fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/mpage.h>
  18 #include <linux/swap.h>
  19 #include <linux/writeback.h>
  20 #include <linux/compat.h>
  21 #include <linux/bit_spinlock.h>
  22 #include <linux/xattr.h>
  23 #include <linux/posix_acl.h>
  24 #include <linux/falloc.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/mount.h>
  28 #include <linux/btrfs.h>
  29 #include <linux/blkdev.h>
  30 #include <linux/posix_acl_xattr.h>
  31 #include <linux/uio.h>
  32 #include <linux/magic.h>
  33 #include <linux/iversion.h>
  34 #include <asm/unaligned.h>
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "ordered-data.h"
  41 #include "xattr.h"
  42 #include "tree-log.h"
  43 #include "volumes.h"
  44 #include "compression.h"
  45 #include "locking.h"
  46 #include "free-space-cache.h"
  47 #include "inode-map.h"
  48 #include "backref.h"
  49 #include "props.h"
  50 #include "qgroup.h"
  51 #include "dedupe.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_dir_ro_inode_operations;
  68 static const struct inode_operations btrfs_special_inode_operations;
  69 static const struct inode_operations btrfs_file_inode_operations;
  70 static const struct address_space_operations btrfs_aops;
  71 static const struct address_space_operations btrfs_symlink_aops;
  72 static const struct file_operations btrfs_dir_file_operations;
  73 static const struct extent_io_ops btrfs_extent_io_ops;
  74
  75 static struct kmem_cache *btrfs_inode_cachep;
  76 struct kmem_cache *btrfs_trans_handle_cachep;
  77 struct kmem_cache *btrfs_path_cachep;
  78 struct kmem_cache *btrfs_free_space_cachep;
  79
  80 #define S_SHIFT 12
  81 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  82         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  83         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  84         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  85         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  86         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  87         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  88         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  89 };
  90
  91 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  92 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  93 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  94 static noinline int cow_file_range(struct inode *inode,
  95                                    struct page *locked_page,
  96                                    u64 start, u64 end, u64 delalloc_end,
  97                                    int *page_started, unsigned long *nr_written,
  98                                    int unlock, struct btrfs_dedupe_hash *hash);
  99 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 100                                        u64 orig_start, u64 block_start,
 101                                        u64 block_len, u64 orig_block_len,
 102                                        u64 ram_bytes, int compress_type,
 103                                        int type);
 104
 105 static void __endio_write_update_ordered(struct inode *inode,
 106                                          const u64 offset, const u64 bytes,
 107                                          const bool uptodate);
 108
 109 /*
 110  * Cleanup all submitted ordered extents in specified range to handle errors
 111  * from the fill_dellaloc() callback.
 112  *
 113  * NOTE: caller must ensure that when an error happens, it can not call
 114  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 115  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 116  * to be released, which we want to happen only when finishing the ordered
 117  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 118  * fill_delalloc() callback already does proper cleanup for the first page of
 119  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 120  * range of the first page.
 121  */
 122 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 123                                                  const u64 offset,
 124                                                  const u64 bytes)
 125 {
 126         unsigned long index = offset >> PAGE_SHIFT;
 127         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 128         struct page *page;
 129
 130         while (index <= end_index) {
 131                 page = find_get_page(inode->i_mapping, index);
 132                 index++;
 133                 if (!page)
 134                         continue;
 135                 ClearPagePrivate2(page);
 136                 put_page(page);
 137         }
 138         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 139                                             bytes - PAGE_SIZE, false);
 140 }
 141
 142 static int btrfs_dirty_inode(struct inode *inode);
 143
 144 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 145 void btrfs_test_inode_set_ops(struct inode *inode)
 146 {
 147         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 148 }
 149 #endif
 150
 151 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 152                                      struct inode *inode,  struct inode *dir,
 153                                      const struct qstr *qstr)
 154 {
 155         int err;
 156
 157         err = btrfs_init_acl(trans, inode, dir);
 158         if (!err)
 159                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 160         return err;
 161 }
 162
 163 /*
 164  * this does all the hard work for inserting an inline extent into
 165  * the btree.  The caller should have done a btrfs_drop_extents so that
 166  * no overlapping inline items exist in the btree
 167  */
 168 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 169                                 struct btrfs_path *path, int extent_inserted,
 170                                 struct btrfs_root *root, struct inode *inode,
 171                                 u64 start, size_t size, size_t compressed_size,
 172                                 int compress_type,
 173                                 struct page **compressed_pages)
 174 {
 175         struct extent_buffer *leaf;
 176         struct page *page = NULL;
 177         char *kaddr;
 178         unsigned long ptr;
 179         struct btrfs_file_extent_item *ei;
 180         int ret;
 181         size_t cur_size = size;
 182         unsigned long offset;
 183
 184         if (compressed_size && compressed_pages)
 185                 cur_size = compressed_size;
 186
 187         inode_add_bytes(inode, size);
 188
 189         if (!extent_inserted) {
 190                 struct btrfs_key key;
 191                 size_t datasize;
 192
 193                 key.objectid = btrfs_ino(BTRFS_I(inode));
 194                 key.offset = start;
 195                 key.type = BTRFS_EXTENT_DATA_KEY;
 196
 197                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 198                 path->leave_spinning = 1;
 199                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 200                                               datasize);
 201                 if (ret)
 202                         goto fail;
 203         }
 204         leaf = path->nodes[0];
 205         ei = btrfs_item_ptr(leaf, path->slots[0],
 206                             struct btrfs_file_extent_item);
 207         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 208         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 209         btrfs_set_file_extent_encryption(leaf, ei, 0);
 210         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 211         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 212         ptr = btrfs_file_extent_inline_start(ei);
 213
 214         if (compress_type != BTRFS_COMPRESS_NONE) {
 215                 struct page *cpage;
 216                 int i = 0;
 217                 while (compressed_size > 0) {
 218                         cpage = compressed_pages[i];
 219                         cur_size = min_t(unsigned long, compressed_size,
 220                                        PAGE_SIZE);
 221
 222                         kaddr = kmap_atomic(cpage);
 223                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 224                         kunmap_atomic(kaddr);
 225
 226                         i++;
 227                         ptr += cur_size;
 228                         compressed_size -= cur_size;
 229                 }
 230                 btrfs_set_file_extent_compression(leaf, ei,
 231                                                   compress_type);
 232         } else {
 233                 page = find_get_page(inode->i_mapping,
 234                                      start >> PAGE_SHIFT);
 235                 btrfs_set_file_extent_compression(leaf, ei, 0);
 236                 kaddr = kmap_atomic(page);
 237                 offset = start & (PAGE_SIZE - 1);
 238                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 239                 kunmap_atomic(kaddr);
 240                 put_page(page);
 241         }
 242         btrfs_mark_buffer_dirty(leaf);
 243         btrfs_release_path(path);
 244
 245         /*
 246          * we're an inline extent, so nobody can
 247          * extend the file past i_size without locking
 248          * a page we already have locked.
 249          *
 250          * We must do any isize and inode updates
 251          * before we unlock the pages.  Otherwise we
 252          * could end up racing with unlink.
 253          */
 254         BTRFS_I(inode)->disk_i_size = inode->i_size;
 255         ret = btrfs_update_inode(trans, root, inode);
 256
 257 fail:
 258         return ret;
 259 }
 260
 261
 262 /*
 263  * conditionally insert an inline extent into the file.  This
 264  * does the checks required to make sure the data is small enough
 265  * to fit as an inline extent.
 266  */
 267 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 268                                           u64 end, size_t compressed_size,
 269                                           int compress_type,
 270                                           struct page **compressed_pages)
 271 {
 272         struct btrfs_root *root = BTRFS_I(inode)->root;
 273         struct btrfs_fs_info *fs_info = root->fs_info;
 274         struct btrfs_trans_handle *trans;
 275         u64 isize = i_size_read(inode);
 276         u64 actual_end = min(end + 1, isize);
 277         u64 inline_len = actual_end - start;
 278         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 279         u64 data_len = inline_len;
 280         int ret;
 281         struct btrfs_path *path;
 282         int extent_inserted = 0;
 283         u32 extent_item_size;
 284
 285         if (compressed_size)
 286                 data_len = compressed_size;
 287
 288         if (start > 0 ||
 289             actual_end > fs_info->sectorsize ||
 290             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 291             (!compressed_size &&
 292             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 293             end + 1 < isize ||
 294             data_len > fs_info->max_inline) {
 295                 return 1;
 296         }
 297
 298         path = btrfs_alloc_path();
 299         if (!path)
 300                 return -ENOMEM;
 301
 302         trans = btrfs_join_transaction(root);
 303         if (IS_ERR(trans)) {
 304                 btrfs_free_path(path);
 305                 return PTR_ERR(trans);
 306         }
 307         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 308
 309         if (compressed_size && compressed_pages)
 310                 extent_item_size = btrfs_file_extent_calc_inline_size(
 311                    compressed_size);
 312         else
 313                 extent_item_size = btrfs_file_extent_calc_inline_size(
 314                     inline_len);
 315
 316         ret = __btrfs_drop_extents(trans, root, inode, path,
 317                                    start, aligned_end, NULL,
 318                                    1, 1, extent_item_size, &extent_inserted);
 319         if (ret) {
 320                 btrfs_abort_transaction(trans, ret);
 321                 goto out;
 322         }
 323
 324         if (isize > actual_end)
 325                 inline_len = min_t(u64, isize, actual_end);
 326         ret = insert_inline_extent(trans, path, extent_inserted,
 327                                    root, inode, start,
 328                                    inline_len, compressed_size,
 329                                    compress_type, compressed_pages);
 330         if (ret && ret != -ENOSPC) {
 331                 btrfs_abort_transaction(trans, ret);
 332                 goto out;
 333         } else if (ret == -ENOSPC) {
 334                 ret = 1;
 335                 goto out;
 336         }
 337
 338         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 339         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 340 out:
 341         /*
 342          * Don't forget to free the reserved space, as for inlined extent
 343          * it won't count as data extent, free them directly here.
 344          * And at reserve time, it's always aligned to page size, so
 345          * just free one page here.
 346          */
 347         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 348         btrfs_free_path(path);
 349         btrfs_end_transaction(trans);
 350         return ret;
 351 }
 352
 353 struct async_extent {
 354         u64 start;
 355         u64 ram_size;
 356         u64 compressed_size;
 357         struct page **pages;
 358         unsigned long nr_pages;
 359         int compress_type;
 360         struct list_head list;
 361 };
 362
 363 struct async_cow {
 364         struct inode *inode;
 365         struct btrfs_root *root;
 366         struct page *locked_page;
 367         u64 start;
 368         u64 end;
 369         unsigned int write_flags;
 370         struct list_head extents;
 371         struct btrfs_work work;
 372 };
 373
 374 static noinline int add_async_extent(struct async_cow *cow,
 375                                      u64 start, u64 ram_size,
 376                                      u64 compressed_size,
 377                                      struct page **pages,
 378                                      unsigned long nr_pages,
 379                                      int compress_type)
 380 {
 381         struct async_extent *async_extent;
 382
 383         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 384         BUG_ON(!async_extent); /* -ENOMEM */
 385         async_extent->start = start;
 386         async_extent->ram_size = ram_size;
 387         async_extent->compressed_size = compressed_size;
 388         async_extent->pages = pages;
 389         async_extent->nr_pages = nr_pages;
 390         async_extent->compress_type = compress_type;
 391         list_add_tail(&async_extent->list, &cow->extents);
 392         return 0;
 393 }
 394
 395 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 396 {
 397         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 398
 399         /* force compress */
 400         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 401                 return 1;
 402         /* defrag ioctl */
 403         if (BTRFS_I(inode)->defrag_compress)
 404                 return 1;
 405         /* bad compression ratios */
 406         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 407                 return 0;
 408         if (btrfs_test_opt(fs_info, COMPRESS) ||
 409             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 410             BTRFS_I(inode)->prop_compress)
 411                 return btrfs_compress_heuristic(inode, start, end);
 412         return 0;
 413 }
 414
 415 static inline void inode_should_defrag(struct btrfs_inode *inode,
 416                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 417 {
 418         /* If this is a small write inside eof, kick off a defrag */
 419         if (num_bytes < small_write &&
 420             (start > 0 || end + 1 < inode->disk_i_size))
 421                 btrfs_add_inode_defrag(NULL, inode);
 422 }
 423
 424 /*
 425  * we create compressed extents in two phases.  The first
 426  * phase compresses a range of pages that have already been
 427  * locked (both pages and state bits are locked).
 428  *
 429  * This is done inside an ordered work queue, and the compression
 430  * is spread across many cpus.  The actual IO submission is step
 431  * two, and the ordered work queue takes care of making sure that
 432  * happens in the same order things were put onto the queue by
 433  * writepages and friends.
 434  *
 435  * If this code finds it can't get good compression, it puts an
 436  * entry onto the work queue to write the uncompressed bytes.  This
 437  * makes sure that both compressed inodes and uncompressed inodes
 438  * are written in the same order that the flusher thread sent them
 439  * down.
 440  */
 441 static noinline void compress_file_range(struct inode *inode,
 442                                         struct page *locked_page,
 443                                         u64 start, u64 end,
 444                                         struct async_cow *async_cow,
 445                                         int *num_added)
 446 {
 447         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 448         u64 blocksize = fs_info->sectorsize;
 449         u64 actual_end;
 450         u64 isize = i_size_read(inode);
 451         int ret = 0;
 452         struct page **pages = NULL;
 453         unsigned long nr_pages;
 454         unsigned long total_compressed = 0;
 455         unsigned long total_in = 0;
 456         int i;
 457         int will_compress;
 458         int compress_type = fs_info->compress_type;
 459         int redirty = 0;
 460
 461         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 462                         SZ_16K);
 463
 464         actual_end = min_t(u64, isize, end + 1);
 465 again:
 466         will_compress = 0;
 467         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 468         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 469         nr_pages = min_t(unsigned long, nr_pages,
 470                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 471
 472         /*
 473          * we don't want to send crud past the end of i_size through
 474          * compression, that's just a waste of CPU time.  So, if the
 475          * end of the file is before the start of our current
 476          * requested range of bytes, we bail out to the uncompressed
 477          * cleanup code that can deal with all of this.
