fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         unsigned long index = offset >> PAGE_SHIFT;
 139         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 140         struct page *page;
 141
 142         while (index <= end_index) {
 143                 page = find_get_page(inode->i_mapping, index);
 144                 index++;
 145                 if (!page)
 146                         continue;
 147                 ClearPagePrivate2(page);
 148                 put_page(page);
 149         }
 150         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 151                                             bytes - PAGE_SIZE, false);
 152 }
 153
 154 static int btrfs_dirty_inode(struct inode *inode);
 155
 156 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 157 void btrfs_test_inode_set_ops(struct inode *inode)
 158 {
 159         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 160 }
 161 #endif
 162
 163 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 164                                      struct inode *inode,  struct inode *dir,
 165                                      const struct qstr *qstr)
 166 {
 167         int err;
 168
 169         err = btrfs_init_acl(trans, inode, dir);
 170         if (!err)
 171                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 172         return err;
 173 }
 174
 175 /*
 176  * this does all the hard work for inserting an inline extent into
 177  * the btree.  The caller should have done a btrfs_drop_extents so that
 178  * no overlapping inline items exist in the btree
 179  */
 180 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 181                                 struct btrfs_path *path, int extent_inserted,
 182                                 struct btrfs_root *root, struct inode *inode,
 183                                 u64 start, size_t size, size_t compressed_size,
 184                                 int compress_type,
 185                                 struct page **compressed_pages)
 186 {
 187         struct extent_buffer *leaf;
 188         struct page *page = NULL;
 189         char *kaddr;
 190         unsigned long ptr;
 191         struct btrfs_file_extent_item *ei;
 192         int ret;
 193         size_t cur_size = size;
 194         unsigned long offset;
 195
 196         if (compressed_size && compressed_pages)
 197                 cur_size = compressed_size;
 198
 199         inode_add_bytes(inode, size);
 200
 201         if (!extent_inserted) {
 202                 struct btrfs_key key;
 203                 size_t datasize;
 204
 205                 key.objectid = btrfs_ino(BTRFS_I(inode));
 206                 key.offset = start;
 207                 key.type = BTRFS_EXTENT_DATA_KEY;
 208
 209                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 210                 path->leave_spinning = 1;
 211                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 212                                               datasize);
 213                 if (ret)
 214                         goto fail;
 215         }
 216         leaf = path->nodes[0];
 217         ei = btrfs_item_ptr(leaf, path->slots[0],
 218                             struct btrfs_file_extent_item);
 219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 220         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 221         btrfs_set_file_extent_encryption(leaf, ei, 0);
 222         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 223         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 224         ptr = btrfs_file_extent_inline_start(ei);
 225
 226         if (compress_type != BTRFS_COMPRESS_NONE) {
 227                 struct page *cpage;
 228                 int i = 0;
 229                 while (compressed_size > 0) {
 230                         cpage = compressed_pages[i];
 231                         cur_size = min_t(unsigned long, compressed_size,
 232                                        PAGE_SIZE);
 233
 234                         kaddr = kmap_atomic(cpage);
 235                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 236                         kunmap_atomic(kaddr);
 237
 238                         i++;
 239                         ptr += cur_size;
 240                         compressed_size -= cur_size;
 241                 }
 242                 btrfs_set_file_extent_compression(leaf, ei,
 243                                                   compress_type);
 244         } else {
 245                 page = find_get_page(inode->i_mapping,
 246                                      start >> PAGE_SHIFT);
 247                 btrfs_set_file_extent_compression(leaf, ei, 0);
 248                 kaddr = kmap_atomic(page);
 249                 offset = start & (PAGE_SIZE - 1);
 250                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 251                 kunmap_atomic(kaddr);
 252                 put_page(page);
 253         }
 254         btrfs_mark_buffer_dirty(leaf);
 255         btrfs_release_path(path);
 256
 257         /*
 258          * we're an inline extent, so nobody can
 259          * extend the file past i_size without locking
 260          * a page we already have locked.
 261          *
 262          * We must do any isize and inode updates
 263          * before we unlock the pages.  Otherwise we
 264          * could end up racing with unlink.
 265          */
 266         BTRFS_I(inode)->disk_i_size = inode->i_size;
 267         ret = btrfs_update_inode(trans, root, inode);
 268
 269 fail:
 270         return ret;
 271 }
 272
 273
 274 /*
 275  * conditionally insert an inline extent into the file.  This
 276  * does the checks required to make sure the data is small enough
 277  * to fit as an inline extent.
 278  */
 279 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 280                                           u64 end, size_t compressed_size,
 281                                           int compress_type,
 282                                           struct page **compressed_pages)
 283 {
 284         struct btrfs_root *root = BTRFS_I(inode)->root;
 285         struct btrfs_fs_info *fs_info = root->fs_info;
 286         struct btrfs_trans_handle *trans;
 287         u64 isize = i_size_read(inode);
 288         u64 actual_end = min(end + 1, isize);
 289         u64 inline_len = actual_end - start;
 290         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 291         u64 data_len = inline_len;
 292         int ret;
 293         struct btrfs_path *path;
 294         int extent_inserted = 0;
 295         u32 extent_item_size;
 296
 297         if (compressed_size)
 298                 data_len = compressed_size;
 299
 300         if (start > 0 ||
 301             actual_end > fs_info->sectorsize ||
 302             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 303             (!compressed_size &&
 304             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 305             end + 1 < isize ||
 306             data_len > fs_info->max_inline) {
 307                 return 1;
 308         }
 309
 310         path = btrfs_alloc_path();
 311         if (!path)
 312                 return -ENOMEM;
 313
 314         trans = btrfs_join_transaction(root);
 315         if (IS_ERR(trans)) {
 316                 btrfs_free_path(path);
 317                 return PTR_ERR(trans);
 318         }
 319         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 320
 321         if (compressed_size && compressed_pages)
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                    compressed_size);
 324         else
 325                 extent_item_size = btrfs_file_extent_calc_inline_size(
 326                     inline_len);
 327
 328         ret = __btrfs_drop_extents(trans, root, inode, path,
 329                                    start, aligned_end, NULL,
 330                                    1, 1, extent_item_size, &extent_inserted);
 331         if (ret) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         }
 335
 336         if (isize > actual_end)
 337                 inline_len = min_t(u64, isize, actual_end);
 338         ret = insert_inline_extent(trans, path, extent_inserted,
 339                                    root, inode, start,
 340                                    inline_len, compressed_size,
 341                                    compress_type, compressed_pages);
 342         if (ret && ret != -ENOSPC) {
 343                 btrfs_abort_transaction(trans, ret);
 344                 goto out;
 345         } else if (ret == -ENOSPC) {
 346                 ret = 1;
 347                 goto out;
 348         }
 349
 350         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 351         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 352 out:
 353         /*
 354          * Don't forget to free the reserved space, as for inlined extent
 355          * it won't count as data extent, free them directly here.
