293 out:
8294         bio_put(bio);
8295 }
8296
8297 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8298                                        u64 first_sector, gfp_t gfp_flags)
8299 {
8300         struct bio *bio;
8301         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8302         if (bio)
8303                 bio_associate_current(bio);
8304         return bio;
8305 }
8306
8307 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8308                                                  struct inode *inode,
8309                                                  struct btrfs_dio_private *dip,
8310                                                  struct bio *bio,
8311                                                  u64 file_offset)
8312 {
8313         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8314         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8315         int ret;
8316
8317         /*
8318          * We load all the csum data we need when we submit
8319          * the first bio to reduce the csum tree search and
8320          * contention.
8321          */
8322         if (dip->logical_offset == file_offset) {
8323                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8324                                                 file_offset);
8325                 if (ret)
8326                         return ret;
8327         }
8328
8329         if (bio == dip->orig_bio)
8330                 return 0;
8331
8332         file_offset -= dip->logical_offset;
8333         file_offset >>= inode->i_sb->s_blocksize_bits;
8334         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8335
8336         return 0;
8337 }
8338
8339 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8340                                          u64 file_offset, int skip_sum,
8341                                          int async_submit)
8342 {
8343         struct btrfs_dio_private *dip = bio->bi_private;
8344         bool write = bio_op(bio) == REQ_OP_WRITE;
8345         struct btrfs_root *root = BTRFS_I(inode)->root;
8346         int ret;
8347
8348         if (async_submit)
8349                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8350
8351         bio_get(bio);
8352
8353         if (!write) {
8354                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8355                                 BTRFS_WQ_ENDIO_DATA);
8356                 if (ret)
8357                         goto err;
8358         }
8359
8360         if (skip_sum)
8361                 goto map;
8362
8363         if (write && async_submit) {
8364                 ret = btrfs_wq_submit_bio(root->fs_info,
8365                                    inode, bio, 0, 0, file_offset,
8366                                    __btrfs_submit_bio_start_direct_io,
8367                                    __btrfs_submit_bio_done);
8368                 goto err;
8369         } else if (write) {
8370                 /*
8371                  * If we aren't doing async submit, calculate the csum of the
8372                  * bio now.
8373                  */
8374                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8375                 if (ret)
8376                         goto err;
8377         } else {
8378                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8379                                                      file_offset);
8380                 if (ret)
8381                         goto err;
8382         }
8383 map:
8384         ret = btrfs_map_bio(root, bio, 0, async_submit);
8385 err:
8386         bio_put(bio);
8387         return ret;
8388 }
8389
8390 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip,
8391                                     int skip_sum)
8392 {
8393         struct inode *inode = dip->inode;
8394         struct btrfs_root *root = BTRFS_I(inode)->root;
8395         struct bio *bio;
8396         struct bio *orig_bio = dip->orig_bio;
8397         struct bio_vec *bvec = orig_bio->bi_io_vec;
8398         u64 start_sector = orig_bio->bi_iter.bi_sector;
8399         u64 file_offset = dip->logical_offset;
8400         u64 submit_len = 0;
8401         u64 map_length;
8402         u32 blocksize = root->sectorsize;
8403         int async_submit = 0;
8404         int nr_sectors;
8405         int ret;
8406         int i;
8407
8408         map_length = orig_bio->bi_iter.bi_size;
8409         ret = btrfs_map_block(root->fs_info, bio_op(orig_bio),
8410                               start_sector << 9, &map_length, NULL, 0);
8411         if (ret)
8412                 return -EIO;
8413
8414         if (map_length >= orig_bio->bi_iter.bi_size) {
8415                 bio = orig_bio;
8416                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8417                 goto submit;
8418         }
8419
8420         /* async crcs make it difficult to collect full stripe writes. */
8421         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8422                 async_submit = 0;
8423         else
8424                 async_submit = 1;
8425
8426         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8427         if (!bio)
8428                 return -ENOMEM;
8429
8430         bio_set_op_attrs(bio, bio_op(orig_bio), bio_flags(orig_bio));
8431         bio->bi_private = dip;
8432         bio->bi_end_io = btrfs_end_dio_bio;
8433         btrfs_io_bio(bio)->logical = file_offset;
8434         atomic_inc(&dip->pending_bios);
8435
8436         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8437                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8438                 i = 0;
8439 next_block:
8440                 if (unlikely(map_length < submit_len + blocksize ||
8441                     bio_add_page(bio, bvec->bv_page, blocksize,
8442                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8443                         /*
8444                          * inc the count before we submit the bio so
8445                          * we know the end IO handler won't happen before
8446                          * we inc the count. Otherwise, the dip might get freed
8447                          * before we're done setting it up
8448                          */
8449                         atomic_inc(&dip->pending_bios);
8450                         ret = __btrfs_submit_dio_bio(bio, inode,
8451                                                      file_offset, skip_sum,
8452                                                      async_submit);
8453                         if (ret) {
8454                                 bio_put(bio);
8455                                 atomic_dec(&dip->pending_bios);
8456                                 goto out_err;
8457                         }
8458
8459                         start_sector += submit_len >> 9;
8460                         file_offset += submit_len;
8461
8462                         submit_len = 0;
8463
8464                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8465                                                   start_sector, GFP_NOFS);
8466                         if (!bio)
8467                                 goto out_err;
8468                         bio_set_op_attrs(bio, bio_op(orig_bio),
8469                                          bio_flags(orig_bio));
8470                         bio->bi_private = dip;
8471                         bio->bi_end_io = btrfs_end_dio_bio;
8472                         btrfs_io_bio(bio)->logical = file_offset;
8473
8474                         map_length = orig_bio->bi_iter.bi_size;
8475                         ret = btrfs_map_block(root->fs_info, bio_op(orig_bio),
8476                                               start_sector << 9,
8477                                               &map_length, NULL, 0);
8478                         if (ret) {
8479                                 bio_put(bio);
8480                                 goto out_err;
8481                         }
8482
8483                         goto next_block;
8484                 } else {
8485                         submit_len += blocksize;
8486                         if (--nr_sectors) {
8487                                 i++;
8488                                 goto next_block;
8489                         }
8490                         bvec++;
8491                 }
8492         }
8493
8494 submit:
8495         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, skip_sum,
8496                                      async_submit);
8497         if (!ret)
8498                 return 0;
8499
8500         bio_put(bio);
8501 out_err:
8502         dip->errors = 1;
8503         /*
8504          * before atomic variable goto zero, we must
8505          * make sure dip->errors is perceived to be set.
8506          */
8507         smp_mb__before_atomic();
8508         if (atomic_dec_and_test(&dip->pending_bios))
8509                 bio_io_error(dip->orig_bio);
8510
8511         /* bio_end_io() will handle error, so we needn't return it */
8512         return 0;
8513 }
8514
8515 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8516                                 loff_t file_offset)
8517 {
8518         struct btrfs_dio_private *dip = NULL;
8519         struct bio *io_bio = NULL;
8520         struct btrfs_io_bio *btrfs_bio;
8521         int skip_sum;
8522         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8523         int ret = 0;
8524
8525         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8526
8527         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8528         if (!io_bio) {
8529                 ret = -ENOMEM;
8530                 goto free_ordered;
8531         }
8532
8533         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8534         if (!dip) {
8535                 ret = -ENOMEM;
8536                 goto free_ordered;
8537         }
8538
8539         dip->private = dio_bio->bi_private;
8540         dip->inode = inode;
8541         dip->logical_offset = file_offset;
8542         dip->bytes = dio_bio->bi_iter.bi_size;
8543         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8544         io_bio->bi_private = dip;
8545         dip->orig_bio = io_bio;
8546         dip->dio_bio = dio_bio;
8547         atomic_set(&dip->pending_bios, 0);
8548         btrfs_bio = btrfs_io_bio(io_bio);
8549         btrfs_bio->logical = file_offset;
8550
8551         if (write) {
8552                 io_bio->bi_end_io = btrfs_endio_direct_write;
8553         } else {
8554                 io_bio->bi_end_io = btrfs_endio_direct_read;
8555                 dip->subio_endio = btrfs_subio_endio_read;
8556         }
8557
8558         /*
8559          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8560          * even if we fail to submit a bio, because in such case we do the
8561          * corresponding error handling below and it must not be done a second
8562          * time by btrfs_direct_IO().
8563          */
8564         if (write) {
8565                 struct btrfs_dio_data *dio_data = current->journal_info;
8566
8567                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8568                         dip->bytes;
8569                 dio_data->unsubmitted_oe_range_start =
8570                         dio_data->unsubmitted_oe_range_end;
8571         }
8572
8573         ret = btrfs_submit_direct_hook(dip, skip_sum);
8574         if (!ret)
8575                 return;
8576
8577         if (btrfs_bio->end_io)
8578                 btrfs_bio->end_io(btrfs_bio, ret);
8579
8580 free_ordered:
8581         /*
8582          * If we arrived here it means either we failed to submit the dip
8583          * or we either failed to clone the dio_bio or failed to allocate the
8584          * dip. If we cloned the dio_bio and allocated the dip, we can just
8585          * call bio_endio against our io_bio so that we get proper resource
8586          * cleanup if we fail to submit the dip, otherwise, we must do the
8587          * same as btrfs_endio_direct_[write|read] because we can't call these
8588          * callbacks - they require an allocated dip and a clone of dio_bio.
8589          */
8590         if (io_bio && dip) {
8591                 io_bio->bi_error = -EIO;
8592                 bio_endio(io_bio);
8593                 /*
8594                  * The end io callbacks free our dip, do the final put on io_bio
8595                  * and all the cleanup and final put for dio_bio (through
8596                  * dio_end_io()).
8597                  */
8598                 dip = NULL;
8599                 io_bio = NULL;
8600         } else {
8601                 if (write)
8602                         btrfs_endio_direct_write_update_ordered(inode,
8603                                                 file_offset,
8604                                                 dio_bio->bi_iter.bi_size,
8605                                                 0);
8606                 else
8607                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8608                               file_offset + dio_bio->bi_iter.bi_size - 1);
8609
8610                 dio_bio->bi_error = -EIO;
8611                 /*
8612                  * Releases and cleans up our dio_bio, no need to bio_put()
8613                  * nor bio_endio()/bio_io_error() against dio_bio.