 478          *
 479          * It isn't really the fastest way to fix things, but this is a
 480          * very uncommon corner.
 481          */
 482         if (actual_end <= start)
 483                 goto cleanup_and_bail_uncompressed;
 484
 485         total_compressed = actual_end - start;
 486
 487         /*
 488          * skip compression for a small file range(<=blocksize) that
 489          * isn't an inline extent, since it doesn't save disk space at all.
 490          */
 491         if (total_compressed <= blocksize &&
 492            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 493                 goto cleanup_and_bail_uncompressed;
 494
 495         total_compressed = min_t(unsigned long, total_compressed,
 496                         BTRFS_MAX_UNCOMPRESSED);
 497         total_in = 0;
 498         ret = 0;
 499
 500         /*
 501          * we do compression for mount -o compress and when the
 502          * inode has not been flagged as nocompress.  This flag can
 503          * change at any time if we discover bad compression ratios.
 504          */
 505         if (inode_need_compress(inode, start, end)) {
 506                 WARN_ON(pages);
 507                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 508                 if (!pages) {
 509                         /* just bail out to the uncompressed code */
 510                         goto cont;
 511                 }
 512
 513                 if (BTRFS_I(inode)->defrag_compress)
 514                         compress_type = BTRFS_I(inode)->defrag_compress;
 515                 else if (BTRFS_I(inode)->prop_compress)
 516                         compress_type = BTRFS_I(inode)->prop_compress;
 517
 518                 /*
 519                  * we need to call clear_page_dirty_for_io on each
 520                  * page in the range.  Otherwise applications with the file
 521                  * mmap'd can wander in and change the page contents while
 522                  * we are compressing them.
 523                  *
 524                  * If the compression fails for any reason, we set the pages
 525                  * dirty again later on.
 526                  *
 527                  * Note that the remaining part is redirtied, the start pointer
 528                  * has moved, the end is the original one.
 529                  */
 530                 if (!redirty) {
 531                         extent_range_clear_dirty_for_io(inode, start, end);
 532                         redirty = 1;
 533                 }
 534
 535                 /* Compression level is applied here and only here */
 536                 ret = btrfs_compress_pages(
 537                         compress_type | (fs_info->compress_level << 4),
 538                                            inode->i_mapping, start,
 539                                            pages,
 540                                            &nr_pages,
 541                                            &total_in,
 542                                            &total_compressed);
 543
 544                 if (!ret) {
 545                         unsigned long offset = total_compressed &
 546                                 (PAGE_SIZE - 1);
 547                         struct page *page = pages[nr_pages - 1];
 548                         char *kaddr;
 549
 550                         /* zero the tail end of the last page, we might be
 551                          * sending it down to disk
 552                          */
 553                         if (offset) {
 554                                 kaddr = kmap_atomic(page);
 555                                 memset(kaddr + offset, 0,
 556                                        PAGE_SIZE - offset);
 557                                 kunmap_atomic(kaddr);
 558                         }
 559                         will_compress = 1;
 560                 }
 561         }
 562 cont:
 563         if (start == 0) {
 564                 /* lets try to make an inline extent */
 565                 if (ret || total_in < actual_end) {
 566                         /* we didn't compress the entire range, try
 567                          * to make an uncompressed inline extent.
 568                          */
 569                         ret = cow_file_range_inline(inode, start, end, 0,
 570                                                     BTRFS_COMPRESS_NONE, NULL);
 571                 } else {
 572                         /* try making a compressed inline extent */
 573                         ret = cow_file_range_inline(inode, start, end,
 574                                                     total_compressed,
 575                                                     compress_type, pages);
 576                 }
 577                 if (ret <= 0) {
 578                         unsigned long clear_flags = EXTENT_DELALLOC |
 579                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 580                                 EXTENT_DO_ACCOUNTING;
 581                         unsigned long page_error_op;
 582
 583                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 584
 585                         /*
 586                          * inline extent creation worked or returned error,
 587                          * we don't need to create any more async work items.
 588                          * Unlock and free up our temp pages.
 589                          *
 590                          * We use DO_ACCOUNTING here because we need the
 591                          * delalloc_release_metadata to be done _after_ we drop
 592                          * our outstanding extent for clearing delalloc for this
 593                          * range.
 594                          */
 595                         extent_clear_unlock_delalloc(inode, start, end, end,
 596                                                      NULL, clear_flags,
 597                                                      PAGE_UNLOCK |
 598                                                      PAGE_CLEAR_DIRTY |
 599                                                      PAGE_SET_WRITEBACK |
 600                                                      page_error_op |
 601                                                      PAGE_END_WRITEBACK);
 602                         goto free_pages_out;
 603                 }
 604         }
 605
 606         if (will_compress) {
 607                 /*
 608                  * we aren't doing an inline extent round the compressed size
 609                  * up to a block size boundary so the allocator does sane
 610                  * things
 611                  */
 612                 total_compressed = ALIGN(total_compressed, blocksize);
 613
 614                 /*
 615                  * one last check to make sure the compression is really a
 616                  * win, compare the page count read with the blocks on disk,
 617                  * compression must free at least one sector size
 618                  */
 619                 total_in = ALIGN(total_in, PAGE_SIZE);
 620                 if (total_compressed + blocksize <= total_in) {
 621                         *num_added += 1;
 622
 623                         /*
 624                          * The async work queues will take care of doing actual
 625                          * allocation on disk for these compressed pages, and
 626                          * will submit them to the elevator.
 627                          */
 628                         add_async_extent(async_cow, start, total_in,
 629                                         total_compressed, pages, nr_pages,
 630                                         compress_type);
 631
 632                         if (start + total_in < end) {
 633                                 start += total_in;
 634                                 pages = NULL;
 635                                 cond_resched();
 636                                 goto again;
 637                         }
 638                         return;
 639                 }
 640         }
 641         if (pages) {
 642                 /*
 643                  * the compression code ran but failed to make things smaller,
 644                  * free any pages it allocated and our page pointer array
 645                  */
 646                 for (i = 0; i < nr_pages; i++) {
 647                         WARN_ON(pages[i]->mapping);
 648                         put_page(pages[i]);
 649                 }
 650                 kfree(pages);
 651                 pages = NULL;
 652                 total_compressed = 0;
 653                 nr_pages = 0;
 654
 655                 /* flag the file so we don't compress in the future */
 656                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 657                     !(BTRFS_I(inode)->prop_compress)) {
 658                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 659                 }
 660         }
 661 cleanup_and_bail_uncompressed:
 662         /*
 663          * No compression, but we still need to write the pages in the file
 664          * we've been given so far.  redirty the locked page if it corresponds
 665          * to our extent and set things up for the async work queue to run
 666          * cow_file_range to do the normal delalloc dance.
 667          */
 668         if (page_offset(locked_page) >= start &&
 669             page_offset(locked_page) <= end)
 670                 __set_page_dirty_nobuffers(locked_page);
 671                 /* unlocked later on in the async handlers */
 672
 673         if (redirty)
 674                 extent_range_redirty_for_io(inode, start, end);
 675         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 676                          BTRFS_COMPRESS_NONE);
 677         *num_added += 1;
 678
 679         return;
 680
 681 free_pages_out:
 682         for (i = 0; i < nr_pages; i++) {
 683                 WARN_ON(pages[i]->mapping);
 684                 put_page(pages[i]);
 685         }
 686         kfree(pages);
 687 }
 688
 689 static void free_async_extent_pages(struct async_extent *async_extent)
 690 {
 691         int i;
 692
 693         if (!async_extent->pages)
 694                 return;
 695
 696         for (i = 0; i < async_extent->nr_pages; i++) {
 697                 WARN_ON(async_extent->pages[i]->mapping);
 698                 put_page(async_extent->pages[i]);
 699         }
 700         kfree(async_extent->pages);
 701         async_extent->nr_pages = 0;
 702         async_extent->pages = NULL;
 703 }
 704
 705 /*
 706  * phase two of compressed writeback.  This is the ordered portion
 707  * of the code, which only gets called in the order the work was
 708  * queued.  We walk all the async extents created by compress_file_range
 709  * and send them down to the disk.
 710  */
 711 static noinline void submit_compressed_extents(struct inode *inode,
 712                                               struct async_cow *async_cow)
 713 {
 714         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 715         struct async_extent *async_extent;
 716         u64 alloc_hint = 0;
 717         struct btrfs_key ins;
 718         struct extent_map *em;
 719         struct btrfs_root *root = BTRFS_I(inode)->root;
 720         struct extent_io_tree *io_tree;
 721         int ret = 0;
 722
 723 again:
 724         while (!list_empty(&async_cow->extents)) {
 725                 async_extent = list_entry(async_cow->extents.next,
 726                                           struct async_extent, list);
 727                 list_del(&async_extent->list);
 728
 729                 io_tree = &BTRFS_I(inode)->io_tree;
 730
 731 retry:
 732                 /* did the compression code fall back to uncompressed IO? */
 733                 if (!async_extent->pages) {
 734                         int page_started = 0;
 735                         unsigned long nr_written = 0;
 736
 737                         lock_extent(io_tree, async_extent->start,
 738                                          async_extent->start +
 739                                          async_extent->ram_size - 1);
 740
 741                         /* allocate blocks */
 742                         ret = cow_file_range(inode, async_cow->locked_page,
 743                                              async_extent->start,
 744                                              async_extent->start +
 745                                              async_extent->ram_size - 1,
 746                                              async_extent->start +
 747                                              async_extent->ram_size - 1,
 748                                              &page_started, &nr_written, 0,
 749                                              NULL);
 750
 751                         /* JDM XXX */
 752
 753                         /*
 754                          * if page_started, cow_file_range inserted an
 755                          * inline extent and took care of all the unlocking
 756                          * and IO for us.  Otherwise, we need to submit
 757                          * all those pages down to the drive.
 758                          */
 759                         if (!page_started && !ret)
 760                                 extent_write_locked_range(inode,
 761                                                   async_extent->start,
 762                                                   async_extent->start +
 763                                                   async_extent->ram_size - 1,
 764                                                   WB_SYNC_ALL);
 765                         else if (ret)
 766                                 unlock_page(async_cow->locked_page);
 767                         kfree(async_extent);
 768                         cond_resched();
 769                         continue;
 770                 }
 771
 772                 lock_extent(io_tree, async_extent->start,
 773                             async_extent->start + async_extent->ram_size - 1);
 774
 775                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 776                                            async_extent->compressed_size,
 777                                            async_extent->compressed_size,
 778                                            0, alloc_hint, &ins, 1, 1);
 779                 if (ret) {
 780                         free_async_extent_pages(async_extent);
 781
 782                         if (ret == -ENOSPC) {
 783                                 unlock_extent(io_tree, async_extent->start,
 784                                               async_extent->start +
 785                                               async_extent->ram_size - 1);
 786
 787                                 /*
 788                                  * we need to redirty the pages if we decide to
 789                                  * fallback to uncompressed IO, otherwise we
 790                                  * will not submit these pages down to lower
 791                                  * layers.
 792                                  */
 793                                 extent_range_redirty_for_io(inode,
 794                                                 async_extent->start,
 795                                                 async_extent->start +
 796                                                 async_extent->ram_size - 1);
 797
 798                                 goto retry;
 799                         }
 800                         goto out_free;
 801                 }
 802                 /*
 803                  * here we're doing allocation and writeback of the
 804                  * compressed pages
 805                  */
 806                 em = create_io_em(inode, async_extent->start,
 807                                   async_extent->ram_size, /* len */
 808                                   async_extent->start, /* orig_start */
 809                                   ins.objectid, /* block_start */
 810                                   ins.offset, /* block_len */
 811                                   ins.offset, /* orig_block_len */
 812                                   async_extent->ram_size, /* ram_bytes */
 813                                   async_extent->compress_type,
 814                                   BTRFS_ORDERED_COMPRESSED);
 815                 if (IS_ERR(em))
 816                         /* ret value is not necessary due to void function */
 817                         goto out_free_reserve;
 818                 free_extent_map(em);
 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(BTRFS_I(inode),
 829                                                 async_extent->start,
 830                                                 async_extent->start +
 831                                                 async_extent->ram_size - 1, 0);
 832                         goto out_free_reserve;
 833                 }
 834                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 835
 836                 /*
 837                  * clear dirty, set writeback and unlock the pages.
 838                  */
 839                 extent_clear_unlock_delalloc(inode, async_extent->start,
 840                                 async_extent->start +
 841                                 async_extent->ram_size - 1,
 842                                 async_extent->start +
 843                                 async_extent->ram_size - 1,
 844                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 845                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 846                                 PAGE_SET_WRITEBACK);
 847                 if (btrfs_submit_compressed_write(inode,
 848                                     async_extent->start,
 849                                     async_extent->ram_size,
 850                                     ins.objectid,
 851                                     ins.offset, async_extent->pages,
 852                                     async_extent->nr_pages,
 853                                     async_cow->write_flags)) {
 854                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 855                         struct page *p = async_extent->pages[0];
 856                         const u64 start = async_extent->start;
 857                         const u64 end = start + async_extent->ram_size - 1;
 858
 859                         p->mapping = inode->i_mapping;
 860                         tree->ops->writepage_end_io_hook(p, start, end,
 861                                                          NULL, 0);
 862                         p->mapping = NULL;
 863                         extent_clear_unlock_delalloc(inode, start, end, end,
 864                                                      NULL, 0,
 865                                                      PAGE_END_WRITEBACK |
 866                                                      PAGE_SET_ERROR);
 867                         free_async_extent_pages(async_extent);
 868                 }
 869                 alloc_hint = ins.objectid + ins.offset;
 870                 kfree(async_extent);
 871                 cond_resched();
 872         }
 873         return;
 874 out_free_reserve:
 875         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 876         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 877 out_free:
 878         extent_clear_unlock_delalloc(inode, async_extent->start,
 879                                      async_extent->start +
 880                                      async_extent->ram_size - 1,
 881                                      async_extent->start +
 882                                      async_extent->ram_size - 1,
 883                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 884                                      EXTENT_DELALLOC_NEW |
 885                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 886                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 887                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 888                                      PAGE_SET_ERROR);
 889         free_async_extent_pages(async_extent);
 890         kfree(async_extent);
 891         goto again;
 892 }
 893
 894 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 895                                       u64 num_bytes)
 896 {
 897         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 898         struct extent_map *em;
 899         u64 alloc_hint = 0;
 900
 901         read_lock(&em_tree->lock);
 902         em = search_extent_mapping(em_tree, start, num_bytes);
 903         if (em) {
 904                 /*
 905                  * if block start isn't an actual block number then find the
 906                  * first block in this inode and use that as a hint.  If that
 907                  * block is also bogus then just don't worry about it.