 356          * And at reserve time, it's always aligned to page size, so
 357          * just free one page here.
 358          */
 359         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 360         btrfs_free_path(path);
 361         btrfs_end_transaction(trans);
 362         return ret;
 363 }
 364
 365 struct async_extent {
 366         u64 start;
 367         u64 ram_size;
 368         u64 compressed_size;
 369         struct page **pages;
 370         unsigned long nr_pages;
 371         int compress_type;
 372         struct list_head list;
 373 };
 374
 375 struct async_cow {
 376         struct inode *inode;
 377         struct btrfs_root *root;
 378         struct page *locked_page;
 379         u64 start;
 380         u64 end;
 381         unsigned int write_flags;
 382         struct list_head extents;
 383         struct btrfs_work work;
 384 };
 385
 386 static noinline int add_async_extent(struct async_cow *cow,
 387                                      u64 start, u64 ram_size,
 388                                      u64 compressed_size,
 389                                      struct page **pages,
 390                                      unsigned long nr_pages,
 391                                      int compress_type)
 392 {
 393         struct async_extent *async_extent;
 394
 395         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 396         BUG_ON(!async_extent); /* -ENOMEM */
 397         async_extent->start = start;
 398         async_extent->ram_size = ram_size;
 399         async_extent->compressed_size = compressed_size;
 400         async_extent->pages = pages;
 401         async_extent->nr_pages = nr_pages;
 402         async_extent->compress_type = compress_type;
 403         list_add_tail(&async_extent->list, &cow->extents);
 404         return 0;
 405 }
 406
 407 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 408 {
 409         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 410
 411         /* force compress */
 412         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 413                 return 1;
 414         /* defrag ioctl */
 415         if (BTRFS_I(inode)->defrag_compress)
 416                 return 1;
 417         /* bad compression ratios */
 418         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 419                 return 0;
 420         if (btrfs_test_opt(fs_info, COMPRESS) ||
 421             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 422             BTRFS_I(inode)->prop_compress)
 423                 return btrfs_compress_heuristic(inode, start, end);
 424         return 0;
 425 }
 426
 427 static inline void inode_should_defrag(struct btrfs_inode *inode,
 428                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 429 {
 430         /* If this is a small write inside eof, kick off a defrag */
 431         if (num_bytes < small_write &&
 432             (start > 0 || end + 1 < inode->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434 }
 435
 436 /*
 437  * we create compressed extents in two phases.  The first
 438  * phase compresses a range of pages that have already been
 439  * locked (both pages and state bits are locked).
 440  *
 441  * This is done inside an ordered work queue, and the compression
 442  * is spread across many cpus.  The actual IO submission is step
 443  * two, and the ordered work queue takes care of making sure that
 444  * happens in the same order things were put onto the queue by
 445  * writepages and friends.
 446  *
 447  * If this code finds it can't get good compression, it puts an
 448  * entry onto the work queue to write the uncompressed bytes.  This
 449  * makes sure that both compressed inodes and uncompressed inodes
 450  * are written in the same order that the flusher thread sent them
 451  * down.
 452  */
 453 static noinline void compress_file_range(struct inode *inode,
 454                                         struct page *locked_page,
 455                                         u64 start, u64 end,
 456                                         struct async_cow *async_cow,
 457                                         int *num_added)
 458 {
 459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 460         u64 blocksize = fs_info->sectorsize;
 461         u64 actual_end;
 462         u64 isize = i_size_read(inode);
 463         int ret = 0;
 464         struct page **pages = NULL;
 465         unsigned long nr_pages;
 466         unsigned long total_compressed = 0;
 467         unsigned long total_in = 0;
 468         int i;
 469         int will_compress;
 470         int compress_type = fs_info->compress_type;
 471         int redirty = 0;
 472
 473         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 474                         SZ_16K);
 475
 476         actual_end = min_t(u64, isize, end + 1);
 477 again:
 478         will_compress = 0;
 479         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 480         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 481         nr_pages = min_t(unsigned long, nr_pages,
 482                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 483
 484         /*
 485          * we don't want to send crud past the end of i_size through
 486          * compression, that's just a waste of CPU time.  So, if the
 487          * end of the file is before the start of our current
 488          * requested range of bytes, we bail out to the uncompressed
 489          * cleanup code that can deal with all of this.
 490          *
 491          * It isn't really the fastest way to fix things, but this is a
 492          * very uncommon corner.
 493          */
 494         if (actual_end <= start)
 495                 goto cleanup_and_bail_uncompressed;
 496
 497         total_compressed = actual_end - start;
 498
 499         /*
 500          * skip compression for a small file range(<=blocksize) that
 501          * isn't an inline extent, since it doesn't save disk space at all.
 502          */
 503         if (total_compressed <= blocksize &&
 504            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 505                 goto cleanup_and_bail_uncompressed;
 506
 507         total_compressed = min_t(unsigned long, total_compressed,
 508                         BTRFS_MAX_UNCOMPRESSED);
 509         total_in = 0;
 510         ret = 0;
 511
 512         /*
 513          * we do compression for mount -o compress and when the
 514          * inode has not been flagged as nocompress.  This flag can
 515          * change at any time if we discover bad compression ratios.
 516          */
 517         if (inode_need_compress(inode, start, end)) {
 518                 WARN_ON(pages);
 519                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 520                 if (!pages) {
 521                         /* just bail out to the uncompressed code */
 522                         goto cont;
 523                 }
 524
 525                 if (BTRFS_I(inode)->defrag_compress)
 526                         compress_type = BTRFS_I(inode)->defrag_compress;
 527                 else if (BTRFS_I(inode)->prop_compress)
 528                         compress_type = BTRFS_I(inode)->prop_compress;
 529
 530                 /*
 531                  * we need to call clear_page_dirty_for_io on each
 532                  * page in the range.  Otherwise applications with the file
 533                  * mmap'd can wander in and change the page contents while
 534                  * we are compressing them.
 535                  *
 536                  * If the compression fails for any reason, we set the pages
 537                  * dirty again later on.
 538                  *
 539                  * Note that the remaining part is redirtied, the start pointer
 540                  * has moved, the end is the original one.