8614                  */
8615                 dio_end_io(dio_bio, ret);
8616         }
8617         if (io_bio)
8618                 bio_put(io_bio);
8619         kfree(dip);
8620 }
8621
8622 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8623                         const struct iov_iter *iter, loff_t offset)
8624 {
8625         int seg;
8626         int i;
8627         unsigned blocksize_mask = root->sectorsize - 1;
8628         ssize_t retval = -EINVAL;
8629
8630         if (offset & blocksize_mask)
8631                 goto out;
8632
8633         if (iov_iter_alignment(iter) & blocksize_mask)
8634                 goto out;
8635
8636         /* If this is a write we don't need to check anymore */
8637         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8638                 return 0;
8639         /*
8640          * Check to make sure we don't have duplicate iov_base's in this
8641          * iovec, if so return EINVAL, otherwise we'll get csum errors
8642          * when reading back.
8643          */
8644         for (seg = 0; seg < iter->nr_segs; seg++) {
8645                 for (i = seg + 1; i < iter->nr_segs; i++) {
8646                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8647                                 goto out;
8648                 }
8649         }
8650         retval = 0;
8651 out:
8652         return retval;
8653 }
8654
8655 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8656 {
8657         struct file *file = iocb->ki_filp;
8658         struct inode *inode = file->f_mapping->host;
8659         struct btrfs_root *root = BTRFS_I(inode)->root;
8660         struct btrfs_dio_data dio_data = { 0 };
8661         loff_t offset = iocb->ki_pos;
8662         size_t count = 0;
8663         int flags = 0;
8664         bool wakeup = true;
8665         bool relock = false;
8666         ssize_t ret;
8667
8668         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8669                 return 0;
8670
8671         inode_dio_begin(inode);
8672         smp_mb__after_atomic();
8673
8674         /*
8675          * The generic stuff only does filemap_write_and_wait_range, which
8676          * isn't enough if we've written compressed pages to this area, so
8677          * we need to flush the dirty pages again to make absolutely sure
8678          * that any outstanding dirty pages are on disk.
8679          */
8680         count = iov_iter_count(iter);
8681         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8682                      &BTRFS_I(inode)->runtime_flags))
8683                 filemap_fdatawrite_range(inode->i_mapping, offset,
8684                                          offset + count - 1);
8685
8686         if (iov_iter_rw(iter) == WRITE) {
8687                 /*
8688                  * If the write DIO is beyond the EOF, we need update
8689                  * the isize, but it is protected by i_mutex. So we can
8690                  * not unlock the i_mutex at this case.
8691                  */
8692                 if (offset + count <= inode->i_size) {
8693                         inode_unlock(inode);
8694                         relock = true;
8695                 }
8696                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8697                 if (ret)
8698                         goto out;
8699                 dio_data.outstanding_extents = div64_u64(count +
8700                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8701                                                 BTRFS_MAX_EXTENT_SIZE);
8702
8703                 /*
8704                  * We need to know how many extents we reserved so that we can
8705                  * do the accounting properly if we go over the number we
8706                  * originally calculated.  Abuse current->journal_info for this.
8707                  */
8708                 dio_data.reserve = round_up(count, root->sectorsize);
8709                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8710                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8711                 current->journal_info = &dio_data;
8712         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8713                                      &BTRFS_I(inode)->runtime_flags)) {
8714                 inode_dio_end(inode);
8715                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8716                 wakeup = false;
8717         }
8718
8719         ret = __blockdev_direct_IO(iocb, inode,
8720                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8721                                    iter, btrfs_get_blocks_direct, NULL,
8722                                    btrfs_submit_direct, flags);
8723         if (iov_iter_rw(iter) == WRITE) {
8724                 current->journal_info = NULL;
8725                 if (ret < 0 && ret != -EIOCBQUEUED) {
8726                         if (dio_data.reserve)
8727                                 btrfs_delalloc_release_space(inode, offset,
8728                                                              dio_data.reserve);
8729                         /*
8730                          * On error we might have left some ordered extents
8731                          * without submitting corresponding bios for them, so
8732                          * cleanup them up to avoid other tasks getting them
8733                          * and waiting for them to complete forever.
8734                          */
8735                         if (dio_data.unsubmitted_oe_range_start <
8736                             dio_data.unsubmitted_oe_range_end)
8737                                 btrfs_endio_direct_write_update_ordered(inode,
8738                                         dio_data.unsubmitted_oe_range_start,
8739                                         dio_data.unsubmitted_oe_range_end -
8740                                         dio_data.unsubmitted_oe_range_start,
8741                                         0);
8742                 } else if (ret >= 0 && (size_t)ret < count)
8743                         btrfs_delalloc_release_space(inode, offset,
8744                                                      count - (size_t)ret);
8745         }
8746 out:
8747         if (wakeup)
8748                 inode_dio_end(inode);
8749         if (relock)
8750                 inode_lock(inode);
8751
8752         return ret;
8753 }
8754
8755 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8756
8757 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8758                 __u64 start, __u64 len)
8759 {
8760         int     ret;
8761
8762         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8763         if (ret)
8764                 return ret;
8765
8766         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8767 }
8768
8769 int btrfs_readpage(struct file *file, struct page *page)
8770 {
8771         struct extent_io_tree *tree;
8772         tree = &BTRFS_I(page->mapping->host)->io_tree;
8773         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8774 }
8775
8776 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8777 {
8778         struct extent_io_tree *tree;
8779         struct inode *inode = page->mapping->host;
8780         int ret;
8781
8782         if (current->flags & PF_MEMALLOC) {
8783                 redirty_page_for_writepage(wbc, page);
8784                 unlock_page(page);
8785                 return 0;
8786         }
8787
8788         /*
8789          * If we are under memory pressure we will call this directly from the
8790          * VM, we need to make sure we have the inode referenced for the ordered
8791          * extent.  If not just return like we didn't do anything.
8792          */
8793         if (!igrab(inode)) {
8794                 redirty_page_for_writepage(wbc, page);
8795                 return AOP_WRITEPAGE_ACTIVATE;
8796         }
8797         tree = &BTRFS_I(page->mapping->host)->io_tree;
8798         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8799         btrfs_add_delayed_iput(inode);
8800         return ret;
8801 }
8802
8803 static int btrfs_writepages(struct address_space *mapping,
8804                             struct writeback_control *wbc)
8805 {
8806         struct extent_io_tree *tree;
8807
8808         tree = &BTRFS_I(mapping->host)->io_tree;
8809         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8810 }
8811
8812 static int
8813 btrfs_readpages(struct file *file, struct address_space *mapping,
8814                 struct list_head *pages, unsigned nr_pages)
8815 {
8816         struct extent_io_tree *tree;
8817         tree = &BTRFS_I(mapping->host)->io_tree;
8818         return extent_readpages(tree, mapping, pages, nr_pages,
8819                                 btrfs_get_extent);
8820 }
8821 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8822 {
8823         struct extent_io_tree *tree;
8824         struct extent_map_tree *map;
8825         int ret;
8826
8827         tree = &BTRFS_I(page->mapping->host)->io_tree;
8828         map = &BTRFS_I(page->mapping->host)->extent_tree;
8829         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8830         if (ret == 1) {
8831                 ClearPagePrivate(page);
8832                 set_page_private(page, 0);
8833                 put_page(page);
8834         }
8835         return ret;
8836 }
8837
8838 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8839 {
8840         if (PageWriteback(page) || PageDirty(page))
8841                 return 0;
8842         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8843 }
8844
8845 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8846                                  unsigned int length)
8847 {
8848         struct inode *inode = page->mapping->host;
8849         struct extent_io_tree *tree;
8850         struct btrfs_ordered_extent *ordered;
8851         struct extent_state *cached_state = NULL;
8852         u64 page_start = page_offset(page);
8853         u64 page_end = page_start + PAGE_SIZE - 1;
8854         u64 start;
8855         u64 end;
8856         int inode_evicting = inode->i_state & I_FREEING;
8857
8858         /*
8859          * we have the page locked, so new writeback can't start,
8860          * and the dirty bit won't be cleared while we are here.
8861          *
8862          * Wait for IO on this page so that we can safely clear
8863          * the PagePrivate2 bit and do ordered accounting
8864          */
8865         wait_on_page_writeback(page);
8866
8867         tree = &BTRFS_I(inode)->io_tree;
8868         if (offset) {
8869                 btrfs_releasepage(page, GFP_NOFS);
8870                 return;
8871         }
8872
8873         if (!inode_evicting)
8874                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8875 again:
8876         start = page_start;
8877         ordered = btrfs_lookup_ordered_range(inode, start,
8878                                         page_end - start + 1);
8879         if (ordered) {
8880                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8881                 /*
8882                  * IO on this page will never be started, so we need
8883                  * to account for any ordered extents now
8884                  */
8885                 if (!inode_evicting)
8886                         clear_extent_bit(tree, start, end,
8887                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8888                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8889                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8890                                          GFP_NOFS);
8891                 /*
8892                  * whoever cleared the private bit is responsible
8893                  * for the finish_ordered_io
8894                  */
8895                 if (TestClearPagePrivate2(page)) {
8896                         struct btrfs_ordered_inode_tree *tree;
8897                         u64 new_len;
8898
8899                         tree = &BTRFS_I(inode)->ordered_tree;
8900
8901                         spin_lock_irq(&tree->lock);
8902                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8903                         new_len = start - ordered->file_offset;
8904                         if (new_len < ordered->truncated_len)
8905                                 ordered->truncated_len = new_len;
8906                         spin_unlock_irq(&tree->lock);
8907
8908                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8909                                                            start,
8910                                                            end - start + 1, 1))
8911                                 btrfs_finish_ordered_io(ordered);
8912                 }
8913                 btrfs_put_ordered_extent(ordered);
8914                 if (!inode_evicting) {
8915                         cached_state = NULL;
8916                         lock_extent_bits(tree, start, end,
8917                                          &cached_state);
8918                 }
8919
8920                 start = end + 1;
8921                 if (start < page_end)
8922                         goto again;
8923         }
8924
8925         /*
8926          * Qgroup reserved space handler
8927          * Page here will be either
8928          * 1) Already written to disk
8929          *    In this case, its reserved space is released from data rsv map
8930          *    and will be freed by delayed_ref handler finally.