 908                  */
 909                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 910                         free_extent_map(em);
 911                         em = search_extent_mapping(em_tree, 0, 0);
 912                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 913                                 alloc_hint = em->block_start;
 914                         if (em)
 915                                 free_extent_map(em);
 916                 } else {
 917                         alloc_hint = em->block_start;
 918                         free_extent_map(em);
 919                 }
 920         }
 921         read_unlock(&em_tree->lock);
 922
 923         return alloc_hint;
 924 }
 925
 926 /*
 927  * when extent_io.c finds a delayed allocation range in the file,
 928  * the call backs end up in this code.  The basic idea is to
 929  * allocate extents on disk for the range, and create ordered data structs
 930  * in ram to track those extents.
 931  *
 932  * locked_page is the page that writepage had locked already.  We use
 933  * it to make sure we don't do extra locks or unlocks.
 934  *
 935  * *page_started is set to one if we unlock locked_page and do everything
 936  * required to start IO on it.  It may be clean and already done with
 937  * IO when we return.
 938  */
 939 static noinline int cow_file_range(struct inode *inode,
 940                                    struct page *locked_page,
 941                                    u64 start, u64 end, u64 delalloc_end,
 942                                    int *page_started, unsigned long *nr_written,
 943                                    int unlock, struct btrfs_dedupe_hash *hash)
 944 {
 945         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 946         struct btrfs_root *root = BTRFS_I(inode)->root;
 947         u64 alloc_hint = 0;
 948         u64 num_bytes;
 949         unsigned long ram_size;
 950         u64 cur_alloc_size = 0;
 951         u64 blocksize = fs_info->sectorsize;
 952         struct btrfs_key ins;
 953         struct extent_map *em;
 954         unsigned clear_bits;
 955         unsigned long page_ops;
 956         bool extent_reserved = false;
 957         int ret = 0;
 958
 959         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 960                 WARN_ON_ONCE(1);
 961                 ret = -EINVAL;
 962                 goto out_unlock;
 963         }
 964
 965         num_bytes = ALIGN(end - start + 1, blocksize);
 966         num_bytes = max(blocksize,  num_bytes);
 967         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 968
 969         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 970
 971         if (start == 0) {
 972                 /* lets try to make an inline extent */
 973                 ret = cow_file_range_inline(inode, start, end, 0,
 974                                             BTRFS_COMPRESS_NONE, NULL);
 975                 if (ret == 0) {
 976                         /*
 977                          * We use DO_ACCOUNTING here because we need the
 978                          * delalloc_release_metadata to be run _after_ we drop
 979                          * our outstanding extent for clearing delalloc for this
 980                          * range.
 981                          */
 982                         extent_clear_unlock_delalloc(inode, start, end,
 983                                      delalloc_end, NULL,
 984                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 985                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 986                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 987                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 988                                      PAGE_END_WRITEBACK);
 989                         *nr_written = *nr_written +
 990                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 991                         *page_started = 1;
 992                         goto out;
 993                 } else if (ret < 0) {
 994                         goto out_unlock;
 995                 }
 996         }
 997
 998         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 999         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1000                         start + num_bytes - 1, 0);
1001
1002         while (num_bytes > 0) {
1003                 cur_alloc_size = num_bytes;
1004                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1005                                            fs_info->sectorsize, 0, alloc_hint,
1006                                            &ins, 1, 1);
1007                 if (ret < 0)
1008                         goto out_unlock;
1009                 cur_alloc_size = ins.offset;
1010                 extent_reserved = true;
1011
1012                 ram_size = ins.offset;
1013                 em = create_io_em(inode, start, ins.offset, /* len */
1014                                   start, /* orig_start */
1015                                   ins.objectid, /* block_start */
1016                                   ins.offset, /* block_len */
1017                                   ins.offset, /* orig_block_len */
1018                                   ram_size, /* ram_bytes */
1019                                   BTRFS_COMPRESS_NONE, /* compress_type */
1020                                   BTRFS_ORDERED_REGULAR /* type */);
1021                 if (IS_ERR(em))
1022                         goto out_reserve;
1023                 free_extent_map(em);
1024
1025                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1026                                                ram_size, cur_alloc_size, 0);
1027                 if (ret)
1028                         goto out_drop_extent_cache;
1029
1030                 if (root->root_key.objectid ==
1031                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1032                         ret = btrfs_reloc_clone_csums(inode, start,
1033                                                       cur_alloc_size);
1034                         /*
1035                          * Only drop cache here, and process as normal.
1036                          *
1037                          * We must not allow extent_clear_unlock_delalloc()
1038                          * at out_unlock label to free meta of this ordered
1039                          * extent, as its meta should be freed by
1040                          * btrfs_finish_ordered_io().
1041                          *
1042                          * So we must continue until @start is increased to
1043                          * skip current ordered extent.
1044                          */
1045                         if (ret)
1046                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1047                                                 start + ram_size - 1, 0);
1048                 }
1049
1050                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1051
1052                 /* we're not doing compressed IO, don't unlock the first
1053                  * page (which the caller expects to stay locked), don't
1054                  * clear any dirty bits and don't set any writeback bits
1055                  *
1056                  * Do set the Private2 bit so we know this page was properly
1057                  * setup for writepage
1058                  */
1059                 page_ops = unlock ? PAGE_UNLOCK : 0;
1060                 page_ops |= PAGE_SET_PRIVATE2;
1061
1062                 extent_clear_unlock_delalloc(inode, start,
1063                                              start + ram_size - 1,
1064                                              delalloc_end, locked_page,
1065                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1066                                              page_ops);
1067                 if (num_bytes < cur_alloc_size)
1068                         num_bytes = 0;
1069                 else
1070                         num_bytes -= cur_alloc_size;
1071                 alloc_hint = ins.objectid + ins.offset;
1072                 start += cur_alloc_size;
1073                 extent_reserved = false;
1074
1075                 /*
1076                  * btrfs_reloc_clone_csums() error, since start is increased
1077                  * extent_clear_unlock_delalloc() at out_unlock label won't
1078                  * free metadata of current ordered extent, we're OK to exit.
1079                  */
1080                 if (ret)
1081                         goto out_unlock;
1082         }
1083 out:
1084         return ret;
1085
1086 out_drop_extent_cache:
1087         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1088 out_reserve:
1089         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1090         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1091 out_unlock:
1092         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1093                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1094         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1095                 PAGE_END_WRITEBACK;
1096         /*
1097          * If we reserved an extent for our delalloc range (or a subrange) and
1098          * failed to create the respective ordered extent, then it means that
1099          * when we reserved the extent we decremented the extent's size from
1100          * the data space_info's bytes_may_use counter and incremented the
1101          * space_info's bytes_reserved counter by the same amount. We must make
1102          * sure extent_clear_unlock_delalloc() does not try to decrement again
1103          * the data space_info's bytes_may_use counter, therefore we do not pass
1104          * it the flag EXTENT_CLEAR_DATA_RESV.
1105          */
1106         if (extent_reserved) {
1107                 extent_clear_unlock_delalloc(inode, start,
1108                                              start + cur_alloc_size,
1109                                              start + cur_alloc_size,
1110                                              locked_page,
1111                                              clear_bits,
1112                                              page_ops);
1113                 start += cur_alloc_size;
1114                 if (start >= end)
1115                         goto out;
1116         }
1117         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1118                                      locked_page,
1119                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1120                                      page_ops);
1121         goto out;
1122 }
1123
1124 /*
1125  * work queue call back to started compression on a file and pages
1126  */
1127 static noinline void async_cow_start(struct btrfs_work *work)
1128 {
1129         struct async_cow *async_cow;
1130         int num_added = 0;
1131         async_cow = container_of(work, struct async_cow, work);
1132
1133         compress_file_range(async_cow->inode, async_cow->locked_page,
1134                             async_cow->start, async_cow->end, async_cow,
1135                             &num_added);
1136         if (num_added == 0) {
1137                 btrfs_add_delayed_iput(async_cow->inode);
1138                 async_cow->inode = NULL;
1139         }
1140 }
1141
1142 /*
1143  * work queue call back to submit previously compressed pages
1144  */
1145 static noinline void async_cow_submit(struct btrfs_work *work)
1146 {
1147         struct btrfs_fs_info *fs_info;
1148         struct async_cow *async_cow;
1149         struct btrfs_root *root;
1150         unsigned long nr_pages;
1151
1152         async_cow = container_of(work, struct async_cow, work);
1153
1154         root = async_cow->root;
1155         fs_info = root->fs_info;
1156         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1157                 PAGE_SHIFT;
1158
1159         /*
1160          * atomic_sub_return implies a barrier for waitqueue_active
1161          */
1162         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1163             5 * SZ_1M &&
1164             waitqueue_active(&fs_info->async_submit_wait))
1165                 wake_up(&fs_info->async_submit_wait);
1166
1167         if (async_cow->inode)
1168                 submit_compressed_extents(async_cow->inode, async_cow);
1169 }
1170
1171 static noinline void async_cow_free(struct btrfs_work *work)
1172 {
1173         struct async_cow *async_cow;
1174         async_cow = container_of(work, struct async_cow, work);
1175         if (async_cow->inode)
1176                 btrfs_add_delayed_iput(async_cow->inode);
1177         kfree(async_cow);
1178 }
1179
1180 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1181                                 u64 start, u64 end, int *page_started,
1182                                 unsigned long *nr_written,
1183                                 unsigned int write_flags)
1184 {
1185         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1186         struct async_cow *async_cow;
1187         struct btrfs_root *root = BTRFS_I(inode)->root;
1188         unsigned long nr_pages;
1189         u64 cur_end;
1190
1191         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1192                          1, 0, NULL);
1193         while (start < end) {
1194                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1195                 BUG_ON(!async_cow); /* -ENOMEM */
1196                 async_cow->inode = igrab(inode);
1197                 async_cow->root = root;
1198                 async_cow->locked_page = locked_page;
1199                 async_cow->start = start;
1200                 async_cow->write_flags = write_flags;
1201
1202                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1203                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1204                         cur_end = end;
1205                 else
1206                         cur_end = min(end, start + SZ_512K - 1);
1207
1208                 async_cow->end = cur_end;
1209                 INIT_LIST_HEAD(&async_cow->extents);
1210
1211                 btrfs_init_work(&async_cow->work,
1212                                 btrfs_delalloc_helper,
1213                                 async_cow_start, async_cow_submit,
1214                                 async_cow_free);
1215
1216                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1217                         PAGE_SHIFT;
1218                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1219
1220                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1221
1222                 *nr_written += nr_pages;
1223                 start = cur_end + 1;
1224         }
1225         *page_started = 1;
1226         return 0;
1227 }
1228
1229 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1230                                         u64 bytenr, u64 num_bytes)
1231 {
1232         int ret;
1233         struct btrfs_ordered_sum *sums;
1234         LIST_HEAD(list);
1235
1236         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1237                                        bytenr + num_bytes - 1, &list, 0);
1238         if (ret == 0 && list_empty(&list))
1239                 return 0;
1240
1241         while (!list_empty(&list)) {
1242                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1243                 list_del(&sums->list);
1244                 kfree(sums);
1245         }
1246         if (ret < 0)
1247                 return ret;
1248         return 1;
1249 }
1250
1251 /*
1252  * when nowcow writeback call back.  This checks for snapshots or COW copies
1253  * of the extents that exist in the file, and COWs the file as required.