 541                  */
 542                 if (!redirty) {
 543                         extent_range_clear_dirty_for_io(inode, start, end);
 544                         redirty = 1;
 545                 }
 546
 547                 /* Compression level is applied here and only here */
 548                 ret = btrfs_compress_pages(
 549                         compress_type | (fs_info->compress_level << 4),
 550                                            inode->i_mapping, start,
 551                                            pages,
 552                                            &nr_pages,
 553                                            &total_in,
 554                                            &total_compressed);
 555
 556                 if (!ret) {
 557                         unsigned long offset = total_compressed &
 558                                 (PAGE_SIZE - 1);
 559                         struct page *page = pages[nr_pages - 1];
 560                         char *kaddr;
 561
 562                         /* zero the tail end of the last page, we might be
 563                          * sending it down to disk
 564                          */
 565                         if (offset) {
 566                                 kaddr = kmap_atomic(page);
 567                                 memset(kaddr + offset, 0,
 568                                        PAGE_SIZE - offset);
 569                                 kunmap_atomic(kaddr);
 570                         }
 571                         will_compress = 1;
 572                 }
 573         }
 574 cont:
 575         if (start == 0) {
 576                 /* lets try to make an inline extent */
 577                 if (ret || total_in < actual_end) {
 578                         /* we didn't compress the entire range, try
 579                          * to make an uncompressed inline extent.
 580                          */
 581                         ret = cow_file_range_inline(inode, start, end, 0,
 582                                                     BTRFS_COMPRESS_NONE, NULL);
 583                 } else {
 584                         /* try making a compressed inline extent */
 585                         ret = cow_file_range_inline(inode, start, end,
 586                                                     total_compressed,
 587                                                     compress_type, pages);
 588                 }
 589                 if (ret <= 0) {
 590                         unsigned long clear_flags = EXTENT_DELALLOC |
 591                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 592                                 EXTENT_DO_ACCOUNTING;
 593                         unsigned long page_error_op;
 594
 595                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 596
 597                         /*
 598                          * inline extent creation worked or returned error,
 599                          * we don't need to create any more async work items.
 600                          * Unlock and free up our temp pages.
 601                          *
 602                          * We use DO_ACCOUNTING here because we need the
 603                          * delalloc_release_metadata to be done _after_ we drop
 604                          * our outstanding extent for clearing delalloc for this
 605                          * range.
 606                          */
 607                         extent_clear_unlock_delalloc(inode, start, end, end,
 608                                                      NULL, clear_flags,
 609                                                      PAGE_UNLOCK |
 610                                                      PAGE_CLEAR_DIRTY |
 611                                                      PAGE_SET_WRITEBACK |
 612                                                      page_error_op |
 613                                                      PAGE_END_WRITEBACK);
 614                         goto free_pages_out;
 615                 }
 616         }
 617
 618         if (will_compress) {
 619                 /*
 620                  * we aren't doing an inline extent round the compressed size
 621                  * up to a block size boundary so the allocator does sane
 622                  * things
 623                  */
 624                 total_compressed = ALIGN(total_compressed, blocksize);
 625
 626                 /*
 627                  * one last check to make sure the compression is really a
 628                  * win, compare the page count read with the blocks on disk,
 629                  * compression must free at least one sector size
 630                  */
 631                 total_in = ALIGN(total_in, PAGE_SIZE);
 632                 if (total_compressed + blocksize <= total_in) {
 633                         *num_added += 1;
 634
 635                         /*
 636                          * The async work queues will take care of doing actual
 637                          * allocation on disk for these compressed pages, and
 638                          * will submit them to the elevator.
 639                          */
 640                         add_async_extent(async_cow, start, total_in,
 641                                         total_compressed, pages, nr_pages,
 642                                         compress_type);
 643
 644                         if (start + total_in < end) {
 645                                 start += total_in;
 646                                 pages = NULL;
 647                                 cond_resched();
 648                                 goto again;
 649                         }
 650                         return;
 651                 }
 652         }
 653         if (pages) {
 654                 /*
 655                  * the compression code ran but failed to make things smaller,
 656                  * free any pages it allocated and our page pointer array
 657                  */
 658                 for (i = 0; i < nr_pages; i++) {
 659                         WARN_ON(pages[i]->mapping);
 660                         put_page(pages[i]);
 661                 }
 662                 kfree(pages);
 663                 pages = NULL;
 664                 total_compressed = 0;
 665                 nr_pages = 0;
 666
 667                 /* flag the file so we don't compress in the future */
 668                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 669                     !(BTRFS_I(inode)->prop_compress)) {
 670                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 671                 }
 672         }
 673 cleanup_and_bail_uncompressed:
 674         /*
 675          * No compression, but we still need to write the pages in the file
 676          * we've been given so far.  redirty the locked page if it corresponds
 677          * to our extent and set things up for the async work queue to run
 678          * cow_file_range to do the normal delalloc dance.
 679          */
 680         if (page_offset(locked_page) >= start &&
 681             page_offset(locked_page) <= end)
 682                 __set_page_dirty_nobuffers(locked_page);
 683                 /* unlocked later on in the async handlers */
 684
 685         if (redirty)
 686                 extent_range_redirty_for_io(inode, start, end);
 687         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 688                          BTRFS_COMPRESS_NONE);
 689         *num_added += 1;
 690
 691         return;
 692
 693 free_pages_out:
 694         for (i = 0; i < nr_pages; i++) {
 695                 WARN_ON(pages[i]->mapping);
 696                 put_page(pages[i]);
 697         }
 698         kfree(pages);
 699 }
 700
 701 static void free_async_extent_pages(struct async_extent *async_extent)
 702 {
 703         int i;
 704
 705         if (!async_extent->pages)
 706                 return;
 707
 708         for (i = 0; i < async_extent->nr_pages; i++) {
 709                 WARN_ON(async_extent->pages[i]->mapping);
 710                 put_page(async_extent->pages[i]);
 711         }
 712         kfree(async_extent->pages);
 713         async_extent->nr_pages = 0;
 714         async_extent->pages = NULL;
 715 }
 716
 717 /*
 718  * phase two of compressed writeback.  This is the ordered portion
 719  * of the code, which only gets called in the order the work was
 720  * queued.  We walk all the async extents created by compress_file_range
 721  * and send them down to the disk.