8931          *    So even we call qgroup_free_data(), it won't decrease reserved
8932          *    space.
8933          * 2) Not written to disk
8934          *    This means the reserved space should be freed here. However,
8935          *    if a truncate invalidates the page (by clearing PageDirty)
8936          *    and the page is accounted for while allocating extent
8937          *    in btrfs_check_data_free_space() we let delayed_ref to
8938          *    free the entire extent.
8939          */
8940         if (PageDirty(page))
8941                 btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8942         if (!inode_evicting) {
8943                 clear_extent_bit(tree, page_start, page_end,
8944                                  EXTENT_LOCKED | EXTENT_DIRTY |
8945                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8946                                  EXTENT_DEFRAG, 1, 1,
8947                                  &cached_state, GFP_NOFS);
8948
8949                 __btrfs_releasepage(page, GFP_NOFS);
8950         }
8951
8952         ClearPageChecked(page);
8953         if (PagePrivate(page)) {
8954                 ClearPagePrivate(page);
8955                 set_page_private(page, 0);
8956                 put_page(page);
8957         }
8958 }
8959
8960 /*
8961  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8962  * called from a page fault handler when a page is first dirtied. Hence we must
8963  * be careful to check for EOF conditions here. We set the page up correctly
8964  * for a written page which means we get ENOSPC checking when writing into
8965  * holes and correct delalloc and unwritten extent mapping on filesystems that
8966  * support these features.
8967  *
8968  * We are not allowed to take the i_mutex here so we have to play games to
8969  * protect against truncate races as the page could now be beyond EOF.  Because
8970  * vmtruncate() writes the inode size before removing pages, once we have the
8971  * page lock we can determine safely if the page is beyond EOF. If it is not
8972  * beyond EOF, then the page is guaranteed safe against truncation until we
8973  * unlock the page.
8974  */
8975 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8976 {
8977         struct page *page = vmf->page;
8978         struct inode *inode = file_inode(vma->vm_file);
8979         struct btrfs_root *root = BTRFS_I(inode)->root;
8980         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8981         struct btrfs_ordered_extent *ordered;
8982         struct extent_state *cached_state = NULL;
8983         char *kaddr;
8984         unsigned long zero_start;
8985         loff_t size;
8986         int ret;
8987         int reserved = 0;
8988         u64 reserved_space;
8989         u64 page_start;
8990         u64 page_end;
8991         u64 end;
8992
8993         reserved_space = PAGE_SIZE;
8994
8995         sb_start_pagefault(inode->i_sb);
8996         page_start = page_offset(page);
8997         page_end = page_start + PAGE_SIZE - 1;
8998         end = page_end;
8999
9000         /*
9001          * Reserving delalloc space after obtaining the page lock can lead to
9002          * deadlock. For example, if a dirty page is locked by this function
9003          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9004          * dirty page write out, then the btrfs_writepage() function could
9005          * end up waiting indefinitely to get a lock on the page currently
9006          * being processed by btrfs_page_mkwrite() function.
9007          */
9008         ret = btrfs_delalloc_reserve_space(inode, page_start,
9009                                            reserved_space);
9010         if (!ret) {
9011                 ret = file_update_time(vma->vm_file);
9012                 reserved = 1;
9013         }
9014         if (ret) {
9015                 if (ret == -ENOMEM)
9016                         ret = VM_FAULT_OOM;
9017                 else /* -ENOSPC, -EIO, etc */
9018                         ret = VM_FAULT_SIGBUS;
9019                 if (reserved)
9020                         goto out;
9021                 goto out_noreserve;
9022         }
9023
9024         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9025 again:
9026         lock_page(page);
9027         size = i_size_read(inode);
9028
9029         if ((page->mapping != inode->i_mapping) ||
9030             (page_start >= size)) {
9031                 /* page got truncated out from underneath us */
9032                 goto out_unlock;
9033         }
9034         wait_on_page_writeback(page);
9035
9036         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9037         set_page_extent_mapped(page);
9038
9039         /*
9040          * we can't set the delalloc bits if there are pending ordered
9041          * extents.  Drop our locks and wait for them to finish
9042          */
9043         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
9044         if (ordered) {
9045                 unlock_extent_cached(io_tree, page_start, page_end,
9046                                      &cached_state, GFP_NOFS);
9047                 unlock_page(page);
9048                 btrfs_start_ordered_extent(inode, ordered, 1);
9049                 btrfs_put_ordered_extent(ordered);
9050                 goto again;
9051         }
9052
9053         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9054                 reserved_space = round_up(size - page_start, root->sectorsize);
9055                 if (reserved_space < PAGE_SIZE) {
9056                         end = page_start + reserved_space - 1;
9057                         spin_lock(&BTRFS_I(inode)->lock);
9058                         BTRFS_I(inode)->outstanding_extents++;
9059                         spin_unlock(&BTRFS_I(inode)->lock);
9060                         btrfs_delalloc_release_space(inode, page_start,
9061                                                 PAGE_SIZE - reserved_space);
9062                 }
9063         }
9064
9065         /*
9066          * XXX - page_mkwrite gets called every time the page is dirtied, even
9067          * if it was already dirty, so for space accounting reasons we need to
9068          * clear any delalloc bits for the range we are fixing to save.  There
9069          * is probably a better way to do this, but for now keep consistent with
9070          * prepare_pages in the normal write path.
9071          */
9072         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9073                           EXTENT_DIRTY | EXTENT_DELALLOC |
9074                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9075                           0, 0, &cached_state, GFP_NOFS);
9076
9077         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9078                                         &cached_state, 0);
9079         if (ret) {
9080                 unlock_extent_cached(io_tree, page_start, page_end,
9081                                      &cached_state, GFP_NOFS);
9082                 ret = VM_FAULT_SIGBUS;
9083                 goto out_unlock;
9084         }
9085         ret = 0;
9086
9087         /* page is wholly or partially inside EOF */
9088         if (page_start + PAGE_SIZE > size)
9089                 zero_start = size & ~PAGE_MASK;
9090         else
9091                 zero_start = PAGE_SIZE;
9092
9093         if (zero_start != PAGE_SIZE) {
9094                 kaddr = kmap(page);
9095                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9096                 flush_dcache_page(page);
9097                 kunmap(page);
9098         }
9099         ClearPageChecked(page);
9100         set_page_dirty(page);
9101         SetPageUptodate(page);
9102
9103         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9104         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9105         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9106
9107         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9108
9109 out_unlock:
9110         if (!ret) {
9111                 sb_end_pagefault(inode->i_sb);
9112                 return VM_FAULT_LOCKED;
9113         }
9114         unlock_page(page);
9115 out:
9116         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9117 out_noreserve:
9118         sb_end_pagefault(inode->i_sb);
9119         return ret;
9120 }
9121
9122 static int btrfs_truncate(struct inode *inode)
9123 {
9124         struct btrfs_root *root = BTRFS_I(inode)->root;
9125         struct btrfs_block_rsv *rsv;
9126         int ret = 0;
9127         int err = 0;
9128         struct btrfs_trans_handle *trans;
9129         u64 mask = root->sectorsize - 1;
9130         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9131
9132         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9133                                        (u64)-1);
9134         if (ret)
9135                 return ret;
9136
9137         /*
9138          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9139          * 3 things going on here
9140          *
9141          * 1) We need to reserve space for our orphan item and the space to
9142          * delete our orphan item.  Lord knows we don't want to have a dangling
9143          * orphan item because we didn't reserve space to remove it.
9144          *
9145          * 2) We need to reserve space to update our inode.
9146          *
9147          * 3) We need to have something to cache all the space that is going to
9148          * be free'd up by the truncate operation, but also have some slack
9149          * space reserved in case it uses space during the truncate (thank you
9150          * very much snapshotting).
9151          *
9152          * And we need these to all be separate.  The fact is we can use a lot of
9153          * space doing the truncate, and we have no earthly idea how much space
9154          * we will use, so we need the truncate reservation to be separate so it
9155          * doesn't end up using space reserved for updating the inode or
9156          * removing the orphan item.  We also need to be able to stop the
9157          * transaction and start a new one, which means we need to be able to
9158          * update the inode several times, and we have no idea of knowing how
9159          * many times that will be, so we can't just reserve 1 item for the
9160          * entirety of the operation, so that has to be done separately as well.
9161          * Then there is the orphan item, which does indeed need to be held on
9162          * to for the whole operation, and we need nobody to touch this reserved
9163          * space except the orphan code.
9164          *
9165          * So that leaves us with
9166          *
9167          * 1) root->orphan_block_rsv - for the orphan deletion.
9168          * 2) rsv - for the truncate reservation, which we will steal from the
9169          * transaction reservation.
9170          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9171          * updating the inode.
9172          */
9173         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9174         if (!rsv)
9175                 return -ENOMEM;
9176         rsv->size = min_size;
9177         rsv->failfast = 1;
9178
9179         /*
9180          * 1 for the truncate slack space
9181          * 1 for updating the inode.
9182          */
9183         trans = btrfs_start_transaction(root, 2);
9184         if (IS_ERR(trans)) {
9185                 err = PTR_ERR(trans);
9186                 goto out;
9187         }
9188
9189         /* Migrate the slack space for the truncate to our reserve */
9190         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9191                                       min_size, 0);
9192         BUG_ON(ret);
9193
9194         /*
9195          * So if we truncate and then write and fsync we normally would just
9196          * write the extents that changed, which is a problem if we need to
9197          * first truncate that entire inode.  So set this flag so we write out
9198          * all of the extents in the inode to the sync log so we're completely
9199          * safe.