1254  *
1255  * If no cow copies or snapshots exist, we write directly to the existing
1256  * blocks on disk
1257  */
1258 static noinline int run_delalloc_nocow(struct inode *inode,
1259                                        struct page *locked_page,
1260                               u64 start, u64 end, int *page_started, int force,
1261                               unsigned long *nr_written)
1262 {
1263         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1264         struct btrfs_root *root = BTRFS_I(inode)->root;
1265         struct extent_buffer *leaf;
1266         struct btrfs_path *path;
1267         struct btrfs_file_extent_item *fi;
1268         struct btrfs_key found_key;
1269         struct extent_map *em;
1270         u64 cow_start;
1271         u64 cur_offset;
1272         u64 extent_end;
1273         u64 extent_offset;
1274         u64 disk_bytenr;
1275         u64 num_bytes;
1276         u64 disk_num_bytes;
1277         u64 ram_bytes;
1278         int extent_type;
1279         int ret, err;
1280         int type;
1281         int nocow;
1282         int check_prev = 1;
1283         bool nolock;
1284         u64 ino = btrfs_ino(BTRFS_I(inode));
1285
1286         path = btrfs_alloc_path();
1287         if (!path) {
1288                 extent_clear_unlock_delalloc(inode, start, end, end,
1289                                              locked_page,
1290                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1291                                              EXTENT_DO_ACCOUNTING |
1292                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1293                                              PAGE_CLEAR_DIRTY |
1294                                              PAGE_SET_WRITEBACK |
1295                                              PAGE_END_WRITEBACK);
1296                 return -ENOMEM;
1297         }
1298
1299         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1300
1301         cow_start = (u64)-1;
1302         cur_offset = start;
1303         while (1) {
1304                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1305                                                cur_offset, 0);
1306                 if (ret < 0)
1307                         goto error;
1308                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1309                         leaf = path->nodes[0];
1310                         btrfs_item_key_to_cpu(leaf, &found_key,
1311                                               path->slots[0] - 1);
1312                         if (found_key.objectid == ino &&
1313                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1314                                 path->slots[0]--;
1315                 }
1316                 check_prev = 0;
1317 next_slot:
1318                 leaf = path->nodes[0];
1319                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1320                         ret = btrfs_next_leaf(root, path);
1321                         if (ret < 0) {
1322                                 if (cow_start != (u64)-1)
1323                                         cur_offset = cow_start;
1324                                 goto error;
1325                         }
1326                         if (ret > 0)
1327                                 break;
1328                         leaf = path->nodes[0];
1329                 }
1330
1331                 nocow = 0;
1332                 disk_bytenr = 0;
1333                 num_bytes = 0;
1334                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1335
1336                 if (found_key.objectid > ino)
1337                         break;
1338                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1339                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1340                         path->slots[0]++;
1341                         goto next_slot;
1342                 }
1343                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1344                     found_key.offset > end)
1345                         break;
1346
1347                 if (found_key.offset > cur_offset) {
1348                         extent_end = found_key.offset;
1349                         extent_type = 0;
1350                         goto out_check;
1351                 }
1352
1353                 fi = btrfs_item_ptr(leaf, path->slots[0],
1354                                     struct btrfs_file_extent_item);
1355                 extent_type = btrfs_file_extent_type(leaf, fi);
1356
1357                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1358                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1359                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1360                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1361                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1362                         extent_end = found_key.offset +
1363                                 btrfs_file_extent_num_bytes(leaf, fi);
1364                         disk_num_bytes =
1365                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1366                         if (extent_end <= start) {
1367                                 path->slots[0]++;
1368                                 goto next_slot;
1369                         }
1370                         if (disk_bytenr == 0)
1371                                 goto out_check;
1372                         if (btrfs_file_extent_compression(leaf, fi) ||
1373                             btrfs_file_extent_encryption(leaf, fi) ||
1374                             btrfs_file_extent_other_encoding(leaf, fi))
1375                                 goto out_check;
1376                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1377                                 goto out_check;
1378                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1379                                 goto out_check;
1380                         ret = btrfs_cross_ref_exist(root, ino,
1381                                                     found_key.offset -
1382                                                     extent_offset, disk_bytenr);
1383                         if (ret) {
1384                                 /*
1385                                  * ret could be -EIO if the above fails to read
1386                                  * metadata.
1387                                  */
1388                                 if (ret < 0) {
1389                                         if (cow_start != (u64)-1)
1390                                                 cur_offset = cow_start;
1391                                         goto error;
1392                                 }
1393
1394                                 WARN_ON_ONCE(nolock);
1395                                 goto out_check;
1396                         }
1397                         disk_bytenr += extent_offset;
1398                         disk_bytenr += cur_offset - found_key.offset;
1399                         num_bytes = min(end + 1, extent_end) - cur_offset;
1400                         /*
1401                          * if there are pending snapshots for this root,
1402                          * we fall into common COW way.
1403                          */
1404                         if (!nolock) {
1405                                 err = btrfs_start_write_no_snapshotting(root);
1406                                 if (!err)
1407                                         goto out_check;
1408                         }
1409                         /*
1410                          * force cow if csum exists in the range.
1411                          * this ensure that csum for a given extent are
1412                          * either valid or do not exist.
1413                          */
1414                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1415                                                   num_bytes);
1416                         if (ret) {
1417                                 if (!nolock)
1418                                         btrfs_end_write_no_snapshotting(root);
1419
1420                                 /*
1421                                  * ret could be -EIO if the above fails to read
1422                                  * metadata.
1423                                  */
1424                                 if (ret < 0) {
1425                                         if (cow_start != (u64)-1)
1426                                                 cur_offset = cow_start;
1427                                         goto error;
1428                                 }
1429                                 WARN_ON_ONCE(nolock);
1430                                 goto out_check;
1431                         }
1432                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1433                                 if (!nolock)
1434                                         btrfs_end_write_no_snapshotting(root);
1435                                 goto out_check;
1436                         }
1437                         nocow = 1;
1438                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1439                         extent_end = found_key.offset +
1440                                 btrfs_file_extent_inline_len(leaf,
1441                                                      path->slots[0], fi);
1442                         extent_end = ALIGN(extent_end,
1443                                            fs_info->sectorsize);
1444                 } else {
1445                         BUG_ON(1);
1446                 }
1447 out_check:
1448                 if (extent_end <= start) {
1449                         path->slots[0]++;
1450                         if (!nolock && nocow)
1451                                 btrfs_end_write_no_snapshotting(root);
1452                         if (nocow)
1453                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1454                         goto next_slot;
1455                 }
1456                 if (!nocow) {
1457                         if (cow_start == (u64)-1)
1458                                 cow_start = cur_offset;
1459                         cur_offset = extent_end;
1460                         if (cur_offset > end)
1461                                 break;
1462                         path->slots[0]++;
1463                         goto next_slot;
1464                 }
1465
1466                 btrfs_release_path(path);
1467                 if (cow_start != (u64)-1) {
1468                         ret = cow_file_range(inode, locked_page,
1469                                              cow_start, found_key.offset - 1,
1470                                              end, page_started, nr_written, 1,
1471                                              NULL);
1472                         if (ret) {
1473                                 if (!nolock && nocow)
1474                                         btrfs_end_write_no_snapshotting(root);
1475                                 if (nocow)
1476                                         btrfs_dec_nocow_writers(fs_info,
1477                                                                 disk_bytenr);
1478                                 goto error;
1479                         }
1480                         cow_start = (u64)-1;
1481                 }
1482
1483                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1484                         u64 orig_start = found_key.offset - extent_offset;
1485
1486                         em = create_io_em(inode, cur_offset, num_bytes,
1487                                           orig_start,
1488                                           disk_bytenr, /* block_start */
1489                                           num_bytes, /* block_len */
1490                                           disk_num_bytes, /* orig_block_len */
1491                                           ram_bytes, BTRFS_COMPRESS_NONE,
1492                                           BTRFS_ORDERED_PREALLOC);
1493                         if (IS_ERR(em)) {
1494                                 if (!nolock && nocow)
1495                                         btrfs_end_write_no_snapshotting(root);
1496                                 if (nocow)
1497                                         btrfs_dec_nocow_writers(fs_info,
1498                                                                 disk_bytenr);
1499                                 ret = PTR_ERR(em);
1500                                 goto error;
1501                         }
1502                         free_extent_map(em);
1503                 }
1504
1505                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1506                         type = BTRFS_ORDERED_PREALLOC;
1507                 } else {
1508                         type = BTRFS_ORDERED_NOCOW;
1509                 }
1510
1511                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1512                                                num_bytes, num_bytes, type);
1513                 if (nocow)
1514                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1515                 BUG_ON(ret); /* -ENOMEM */
1516
1517                 if (root->root_key.objectid ==
1518                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1519                         /*
1520                          * Error handled later, as we must prevent
1521                          * extent_clear_unlock_delalloc() in error handler
1522                          * from freeing metadata of created ordered extent.
1523                          */
1524                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1525                                                       num_bytes);
1526
1527                 extent_clear_unlock_delalloc(inode, cur_offset,
1528                                              cur_offset + num_bytes - 1, end,
1529                                              locked_page, EXTENT_LOCKED |
1530                                              EXTENT_DELALLOC |
1531                                              EXTENT_CLEAR_DATA_RESV,
1532                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1533
1534                 if (!nolock && nocow)
1535                         btrfs_end_write_no_snapshotting(root);
1536                 cur_offset = extent_end;
1537
1538                 /*
1539                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1540                  * handler, as metadata for created ordered extent will only
1541                  * be freed by btrfs_finish_ordered_io().
1542                  */
1543                 if (ret)
1544                         goto error;
1545                 if (cur_offset > end)
1546                         break;
1547         }
1548         btrfs_release_path(path);
1549
1550         if (cur_offset <= end && cow_start == (u64)-1) {
1551                 cow_start = cur_offset;
1552                 cur_offset = end;
1553         }
1554
1555         if (cow_start != (u64)-1) {
1556                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1557                                      page_started, nr_written, 1, NULL);
1558                 if (ret)
1559                         goto error;
1560         }
1561
1562 error:
1563         if (ret && cur_offset < end)
1564                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1565                                              locked_page, EXTENT_LOCKED |
1566                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1567                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1568                                              PAGE_CLEAR_DIRTY |
1569                                              PAGE_SET_WRITEBACK |
1570                                              PAGE_END_WRITEBACK);
1571         btrfs_free_path(path);
1572         return ret;
1573 }
1574
1575 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1576 {
1577
1578         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1579             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1580                 return 0;
1581
1582         /*
1583          * @defrag_bytes is a hint value, no spinlock held here,
1584          * if is not zero, it means the file is defragging.
1585          * Force cow if given extent needs to be defragged.
1586          */
1587         if (BTRFS_I(inode)->defrag_bytes &&
1588             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1589                            EXTENT_DEFRAG, 0, NULL))
1590                 return 1;
1591
1592         return 0;
1593 }
1594
1595 /*
1596  * extent_io.c call back to do delayed allocation processing
1597  */
1598 static int run_delalloc_range(void *private_data, struct page *locked_page,
1599                               u64 start, u64 end, int *page_started,
1600                               unsigned long *nr_written,
1601                               struct writeback_control *wbc)
1602 {
1603         struct inode *inode = private_data;
1604         int ret;
1605         int force_cow = need_force_cow(inode, start, end);
1606         unsigned int write_flags = wbc_to_write_flags(wbc);
1607
1608         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1609                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1610                                          page_started, 1, nr_written);
1611         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1612                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1613                                          page_started, 0, nr_written);
1614         } else if (!inode_need_compress(inode, start, end)) {
1615                 ret = cow_file_range(inode, locked_page, start, end, end,
1616                                       page_started, nr_written, 1, NULL);
1617         } else {
1618                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1619                         &BTRFS_I(inode)->runtime_flags);
1620                 ret = cow_file_range_async(inode, locked_page, start, end,
1621                                            page_started, nr_written,
1622                                            write_flags);
1623         }
1624         if (ret)
1625                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1626         return ret;
1627 }
1628
1629 static void btrfs_split_extent_hook(void *private_data,
1630                                     struct extent_state *orig, u64 split)
1631 {
1632         struct inode *inode = private_data;
1633         u64 size;
1634
1635         /* not delalloc, ignore it */
1636         if (!(orig->state & EXTENT_DELALLOC))
1637                 return;
1638
1639         size = orig->end - orig->start + 1;
1640         if (size > BTRFS_MAX_EXTENT_SIZE) {
1641                 u32 num_extents;
1642                 u64 new_size;
1643
1644                 /*
1645                  * See the explanation in btrfs_merge_extent_hook, the same
1646                  * applies here, just in reverse.
1647                  */
1648                 new_size = orig->end - split + 1;
1649                 num_extents = count_max_extents(new_size);
1650                 new_size = split - orig->start;
1651                 num_extents += count_max_extents(new_size);
1652                 if (count_max_extents(size) >= num_extents)
1653                         return;
1654         }
1655
1656         spin_lock(&BTRFS_I(inode)->lock);
1657         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1658         spin_unlock(&BTRFS_I(inode)->lock);
1659 }
1660
1661 /*
1662  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1663  * extents so we can keep track of new extents that are just merged onto old
1664  * extents, such as when we are doing sequential writes, so we can properly
1665  * account for the metadata space we'll need.
1666  */
1667 static void btrfs_merge_extent_hook(void *private_data,
1668                                     struct extent_state *new,
1669                                     struct extent_state *other)
1670 {
1671         struct inode *inode = private_data;
1672         u64 new_size, old_size;
1673         u32 num_extents;
1674
1675         /* not delalloc, ignore it */
1676         if (!(other->state & EXTENT_DELALLOC))
1677                 return;
1678
1679         if (new->start > other->start)
1680                 new_size = new->end - other->start + 1;
1681         else
1682                 new_size = other->end - new->start + 1;
1683
1684         /* we're not bigger than the max, unreserve the space and go */
1685         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1686                 spin_lock(&BTRFS_I(inode)->lock);
1687                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1688                 spin_unlock(&BTRFS_I(inode)->lock);
1689                 return;
1690         }
1691
1692         /*
1693          * We have to add up either side to figure out how many extents were
1694          * accounted for before we merged into one big extent.  If the number of
1695          * extents we accounted for is <= the amount we need for the new range
1696          * then we can return, otherwise drop.  Think of it like this
1697          *
1698          * [ 4k][MAX_SIZE]
1699          *
1700          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1701          * need 2 outstanding extents, on one side we have 1 and the other side
1702          * we have 1 so they are == and we can return.  But in this case
1703          *
1704          * [MAX_SIZE+4k][MAX_SIZE+4k]
1705          *
1706          * Each range on their own accounts for 2 extents, but merged together
1707          * they are only 3 extents worth of accounting, so we need to drop in
1708          * this case.