 722  */
 723 static noinline void submit_compressed_extents(struct inode *inode,
 724                                               struct async_cow *async_cow)
 725 {
 726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 727         struct async_extent *async_extent;
 728         u64 alloc_hint = 0;
 729         struct btrfs_key ins;
 730         struct extent_map *em;
 731         struct btrfs_root *root = BTRFS_I(inode)->root;
 732         struct extent_io_tree *io_tree;
 733         int ret = 0;
 734
 735 again:
 736         while (!list_empty(&async_cow->extents)) {
 737                 async_extent = list_entry(async_cow->extents.next,
 738                                           struct async_extent, list);
 739                 list_del(&async_extent->list);
 740
 741                 io_tree = &BTRFS_I(inode)->io_tree;
 742
 743 retry:
 744                 /* did the compression code fall back to uncompressed IO? */
 745                 if (!async_extent->pages) {
 746                         int page_started = 0;
 747                         unsigned long nr_written = 0;
 748
 749                         lock_extent(io_tree, async_extent->start,
 750                                          async_extent->start +
 751                                          async_extent->ram_size - 1);
 752
 753                         /* allocate blocks */
 754                         ret = cow_file_range(inode, async_cow->locked_page,
 755                                              async_extent->start,
 756                                              async_extent->start +
 757                                              async_extent->ram_size - 1,
 758                                              async_extent->start +
 759                                              async_extent->ram_size - 1,
 760                                              &page_started, &nr_written, 0,
 761                                              NULL);
 762
 763                         /* JDM XXX */
 764
 765                         /*
 766                          * if page_started, cow_file_range inserted an
 767                          * inline extent and took care of all the unlocking
 768                          * and IO for us.  Otherwise, we need to submit
 769                          * all those pages down to the drive.
 770                          */
 771                         if (!page_started && !ret)
 772                                 extent_write_locked_range(inode,
 773                                                   async_extent->start,
 774                                                   async_extent->start +
 775                                                   async_extent->ram_size - 1,
 776                                                   WB_SYNC_ALL);
 777                         else if (ret)
 778                                 unlock_page(async_cow->locked_page);
 779                         kfree(async_extent);
 780                         cond_resched();
 781                         continue;
 782                 }
 783
 784                 lock_extent(io_tree, async_extent->start,
 785                             async_extent->start + async_extent->ram_size - 1);
 786
 787                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 788                                            async_extent->compressed_size,
 789                                            async_extent->compressed_size,
 790                                            0, alloc_hint, &ins, 1, 1);
 791                 if (ret) {
 792                         free_async_extent_pages(async_extent);
 793
 794                         if (ret == -ENOSPC) {
 795                                 unlock_extent(io_tree, async_extent->start,
 796                                               async_extent->start +
 797                                               async_extent->ram_size - 1);
 798
 799                                 /*
 800                                  * we need to redirty the pages if we decide to
 801                                  * fallback to uncompressed IO, otherwise we
 802                                  * will not submit these pages down to lower
 803                                  * layers.
 804                                  */
 805                                 extent_range_redirty_for_io(inode,
 806                                                 async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1);
 809
 810                                 goto retry;
 811                         }
 812                         goto out_free;
 813                 }
 814                 /*
 815                  * here we're doing allocation and writeback of the
 816                  * compressed pages
 817                  */
 818                 em = create_io_em(inode, async_extent->start,
 819                                   async_extent->ram_size, /* len */
 820                                   async_extent->start, /* orig_start */
 821                                   ins.objectid, /* block_start */
 822                                   ins.offset, /* block_len */
 823                                   ins.offset, /* orig_block_len */
 824                                   async_extent->ram_size, /* ram_bytes */
 825                                   async_extent->compress_type,
 826                                   BTRFS_ORDERED_COMPRESSED);
 827                 if (IS_ERR(em))
 828                         /* ret value is not necessary due to void function */
 829                         goto out_free_reserve;
 830                 free_extent_map(em);
 831
 832                 ret = btrfs_add_ordered_extent_compress(inode,
 833                                                 async_extent->start,
 834                                                 ins.objectid,
 835                                                 async_extent->ram_size,
 836                                                 ins.offset,
 837                                                 BTRFS_ORDERED_COMPRESSED,
 838                                                 async_extent->compress_type);
 839                 if (ret) {
 840                         btrfs_drop_extent_cache(BTRFS_I(inode),
 841                                                 async_extent->start,
 842                                                 async_extent->start +
 843                                                 async_extent->ram_size - 1, 0);
 844                         goto out_free_reserve;
 845                 }
 846                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 847
 848                 /*
 849                  * clear dirty, set writeback and unlock the pages.
 850                  */
 851                 extent_clear_unlock_delalloc(inode, async_extent->start,
 852                                 async_extent->start +
 853                                 async_extent->ram_size - 1,
 854                                 async_extent->start +
 855                                 async_extent->ram_size - 1,
 856                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 857                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 858                                 PAGE_SET_WRITEBACK);
 859                 if (btrfs_submit_compressed_write(inode,
 860                                     async_extent->start,
 861                                     async_extent->ram_size,
 862                                     ins.objectid,
 863                                     ins.offset, async_extent->pages,
 864                                     async_extent->nr_pages,
 865                                     async_cow->write_flags)) {
 866                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 867                         struct page *p = async_extent->pages[0];
 868                         const u64 start = async_extent->start;
 869                         const u64 end = start + async_extent->ram_size - 1;
 870
 871                         p->mapping = inode->i_mapping;
 872                         tree->ops->writepage_end_io_hook(p, start, end,
 873                                                          NULL, 0);
 874                         p->mapping = NULL;
 875                         extent_clear_unlock_delalloc(inode, start, end, end,
 876                                                      NULL, 0,
 877                                                      PAGE_END_WRITEBACK |
 878                                                      PAGE_SET_ERROR);
 879                         free_async_extent_pages(async_extent);
 880                 }
 881                 alloc_hint = ins.objectid + ins.offset;
 882                 kfree(async_extent);
 883                 cond_resched();
 884         }
 885         return;
 886 out_free_reserve:
 887         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 888         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 889 out_free:
 890         extent_clear_unlock_delalloc(inode, async_extent->start,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      async_extent->start +
 894                                      async_extent->ram_size - 1,
 895                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 896                                      EXTENT_DELALLOC_NEW |
 897                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 898                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 899                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 900                                      PAGE_SET_ERROR);
 901         free_async_extent_pages(async_extent);
 902         kfree(async_extent);
 903         goto again;
 904 }
 905
 906 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 907                                       u64 num_bytes)
 908 {
 909         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 910         struct extent_map *em;
 911         u64 alloc_hint = 0;
 912
 913         read_lock(&em_tree->lock);
 914         em = search_extent_mapping(em_tree, start, num_bytes);
 915         if (em) {
 916                 /*
 917                  * if block start isn't an actual block number then find the
 918                  * first block in this inode and use that as a hint.  If that
 919                  * block is also bogus then just don't worry about it.