9200          */
9201         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9202         trans->block_rsv = rsv;
9203
9204         while (1) {
9205                 ret = btrfs_truncate_inode_items(trans, root, inode,
9206                                                  inode->i_size,
9207                                                  BTRFS_EXTENT_DATA_KEY);
9208                 if (ret != -ENOSPC && ret != -EAGAIN) {
9209                         err = ret;
9210                         break;
9211                 }
9212
9213                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9214                 ret = btrfs_update_inode(trans, root, inode);
9215                 if (ret) {
9216                         err = ret;
9217                         break;
9218                 }
9219
9220                 btrfs_end_transaction(trans, root);
9221                 btrfs_btree_balance_dirty(root);
9222
9223                 trans = btrfs_start_transaction(root, 2);
9224                 if (IS_ERR(trans)) {
9225                         ret = err = PTR_ERR(trans);
9226                         trans = NULL;
9227                         break;
9228                 }
9229
9230                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9231                                               rsv, min_size, 0);
9232                 BUG_ON(ret);    /* shouldn't happen */
9233                 trans->block_rsv = rsv;
9234         }
9235
9236         if (ret == 0 && inode->i_nlink > 0) {
9237                 trans->block_rsv = root->orphan_block_rsv;
9238                 ret = btrfs_orphan_del(trans, inode);
9239                 if (ret)
9240                         err = ret;
9241         }
9242
9243         if (trans) {
9244                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9245                 ret = btrfs_update_inode(trans, root, inode);
9246                 if (ret && !err)
9247                         err = ret;
9248
9249                 ret = btrfs_end_transaction(trans, root);
9250                 btrfs_btree_balance_dirty(root);
9251         }
9252 out:
9253         btrfs_free_block_rsv(root, rsv);
9254
9255         if (ret && !err)
9256                 err = ret;
9257
9258         return err;
9259 }
9260
9261 /*
9262  * create a new subvolume directory/inode (helper for the ioctl).
9263  */
9264 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9265                              struct btrfs_root *new_root,
9266                              struct btrfs_root *parent_root,
9267                              u64 new_dirid)
9268 {
9269         struct inode *inode;
9270         int err;
9271         u64 index = 0;
9272
9273         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9274                                 new_dirid, new_dirid,
9275                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9276                                 &index);
9277         if (IS_ERR(inode))
9278                 return PTR_ERR(inode);
9279         inode->i_op = &btrfs_dir_inode_operations;
9280         inode->i_fop = &btrfs_dir_file_operations;
9281
9282         set_nlink(inode, 1);
9283         btrfs_i_size_write(inode, 0);
9284         unlock_new_inode(inode);
9285
9286         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9287         if (err)
9288                 btrfs_err(new_root->fs_info,
9289                           "error inheriting subvolume %llu properties: %d",
9290                           new_root->root_key.objectid, err);
9291
9292         err = btrfs_update_inode(trans, new_root, inode);
9293
9294         iput(inode);
9295         return err;
9296 }
9297
9298 struct inode *btrfs_alloc_inode(struct super_block *sb)
9299 {
9300         struct btrfs_inode *ei;
9301         struct inode *inode;
9302
9303         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9304         if (!ei)
9305                 return NULL;
9306
9307         ei->root = NULL;
9308         ei->generation = 0;
9309         ei->last_trans = 0;
9310         ei->last_sub_trans = 0;
9311         ei->logged_trans = 0;
9312         ei->delalloc_bytes = 0;
9313         ei->defrag_bytes = 0;
9314         ei->disk_i_size = 0;
9315         ei->flags = 0;
9316         ei->csum_bytes = 0;
9317         ei->index_cnt = (u64)-1;
9318         ei->dir_index = 0;
9319         ei->last_unlink_trans = 0;
9320         ei->last_log_commit = 0;
9321         ei->delayed_iput_count = 0;
9322
9323         spin_lock_init(&ei->lock);
9324         ei->outstanding_extents = 0;
9325         ei->reserved_extents = 0;
9326
9327         ei->runtime_flags = 0;
9328         ei->force_compress = BTRFS_COMPRESS_NONE;
9329
9330         ei->delayed_node = NULL;
9331
9332         ei->i_otime.tv_sec = 0;
9333         ei->i_otime.tv_nsec = 0;
9334
9335         inode = &ei->vfs_inode;
9336         extent_map_tree_init(&ei->extent_tree);
9337         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9338         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9339         ei->io_tree.track_uptodate = 1;
9340         ei->io_failure_tree.track_uptodate = 1;
9341         atomic_set(&ei->sync_writers, 0);
9342         mutex_init(&ei->log_mutex);
9343         mutex_init(&ei->delalloc_mutex);
9344         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9345         INIT_LIST_HEAD(&ei->delalloc_inodes);
9346         INIT_LIST_HEAD(&ei->delayed_iput);
9347         RB_CLEAR_NODE(&ei->rb_node);
9348         init_rwsem(&ei->dio_sem);
9349
9350         return inode;
9351 }
9352
9353 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9354 void btrfs_test_destroy_inode(struct inode *inode)
9355 {
9356         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9357         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9358 }
9359 #endif
9360
9361 static void btrfs_i_callback(struct rcu_head *head)
9362 {
9363         struct inode *inode = container_of(head, struct inode, i_rcu);
9364         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9365 }
9366
9367 void btrfs_destroy_inode(struct inode *inode)
9368 {
9369         struct btrfs_ordered_extent *ordered;
9370         struct btrfs_root *root = BTRFS_I(inode)->root;
9371
9372         WARN_ON(!hlist_empty(&inode->i_dentry));
9373         WARN_ON(inode->i_data.nrpages);
9374         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9375         WARN_ON(BTRFS_I(inode)->reserved_extents);
9376         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9377         WARN_ON(BTRFS_I(inode)->csum_bytes);
9378         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9379
9380         /*
9381          * This can happen where we create an inode, but somebody else also
9382          * created the same inode and we need to destroy the one we already
9383          * created.
9384          */
9385         if (!root)
9386                 goto free;
9387
9388         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9389                      &BTRFS_I(inode)->runtime_flags)) {
9390                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9391                         btrfs_ino(inode));
9392                 atomic_dec(&root->orphan_inodes);
9393         }
9394
9395         while (1) {
9396                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9397                 if (!ordered)
9398                         break;
9399                 else {
9400                         btrfs_err(root->fs_info,
9401                                   "found ordered extent %llu %llu on inode cleanup",
9402                                   ordered->file_offset, ordered->len);
9403                         btrfs_remove_ordered_extent(inode, ordered);
9404                         btrfs_put_ordered_extent(ordered);
9405                         btrfs_put_ordered_extent(ordered);
9406                 }
9407         }
9408         btrfs_qgroup_check_reserved_leak(inode);
9409         inode_tree_del(inode);
9410         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9411 free:
9412         call_rcu(&inode->i_rcu, btrfs_i_callback);
9413 }
9414
9415 int btrfs_drop_inode(struct inode *inode)
9416 {
9417         struct btrfs_root *root = BTRFS_I(inode)->root;
9418
9419         if (root == NULL)
9420                 return 1;
9421
9422         /* the snap/subvol tree is on deleting */
9423         if (btrfs_root_refs(&root->root_item) == 0)
9424                 return 1;
9425         else
9426                 return generic_drop_inode(inode);
9427 }
9428
9429 static void init_once(void *foo)
9430 {
9431         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9432
9433         inode_init_once(&ei->vfs_inode);
9434 }
9435
9436 void btrfs_destroy_cachep(void)
9437 {
9438         /*
9439          * Make sure all delayed rcu free inodes are flushed before we
9440          * destroy cache.
9441          */
9442         rcu_barrier();
9443         kmem_cache_destroy(btrfs_inode_cachep);
9444         kmem_cache_destroy(btrfs_trans_handle_cachep);
9445         kmem_cache_destroy(btrfs_transaction_cachep);
9446         kmem_cache_destroy(btrfs_path_cachep);
9447         kmem_cache_destroy(btrfs_free_space_cachep);
9448 }
9449
9450 int btrfs_init_cachep(void)
9451 {
9452         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9453                         sizeof(struct btrfs_inode), 0,
9454                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9455                         init_once);
9456         if (!btrfs_inode_cachep)
9457                 goto fail;
9458
9459         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9460                         sizeof(struct btrfs_trans_handle), 0,
9461                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9462         if (!btrfs_trans_handle_cachep)
9463                 goto fail;
9464
9465         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9466                         sizeof(struct btrfs_transaction), 0,
9467                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9468         if (!btrfs_transaction_cachep)
9469                 goto fail;
9470
9471         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9472                         sizeof(struct btrfs_path), 0,
9473                         SLAB_MEM_SPREAD, NULL);
9474         if (!btrfs_path_cachep)
9475                 goto fail;
9476
9477         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9478                         sizeof(struct btrfs_free_space), 0,
9479                         SLAB_MEM_SPREAD, NULL);
9480         if (!btrfs_free_space_cachep)
9481                 goto fail;
9482
9483         return 0;
9484 fail:
9485         btrfs_destroy_cachep();
9486         return -ENOMEM;
9487 }
9488
9489 static int btrfs_getattr(struct vfsmount *mnt,
9490                          struct dentry *dentry, struct kstat *stat)
9491 {
9492         u64 delalloc_bytes;
9493         struct inode *inode = d_inode(dentry);
9494         u32 blocksize = inode->i_sb->s_blocksize;
9495
9496         generic_fillattr(inode, stat);
9497         stat->dev = BTRFS_I(inode)->root->anon_dev;
9498
9499         spin_lock(&BTRFS_I(inode)->lock);
9500         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9501         spin_unlock(&BTRFS_I(inode)->lock);
9502         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9503                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9504         return 0;
9505 }
9506
9507 static int btrfs_rename_exchange(struct inode *old_dir,
9508                               struct dentry *old_dentry,
9509                               struct inode *new_dir,
9510                               struct dentry *new_dentry)
9511 {
9512         struct btrfs_trans_handle *trans;
9513         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9514         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9515         struct inode *new_inode = new_dentry->d_inode;
9516         struct inode *old_inode = old_dentry->d_inode;
9517         struct timespec ctime = current_time(old_inode);
9518         struct dentry *parent;
9519         u64 old_ino = btrfs_ino(old_inode);
9520         u64 new_ino = btrfs_ino(new_inode);
9521         u64 old_idx = 0;
9522         u64 new_idx = 0;
9523         u64 root_objectid;
9524         int ret;
9525         bool root_log_pinned = false;
9526         bool dest_log_pinned = false;
9527
9528         /* we only allow rename subvolume link between subvolumes */
9529         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9530                 return -EXDEV;
9531
9532         /* close the race window with snapshot create/destroy ioctl */
9533         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9534                 down_read(&root->fs_info->subvol_sem);
9535         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9536                 down_read(&dest->fs_info->subvol_sem);
9537
9538         /*
9539          * We want to reserve the absolute worst case amount of items.  So if
9540          * both inodes are subvols and we need to unlink them then that would
9541          * require 4 item modifications, but if they are both normal inodes it
9542          * would require 5 item modifications, so we'll assume their normal
9543          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9544          * should cover the worst case number of items we'll modify.