1709          */
1710         old_size = other->end - other->start + 1;
1711         num_extents = count_max_extents(old_size);
1712         old_size = new->end - new->start + 1;
1713         num_extents += count_max_extents(old_size);
1714         if (count_max_extents(new_size) >= num_extents)
1715                 return;
1716
1717         spin_lock(&BTRFS_I(inode)->lock);
1718         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1719         spin_unlock(&BTRFS_I(inode)->lock);
1720 }
1721
1722 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1723                                       struct inode *inode)
1724 {
1725         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1726
1727         spin_lock(&root->delalloc_lock);
1728         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1729                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1730                               &root->delalloc_inodes);
1731                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1732                         &BTRFS_I(inode)->runtime_flags);
1733                 root->nr_delalloc_inodes++;
1734                 if (root->nr_delalloc_inodes == 1) {
1735                         spin_lock(&fs_info->delalloc_root_lock);
1736                         BUG_ON(!list_empty(&root->delalloc_root));
1737                         list_add_tail(&root->delalloc_root,
1738                                       &fs_info->delalloc_roots);
1739                         spin_unlock(&fs_info->delalloc_root_lock);
1740                 }
1741         }
1742         spin_unlock(&root->delalloc_lock);
1743 }
1744
1745 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1746                                      struct btrfs_inode *inode)
1747 {
1748         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1749
1750         spin_lock(&root->delalloc_lock);
1751         if (!list_empty(&inode->delalloc_inodes)) {
1752                 list_del_init(&inode->delalloc_inodes);
1753                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1754                           &inode->runtime_flags);
1755                 root->nr_delalloc_inodes--;
1756                 if (!root->nr_delalloc_inodes) {
1757                         spin_lock(&fs_info->delalloc_root_lock);
1758                         BUG_ON(list_empty(&root->delalloc_root));
1759                         list_del_init(&root->delalloc_root);
1760                         spin_unlock(&fs_info->delalloc_root_lock);
1761                 }
1762         }
1763         spin_unlock(&root->delalloc_lock);
1764 }
1765
1766 /*
1767  * extent_io.c set_bit_hook, used to track delayed allocation
1768  * bytes in this file, and to maintain the list of inodes that
1769  * have pending delalloc work to be done.
1770  */
1771 static void btrfs_set_bit_hook(void *private_data,
1772                                struct extent_state *state, unsigned *bits)
1773 {
1774         struct inode *inode = private_data;
1775
1776         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1777
1778         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1779                 WARN_ON(1);
1780         /*
1781          * set_bit and clear bit hooks normally require _irqsave/restore
1782          * but in this case, we are only testing for the DELALLOC
1783          * bit, which is only set or cleared with irqs on
1784          */
1785         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1786                 struct btrfs_root *root = BTRFS_I(inode)->root;
1787                 u64 len = state->end + 1 - state->start;
1788                 u32 num_extents = count_max_extents(len);
1789                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1790
1791                 spin_lock(&BTRFS_I(inode)->lock);
1792                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1793                 spin_unlock(&BTRFS_I(inode)->lock);
1794
1795                 /* For sanity tests */
1796                 if (btrfs_is_testing(fs_info))
1797                         return;
1798
1799                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1800                                          fs_info->delalloc_batch);
1801                 spin_lock(&BTRFS_I(inode)->lock);
1802                 BTRFS_I(inode)->delalloc_bytes += len;
1803                 if (*bits & EXTENT_DEFRAG)
1804                         BTRFS_I(inode)->defrag_bytes += len;
1805                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1806                                          &BTRFS_I(inode)->runtime_flags))
1807                         btrfs_add_delalloc_inodes(root, inode);
1808                 spin_unlock(&BTRFS_I(inode)->lock);
1809         }
1810
1811         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1812             (*bits & EXTENT_DELALLOC_NEW)) {
1813                 spin_lock(&BTRFS_I(inode)->lock);
1814                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1815                         state->start;
1816                 spin_unlock(&BTRFS_I(inode)->lock);
1817         }
1818 }
1819
1820 /*
1821  * extent_io.c clear_bit_hook, see set_bit_hook for why
1822  */
1823 static void btrfs_clear_bit_hook(void *private_data,
1824                                  struct extent_state *state,
1825                                  unsigned *bits)
1826 {
1827         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1828         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1829         u64 len = state->end + 1 - state->start;
1830         u32 num_extents = count_max_extents(len);
1831
1832         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1833                 spin_lock(&inode->lock);
1834                 inode->defrag_bytes -= len;
1835                 spin_unlock(&inode->lock);
1836         }
1837
1838         /*
1839          * set_bit and clear bit hooks normally require _irqsave/restore
1840          * but in this case, we are only testing for the DELALLOC
1841          * bit, which is only set or cleared with irqs on
1842          */
1843         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1844                 struct btrfs_root *root = inode->root;
1845                 bool do_list = !btrfs_is_free_space_inode(inode);
1846
1847                 spin_lock(&inode->lock);
1848                 btrfs_mod_outstanding_extents(inode, -num_extents);
1849                 spin_unlock(&inode->lock);
1850
1851                 /*
1852                  * We don't reserve metadata space for space cache inodes so we
1853                  * don't need to call dellalloc_release_metadata if there is an
1854                  * error.
1855                  */
1856                 if (*bits & EXTENT_CLEAR_META_RESV &&
1857                     root != fs_info->tree_root)
1858                         btrfs_delalloc_release_metadata(inode, len, false);
1859
1860                 /* For sanity tests. */
1861                 if (btrfs_is_testing(fs_info))
1862                         return;
1863
1864                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1865                     do_list && !(state->state & EXTENT_NORESERVE) &&
1866                     (*bits & EXTENT_CLEAR_DATA_RESV))
1867                         btrfs_free_reserved_data_space_noquota(
1868                                         &inode->vfs_inode,
1869                                         state->start, len);
1870
1871                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1872                                          fs_info->delalloc_batch);
1873                 spin_lock(&inode->lock);
1874                 inode->delalloc_bytes -= len;
1875                 if (do_list && inode->delalloc_bytes == 0 &&
1876                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1877                                         &inode->runtime_flags))
1878                         btrfs_del_delalloc_inode(root, inode);
1879                 spin_unlock(&inode->lock);
1880         }
1881
1882         if ((state->state & EXTENT_DELALLOC_NEW) &&
1883             (*bits & EXTENT_DELALLOC_NEW)) {
1884                 spin_lock(&inode->lock);
1885                 ASSERT(inode->new_delalloc_bytes >= len);
1886                 inode->new_delalloc_bytes -= len;
1887                 spin_unlock(&inode->lock);
1888         }
1889 }
1890
1891 /*
1892  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1893  * we don't create bios that span stripes or chunks
1894  *
1895  * return 1 if page cannot be merged to bio
1896  * return 0 if page can be merged to bio
1897  * return error otherwise
1898  */
1899 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1900                          size_t size, struct bio *bio,
1901                          unsigned long bio_flags)
1902 {
1903         struct inode *inode = page->mapping->host;
1904         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1905         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1906         u64 length = 0;
1907         u64 map_length;
1908         int ret;
1909
1910         if (bio_flags & EXTENT_BIO_COMPRESSED)
1911                 return 0;
1912
1913         length = bio->bi_iter.bi_size;
1914         map_length = length;
1915         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1916                               NULL, 0);
1917         if (ret < 0)
1918                 return ret;
1919         if (map_length < length + size)
1920                 return 1;
1921         return 0;
1922 }
1923
1924 /*
1925  * in order to insert checksums into the metadata in large chunks,
1926  * we wait until bio submission time.   All the pages in the bio are
1927  * checksummed and sums are attached onto the ordered extent record.
1928  *
1929  * At IO completion time the cums attached on the ordered extent record
1930  * are inserted into the btree
1931  */
1932 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1933                                     u64 bio_offset)
1934 {
1935         struct inode *inode = private_data;
1936         blk_status_t ret = 0;
1937
1938         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1939         BUG_ON(ret); /* -ENOMEM */
1940         return 0;
1941 }
1942
1943 /*
1944  * in order to insert checksums into the metadata in large chunks,
1945  * we wait until bio submission time.   All the pages in the bio are
1946  * checksummed and sums are attached onto the ordered extent record.
1947  *
1948  * At IO completion time the cums attached on the ordered extent record
1949  * are inserted into the btree
1950  */
1951 static blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1952                           int mirror_num)
1953 {
1954         struct inode *inode = private_data;
1955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1956         blk_status_t ret;
1957
1958         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1959         if (ret) {
1960                 bio->bi_status = ret;
1961                 bio_endio(bio);
1962         }
1963         return ret;
1964 }
1965
1966 /*
1967  * extent_io.c submission hook. This does the right thing for csum calculation
1968  * on write, or reading the csums from the tree before a read.
1969  *
1970  * Rules about async/sync submit,
1971  * a) read:                             sync submit
1972  *
1973  * b) write without checksum:           sync submit
1974  *
1975  * c) write with checksum:
1976  *    c-1) if bio is issued by fsync:   sync submit
1977  *         (sync_writers != 0)
1978  *
1979  *    c-2) if root is reloc root:       sync submit
1980  *         (only in case of buffered IO)
1981  *
1982  *    c-3) otherwise:                   async submit
1983  */
1984 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1985                                  int mirror_num, unsigned long bio_flags,
1986                                  u64 bio_offset)
1987 {
1988         struct inode *inode = private_data;
1989         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1990         struct btrfs_root *root = BTRFS_I(inode)->root;
1991         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1992         blk_status_t ret = 0;
1993         int skip_sum;
1994         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1995
1996         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1997
1998         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1999                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2000
2001         if (bio_op(bio) != REQ_OP_WRITE) {
2002                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2003                 if (ret)
2004                         goto out;
2005
2006                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2007                         ret = btrfs_submit_compressed_read(inode, bio,
2008                                                            mirror_num,
2009                                                            bio_flags);
2010                         goto out;
2011                 } else if (!skip_sum) {
2012                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2013                         if (ret)
2014                                 goto out;
2015                 }
2016                 goto mapit;
2017         } else if (async && !skip_sum) {
2018                 /* csum items have already been cloned */
2019                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2020                         goto mapit;
2021                 /* we're doing a write, do the async checksumming */
2022                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2023                                           bio_offset, inode,
2024                                           btrfs_submit_bio_start,
2025                                           btrfs_submit_bio_done);
2026                 goto out;
2027         } else if (!skip_sum) {
2028                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2029                 if (ret)
2030                         goto out;
2031         }
2032
2033 mapit:
2034         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2035
2036 out:
2037         if (ret) {
2038                 bio->bi_status = ret;
2039                 bio_endio(bio);
2040         }
2041         return ret;
2042 }
2043
2044 /*
2045  * given a list of ordered sums record them in the inode.  This happens
2046  * at IO completion time based on sums calculated at bio submission time.
2047  */
2048 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2049                              struct inode *inode, struct list_head *list)
2050 {
2051         struct btrfs_ordered_sum *sum;
2052         int ret;
2053
2054         list_for_each_entry(sum, list, list) {
2055                 trans->adding_csums = true;
2056                 ret = btrfs_csum_file_blocks(trans,
2057                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2058                 trans->adding_csums = false;
2059                 if (ret)
2060                         return ret;
2061         }
2062         return 0;
2063 }
2064
2065 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2066                               unsigned int extra_bits,
2067                               struct extent_state **cached_state, int dedupe)
2068 {
2069         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2070         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2071                                    extra_bits, cached_state);
2072 }
2073
2074 /* see btrfs_writepage_start_hook for details on why this is required */
2075 struct btrfs_writepage_fixup {
2076         struct page *page;
2077         struct btrfs_work work;
2078 };
2079
2080 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2081 {
2082         struct btrfs_writepage_fixup *fixup;
2083         struct btrfs_ordered_extent *ordered;
2084         struct extent_state *cached_state = NULL;
2085         struct extent_changeset *data_reserved = NULL;
2086         struct page *page;
2087         struct inode *inode;
2088         u64 page_start;
2089         u64 page_end;
2090         int ret;
2091
2092         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2093         page = fixup->page;
2094 again:
2095         lock_page(page);
2096         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2097                 ClearPageChecked(page);
2098                 goto out_page;
2099         }
2100
2101         inode = page->mapping->host;
2102         page_start = page_offset(page);
2103         page_end = page_offset(page) + PAGE_SIZE - 1;
2104
2105         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2106                          &cached_state);
2107
2108         /* already ordered? We're done */
2109         if (PagePrivate2(page))
2110                 goto out;
2111
2112         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2113                                         PAGE_SIZE);
2114         if (ordered) {
2115                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2116                                      page_end, &cached_state);
2117                 unlock_page(page);
2118                 btrfs_start_ordered_extent(inode, ordered, 1);
2119                 btrfs_put_ordered_extent(ordered);
2120                 goto again;
2121         }
2122
2123         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2124                                            PAGE_SIZE);
2125         if (ret) {
2126                 mapping_set_error(page->mapping, ret);
2127                 end_extent_writepage(page, ret, page_start, page_end);
2128                 ClearPageChecked(page);
2129                 goto out;
2130          }
2131
2132         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2133                                         &cached_state, 0);
2134         if (ret) {
2135                 mapping_set_error(page->mapping, ret);
2136                 end_extent_writepage(page, ret, page_start, page_end);
2137                 ClearPageChecked(page);
2138                 goto out;
2139         }
2140
2141         ClearPageChecked(page);
2142         set_page_dirty(page);
2143         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2144 out:
2145         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2146                              &cached_state);
2147 out_page:
2148         unlock_page(page);
2149         put_page(page);
2150         kfree(fixup);
2151         extent_changeset_free(data_reserved);
2152 }
2153
2154 /*
2155  * There are a few paths in the higher layers of the kernel that directly
2156  * set the page dirty bit without asking the filesystem if it is a
2157  * good idea.  This causes problems because we want to make sure COW
2158  * properly happens and the data=ordered rules are followed.
2159  *
2160  * In our case any range that doesn't have the ORDERED bit set
2161  * hasn't been properly setup for IO.  We kick off an async process
2162  * to fix it up.  The async helper will wait for ordered extents, set
2163  * the delalloc bit and make it safe to write the page.