 920                  */
 921                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 922                         free_extent_map(em);
 923                         em = search_extent_mapping(em_tree, 0, 0);
 924                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 925                                 alloc_hint = em->block_start;
 926                         if (em)
 927                                 free_extent_map(em);
 928                 } else {
 929                         alloc_hint = em->block_start;
 930                         free_extent_map(em);
 931                 }
 932         }
 933         read_unlock(&em_tree->lock);
 934
 935         return alloc_hint;
 936 }
 937
 938 /*
 939  * when extent_io.c finds a delayed allocation range in the file,
 940  * the call backs end up in this code.  The basic idea is to
 941  * allocate extents on disk for the range, and create ordered data structs
 942  * in ram to track those extents.
 943  *
 944  * locked_page is the page that writepage had locked already.  We use
 945  * it to make sure we don't do extra locks or unlocks.
 946  *
 947  * *page_started is set to one if we unlock locked_page and do everything
 948  * required to start IO on it.  It may be clean and already done with
 949  * IO when we return.
 950  */
 951 static noinline int cow_file_range(struct inode *inode,
 952                                    struct page *locked_page,
 953                                    u64 start, u64 end, u64 delalloc_end,
 954                                    int *page_started, unsigned long *nr_written,
 955                                    int unlock, struct btrfs_dedupe_hash *hash)
 956 {
 957         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 958         struct btrfs_root *root = BTRFS_I(inode)->root;
 959         u64 alloc_hint = 0;
 960         u64 num_bytes;
 961         unsigned long ram_size;
 962         u64 cur_alloc_size = 0;
 963         u64 blocksize = fs_info->sectorsize;
 964         struct btrfs_key ins;
 965         struct extent_map *em;
 966         unsigned clear_bits;
 967         unsigned long page_ops;
 968         bool extent_reserved = false;
 969         int ret = 0;
 970
 971         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 972                 WARN_ON_ONCE(1);
 973                 ret = -EINVAL;
 974                 goto out_unlock;
 975         }
 976
 977         num_bytes = ALIGN(end - start + 1, blocksize);
 978         num_bytes = max(blocksize,  num_bytes);
 979         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 980
 981         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 982
 983         if (start == 0) {
 984                 /* lets try to make an inline extent */
 985                 ret = cow_file_range_inline(inode, start, end, 0,
 986                                             BTRFS_COMPRESS_NONE, NULL);
 987                 if (ret == 0) {
 988                         /*
 989                          * We use DO_ACCOUNTING here because we need the
 990                          * delalloc_release_metadata to be run _after_ we drop
 991                          * our outstanding extent for clearing delalloc for this
 992                          * range.
 993                          */
 994                         extent_clear_unlock_delalloc(inode, start, end,
 995                                      delalloc_end, NULL,
 996                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 997                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 998                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 999                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1000                                      PAGE_END_WRITEBACK);
1001                         *nr_written = *nr_written +
1002                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1003                         *page_started = 1;
1004                         goto out;
1005                 } else if (ret < 0) {
1006                         goto out_unlock;
1007                 }
1008         }
1009
1010         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1011         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1012                         start + num_bytes - 1, 0);
1013
1014         while (num_bytes > 0) {
1015                 cur_alloc_size = num_bytes;
1016                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1017                                            fs_info->sectorsize, 0, alloc_hint,
1018                                            &ins, 1, 1);
1019                 if (ret < 0)
1020                         goto out_unlock;
1021                 cur_alloc_size = ins.offset;
1022                 extent_reserved = true;
1023
1024                 ram_size = ins.offset;
1025                 em = create_io_em(inode, start, ins.offset, /* len */
1026                                   start, /* orig_start */
1027                                   ins.objectid, /* block_start */
1028                                   ins.offset, /* block_len */
1029                                   ins.offset, /* orig_block_len */
1030                                   ram_size, /* ram_bytes */
1031                                   BTRFS_COMPRESS_NONE, /* compress_type */
1032                                   BTRFS_ORDERED_REGULAR /* type */);
1033                 if (IS_ERR(em))
1034                         goto out_reserve;
1035                 free_extent_map(em);
1036
1037                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1038                                                ram_size, cur_alloc_size, 0);
1039                 if (ret)
1040                         goto out_drop_extent_cache;
1041
1042                 if (root->root_key.objectid ==
1043                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1044                         ret = btrfs_reloc_clone_csums(inode, start,
1045                                                       cur_alloc_size);
1046                         /*
1047                          * Only drop cache here, and process as normal.
1048                          *
1049                          * We must not allow extent_clear_unlock_delalloc()
1050                          * at out_unlock label to free meta of this ordered
1051                          * extent, as its meta should be freed by
1052                          * btrfs_finish_ordered_io().
1053                          *
1054                          * So we must continue until @start is increased to
1055                          * skip current ordered extent.
1056                          */
1057                         if (ret)
1058                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1059                                                 start + ram_size - 1, 0);
1060                 }
1061
1062                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1063
1064                 /* we're not doing compressed IO, don't unlock the first
1065                  * page (which the caller expects to stay locked), don't
1066                  * clear any dirty bits and don't set any writeback bits
1067                  *
1068                  * Do set the Private2 bit so we know this page was properly
1069                  * setup for writepage
1070                  */
1071                 page_ops = unlock ? PAGE_UNLOCK : 0;
1072                 page_ops |= PAGE_SET_PRIVATE2;
1073
1074                 extent_clear_unlock_delalloc(inode, start,
1075                                              start + ram_size - 1,
1076                                              delalloc_end, locked_page,
1077                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1078                                              page_ops);
1079                 if (num_bytes < cur_alloc_size)
1080                         num_bytes = 0;
1081                 else
1082                         num_bytes -= cur_alloc_size;
1083                 alloc_hint = ins.objectid + ins.offset;
1084                 start += cur_alloc_size;
1085                 extent_reserved = false;
1086
1087                 /*
1088                  * btrfs_reloc_clone_csums() error, since start is increased
1089                  * extent_clear_unlock_delalloc() at out_unlock label won't
1090                  * free metadata of current ordered extent, we're OK to exit.