9545          */
9546         trans = btrfs_start_transaction(root, 12);
9547         if (IS_ERR(trans)) {
9548                 ret = PTR_ERR(trans);
9549                 goto out_notrans;
9550         }
9551
9552         /*
9553          * We need to find a free sequence number both in the source and
9554          * in the destination directory for the exchange.
9555          */
9556         ret = btrfs_set_inode_index(new_dir, &old_idx);
9557         if (ret)
9558                 goto out_fail;
9559         ret = btrfs_set_inode_index(old_dir, &new_idx);
9560         if (ret)
9561                 goto out_fail;
9562
9563         BTRFS_I(old_inode)->dir_index = 0ULL;
9564         BTRFS_I(new_inode)->dir_index = 0ULL;
9565
9566         /* Reference for the source. */
9567         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9568                 /* force full log commit if subvolume involved. */
9569                 btrfs_set_log_full_commit(root->fs_info, trans);
9570         } else {
9571                 btrfs_pin_log_trans(root);
9572                 root_log_pinned = true;
9573                 ret = btrfs_insert_inode_ref(trans, dest,
9574                                              new_dentry->d_name.name,
9575                                              new_dentry->d_name.len,
9576                                              old_ino,
9577                                              btrfs_ino(new_dir), old_idx);
9578                 if (ret)
9579                         goto out_fail;
9580         }
9581
9582         /* And now for the dest. */
9583         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9584                 /* force full log commit if subvolume involved. */
9585                 btrfs_set_log_full_commit(dest->fs_info, trans);
9586         } else {
9587                 btrfs_pin_log_trans(dest);
9588                 dest_log_pinned = true;
9589                 ret = btrfs_insert_inode_ref(trans, root,
9590                                              old_dentry->d_name.name,
9591                                              old_dentry->d_name.len,
9592                                              new_ino,
9593                                              btrfs_ino(old_dir), new_idx);
9594                 if (ret)
9595                         goto out_fail;
9596         }
9597
9598         /* Update inode version and ctime/mtime. */
9599         inode_inc_iversion(old_dir);
9600         inode_inc_iversion(new_dir);
9601         inode_inc_iversion(old_inode);
9602         inode_inc_iversion(new_inode);
9603         old_dir->i_ctime = old_dir->i_mtime = ctime;
9604         new_dir->i_ctime = new_dir->i_mtime = ctime;
9605         old_inode->i_ctime = ctime;
9606         new_inode->i_ctime = ctime;
9607
9608         if (old_dentry->d_parent != new_dentry->d_parent) {
9609                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9610                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9611         }
9612
9613         /* src is a subvolume */
9614         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9615                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9616                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9617                                           root_objectid,
9618                                           old_dentry->d_name.name,
9619                                           old_dentry->d_name.len);
9620         } else { /* src is an inode */
9621                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9622                                            old_dentry->d_inode,
9623                                            old_dentry->d_name.name,
9624                                            old_dentry->d_name.len);
9625                 if (!ret)
9626                         ret = btrfs_update_inode(trans, root, old_inode);
9627         }
9628         if (ret) {
9629                 btrfs_abort_transaction(trans, ret);
9630                 goto out_fail;
9631         }
9632
9633         /* dest is a subvolume */
9634         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9635                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9636                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9637                                           root_objectid,
9638                                           new_dentry->d_name.name,
9639                                           new_dentry->d_name.len);
9640         } else { /* dest is an inode */
9641                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9642                                            new_dentry->d_inode,
9643                                            new_dentry->d_name.name,
9644                                            new_dentry->d_name.len);
9645                 if (!ret)
9646                         ret = btrfs_update_inode(trans, dest, new_inode);
9647         }
9648         if (ret) {
9649                 btrfs_abort_transaction(trans, ret);
9650                 goto out_fail;
9651         }
9652
9653         ret = btrfs_add_link(trans, new_dir, old_inode,
9654                              new_dentry->d_name.name,
9655                              new_dentry->d_name.len, 0, old_idx);
9656         if (ret) {
9657                 btrfs_abort_transaction(trans, ret);
9658                 goto out_fail;
9659         }
9660
9661         ret = btrfs_add_link(trans, old_dir, new_inode,
9662                              old_dentry->d_name.name,
9663                              old_dentry->d_name.len, 0, new_idx);
9664         if (ret) {
9665                 btrfs_abort_transaction(trans, ret);
9666                 goto out_fail;
9667         }
9668
9669         if (old_inode->i_nlink == 1)
9670                 BTRFS_I(old_inode)->dir_index = old_idx;
9671         if (new_inode->i_nlink == 1)
9672                 BTRFS_I(new_inode)->dir_index = new_idx;
9673
9674         if (root_log_pinned) {
9675                 parent = new_dentry->d_parent;
9676                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9677                 btrfs_end_log_trans(root);
9678                 root_log_pinned = false;
9679         }
9680         if (dest_log_pinned) {
9681                 parent = old_dentry->d_parent;
9682                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9683                 btrfs_end_log_trans(dest);
9684                 dest_log_pinned = false;
9685         }
9686 out_fail:
9687         /*
9688          * If we have pinned a log and an error happened, we unpin tasks
9689          * trying to sync the log and force them to fallback to a transaction
9690          * commit if the log currently contains any of the inodes involved in
9691          * this rename operation (to ensure we do not persist a log with an
9692          * inconsistent state for any of these inodes or leading to any
9693          * inconsistencies when replayed). If the transaction was aborted, the
9694          * abortion reason is propagated to userspace when attempting to commit
9695          * the transaction. If the log does not contain any of these inodes, we
9696          * allow the tasks to sync it.
9697          */
9698         if (ret && (root_log_pinned || dest_log_pinned)) {
9699                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9700                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9701                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9702                     (new_inode &&
9703                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9704                     btrfs_set_log_full_commit(root->fs_info, trans);
9705
9706                 if (root_log_pinned) {
9707                         btrfs_end_log_trans(root);
9708                         root_log_pinned = false;
9709                 }
9710                 if (dest_log_pinned) {
9711                         btrfs_end_log_trans(dest);
9712                         dest_log_pinned = false;
9713                 }
9714         }
9715         ret = btrfs_end_transaction(trans, root);
9716 out_notrans:
9717         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9718                 up_read(&dest->fs_info->subvol_sem);
9719         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9720                 up_read(&root->fs_info->subvol_sem);
9721
9722         return ret;
9723 }
9724
9725 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9726                                      struct btrfs_root *root,
9727                                      struct inode *dir,
9728                                      struct dentry *dentry)
9729 {
9730         int ret;
9731         struct inode *inode;
9732         u64 objectid;
9733         u64 index;
9734
9735         ret = btrfs_find_free_ino(root, &objectid);
9736         if (ret)
9737                 return ret;
9738
9739         inode = btrfs_new_inode(trans, root, dir,
9740                                 dentry->d_name.name,
9741                                 dentry->d_name.len,
9742                                 btrfs_ino(dir),
9743                                 objectid,
9744                                 S_IFCHR | WHITEOUT_MODE,
9745                                 &index);
9746
9747         if (IS_ERR(inode)) {
9748                 ret = PTR_ERR(inode);
9749                 return ret;
9750         }
9751
9752         inode->i_op = &btrfs_special_inode_operations;
9753         init_special_inode(inode, inode->i_mode,
9754                 WHITEOUT_DEV);
9755
9756         ret = btrfs_init_inode_security(trans, inode, dir,
9757                                 &dentry->d_name);
9758         if (ret)
9759                 goto out;
9760
9761         ret = btrfs_add_nondir(trans, dir, dentry,
9762                                 inode, 0, index);
9763         if (ret)
9764                 goto out;
9765
9766         ret = btrfs_update_inode(trans, root, inode);
9767 out:
9768         unlock_new_inode(inode);
9769         if (ret)
9770                 inode_dec_link_count(inode);
9771         iput(inode);
9772
9773         return ret;
9774 }
9775
9776 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9777                            struct inode *new_dir, struct dentry *new_dentry,
9778                            unsigned int flags)
9779 {
9780         struct btrfs_trans_handle *trans;
9781         unsigned int trans_num_items;
9782         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9783         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9784         struct inode *new_inode = d_inode(new_dentry);
9785         struct inode *old_inode = d_inode(old_dentry);
9786         u64 index = 0;
9787         u64 root_objectid;
9788         int ret;
9789         u64 old_ino = btrfs_ino(old_inode);
9790         bool log_pinned = false;
9791
9792         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9793                 return -EPERM;
9794
9795         /* we only allow rename subvolume link between subvolumes */
9796         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9797                 return -EXDEV;
9798
9799         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9800             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9801                 return -ENOTEMPTY;
9802
9803         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9804             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9805                 return -ENOTEMPTY;
9806
9807
9808         /* check for collisions, even if the  name isn't there */
9809         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9810                              new_dentry->d_name.name,
9811                              new_dentry->d_name.len);
9812
9813         if (ret) {
9814                 if (ret == -EEXIST) {
9815                         /* we shouldn't get
9816                          * eexist without a new_inode */
9817                         if (WARN_ON(!new_inode)) {
9818                                 return ret;
9819                         }
9820                 } else {
9821                         /* maybe -EOVERFLOW */
9822                         return ret;
9823                 }
9824         }
9825         ret = 0;
9826
9827         /*
9828          * we're using rename to replace one file with another.  Start IO on it
9829          * now so  we don't add too much work to the end of the transaction
9830          */
9831         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9832                 filemap_flush(old_inode->i_mapping);
9833
9834         /* close the racy window with snapshot create/destroy ioctl */
9835         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9836                 down_read(&root->fs_info->subvol_sem);
9837         /*
9838          * We want to reserve the absolute worst case amount of items.  So if
9839          * both inodes are subvols and we need to unlink them then that would
9840          * require 4 item modifications, but if they are both normal inodes it
9841          * would require 5 item modifications, so we'll assume they are normal
9842          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9843          * should cover the worst case number of items we'll modify.