2164  */
2165 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2166 {
2167         struct inode *inode = page->mapping->host;
2168         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2169         struct btrfs_writepage_fixup *fixup;
2170
2171         /* this page is properly in the ordered list */
2172         if (TestClearPagePrivate2(page))
2173                 return 0;
2174
2175         if (PageChecked(page))
2176                 return -EAGAIN;
2177
2178         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2179         if (!fixup)
2180                 return -EAGAIN;
2181
2182         SetPageChecked(page);
2183         get_page(page);
2184         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2185                         btrfs_writepage_fixup_worker, NULL, NULL);
2186         fixup->page = page;
2187         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2188         return -EBUSY;
2189 }
2190
2191 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2192                                        struct inode *inode, u64 file_pos,
2193                                        u64 disk_bytenr, u64 disk_num_bytes,
2194                                        u64 num_bytes, u64 ram_bytes,
2195                                        u8 compression, u8 encryption,
2196                                        u16 other_encoding, int extent_type)
2197 {
2198         struct btrfs_root *root = BTRFS_I(inode)->root;
2199         struct btrfs_file_extent_item *fi;
2200         struct btrfs_path *path;
2201         struct extent_buffer *leaf;
2202         struct btrfs_key ins;
2203         u64 qg_released;
2204         int extent_inserted = 0;
2205         int ret;
2206
2207         path = btrfs_alloc_path();
2208         if (!path)
2209                 return -ENOMEM;
2210
2211         /*
2212          * we may be replacing one extent in the tree with another.
2213          * The new extent is pinned in the extent map, and we don't want
2214          * to drop it from the cache until it is completely in the btree.
2215          *
2216          * So, tell btrfs_drop_extents to leave this extent in the cache.
2217          * the caller is expected to unpin it and allow it to be merged
2218          * with the others.
2219          */
2220         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2221                                    file_pos + num_bytes, NULL, 0,
2222                                    1, sizeof(*fi), &extent_inserted);
2223         if (ret)
2224                 goto out;
2225
2226         if (!extent_inserted) {
2227                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2228                 ins.offset = file_pos;
2229                 ins.type = BTRFS_EXTENT_DATA_KEY;
2230
2231                 path->leave_spinning = 1;
2232                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2233                                               sizeof(*fi));
2234                 if (ret)
2235                         goto out;
2236         }
2237         leaf = path->nodes[0];
2238         fi = btrfs_item_ptr(leaf, path->slots[0],
2239                             struct btrfs_file_extent_item);
2240         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2241         btrfs_set_file_extent_type(leaf, fi, extent_type);
2242         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2243         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2244         btrfs_set_file_extent_offset(leaf, fi, 0);
2245         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2246         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2247         btrfs_set_file_extent_compression(leaf, fi, compression);
2248         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2249         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2250
2251         btrfs_mark_buffer_dirty(leaf);
2252         btrfs_release_path(path);
2253
2254         inode_add_bytes(inode, num_bytes);
2255
2256         ins.objectid = disk_bytenr;
2257         ins.offset = disk_num_bytes;
2258         ins.type = BTRFS_EXTENT_ITEM_KEY;
2259
2260         /*
2261          * Release the reserved range from inode dirty range map, as it is
2262          * already moved into delayed_ref_head
2263          */
2264         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2265         if (ret < 0)
2266                 goto out;
2267         qg_released = ret;
2268         ret = btrfs_alloc_reserved_file_extent(trans, root,
2269                                                btrfs_ino(BTRFS_I(inode)),
2270                                                file_pos, qg_released, &ins);
2271 out:
2272         btrfs_free_path(path);
2273
2274         return ret;
2275 }
2276
2277 /* snapshot-aware defrag */
2278 struct sa_defrag_extent_backref {
2279         struct rb_node node;
2280         struct old_sa_defrag_extent *old;
2281         u64 root_id;
2282         u64 inum;
2283         u64 file_pos;
2284         u64 extent_offset;
2285         u64 num_bytes;
2286         u64 generation;
2287 };
2288
2289 struct old_sa_defrag_extent {
2290         struct list_head list;
2291         struct new_sa_defrag_extent *new;
2292
2293         u64 extent_offset;
2294         u64 bytenr;
2295         u64 offset;
2296         u64 len;
2297         int count;
2298 };
2299
2300 struct new_sa_defrag_extent {
2301         struct rb_root root;
2302         struct list_head head;
2303         struct btrfs_path *path;
2304         struct inode *inode;
2305         u64 file_pos;
2306         u64 len;
2307         u64 bytenr;
2308         u64 disk_len;
2309         u8 compress_type;
2310 };
2311
2312 static int backref_comp(struct sa_defrag_extent_backref *b1,
2313                         struct sa_defrag_extent_backref *b2)
2314 {
2315         if (b1->root_id < b2->root_id)
2316                 return -1;
2317         else if (b1->root_id > b2->root_id)
2318                 return 1;
2319
2320         if (b1->inum < b2->inum)
2321                 return -1;
2322         else if (b1->inum > b2->inum)
2323                 return 1;
2324
2325         if (b1->file_pos < b2->file_pos)
2326                 return -1;
2327         else if (b1->file_pos > b2->file_pos)
2328                 return 1;
2329
2330         /*
2331          * [------------------------------] ===> (a range of space)
2332          *     |<--->|   |<---->| =============> (fs/file tree A)
2333          * |<---------------------------->| ===> (fs/file tree B)
2334          *
2335          * A range of space can refer to two file extents in one tree while
2336          * refer to only one file extent in another tree.
2337          *
2338          * So we may process a disk offset more than one time(two extents in A)
2339          * and locate at the same extent(one extent in B), then insert two same
2340          * backrefs(both refer to the extent in B).
2341          */
2342         return 0;
2343 }
2344
2345 static void backref_insert(struct rb_root *root,
2346                            struct sa_defrag_extent_backref *backref)
2347 {
2348         struct rb_node **p = &root->rb_node;
2349         struct rb_node *parent = NULL;
2350         struct sa_defrag_extent_backref *entry;
2351         int ret;
2352
2353         while (*p) {
2354                 parent = *p;
2355                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2356
2357                 ret = backref_comp(backref, entry);
2358                 if (ret < 0)
2359                         p = &(*p)->rb_left;
2360                 else
2361                         p = &(*p)->rb_right;
2362         }
2363
2364         rb_link_node(&backref->node, parent, p);
2365         rb_insert_color(&backref->node, root);
2366 }
2367
2368 /*
2369  * Note the backref might has changed, and in this case we just return 0.
2370  */
2371 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2372                                        void *ctx)
2373 {
2374         struct btrfs_file_extent_item *extent;
2375         struct old_sa_defrag_extent *old = ctx;
2376         struct new_sa_defrag_extent *new = old->new;
2377         struct btrfs_path *path = new->path;
2378         struct btrfs_key key;
2379         struct btrfs_root *root;
2380         struct sa_defrag_extent_backref *backref;
2381         struct extent_buffer *leaf;
2382         struct inode *inode = new->inode;
2383         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2384         int slot;
2385         int ret;
2386         u64 extent_offset;
2387         u64 num_bytes;
2388
2389         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2390             inum == btrfs_ino(BTRFS_I(inode)))
2391                 return 0;
2392
2393         key.objectid = root_id;
2394         key.type = BTRFS_ROOT_ITEM_KEY;
2395         key.offset = (u64)-1;
2396
2397         root = btrfs_read_fs_root_no_name(fs_info, &key);
2398         if (IS_ERR(root)) {
2399                 if (PTR_ERR(root) == -ENOENT)
2400                         return 0;
2401                 WARN_ON(1);
2402                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2403                          inum, offset, root_id);
2404                 return PTR_ERR(root);
2405         }
2406
2407         key.objectid = inum;
2408         key.type = BTRFS_EXTENT_DATA_KEY;
2409         if (offset > (u64)-1 << 32)
2410                 key.offset = 0;
2411         else
2412                 key.offset = offset;
2413
2414         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2415         if (WARN_ON(ret < 0))
2416                 return ret;
2417         ret = 0;
2418
2419         while (1) {
2420                 cond_resched();
2421
2422                 leaf = path->nodes[0];
2423                 slot = path->slots[0];
2424
2425                 if (slot >= btrfs_header_nritems(leaf)) {
2426                         ret = btrfs_next_leaf(root, path);
2427                         if (ret < 0) {
2428                                 goto out;
2429                         } else if (ret > 0) {
2430                                 ret = 0;
2431                                 goto out;
2432                         }
2433                         continue;
2434                 }
2435
2436                 path->slots[0]++;
2437
2438                 btrfs_item_key_to_cpu(leaf, &key, slot);
2439
2440                 if (key.objectid > inum)
2441                         goto out;
2442
2443                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2444                         continue;
2445
2446                 extent = btrfs_item_ptr(leaf, slot,
2447                                         struct btrfs_file_extent_item);
2448
2449                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2450                         continue;
2451
2452                 /*
2453                  * 'offset' refers to the exact key.offset,
2454                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2455                  * (key.offset - extent_offset).
2456                  */
2457                 if (key.offset != offset)
2458                         continue;
2459
2460                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2461                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2462
2463                 if (extent_offset >= old->extent_offset + old->offset +
2464                     old->len || extent_offset + num_bytes <=
2465                     old->extent_offset + old->offset)
2466                         continue;
2467                 break;
2468         }
2469
2470         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2471         if (!backref) {
2472                 ret = -ENOENT;
2473                 goto out;
2474         }
2475
2476         backref->root_id = root_id;
2477         backref->inum = inum;
2478         backref->file_pos = offset;
2479         backref->num_bytes = num_bytes;
2480         backref->extent_offset = extent_offset;
2481         backref->generation = btrfs_file_extent_generation(leaf, extent);
2482         backref->old = old;
2483         backref_insert(&new->root, backref);
2484         old->count++;
2485 out:
2486         btrfs_release_path(path);
2487         WARN_ON(ret);
2488         return ret;
2489 }
2490
2491 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2492                                    struct new_sa_defrag_extent *new)
2493 {
2494         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2495         struct old_sa_defrag_extent *old, *tmp;
2496         int ret;
2497
2498         new->path = path;
2499
2500         list_for_each_entry_safe(old, tmp, &new->head, list) {
2501                 ret = iterate_inodes_from_logical(old->bytenr +
2502                                                   old->extent_offset, fs_info,
2503                                                   path, record_one_backref,
2504                                                   old, false);
2505                 if (ret < 0 && ret != -ENOENT)
2506                         return false;
2507
2508                 /* no backref to be processed for this extent */
2509                 if (!old->count) {
2510                         list_del(&old->list);
2511                         kfree(old);
2512                 }
2513         }
2514
2515         if (list_empty(&new->head))
2516                 return false;
2517
2518         return true;
2519 }
2520
2521 static int relink_is_mergable(struct extent_buffer *leaf,
2522                               struct btrfs_file_extent_item *fi,
2523                               struct new_sa_defrag_extent *new)
2524 {
2525         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2526                 return 0;
2527
2528         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2529                 return 0;
2530
2531         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2532                 return 0;
2533
2534         if (btrfs_file_extent_encryption(leaf, fi) ||
2535             btrfs_file_extent_other_encoding(leaf, fi))
2536                 return 0;
2537
2538         return 1;
2539 }
2540
2541 /*
2542  * Note the backref might has changed, and in this case we just return 0.