1091                  */
1092                 if (ret)
1093                         goto out_unlock;
1094         }
1095 out:
1096         return ret;
1097
1098 out_drop_extent_cache:
1099         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1100 out_reserve:
1101         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1102         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1103 out_unlock:
1104         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1105                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1106         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1107                 PAGE_END_WRITEBACK;
1108         /*
1109          * If we reserved an extent for our delalloc range (or a subrange) and
1110          * failed to create the respective ordered extent, then it means that
1111          * when we reserved the extent we decremented the extent's size from
1112          * the data space_info's bytes_may_use counter and incremented the
1113          * space_info's bytes_reserved counter by the same amount. We must make
1114          * sure extent_clear_unlock_delalloc() does not try to decrement again
1115          * the data space_info's bytes_may_use counter, therefore we do not pass
1116          * it the flag EXTENT_CLEAR_DATA_RESV.
1117          */
1118         if (extent_reserved) {
1119                 extent_clear_unlock_delalloc(inode, start,
1120                                              start + cur_alloc_size,
1121                                              start + cur_alloc_size,
1122                                              locked_page,
1123                                              clear_bits,
1124                                              page_ops);
1125                 start += cur_alloc_size;
1126                 if (start >= end)
1127                         goto out;
1128         }
1129         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1130                                      locked_page,
1131                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1132                                      page_ops);
1133         goto out;
1134 }
1135
1136 /*
1137  * work queue call back to started compression on a file and pages
1138  */
1139 static noinline void async_cow_start(struct btrfs_work *work)
1140 {
1141         struct async_cow *async_cow;
1142         int num_added = 0;
1143         async_cow = container_of(work, struct async_cow, work);
1144
1145         compress_file_range(async_cow->inode, async_cow->locked_page,
1146                             async_cow->start, async_cow->end, async_cow,
1147                             &num_added);
1148         if (num_added == 0) {
1149                 btrfs_add_delayed_iput(async_cow->inode);
1150                 async_cow->inode = NULL;
1151         }
1152 }
1153
1154 /*
1155  * work queue call back to submit previously compressed pages
1156  */
1157 static noinline void async_cow_submit(struct btrfs_work *work)
1158 {
1159         struct btrfs_fs_info *fs_info;
1160         struct async_cow *async_cow;
1161         struct btrfs_root *root;
1162         unsigned long nr_pages;
1163
1164         async_cow = container_of(work, struct async_cow, work);
1165
1166         root = async_cow->root;
1167         fs_info = root->fs_info;
1168         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1169                 PAGE_SHIFT;
1170
1171         /*
1172          * atomic_sub_return implies a barrier for waitqueue_active
1173          */
1174         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1175             5 * SZ_1M &&
1176             waitqueue_active(&fs_info->async_submit_wait))
1177                 wake_up(&fs_info->async_submit_wait);
1178
1179         if (async_cow->inode)
1180                 submit_compressed_extents(async_cow->inode, async_cow);
1181 }
1182
1183 static noinline void async_cow_free(struct btrfs_work *work)
1184 {
1185         struct async_cow *async_cow;
1186         async_cow = container_of(work, struct async_cow, work);
1187         if (async_cow->inode)
1188                 btrfs_add_delayed_iput(async_cow->inode);
1189         kfree(async_cow);
1190 }
1191
1192 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1193                                 u64 start, u64 end, int *page_started,
1194                                 unsigned long *nr_written,
1195                                 unsigned int write_flags)
1196 {
1197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1198         struct async_cow *async_cow;
1199         struct btrfs_root *root = BTRFS_I(inode)->root;
1200         unsigned long nr_pages;
1201         u64 cur_end;
1202
1203         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1204                          1, 0, NULL);
1205         while (start < end) {
1206                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1207                 BUG_ON(!async_cow); /* -ENOMEM */
1208                 async_cow->inode = igrab(inode);
1209                 async_cow->root = root;
1210                 async_cow->locked_page = locked_page;
1211                 async_cow->start = start;
1212                 async_cow->write_flags = write_flags;
1213
1214                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1215                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1216                         cur_end = end;
1217                 else
1218                         cur_end = min(end, start + SZ_512K - 1);
1219
1220                 async_cow->end = cur_end;
1221                 INIT_LIST_HEAD(&async_cow->extents);
1222
1223                 btrfs_init_work(&async_cow->work,
1224                                 btrfs_delalloc_helper,
1225                                 async_cow_start, async_cow_submit,
1226                                 async_cow_free);
1227
1228                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1229                         PAGE_SHIFT;
1230                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1231
1232                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1233
1234                 *nr_written += nr_pages;
1235                 start = cur_end + 1;
1236         }
1237         *page_started = 1;
1238         return 0;
1239 }
1240
1241 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1242                                         u64 bytenr, u64 num_bytes)
1243 {
1244         int ret;
1245         struct btrfs_ordered_sum *sums;
1246         LIST_HEAD(list);
1247
1248         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1249                                        bytenr + num_bytes - 1, &list, 0);
1250         if (ret == 0 && list_empty(&list))
1251                 return 0;
1252
1253         while (!list_empty(&list)) {
1254                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1255                 list_del(&sums->list);
1256                 kfree(sums);
1257         }
1258         if (ret < 0)
1259                 return ret;
1260         return 1;
1261 }
1262
1263 /*
1264  * when nowcow writeback call back.  This checks for snapshots or COW copies
1265  * of the extents that exist in the file, and COWs the file as required.