9844          * If our rename has the whiteout flag, we need more 5 units for the
9845          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9846          * when selinux is enabled).
9847          */
9848         trans_num_items = 11;
9849         if (flags & RENAME_WHITEOUT)
9850                 trans_num_items += 5;
9851         trans = btrfs_start_transaction(root, trans_num_items);
9852         if (IS_ERR(trans)) {
9853                 ret = PTR_ERR(trans);
9854                 goto out_notrans;
9855         }
9856
9857         if (dest != root)
9858                 btrfs_record_root_in_trans(trans, dest);
9859
9860         ret = btrfs_set_inode_index(new_dir, &index);
9861         if (ret)
9862                 goto out_fail;
9863
9864         BTRFS_I(old_inode)->dir_index = 0ULL;
9865         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9866                 /* force full log commit if subvolume involved. */
9867                 btrfs_set_log_full_commit(root->fs_info, trans);
9868         } else {
9869                 btrfs_pin_log_trans(root);
9870                 log_pinned = true;
9871                 ret = btrfs_insert_inode_ref(trans, dest,
9872                                              new_dentry->d_name.name,
9873                                              new_dentry->d_name.len,
9874                                              old_ino,
9875                                              btrfs_ino(new_dir), index);
9876                 if (ret)
9877                         goto out_fail;
9878         }
9879
9880         inode_inc_iversion(old_dir);
9881         inode_inc_iversion(new_dir);
9882         inode_inc_iversion(old_inode);
9883         old_dir->i_ctime = old_dir->i_mtime =
9884         new_dir->i_ctime = new_dir->i_mtime =
9885         old_inode->i_ctime = current_time(old_dir);
9886
9887         if (old_dentry->d_parent != new_dentry->d_parent)
9888                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9889
9890         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9891                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9892                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9893                                         old_dentry->d_name.name,
9894                                         old_dentry->d_name.len);
9895         } else {
9896                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9897                                         d_inode(old_dentry),
9898                                         old_dentry->d_name.name,
9899                                         old_dentry->d_name.len);
9900                 if (!ret)
9901                         ret = btrfs_update_inode(trans, root, old_inode);
9902         }
9903         if (ret) {
9904                 btrfs_abort_transaction(trans, ret);
9905                 goto out_fail;
9906         }
9907
9908         if (new_inode) {
9909                 inode_inc_iversion(new_inode);
9910                 new_inode->i_ctime = current_time(new_inode);
9911                 if (unlikely(btrfs_ino(new_inode) ==
9912                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9913                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9914                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9915                                                 root_objectid,
9916                                                 new_dentry->d_name.name,
9917                                                 new_dentry->d_name.len);
9918                         BUG_ON(new_inode->i_nlink == 0);
9919                 } else {
9920                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9921                                                  d_inode(new_dentry),
9922                                                  new_dentry->d_name.name,
9923                                                  new_dentry->d_name.len);
9924                 }
9925                 if (!ret && new_inode->i_nlink == 0)
9926                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9927                 if (ret) {
9928                         btrfs_abort_transaction(trans, ret);
9929                         goto out_fail;
9930                 }
9931         }
9932
9933         ret = btrfs_add_link(trans, new_dir, old_inode,
9934                              new_dentry->d_name.name,
9935                              new_dentry->d_name.len, 0, index);
9936         if (ret) {
9937                 btrfs_abort_transaction(trans, ret);
9938                 goto out_fail;
9939         }
9940
9941         if (old_inode->i_nlink == 1)
9942                 BTRFS_I(old_inode)->dir_index = index;
9943
9944         if (log_pinned) {
9945                 struct dentry *parent = new_dentry->d_parent;
9946
9947                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9948                 btrfs_end_log_trans(root);
9949                 log_pinned = false;
9950         }
9951
9952         if (flags & RENAME_WHITEOUT) {
9953                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9954                                                 old_dentry);
9955
9956                 if (ret) {
9957                         btrfs_abort_transaction(trans, ret);
9958                         goto out_fail;
9959                 }
9960         }
9961 out_fail:
9962         /*
9963          * If we have pinned the log and an error happened, we unpin tasks
9964          * trying to sync the log and force them to fallback to a transaction
9965          * commit if the log currently contains any of the inodes involved in
9966          * this rename operation (to ensure we do not persist a log with an
9967          * inconsistent state for any of these inodes or leading to any
9968          * inconsistencies when replayed). If the transaction was aborted, the
9969          * abortion reason is propagated to userspace when attempting to commit
9970          * the transaction. If the log does not contain any of these inodes, we
9971          * allow the tasks to sync it.
9972          */
9973         if (ret && log_pinned) {
9974                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9975                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9976                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9977                     (new_inode &&
9978                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9979                     btrfs_set_log_full_commit(root->fs_info, trans);
9980
9981                 btrfs_end_log_trans(root);
9982                 log_pinned = false;
9983         }
9984         btrfs_end_transaction(trans, root);
9985 out_notrans:
9986         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9987                 up_read(&root->fs_info->subvol_sem);
9988
9989         return ret;
9990 }
9991
9992 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9993                          struct inode *new_dir, struct dentry *new_dentry,
9994                          unsigned int flags)
9995 {
9996         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9997                 return -EINVAL;
9998
9999         if (flags & RENAME_EXCHANGE)
10000                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10001                                           new_dentry);
10002
10003         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10004 }
10005
10006 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10007 {
10008         struct btrfs_delalloc_work *delalloc_work;
10009         struct inode *inode;
10010
10011         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10012                                      work);
10013         inode = delalloc_work->inode;
10014         filemap_flush(inode->i_mapping);
10015         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10016                                 &BTRFS_I(inode)->runtime_flags))
10017                 filemap_flush(inode->i_mapping);
10018
10019         if (delalloc_work->delay_iput)
10020                 btrfs_add_delayed_iput(inode);
10021         else
10022                 iput(inode);
10023         complete(&delalloc_work->completion);
10024 }
10025
10026 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10027                                                     int delay_iput)
10028 {
10029         struct btrfs_delalloc_work *work;
10030
10031         work = kmalloc(sizeof(*work), GFP_NOFS);
10032         if (!work)
10033                 return NULL;
10034
10035         init_completion(&work->completion);
10036         INIT_LIST_HEAD(&work->list);
10037         work->inode = inode;
10038         work->delay_iput = delay_iput;
10039         WARN_ON_ONCE(!inode);
10040         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10041                         btrfs_run_delalloc_work, NULL, NULL);
10042
10043         return work;
10044 }
10045
10046 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10047 {
10048         wait_for_completion(&work->completion);
10049         kfree(work);
10050 }
10051
10052 /*
10053  * some fairly slow code that needs optimization. This walks the list
10054  * of all the inodes with pending delalloc and forces them to disk.