2543  */
2544 static noinline int relink_extent_backref(struct btrfs_path *path,
2545                                  struct sa_defrag_extent_backref *prev,
2546                                  struct sa_defrag_extent_backref *backref)
2547 {
2548         struct btrfs_file_extent_item *extent;
2549         struct btrfs_file_extent_item *item;
2550         struct btrfs_ordered_extent *ordered;
2551         struct btrfs_trans_handle *trans;
2552         struct btrfs_root *root;
2553         struct btrfs_key key;
2554         struct extent_buffer *leaf;
2555         struct old_sa_defrag_extent *old = backref->old;
2556         struct new_sa_defrag_extent *new = old->new;
2557         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2558         struct inode *inode;
2559         struct extent_state *cached = NULL;
2560         int ret = 0;
2561         u64 start;
2562         u64 len;
2563         u64 lock_start;
2564         u64 lock_end;
2565         bool merge = false;
2566         int index;
2567
2568         if (prev && prev->root_id == backref->root_id &&
2569             prev->inum == backref->inum &&
2570             prev->file_pos + prev->num_bytes == backref->file_pos)
2571                 merge = true;
2572
2573         /* step 1: get root */
2574         key.objectid = backref->root_id;
2575         key.type = BTRFS_ROOT_ITEM_KEY;
2576         key.offset = (u64)-1;
2577
2578         index = srcu_read_lock(&fs_info->subvol_srcu);
2579
2580         root = btrfs_read_fs_root_no_name(fs_info, &key);
2581         if (IS_ERR(root)) {
2582                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2583                 if (PTR_ERR(root) == -ENOENT)
2584                         return 0;
2585                 return PTR_ERR(root);
2586         }
2587
2588         if (btrfs_root_readonly(root)) {
2589                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2590                 return 0;
2591         }
2592
2593         /* step 2: get inode */
2594         key.objectid = backref->inum;
2595         key.type = BTRFS_INODE_ITEM_KEY;
2596         key.offset = 0;
2597
2598         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2599         if (IS_ERR(inode)) {
2600                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2601                 return 0;
2602         }
2603
2604         srcu_read_unlock(&fs_info->subvol_srcu, index);
2605
2606         /* step 3: relink backref */
2607         lock_start = backref->file_pos;
2608         lock_end = backref->file_pos + backref->num_bytes - 1;
2609         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2610                          &cached);
2611
2612         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2613         if (ordered) {
2614                 btrfs_put_ordered_extent(ordered);
2615                 goto out_unlock;
2616         }
2617
2618         trans = btrfs_join_transaction(root);
2619         if (IS_ERR(trans)) {
2620                 ret = PTR_ERR(trans);
2621                 goto out_unlock;
2622         }
2623
2624         key.objectid = backref->inum;
2625         key.type = BTRFS_EXTENT_DATA_KEY;
2626         key.offset = backref->file_pos;
2627
2628         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2629         if (ret < 0) {
2630                 goto out_free_path;
2631         } else if (ret > 0) {
2632                 ret = 0;
2633                 goto out_free_path;
2634         }
2635
2636         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2637                                 struct btrfs_file_extent_item);
2638
2639         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2640             backref->generation)
2641                 goto out_free_path;
2642
2643         btrfs_release_path(path);
2644
2645         start = backref->file_pos;
2646         if (backref->extent_offset < old->extent_offset + old->offset)
2647                 start += old->extent_offset + old->offset -
2648                          backref->extent_offset;
2649
2650         len = min(backref->extent_offset + backref->num_bytes,
2651                   old->extent_offset + old->offset + old->len);
2652         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2653
2654         ret = btrfs_drop_extents(trans, root, inode, start,
2655                                  start + len, 1);
2656         if (ret)
2657                 goto out_free_path;
2658 again:
2659         key.objectid = btrfs_ino(BTRFS_I(inode));
2660         key.type = BTRFS_EXTENT_DATA_KEY;
2661         key.offset = start;
2662
2663         path->leave_spinning = 1;
2664         if (merge) {
2665                 struct btrfs_file_extent_item *fi;
2666                 u64 extent_len;
2667                 struct btrfs_key found_key;
2668
2669                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2670                 if (ret < 0)
2671                         goto out_free_path;
2672
2673                 path->slots[0]--;
2674                 leaf = path->nodes[0];
2675                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2676
2677                 fi = btrfs_item_ptr(leaf, path->slots[0],
2678                                     struct btrfs_file_extent_item);
2679                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2680
2681                 if (extent_len + found_key.offset == start &&
2682                     relink_is_mergable(leaf, fi, new)) {
2683                         btrfs_set_file_extent_num_bytes(leaf, fi,
2684                                                         extent_len + len);
2685                         btrfs_mark_buffer_dirty(leaf);
2686                         inode_add_bytes(inode, len);
2687
2688                         ret = 1;
2689                         goto out_free_path;
2690                 } else {
2691                         merge = false;
2692                         btrfs_release_path(path);
2693                         goto again;
2694                 }
2695         }
2696
2697         ret = btrfs_insert_empty_item(trans, root, path, &key,
2698                                         sizeof(*extent));
2699         if (ret) {
2700                 btrfs_abort_transaction(trans, ret);
2701                 goto out_free_path;
2702         }
2703
2704         leaf = path->nodes[0];
2705         item = btrfs_item_ptr(leaf, path->slots[0],
2706                                 struct btrfs_file_extent_item);
2707         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2708         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2709         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2710         btrfs_set_file_extent_num_bytes(leaf, item, len);
2711         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2712         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2713         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2714         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2715         btrfs_set_file_extent_encryption(leaf, item, 0);
2716         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2717
2718         btrfs_mark_buffer_dirty(leaf);
2719         inode_add_bytes(inode, len);
2720         btrfs_release_path(path);
2721
2722         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2723                         new->disk_len, 0,
2724                         backref->root_id, backref->inum,
2725                         new->file_pos); /* start - extent_offset */
2726         if (ret) {
2727                 btrfs_abort_transaction(trans, ret);
2728                 goto out_free_path;
2729         }
2730
2731         ret = 1;
2732 out_free_path:
2733         btrfs_release_path(path);
2734         path->leave_spinning = 0;
2735         btrfs_end_transaction(trans);
2736 out_unlock:
2737         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2738                              &cached);
2739         iput(inode);
2740         return ret;
2741 }
2742
2743 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2744 {
2745         struct old_sa_defrag_extent *old, *tmp;
2746
2747         if (!new)
2748                 return;
2749
2750         list_for_each_entry_safe(old, tmp, &new->head, list) {
2751                 kfree(old);
2752         }
2753         kfree(new);
2754 }
2755
2756 static void relink_file_extents(struct new_sa_defrag_extent *new)
2757 {
2758         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2759         struct btrfs_path *path;
2760         struct sa_defrag_extent_backref *backref;
2761         struct sa_defrag_extent_backref *prev = NULL;
2762         struct inode *inode;
2763         struct rb_node *node;
2764         int ret;
2765
2766         inode = new->inode;
2767
2768         path = btrfs_alloc_path();
2769         if (!path)
2770                 return;
2771
2772         if (!record_extent_backrefs(path, new)) {
2773                 btrfs_free_path(path);
2774                 goto out;
2775         }
2776         btrfs_release_path(path);
2777
2778         while (1) {
2779                 node = rb_first(&new->root);
2780                 if (!node)
2781                         break;
2782                 rb_erase(node, &new->root);
2783
2784                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2785
2786                 ret = relink_extent_backref(path, prev, backref);
2787                 WARN_ON(ret < 0);
2788
2789                 kfree(prev);
2790
2791                 if (ret == 1)
2792                         prev = backref;
2793                 else
2794                         prev = NULL;
2795                 cond_resched();
2796         }
2797         kfree(prev);
2798
2799         btrfs_free_path(path);
2800 out:
2801         free_sa_defrag_extent(new);
2802
2803         atomic_dec(&fs_info->defrag_running);
2804         wake_up(&fs_info->transaction_wait);
2805 }
2806
2807 static struct new_sa_defrag_extent *
2808 record_old_file_extents(struct inode *inode,
2809                         struct btrfs_ordered_extent *ordered)
2810 {
2811         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2812         struct btrfs_root *root = BTRFS_I(inode)->root;
2813         struct btrfs_path *path;
2814         struct btrfs_key key;
2815         struct old_sa_defrag_extent *old;
2816         struct new_sa_defrag_extent *new;
2817         int ret;
2818
2819         new = kmalloc(sizeof(*new), GFP_NOFS);
2820         if (!new)
2821                 return NULL;
2822
2823         new->inode = inode;
2824         new->file_pos = ordered->file_offset;
2825         new->len = ordered->len;
2826         new->bytenr = ordered->start;
2827         new->disk_len = ordered->disk_len;
2828         new->compress_type = ordered->compress_type;
2829         new->root = RB_ROOT;
2830         INIT_LIST_HEAD(&new->head);
2831
2832         path = btrfs_alloc_path();
2833         if (!path)
2834                 goto out_kfree;
2835
2836         key.objectid = btrfs_ino(BTRFS_I(inode));
2837         key.type = BTRFS_EXTENT_DATA_KEY;
2838         key.offset = new->file_pos;
2839
2840         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2841         if (ret < 0)
2842                 goto out_free_path;
2843         if (ret > 0 && path->slots[0] > 0)
2844                 path->slots[0]--;
2845
2846         /* find out all the old extents for the file range */
2847         while (1) {
2848                 struct btrfs_file_extent_item *extent;
2849                 struct extent_buffer *l;
2850                 int slot;
2851                 u64 num_bytes;
2852                 u64 offset;
2853                 u64 end;
2854                 u64 disk_bytenr;
2855                 u64 extent_offset;
2856
2857                 l = path->nodes[0];
2858                 slot = path->slots[0];
2859
2860                 if (slot >= btrfs_header_nritems(l)) {
2861                         ret = btrfs_next_leaf(root, path);
2862                         if (ret < 0)
2863                                 goto out_free_path;
2864                         else if (ret > 0)
2865                                 break;
2866                         continue;
2867                 }
2868
2869                 btrfs_item_key_to_cpu(l, &key, slot);
2870
2871                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2872                         break;
2873                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2874                         break;
2875                 if (key.offset >= new->file_pos + new->len)
2876                         break;
2877
2878                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2879
2880                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2881                 if (key.offset + num_bytes < new->file_pos)
2882                         goto next;
2883
2884                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2885                 if (!disk_bytenr)
2886                         goto next;
2887
2888                 extent_offset = btrfs_file_extent_offset(l, extent);
2889
2890                 old = kmalloc(sizeof(*old), GFP_NOFS);
2891                 if (!old)
2892                         goto out_free_path;
2893
2894                 offset = max(new->file_pos, key.offset);
2895                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2896
2897                 old->bytenr = disk_bytenr;
2898                 old->extent_offset = extent_offset;
2899                 old->offset = offset - key.offset;
2900                 old->len = end - offset;
2901                 old->new = new;
2902                 old->count = 0;
2903                 list_add_tail(&old->list, &new->head);
2904 next:
2905                 path->slots[0]++;
2906                 cond_resched();
2907         }
2908
2909         btrfs_free_path(path);
2910         atomic_inc(&fs_info->defrag_running);
2911
2912         return new;
2913
2914 out_free_path:
2915         btrfs_free_path(path);
2916 out_kfree:
2917         free_sa_defrag_extent(new);
2918         return NULL;
2919 }
2920
2921 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2922                                          u64 start, u64 len)
2923 {
2924         struct btrfs_block_group_cache *cache;
2925
2926         cache = btrfs_lookup_block_group(fs_info, start);
2927         ASSERT(cache);
2928
2929         spin_lock(&cache->lock);
2930         cache->delalloc_bytes -= len;
2931         spin_unlock(&cache->lock);
2932
2933         btrfs_put_block_group(cache);
2934 }
2935
2936 /* as ordered data IO finishes, this gets called so we can finish
2937  * an ordered extent if the range of bytes in the file it covers are
2938  * fully written.
2939  */
2940 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2941 {
2942         struct inode *inode = ordered_extent->inode;
2943         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2944         struct btrfs_root *root = BTRFS_I(inode)->root;
2945         struct btrfs_trans_handle *trans = NULL;
2946         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2947         struct extent_state *cached_state = NULL;
2948         struct new_sa_defrag_extent *new = NULL;
2949         int compress_type = 0;
2950         int ret = 0;
2951         u64 logical_len = ordered_extent->len;
2952         bool nolock;
2953         bool truncated = false;
2954         bool range_locked = false;
2955         bool clear_new_delalloc_bytes = false;
2956
2957         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2958             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2959             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2960                 clear_new_delalloc_bytes = true;
2961
2962         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2963
2964         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2965                 ret = -EIO;
2966                 goto out;
2967         }
2968
2969         btrfs_free_io_failure_record(BTRFS_I(inode),
2970                         ordered_extent->file_offset,
2971                         ordered_extent->file_offset +
2972                         ordered_extent->len - 1);
2973
2974         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2975                 truncated = true;
2976                 logical_len = ordered_extent->truncated_len;
2977                 /* Truncated the entire extent, don't bother adding */
2978                 if (!logical_len)
2979                         goto out;
2980         }
2981
2982         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2983                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2984
2985                 /*
2986                  * For mwrite(mmap + memset to write) case, we still reserve
2987                  * space for NOCOW range.
2988                  * As NOCOW won't cause a new delayed ref, just free the space
2989                  */
2990                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2991                                        ordered_extent->len);
2992                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2993                 if (nolock)
2994                         trans = btrfs_join_transaction_nolock(root);
2995                 else
2996                         trans = btrfs_join_transaction(root);
2997                 if (IS_ERR(trans)) {
2998                         ret = PTR_ERR(trans);
2999                         trans = NULL;
3000                         goto out;
3001                 }
3002                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3003                 ret = btrfs_update_inode_fallback(trans, root, inode);
3004                 if (ret) /* -ENOMEM or corruption */
3005                         btrfs_abort_transaction(trans, ret);
3006                 goto out;
3007         }
3008
3009         range_locked = true;
3010         lock_extent_bits(io_tree, ordered_extent->file_offset,
3011                          ordered_extent->file_offset + ordered_extent->len - 1,
3012                          &cached_state);
3013
3014         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3015                         ordered_extent->file_offset + ordered_extent->len - 1,
3016                         EXTENT_DEFRAG, 0, cached_state);
3017         if (ret) {
3018                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3019                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3020                         /* the inode is shared */
3021                         new = record_old_file_extents(inode, ordered_extent);
3022
3023                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3024                         ordered_extent->file_offset + ordered_extent->len - 1,
3025                         EXTENT_DEFRAG, 0, 0, &cached_state);
3026         }
3027
3028         if (nolock)
3029                 trans = btrfs_join_transaction_nolock(root);
3030         else
3031                 trans = btrfs_join_transaction(root);
3032         if (IS_ERR(trans)) {
3033                 ret = PTR_ERR(trans);
3034                 trans = NULL;
3035                 goto out;
3036         }
3037
3038         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3039
3040         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3041                 compress_type = ordered_extent->compress_type;
3042         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3043                 BUG_ON(compress_type);
3044                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3045                                        ordered_extent->len);
3046                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3047                                                 ordered_extent->file_offset,
3048                                                 ordered_extent->file_offset +
3049                                                 logical_len);
3050         } else {
3051                 BUG_ON(root == fs_info->tree_root);
3052                 ret = insert_reserved_file_extent(trans, inode,
3053                                                 ordered_extent->file_offset,
3054                                                 ordered_extent->start,
3055                                                 ordered_extent->disk_len,
3056                                                 logical_len, logical_len,
3057                                                 compress_type, 0, 0,
3058                                                 BTRFS_FILE_EXTENT_REG);
3059                 if (!ret)
3060                         btrfs_release_delalloc_bytes(fs_info,
3061                                                      ordered_extent->start,
3062                                                      ordered_extent->disk_len);
3063         }
3064         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3065                            ordered_extent->file_offset, ordered_extent->len,
3066                            trans->transid);
3067         if (ret < 0) {
3068                 btrfs_abort_transaction(trans, ret);
3069                 goto out;
3070         }
3071
3072         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3073         if (ret) {
3074                 btrfs_abort_transaction(trans, ret);
3075                 goto out;
3076         }
3077
3078         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3079         ret = btrfs_update_inode_fallback(trans, root, inode);
3080         if (ret) { /* -ENOMEM or corruption */
3081                 btrfs_abort_transaction(trans, ret);
3082                 goto out;
3083         }
3084         ret = 0;
3085 out:
3086         if (range_locked || clear_new_delalloc_bytes) {
3087                 unsigned int clear_bits = 0;
3088
3089                 if (range_locked)
3090                         clear_bits |= EXTENT_LOCKED;
3091                 if (clear_new_delalloc_bytes)
3092                         clear_bits |= EXTENT_DELALLOC_NEW;
3093                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3094                                  ordered_extent->file_offset,
3095                                  ordered_extent->file_offset +
3096                                  ordered_extent->len - 1,
3097                                  clear_bits,
3098                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3099                                  0, &cached_state);
3100         }
3101
3102         if (trans)
3103                 btrfs_end_transaction(trans);
3104
3105         if (ret || truncated) {
3106                 u64 start, end;
3107
3108                 if (truncated)
3109                         start = ordered_extent->file_offset + logical_len;
3110                 else
3111                         start = ordered_extent->file_offset;
3112                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3113                 clear_extent_uptodate(io_tree, start, end, NULL);
3114
3115                 /* Drop the cache for the part of the extent we didn't write. */
3116                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3117
3118                 /*
3119                  * If the ordered extent had an IOERR or something else went
3120                  * wrong we need to return the space for this ordered extent
3121                  * back to the allocator.  We only free the extent in the
3122                  * truncated case if we didn't write out the extent at all.