1266  *
1267  * If no cow copies or snapshots exist, we write directly to the existing
1268  * blocks on disk
1269  */
1270 static noinline int run_delalloc_nocow(struct inode *inode,
1271                                        struct page *locked_page,
1272                               u64 start, u64 end, int *page_started, int force,
1273                               unsigned long *nr_written)
1274 {
1275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1276         struct btrfs_root *root = BTRFS_I(inode)->root;
1277         struct extent_buffer *leaf;
1278         struct btrfs_path *path;
1279         struct btrfs_file_extent_item *fi;
1280         struct btrfs_key found_key;
1281         struct extent_map *em;
1282         u64 cow_start;
1283         u64 cur_offset;
1284         u64 extent_end;
1285         u64 extent_offset;
1286         u64 disk_bytenr;
1287         u64 num_bytes;
1288         u64 disk_num_bytes;
1289         u64 ram_bytes;
1290         int extent_type;
1291         int ret, err;
1292         int type;
1293         int nocow;
1294         int check_prev = 1;
1295         bool nolock;
1296         u64 ino = btrfs_ino(BTRFS_I(inode));
1297
1298         path = btrfs_alloc_path();
1299         if (!path) {
1300                 extent_clear_unlock_delalloc(inode, start, end, end,
1301                                              locked_page,
1302                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1303                                              EXTENT_DO_ACCOUNTING |
1304                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1305                                              PAGE_CLEAR_DIRTY |
1306                                              PAGE_SET_WRITEBACK |
1307                                              PAGE_END_WRITEBACK);
1308                 return -ENOMEM;
1309         }
1310
1311         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1312
1313         cow_start = (u64)-1;
1314         cur_offset = start;
1315         while (1) {
1316                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1317                                                cur_offset, 0);
1318                 if (ret < 0)
1319                         goto error;
1320                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1321                         leaf = path->nodes[0];
1322                         btrfs_item_key_to_cpu(leaf, &found_key,
1323                                               path->slots[0] - 1);
1324                         if (found_key.objectid == ino &&
1325                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1326                                 path->slots[0]--;
1327                 }
1328                 check_prev = 0;
1329 next_slot:
1330                 leaf = path->nodes[0];
1331                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1332                         ret = btrfs_next_leaf(root, path);
1333                         if (ret < 0) {
1334                                 if (cow_start != (u64)-1)
1335                                         cur_offset = cow_start;
1336                                 goto error;
1337                         }
1338                         if (ret > 0)
1339                                 break;
1340                         leaf = path->nodes[0];
1341                 }
1342
1343                 nocow = 0;
1344                 disk_bytenr = 0;
1345                 num_bytes = 0;
1346                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1347
1348                 if (found_key.objectid > ino)
1349                         break;
1350                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1351                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1352                         path->slots[0]++;
1353                         goto next_slot;
1354                 }
1355                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1356                     found_key.offset > end)
1357                         break;
1358
1359                 if (found_key.offset > cur_offset) {
1360                         extent_end = found_key.offset;
1361                         extent_type = 0;
1362                         goto out_check;
1363                 }
1364
1365                 fi = btrfs_item_ptr(leaf, path->slots[0],
1366                                     struct btrfs_file_extent_item);
1367                 extent_type = btrfs_file_extent_type(leaf, fi);
1368
1369                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1370                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1371                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1372                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1373                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1374                         extent_end = found_key.offset +
1375                                 btrfs_file_extent_num_bytes(leaf, fi);
1376                         disk_num_bytes =
1377                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1378                         if (extent_end <= start) {
1379                                 path->slots[0]++;
1380                                 goto next_slot;
1381                         }
1382                         if (disk_bytenr == 0)
1383                                 goto out_check;
1384                         if (btrfs_file_extent_compression(leaf, fi) ||
1385                             btrfs_file_extent_encryption(leaf, fi) ||
1386                             btrfs_file_extent_other_encoding(leaf, fi))
1387                                 goto out_check;
1388                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1389                                 goto out_check;
1390                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1391                                 goto out_check;
1392                         ret = btrfs_cross_ref_exist(root, ino,
1393                                                     found_key.offset -
1394                                                     extent_offset, disk_bytenr);
1395                         if (ret) {
1396                                 /*
1397                                  * ret could be -EIO if the above fails to read
1398                                  * metadata.
1399                                  */
1400                                 if (ret < 0) {
1401                                         if (cow_start != (u64)-1)
1402                                                 cur_offset = cow_start;
1403                                         goto error;
1404                                 }
1405
1406                                 WARN_ON_ONCE(nolock);
1407                                 goto out_check;
1408                         }
1409                         disk_bytenr += extent_offset;
1410                         disk_bytenr += cur_offset - found_key.offset;
1411                         num_bytes = min(end + 1, extent_end) - cur_offset;
1412                         /*
1413                          * if there are pending snapshots for this root,
1414                          * we fall into common COW way.
1415                          */
1416                         if (!nolock) {
1417                                 err = btrfs_start_write_no_snapshotting(root);
1418                                 if (!err)
1419                                         goto out_check;
1420                         }
1421                         /*
1422                          * force cow if csum exists in the range.
1423                          * this ensure that csum for a given extent are
1424                          * either valid or do not exist.
1425                          */
1426                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1427                                                   num_bytes);
1428                         if (ret) {
1429                                 if (!nolock)
1430                                         btrfs_end_write_no_snapshotting(root);
1431
1432                                 /*
1433                                  * ret could be -EIO if the above fails to read
1434                                  * metadata.
1435                                  */
1436                                 if (ret < 0) {
1437                                         if (cow_start != (u64)-1)
1438                                                 cur_offset = cow_start;
1439                                         goto error;
1440                                 }
1441                                 WARN_ON_ONCE(nolock);
1442                                 goto out_check;
1443                         }
1444                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1445                                 if (!nolock)
1446                                         btrfs_end_write_no_snapshotting(root);
1447                                 goto out_check;
1448                         }
1449                         nocow = 1;
1450                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1451                         extent_end = found_key.offset +
1452                                 btrfs_file_extent_inline_len(leaf,
1453                                                      path->slots[0], fi);
1454                         extent_end = ALIGN(extent_end,
1455                                            fs_info->sectorsize);
1456                 } else {
1457                         BUG_ON(1);
1458                 }
1459 out_check:
1460                 if (extent_end <= start) {
1461                         path->slots[0]++;
1462                         if (!nolock && nocow)
1463                                 btrfs_end_write_no_snapshotting(root);
1464                         if (nocow)
1465                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1466                         goto next_slot;
1467                 }
1468                 if (!nocow) {
1469                         if (cow_start == (u64)-1)
1470                                 cow_start = cur_offset;
1471                         cur_offset = extent_end;
1472                         if (cur_offset > end)
1473                                 break;
1474                         path->slots[0]++;
1475                         goto next_slot;
1476                 }
1477
1478                 btrfs_release_path(path);
1479                 if (cow_start != (u64)-1) {
1480                         ret = cow_file_range(inode, locked_page,
1481                                              cow_start, found_key.offset - 1,
1482                                              end, page_started, nr_written, 1,
1483                                              NULL);
1484                         if (ret) {
1485                                 if (!nolock && nocow)
1486                                         btrfs_end_write_no_snapshotting(root);
1487                                 if (nocow)
1488                                         btrfs_dec_nocow_writers(fs_info,
1489                                                                 disk_bytenr);
1490                                 goto error;
1491                         }
1492                         cow_start = (u64)-1;
1493                 }
1494
1495                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1496                         u64 orig_start = found_key.offset - extent_offset;
1497
1498                         em = create_io_em(inode, cur_offset, num_bytes,
1499                                           orig_start,
1500                                           disk_bytenr, /* block_start */
1501                                           num_bytes, /* block_len */
1502                                           disk_num_bytes, /* orig_block_len */
1503                                           ram_bytes, BTRFS_COMPRESS_NONE,
1504                                           BTRFS_ORDERED_PREALLOC);
1505                         if (IS_ERR(em)) {
1506                                 if (!nolock && nocow)
1507                                         btrfs_end_write_no_snapshotting(root);
1508                                 if (nocow)
1509                                         btrfs_dec_nocow_writers(fs_info,
1510                                                                 disk_bytenr);
1511                                 ret = PTR_ERR(em);
1512                                 goto error;
1513                         }
1514                         free_extent_map(em);
1515                 }
1516
1517                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1518                         type = BTRFS_ORDERED_PREALLOC;
1519                 } else {
1520                         type = BTRFS_ORDERED_NOCOW;
1521                 }
1522
1523                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1524                                                num_bytes, num_bytes, type);
1525                 if (nocow)
1526                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1527                 BUG_ON(ret); /* -ENOMEM */
1528
1529                 if (root->root_key.objectid ==
1530                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1531                         /*
1532                          * Error handled later, as we must prevent
1533                          * extent_clear_unlock_delalloc() in error handler
1534                          * from freeing metadata of created ordered extent.