10055  */
10056 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10057                                    int nr)
10058 {
10059         struct btrfs_inode *binode;
10060         struct inode *inode;
10061         struct btrfs_delalloc_work *work, *next;
10062         struct list_head works;
10063         struct list_head splice;
10064         int ret = 0;
10065
10066         INIT_LIST_HEAD(&works);
10067         INIT_LIST_HEAD(&splice);
10068
10069         mutex_lock(&root->delalloc_mutex);
10070         spin_lock(&root->delalloc_lock);
10071         list_splice_init(&root->delalloc_inodes, &splice);
10072         while (!list_empty(&splice)) {
10073                 binode = list_entry(splice.next, struct btrfs_inode,
10074                                     delalloc_inodes);
10075
10076                 list_move_tail(&binode->delalloc_inodes,
10077                                &root->delalloc_inodes);
10078                 inode = igrab(&binode->vfs_inode);
10079                 if (!inode) {
10080                         cond_resched_lock(&root->delalloc_lock);
10081                         continue;
10082                 }
10083                 spin_unlock(&root->delalloc_lock);
10084
10085                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10086                 if (!work) {
10087                         if (delay_iput)
10088                                 btrfs_add_delayed_iput(inode);
10089                         else
10090                                 iput(inode);
10091                         ret = -ENOMEM;
10092                         goto out;
10093                 }
10094                 list_add_tail(&work->list, &works);
10095                 btrfs_queue_work(root->fs_info->flush_workers,
10096                                  &work->work);
10097                 ret++;
10098                 if (nr != -1 && ret >= nr)
10099                         goto out;
10100                 cond_resched();
10101                 spin_lock(&root->delalloc_lock);
10102         }
10103         spin_unlock(&root->delalloc_lock);
10104
10105 out:
10106         list_for_each_entry_safe(work, next, &works, list) {
10107                 list_del_init(&work->list);
10108                 btrfs_wait_and_free_delalloc_work(work);
10109         }
10110
10111         if (!list_empty_careful(&splice)) {
10112                 spin_lock(&root->delalloc_lock);
10113                 list_splice_tail(&splice, &root->delalloc_inodes);
10114                 spin_unlock(&root->delalloc_lock);
10115         }
10116         mutex_unlock(&root->delalloc_mutex);
10117         return ret;
10118 }
10119
10120 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10121 {
10122         int ret;
10123
10124         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10125                 return -EROFS;
10126
10127         ret = __start_delalloc_inodes(root, delay_iput, -1);
10128         if (ret > 0)
10129                 ret = 0;
10130         /*
10131          * the filemap_flush will queue IO into the worker threads, but
10132          * we have to make sure the IO is actually started and that
10133          * ordered extents get created before we return
10134          */
10135         atomic_inc(&root->fs_info->async_submit_draining);
10136         while (atomic_read(&root->fs_info->nr_async_submits) ||
10137               atomic_read(&root->fs_info->async_delalloc_pages)) {
10138                 wait_event(root->fs_info->async_submit_wait,
10139                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10140                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10141         }
10142         atomic_dec(&root->fs_info->async_submit_draining);
10143         return ret;
10144 }
10145
10146 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10147                                int nr)
10148 {
10149         struct btrfs_root *root;
10150         struct list_head splice;
10151         int ret;
10152
10153         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10154                 return -EROFS;
10155
10156         INIT_LIST_HEAD(&splice);
10157
10158         mutex_lock(&fs_info->delalloc_root_mutex);
10159         spin_lock(&fs_info->delalloc_root_lock);
10160         list_splice_init(&fs_info->delalloc_roots, &splice);
10161         while (!list_empty(&splice) && nr) {
10162                 root = list_first_entry(&splice, struct btrfs_root,
10163                                         delalloc_root);
10164                 root = btrfs_grab_fs_root(root);
10165                 BUG_ON(!root);
10166                 list_move_tail(&root->delalloc_root,
10167                                &fs_info->delalloc_roots);
10168                 spin_unlock(&fs_info->delalloc_root_lock);
10169
10170                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10171                 btrfs_put_fs_root(root);
10172                 if (ret < 0)
10173                         goto out;
10174
10175                 if (nr != -1) {
10176                         nr -= ret;
10177                         WARN_ON(nr < 0);
10178                 }
10179                 spin_lock(&fs_info->delalloc_root_lock);
10180         }
10181         spin_unlock(&fs_info->delalloc_root_lock);
10182
10183         ret = 0;
10184         atomic_inc(&fs_info->async_submit_draining);
10185         while (atomic_read(&fs_info->nr_async_submits) ||
10186               atomic_read(&fs_info->async_delalloc_pages)) {
10187                 wait_event(fs_info->async_submit_wait,
10188                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10189                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10190         }
10191         atomic_dec(&fs_info->async_submit_draining);
10192 out:
10193         if (!list_empty_careful(&splice)) {
10194                 spin_lock(&fs_info->delalloc_root_lock);
10195                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10196                 spin_unlock(&fs_info->delalloc_root_lock);
10197         }
10198         mutex_unlock(&fs_info->delalloc_root_mutex);
10199         return ret;
10200 }
10201
10202 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10203                          const char *symname)
10204 {
10205         struct btrfs_trans_handle *trans;
10206         struct btrfs_root *root = BTRFS_I(dir)->root;
10207         struct btrfs_path *path;
10208         struct btrfs_key key;
10209         struct inode *inode = NULL;
10210         int err;
10211         int drop_inode = 0;
10212         u64 objectid;
10213         u64 index = 0;
10214         int name_len;
10215         int datasize;
10216         unsigned long ptr;
10217         struct btrfs_file_extent_item *ei;
10218         struct extent_buffer *leaf;
10219
10220         name_len = strlen(symname);
10221         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10222                 return -ENAMETOOLONG;
10223
10224         /*
10225          * 2 items for inode item and ref
10226          * 2 items for dir items
10227          * 1 item for updating parent inode item
10228          * 1 item for the inline extent item
10229          * 1 item for xattr if selinux is on
10230          */
10231         trans = btrfs_start_transaction(root, 7);
10232         if (IS_ERR(trans))
10233                 return PTR_ERR(trans);
10234
10235         err = btrfs_find_free_ino(root, &objectid);
10236         if (err)
10237                 goto out_unlock;
10238
10239         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10240                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10241                                 S_IFLNK|S_IRWXUGO, &index);
10242         if (IS_ERR(inode)) {
10243                 err = PTR_ERR(inode);
10244                 goto out_unlock;
10245         }
10246
10247         /*
10248         * If the active LSM wants to access the inode during
10249         * d_instantiate it needs these. Smack checks to see
10250         * if the filesystem supports xattrs by looking at the
10251         * ops vector.
10252         */
10253         inode->i_fop = &btrfs_file_operations;
10254         inode->i_op = &btrfs_file_inode_operations;
10255         inode->i_mapping->a_ops = &btrfs_aops;
10256         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10257
10258         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10259         if (err)
10260                 goto out_unlock_inode;
10261
10262         path = btrfs_alloc_path();
10263         if (!path) {
10264                 err = -ENOMEM;
10265                 goto out_unlock_inode;
10266         }
10267         key.objectid = btrfs_ino(inode);
10268         key.offset = 0;
10269         key.type = BTRFS_EXTENT_DATA_KEY;
10270         datasize = btrfs_file_extent_calc_inline_size(name_len);
10271         err = btrfs_insert_empty_item(trans, root, path, &key,
10272                                       datasize);
10273         if (err) {
10274                 btrfs_free_path(path);
10275                 goto out_unlock_inode;
10276         }
10277         leaf = path->nodes[0];
10278         ei = btrfs_item_ptr(leaf, path->slots[0],
10279                             struct btrfs_file_extent_item);
10280         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10281         btrfs_set_file_extent_type(leaf, ei,
10282                                    BTRFS_FILE_EXTENT_INLINE);
10283         btrfs_set_file_extent_encryption(leaf, ei, 0);
10284         btrfs_set_file_extent_compression(leaf, ei, 0);
10285         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10286         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10287
10288         ptr = btrfs_file_extent_inline_start(ei);
10289         write_extent_buffer(leaf, symname, ptr, name_len);
10290         btrfs_mark_buffer_dirty(leaf);
10291         btrfs_free_path(path);
10292
10293         inode->i_op = &btrfs_symlink_inode_operations;
10294         inode_nohighmem(inode);
10295         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10296         inode_set_bytes(inode, name_len);
10297         btrfs_i_size_write(inode, name_len);
10298         err = btrfs_update_inode(trans, root, inode);
10299         /*
10300          * Last step, add directory indexes for our symlink inode. This is the
10301          * last step to avoid extra cleanup of these indexes if an error happens
10302          * elsewhere above.
10303          */
10304         if (!err)
10305                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10306         if (err) {
10307                 drop_inode = 1;
10308                 goto out_unlock_inode;
10309         }
10310
10311         unlock_new_inode(inode);
10312         d_instantiate(dentry, inode);
10313
10314 out_unlock:
10315         btrfs_end_transaction(trans, root);
10316         if (drop_inode) {
10317                 inode_dec_link_count(inode);
10318                 iput(inode);
10319         }
10320         btrfs_btree_balance_dirty(root);
10321         return err;
10322
10323 out_unlock_inode:
10324         drop_inode = 1;
10325         unlock_new_inode(inode);
10326         goto out_unlock;
10327 }
10328
10329 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10330                                        u64 start, u64 num_bytes, u64 min_size,
10331                                        loff_t actual_len, u64 *alloc_hint,
10332                                        struct btrfs_trans_handle *trans)
10333 {
10334         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10335         struct extent_map *em;
10336         struct btrfs_root *root = BTRFS_I(inode)->root;
10337         struct btrfs_key ins;
10338         u64 cur_offset = start;
10339         u64 i_size;
10340         u64 cur_bytes;
10341         u64 last_alloc = (u64)-1;
10342         int ret = 0;
10343         bool own_trans = true;
10344         u64 end = start + num_bytes - 1;
10345
10346         if (trans)
10347                 own_trans = false;
10348         while (num_bytes > 0) {
10349                 if (own_trans) {
10350                         trans = btrfs_start_transaction(root, 3);
10351                         if (IS_ERR(trans)) {
10352                                 ret = PTR_ERR(trans);
10353                                 break;
10354                         }
10355                 }
10356
10357                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10358                 cur_bytes = max(cur_bytes, min_size);
10359                 /*
10360                  * If we are severely fragmented we could end up with really
10361                  * small allocations, so if the allocator is returning small
10362                  * chunks lets make its job easier by only searching for those
10363                  * sized chunks.