3123                  */
3124                 if ((ret || !logical_len) &&
3125                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3126                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3127                         btrfs_free_reserved_extent(fs_info,
3128                                                    ordered_extent->start,
3129                                                    ordered_extent->disk_len, 1);
3130         }
3131
3132
3133         /*
3134          * This needs to be done to make sure anybody waiting knows we are done
3135          * updating everything for this ordered extent.
3136          */
3137         btrfs_remove_ordered_extent(inode, ordered_extent);
3138
3139         /* for snapshot-aware defrag */
3140         if (new) {
3141                 if (ret) {
3142                         free_sa_defrag_extent(new);
3143                         atomic_dec(&fs_info->defrag_running);
3144                 } else {
3145                         relink_file_extents(new);
3146                 }
3147         }
3148
3149         /* once for us */
3150         btrfs_put_ordered_extent(ordered_extent);
3151         /* once for the tree */
3152         btrfs_put_ordered_extent(ordered_extent);
3153
3154         return ret;
3155 }
3156
3157 static void finish_ordered_fn(struct btrfs_work *work)
3158 {
3159         struct btrfs_ordered_extent *ordered_extent;
3160         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3161         btrfs_finish_ordered_io(ordered_extent);
3162 }
3163
3164 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3165                                 struct extent_state *state, int uptodate)
3166 {
3167         struct inode *inode = page->mapping->host;
3168         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3169         struct btrfs_ordered_extent *ordered_extent = NULL;
3170         struct btrfs_workqueue *wq;
3171         btrfs_work_func_t func;
3172
3173         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3174
3175         ClearPagePrivate2(page);
3176         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3177                                             end - start + 1, uptodate))
3178                 return;
3179
3180         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3181                 wq = fs_info->endio_freespace_worker;
3182                 func = btrfs_freespace_write_helper;
3183         } else {
3184                 wq = fs_info->endio_write_workers;
3185                 func = btrfs_endio_write_helper;
3186         }
3187
3188         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3189                         NULL);
3190         btrfs_queue_work(wq, &ordered_extent->work);
3191 }
3192
3193 static int __readpage_endio_check(struct inode *inode,
3194                                   struct btrfs_io_bio *io_bio,
3195                                   int icsum, struct page *page,
3196                                   int pgoff, u64 start, size_t len)
3197 {
3198         char *kaddr;
3199         u32 csum_expected;
3200         u32 csum = ~(u32)0;
3201
3202         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3203
3204         kaddr = kmap_atomic(page);
3205         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3206         btrfs_csum_final(csum, (u8 *)&csum);
3207         if (csum != csum_expected)
3208                 goto zeroit;
3209
3210         kunmap_atomic(kaddr);
3211         return 0;
3212 zeroit:
3213         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3214                                     io_bio->mirror_num);
3215         memset(kaddr + pgoff, 1, len);
3216         flush_dcache_page(page);
3217         kunmap_atomic(kaddr);
3218         return -EIO;
3219 }
3220
3221 /*
3222  * when reads are done, we need to check csums to verify the data is correct
3223  * if there's a match, we allow the bio to finish.  If not, the code in
3224  * extent_io.c will try to find good copies for us.
3225  */
3226 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3227                                       u64 phy_offset, struct page *page,
3228                                       u64 start, u64 end, int mirror)
3229 {
3230         size_t offset = start - page_offset(page);
3231         struct inode *inode = page->mapping->host;
3232         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3233         struct btrfs_root *root = BTRFS_I(inode)->root;
3234
3235         if (PageChecked(page)) {
3236                 ClearPageChecked(page);
3237                 return 0;
3238         }
3239
3240         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3241                 return 0;
3242
3243         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3244             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3245                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3246                 return 0;
3247         }
3248
3249         phy_offset >>= inode->i_sb->s_blocksize_bits;
3250         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3251                                       start, (size_t)(end - start + 1));
3252 }
3253
3254 /*
3255  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3256  *
3257  * @inode: The inode we want to perform iput on
3258  *
3259  * This function uses the generic vfs_inode::i_count to track whether we should
3260  * just decrement it (in case it's > 1) or if this is the last iput then link
3261  * the inode to the delayed iput machinery. Delayed iputs are processed at
3262  * transaction commit time/superblock commit/cleaner kthread.
3263  */
3264 void btrfs_add_delayed_iput(struct inode *inode)
3265 {
3266         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3267         struct btrfs_inode *binode = BTRFS_I(inode);
3268
3269         if (atomic_add_unless(&inode->i_count, -1, 1))
3270                 return;
3271
3272         spin_lock(&fs_info->delayed_iput_lock);
3273         ASSERT(list_empty(&binode->delayed_iput));
3274         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3275         spin_unlock(&fs_info->delayed_iput_lock);
3276 }
3277
3278 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3279 {
3280
3281         spin_lock(&fs_info->delayed_iput_lock);
3282         while (!list_empty(&fs_info->delayed_iputs)) {
3283                 struct btrfs_inode *inode;
3284
3285                 inode = list_first_entry(&fs_info->delayed_iputs,
3286                                 struct btrfs_inode, delayed_iput);
3287                 list_del_init(&inode->delayed_iput);
3288                 spin_unlock(&fs_info->delayed_iput_lock);
3289                 iput(&inode->vfs_inode);
3290                 spin_lock(&fs_info->delayed_iput_lock);
3291         }
3292         spin_unlock(&fs_info->delayed_iput_lock);
3293 }
3294
3295 /*
3296  * This is called in transaction commit time. If there are no orphan
3297  * files in the subvolume, it removes orphan item and frees block_rsv
3298  * structure.
3299  */
3300 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3301                               struct btrfs_root *root)
3302 {
3303         struct btrfs_fs_info *fs_info = root->fs_info;
3304         struct btrfs_block_rsv *block_rsv;
3305         int ret;
3306
3307         if (atomic_read(&root->orphan_inodes) ||
3308             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3309                 return;
3310
3311         spin_lock(&root->orphan_lock);
3312         if (atomic_read(&root->orphan_inodes)) {
3313                 spin_unlock(&root->orphan_lock);
3314                 return;
3315         }
3316
3317         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3318                 spin_unlock(&root->orphan_lock);
3319                 return;
3320         }
3321
3322         block_rsv = root->orphan_block_rsv;
3323         root->orphan_block_rsv = NULL;
3324         spin_unlock(&root->orphan_lock);
3325
3326         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3327             btrfs_root_refs(&root->root_item) > 0) {
3328                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3329                                             root->root_key.objectid);
3330                 if (ret)
3331                         btrfs_abort_transaction(trans, ret);
3332                 else
3333                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3334                                   &root->state);
3335         }
3336
3337         if (block_rsv) {
3338                 WARN_ON(block_rsv->size > 0);
3339                 btrfs_free_block_rsv(fs_info, block_rsv);
3340         }
3341 }
3342
3343 /*
3344  * This creates an orphan entry for the given inode in case something goes
3345  * wrong in the middle of an unlink/truncate.
3346  *
3347  * NOTE: caller of this function should reserve 5 units of metadata for
3348  *       this function.
3349  */
3350 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3351                 struct btrfs_inode *inode)
3352 {
3353         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3354         struct btrfs_root *root = inode->root;
3355         struct btrfs_block_rsv *block_rsv = NULL;
3356         int reserve = 0;
3357         bool insert = false;
3358         int ret;
3359
3360         if (!root->orphan_block_rsv) {
3361                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3362                                                   BTRFS_BLOCK_RSV_TEMP);
3363                 if (!block_rsv)
3364                         return -ENOMEM;
3365         }
3366
3367         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3368                               &inode->runtime_flags))
3369                 insert = true;
3370
3371         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3372                               &inode->runtime_flags))
3373                 reserve = 1;
3374
3375         spin_lock(&root->orphan_lock);
3376         /* If someone has created ->orphan_block_rsv, be happy to use it. */
3377         if (!root->orphan_block_rsv) {
3378                 root->orphan_block_rsv = block_rsv;
3379         } else if (block_rsv) {
3380                 btrfs_free_block_rsv(fs_info, block_rsv);
3381                 block_rsv = NULL;
3382         }
3383
3384         if (insert)
3385                 atomic_inc(&root->orphan_inodes);
3386         spin_unlock(&root->orphan_lock);
3387
3388         /* grab metadata reservation from transaction handle */
3389         if (reserve) {
3390                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3391                 ASSERT(!ret);
3392                 if (ret) {
3393                         /*
3394                          * dec doesn't need spin_lock as ->orphan_block_rsv
3395                          * would be released only if ->orphan_inodes is
3396                          * zero.
3397                          */
3398                         atomic_dec(&root->orphan_inodes);
3399                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3400                                   &inode->runtime_flags);
3401                         if (insert)
3402                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3403                                           &inode->runtime_flags);
3404                         return ret;
3405                 }
3406         }
3407
3408         /* insert an orphan item to track this unlinked/truncated file */
3409         if (insert) {
3410                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3411                 if (ret) {
3412                         if (reserve) {
3413                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3414                                           &inode->runtime_flags);
3415                                 btrfs_orphan_release_metadata(inode);
3416                         }
3417                         /*
3418                          * btrfs_orphan_commit_root may race with us and set
3419                          * ->orphan_block_rsv to zero, in order to avoid that,
3420                          * decrease ->orphan_inodes after everything is done.
3421                          */
3422                         atomic_dec(&root->orphan_inodes);
3423                         if (ret != -EEXIST) {
3424                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3425                                           &inode->runtime_flags);
3426                                 btrfs_abort_transaction(trans, ret);
3427                                 return ret;
3428                         }
3429                 }
3430                 ret = 0;
3431         }
3432
3433         return 0;
3434 }
3435
3436 /*
3437  * We have done the truncate/delete so we can go ahead and remove the orphan
3438  * item for this particular inode.
3439  */
3440 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3441                             struct btrfs_inode *inode)
3442 {
3443         struct btrfs_root *root = inode->root;
3444         int delete_item = 0;
3445         int ret = 0;
3446
3447         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3448                                &inode->runtime_flags))
3449                 delete_item = 1;
3450
3451         if (delete_item && trans)
3452                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3453
3454         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3455                                &inode->runtime_flags))
3456                 btrfs_orphan_release_metadata(inode);
3457
3458         /*
3459          * btrfs_orphan_commit_root may race with us and set ->orphan_block_rsv
3460          * to zero, in order to avoid that, decrease ->orphan_inodes after
3461          * everything is done.
3462          */
3463         if (delete_item)
3464                 atomic_dec(&root->orphan_inodes);
3465
3466         return ret;
3467 }
3468
3469 /*
3470  * this cleans up any orphans that may be left on the list from the last use
3471  * of this root.
3472  */
3473 int btrfs_orphan_cleanup(struct btrfs_root *root)
3474 {
3475         struct btrfs_fs_info *fs_info = root->fs_info;
3476         struct btrfs_path *path;
3477         struct extent_buffer *leaf;
3478         struct btrfs_key key, found_key;
3479         struct btrfs_trans_handle *trans;
3480         struct inode *inode;
3481         u64 last_objectid = 0;
3482         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3483
3484         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3485                 return 0;
3486
3487         path = btrfs_alloc_path();
3488         if (!path) {
3489                 ret = -ENOMEM;
3490                 goto out;
3491         }
3492         path->reada = READA_BACK;
3493
3494         key.objectid = BTRFS_ORPHAN_OBJECTID;
3495         key.type = BTRFS_ORPHAN_ITEM_KEY;
3496         key.offset = (u64)-1;
3497
3498         while (1) {
3499                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3500                 if (ret < 0)
3501                         goto out;
3502
3503                 /*
3504                  * if ret == 0 means we found what we were searching for, which
3505                  * is weird, but possible, so only screw with path if we didn't
3506                  * find the key and see if we have stuff that matches
3507                  */
3508                 if (ret > 0) {
3509                         ret = 0;
3510                         if (path->slots[0] == 0)
3511                                 break;
3512                         path->slots[0]--;
3513                 }
3514
3515                 /* pull out the item */
3516                 leaf = path->nodes[0];
3517                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3518
3519                 /* make sure the item matches what we want */
3520                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3521                         break;
3522                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3523                         break;
3524
3525                 /* release the path since we're done with it */
3526                 btrfs_release_path(path);
3527
3528                 /*
3529                  * this is where we are basically btrfs_lookup, without the
3530                  * crossing root thing.  we store the inode number in the
3531                  * offset of the orphan item.
3532                  */
3533
3534                 if (found_key.offset == last_objectid) {
3535                         btrfs_err(fs_info,
3536                                   "Error removing orphan entry, stopping orphan cleanup");
3537                         ret = -EINVAL;
3538                         goto out;
3539                 }
3540
3541                 last_objectid = found_key.offset;
3542
3543                 found_key.objectid = found_key.offset;
3544                 found_key.type = BTRFS_INODE_ITEM_KEY;
3545                 found_key.offset = 0;
3546                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);