10364                  */
10365                 cur_bytes = min(cur_bytes, last_alloc);
10366                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10367                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10368                 if (ret) {
10369                         if (own_trans)
10370                                 btrfs_end_transaction(trans, root);
10371                         break;
10372                 }
10373                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10374
10375                 last_alloc = ins.offset;
10376                 ret = insert_reserved_file_extent(trans, inode,
10377                                                   cur_offset, ins.objectid,
10378                                                   ins.offset, ins.offset,
10379                                                   ins.offset, 0, 0, 0,
10380                                                   BTRFS_FILE_EXTENT_PREALLOC);
10381                 if (ret) {
10382                         btrfs_free_reserved_extent(root, ins.objectid,
10383                                                    ins.offset, 0);
10384                         btrfs_abort_transaction(trans, ret);
10385                         if (own_trans)
10386                                 btrfs_end_transaction(trans, root);
10387                         break;
10388                 }
10389
10390                 btrfs_drop_extent_cache(inode, cur_offset,
10391                                         cur_offset + ins.offset -1, 0);
10392
10393                 em = alloc_extent_map();
10394                 if (!em) {
10395                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10396                                 &BTRFS_I(inode)->runtime_flags);
10397                         goto next;
10398                 }
10399
10400                 em->start = cur_offset;
10401                 em->orig_start = cur_offset;
10402                 em->len = ins.offset;
10403                 em->block_start = ins.objectid;
10404                 em->block_len = ins.offset;
10405                 em->orig_block_len = ins.offset;
10406                 em->ram_bytes = ins.offset;
10407                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10408                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10409                 em->generation = trans->transid;
10410
10411                 while (1) {
10412                         write_lock(&em_tree->lock);
10413                         ret = add_extent_mapping(em_tree, em, 1);
10414                         write_unlock(&em_tree->lock);
10415                         if (ret != -EEXIST)
10416                                 break;
10417                         btrfs_drop_extent_cache(inode, cur_offset,
10418                                                 cur_offset + ins.offset - 1,
10419                                                 0);
10420                 }
10421                 free_extent_map(em);
10422 next:
10423                 num_bytes -= ins.offset;
10424                 cur_offset += ins.offset;
10425                 *alloc_hint = ins.objectid + ins.offset;
10426
10427                 inode_inc_iversion(inode);
10428                 inode->i_ctime = current_time(inode);
10429                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10430                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10431                     (actual_len > inode->i_size) &&
10432                     (cur_offset > inode->i_size)) {
10433                         if (cur_offset > actual_len)
10434                                 i_size = actual_len;
10435                         else
10436                                 i_size = cur_offset;
10437                         i_size_write(inode, i_size);
10438                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10439                 }
10440
10441                 ret = btrfs_update_inode(trans, root, inode);
10442
10443                 if (ret) {
10444                         btrfs_abort_transaction(trans, ret);
10445                         if (own_trans)
10446                                 btrfs_end_transaction(trans, root);
10447                         break;
10448                 }
10449
10450                 if (own_trans)
10451                         btrfs_end_transaction(trans, root);
10452         }
10453         if (cur_offset < end)
10454                 btrfs_free_reserved_data_space(inode, cur_offset,
10455                         end - cur_offset + 1);
10456         return ret;
10457 }
10458
10459 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10460                               u64 start, u64 num_bytes, u64 min_size,
10461                               loff_t actual_len, u64 *alloc_hint)
10462 {
10463         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10464                                            min_size, actual_len, alloc_hint,
10465                                            NULL);
10466 }
10467
10468 int btrfs_prealloc_file_range_trans(struct inode *inode,
10469                                     struct btrfs_trans_handle *trans, int mode,
10470                                     u64 start, u64 num_bytes, u64 min_size,
10471                                     loff_t actual_len, u64 *alloc_hint)
10472 {
10473         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10474                                            min_size, actual_len, alloc_hint, trans);
10475 }
10476
10477 static int btrfs_set_page_dirty(struct page *page)
10478 {
10479         return __set_page_dirty_nobuffers(page);
10480 }
10481
10482 static int btrfs_permission(struct inode *inode, int mask)
10483 {
10484         struct btrfs_root *root = BTRFS_I(inode)->root;
10485         umode_t mode = inode->i_mode;
10486
10487         if (mask & MAY_WRITE &&
10488             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10489                 if (btrfs_root_readonly(root))
10490                         return -EROFS;
10491                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10492                         return -EACCES;
10493         }
10494         return generic_permission(inode, mask);
10495 }
10496
10497 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10498 {
10499         struct btrfs_trans_handle *trans;
10500         struct btrfs_root *root = BTRFS_I(dir)->root;
10501         struct inode *inode = NULL;
10502         u64 objectid;
10503         u64 index;
10504         int ret = 0;
10505
10506         /*
10507          * 5 units required for adding orphan entry
10508          */
10509         trans = btrfs_start_transaction(root, 5);
10510         if (IS_ERR(trans))
10511                 return PTR_ERR(trans);
10512
10513         ret = btrfs_find_free_ino(root, &objectid);
10514         if (ret)
10515                 goto out;
10516
10517         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10518                                 btrfs_ino(dir), objectid, mode, &index);
10519         if (IS_ERR(inode)) {
10520                 ret = PTR_ERR(inode);
10521                 inode = NULL;
10522                 goto out;
10523         }
10524
10525         inode->i_fop = &btrfs_file_operations;
10526         inode->i_op = &btrfs_file_inode_operations;
10527
10528         inode->i_mapping->a_ops = &btrfs_aops;
10529         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10530
10531         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10532         if (ret)
10533                 goto out_inode;
10534
10535         ret = btrfs_update_inode(trans, root, inode);
10536         if (ret)
10537                 goto out_inode;
10538         ret = btrfs_orphan_add(trans, inode);
10539         if (ret)
10540                 goto out_inode;
10541
10542         /*
10543          * We set number of links to 0 in btrfs_new_inode(), and here we set
10544          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10545          * through:
10546          *
10547          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10548          */
10549         set_nlink(inode, 1);
10550         unlock_new_inode(inode);
10551         d_tmpfile(dentry, inode);
10552         mark_inode_dirty(inode);
10553
10554 out:
10555         btrfs_end_transaction(trans, root);
10556         if (ret)
10557                 iput(inode);
10558         btrfs_balance_delayed_items(root);
10559         btrfs_btree_balance_dirty(root);
10560         return ret;
10561
10562 out_inode:
10563         unlock_new_inode(inode);
10564         goto out;
10565
10566 }
10567
10568 static const struct inode_operations btrfs_dir_inode_operations = {
10569         .getattr        = btrfs_getattr,
10570         .lookup         = btrfs_lookup,
10571         .create         = btrfs_create,
10572         .unlink         = btrfs_unlink,
10573         .link           = btrfs_link,
10574         .mkdir          = btrfs_mkdir,
10575         .rmdir          = btrfs_rmdir,
10576         .rename         = btrfs_rename2,
10577         .symlink        = btrfs_symlink,
10578         .setattr        = btrfs_setattr,
10579         .mknod          = btrfs_mknod,
10580         .listxattr      = btrfs_listxattr,
10581         .permission     = btrfs_permission,
10582         .get_acl        = btrfs_get_acl,
10583         .set_acl        = btrfs_set_acl,
10584         .update_time    = btrfs_update_time,
10585         .tmpfile        = btrfs_tmpfile,
10586 };
10587 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10588         .lookup         = btrfs_lookup,
10589         .permission     = btrfs_permission,
10590         .get_acl        = btrfs_get_acl,
10591         .set_acl        = btrfs_set_acl,
10592         .update_time    = btrfs_update_time,
10593 };
10594
10595 static const struct file_operations btrfs_dir_file_operations = {
10596         .llseek         = generic_file_llseek,
10597         .read           = generic_read_dir,
10598         .iterate_shared = btrfs_real_readdir,
10599         .unlocked_ioctl = btrfs_ioctl,
10600 #ifdef CONFIG_COMPAT
10601         .compat_ioctl   = btrfs_compat_ioctl,
10602 #endif
10603         .release        = btrfs_release_file,
10604         .fsync          = btrfs_sync_file,
10605 };
10606
10607 static const struct extent_io_ops btrfs_extent_io_ops = {
10608         .fill_delalloc = run_delalloc_range,
10609         .submit_bio_hook = btrfs_submit_bio_hook,
10610         .merge_bio_hook = btrfs_merge_bio_hook,
10611         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10612         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10613         .writepage_start_hook = btrfs_writepage_start_hook,
10614         .set_bit_hook = btrfs_set_bit_hook,
10615         .clear_bit_hook = btrfs_clear_bit_hook,
10616         .merge_extent_hook = btrfs_merge_extent_hook,
10617         .split_extent_hook = btrfs_split_extent_hook,
10618 };
10619
10620 /*
10621  * btrfs doesn't support the bmap operation because swapfiles
10622  * use bmap to make a mapping of extents in the file.  They assume
10623  * these extents won't change over the life of the file and they
10624  * use the bmap result to do IO directly to the drive.
10625  *
10626  * the btrfs bmap call would return logical addresses that aren't
10627  * suitable for IO and they also will change frequently as COW
10628  * operations happen.  So, swapfile + btrfs == corruption.
10629  *
10630  * For now we're avoiding this by dropping bmap.
10631  */
10632 static const struct address_space_operations btrfs_aops = {
10633         .readpage       = btrfs_readpage,
10634         .writepage      = btrfs_writepage,
10635         .writepages     = btrfs_writepages,
10636         .readpages      = btrfs_readpages,
10637         .direct_IO      = btrfs_direct_IO,
10638         .invalidatepage = btrfs_invalidatepage,
10639         .releasepage    = btrfs_releasepage,
10640         .set_page_dirty = btrfs_set_page_dirty,
10641         .error_remove_page = generic_error_remove_page,
10642 };
10643
10644 static const struct address_space_operations btrfs_symlink_aops = {
10645         .readpage       = btrfs_readpage,
10646         .writepage      = btrfs_writepage,
10647         .invalidatepage = btrfs_invalidatepage,
10648         .releasepage    = btrfs_releasepage,
10649 };
10650
10651 static const struct inode_operations btrfs_file_inode_operations = {
10652         .getattr        = btrfs_getattr,
10653         .setattr        = btrfs_setattr,
10654         .listxattr      = btrfs_listxattr,
10655         .permission     = btrfs_permission,
10656         .fiemap         = btrfs_fiemap,
10657         .get_acl        = btrfs_get_acl,
10658         .set_acl        = btrfs_set_acl,
10659         .update_time    = btrfs_update_time,
10660 };
10661 static const struct inode_operations btrfs_special_inode_operations = {
10662         .getattr        = btrfs_getattr,
10663         .setattr        = btrfs_setattr,
10664         .permission     = btrfs_permission,
10665         .listxattr      = btrfs_listxattr,
10666         .get_acl        = btrfs_get_acl,
10667         .set_acl        = btrfs_set_acl,
10668         .update_time    = btrfs_update_time,
10669 };
10670 static const struct inode_operations btrfs_symlink_inode_operations = {
10671         .readlink       = generic_readlink,
10672         .get_link       = page_get_link,
10673         .getattr        = btrfs_getattr,
10674         .setattr        = btrfs_setattr,
10675         .permission     = btrfs_permission,
10676         .listxattr      = btrfs_listxattr,
10677         .update_time    = btrfs_update_time,
10678 };
10679
10680 const struct dentry_operations btrfs_dentry_operations = {
10681         .d_delete       = btrfs_dentry_delete,
10682         .d_release      = btrfs_dentry_release,
10683 };