1535                          */
1536                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1537                                                       num_bytes);
1538
1539                 extent_clear_unlock_delalloc(inode, cur_offset,
1540                                              cur_offset + num_bytes - 1, end,
1541                                              locked_page, EXTENT_LOCKED |
1542                                              EXTENT_DELALLOC |
1543                                              EXTENT_CLEAR_DATA_RESV,
1544                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1545
1546                 if (!nolock && nocow)
1547                         btrfs_end_write_no_snapshotting(root);
1548                 cur_offset = extent_end;
1549
1550                 /*
1551                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1552                  * handler, as metadata for created ordered extent will only
1553                  * be freed by btrfs_finish_ordered_io().
1554                  */
1555                 if (ret)
1556                         goto error;
1557                 if (cur_offset > end)
1558                         break;
1559         }
1560         btrfs_release_path(path);
1561
1562         if (cur_offset <= end && cow_start == (u64)-1) {
1563                 cow_start = cur_offset;
1564                 cur_offset = end;
1565         }
1566
1567         if (cow_start != (u64)-1) {
1568                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1569                                      page_started, nr_written, 1, NULL);
1570                 if (ret)
1571                         goto error;
1572         }
1573
1574 error:
1575         if (ret && cur_offset < end)
1576                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1577                                              locked_page, EXTENT_LOCKED |
1578                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1579                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1580                                              PAGE_CLEAR_DIRTY |
1581                                              PAGE_SET_WRITEBACK |
1582                                              PAGE_END_WRITEBACK);
1583         btrfs_free_path(path);
1584         return ret;
1585 }
1586
1587 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1588 {
1589
1590         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1591             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1592                 return 0;
1593
1594         /*
1595          * @defrag_bytes is a hint value, no spinlock held here,
1596          * if is not zero, it means the file is defragging.
1597          * Force cow if given extent needs to be defragged.
1598          */
1599         if (BTRFS_I(inode)->defrag_bytes &&
1600             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1601                            EXTENT_DEFRAG, 0, NULL))
1602                 return 1;
1603
1604         return 0;
1605 }
1606
1607 /*
1608  * extent_io.c call back to do delayed allocation processing
1609  */
1610 static int run_delalloc_range(void *private_data, struct page *locked_page,
1611                               u64 start, u64 end, int *page_started,
1612                               unsigned long *nr_written,
1613                               struct writeback_control *wbc)
1614 {
1615         struct inode *inode = private_data;
1616         int ret;
1617         int force_cow = need_force_cow(inode, start, end);
1618         unsigned int write_flags = wbc_to_write_flags(wbc);
1619
1620         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1621                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1622                                          page_started, 1, nr_written);
1623         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1624                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1625                                          page_started, 0, nr_written);
1626         } else if (!inode_need_compress(inode, start, end)) {
1627                 ret = cow_file_range(inode, locked_page, start, end, end,
1628                                       page_started, nr_written, 1, NULL);
1629         } else {
1630                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1631                         &BTRFS_I(inode)->runtime_flags);
1632                 ret = cow_file_range_async(inode, locked_page, start, end,
1633                                            page_started, nr_written,
1634                                            write_flags);
1635         }
1636         if (ret)
1637                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1638         return ret;
1639 }
1640
1641 static void btrfs_split_extent_hook(void *private_data,
1642                                     struct extent_state *orig, u64 split)
1643 {
1644         struct inode *inode = private_data;
1645         u64 size;
1646
1647         /* not delalloc, ignore it */
1648         if (!(orig->state & EXTENT_DELALLOC))
1649                 return;
1650
1651         size = orig->end - orig->start + 1;
1652         if (size > BTRFS_MAX_EXTENT_SIZE) {
1653                 u32 num_extents;
1654                 u64 new_size;
1655
1656                 /*
1657                  * See the explanation in btrfs_merge_extent_hook, the same
1658                  * applies here, just in reverse.
1659                  */
1660                 new_size = orig->end - split + 1;
1661                 num_extents = count_max_extents(new_size);
1662                 new_size = split - orig->start;
1663                 num_extents += count_max_extents(new_size);
1664                 if (count_max_extents(size) >= num_extents)
1665                         return;
1666         }
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 /*
1674  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1675  * extents so we can keep track of new extents that are just merged onto old
1676  * extents, such as when we are doing sequential writes, so we can properly
1677  * account for the metadata space we'll need.
1678  */
1679 static void btrfs_merge_extent_hook(void *private_data,
1680                                     struct extent_state *new,
1681                                     struct extent_state *other)
1682 {
1683         struct inode *inode = private_data;
1684         u64 new_size, old_size;
1685         u32 num_extents;
1686
1687         /* not delalloc, ignore it */
1688         if (!(other->state & EXTENT_DELALLOC))
1689                 return;
1690
1691         if (new->start > other->start)
1692                 new_size = new->end - other->start + 1;
1693         else
1694                 new_size = other->end - new->start + 1;
1695
1696         /* we're not bigger than the max, unreserve the space and go */
1697         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1698                 spin_lock(&BTRFS_I(inode)->lock);
1699                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1700                 spin_unlock(&BTRFS_I(inode)->lock);
1701                 return;
1702         }
1703
1704         /*
1705          * We have to add up either side to figure out how many extents were
1706          * accounted for before we merged into one big extent.  If the number of
1707          * extents we accounted for is <= the amount we need for the new range
1708          * then we can return, otherwise drop.  Think of it like this
1709          *
1710          * [ 4k][MAX_SIZE]
1711          *
1712          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1713          * need 2 outstanding extents, on one side we have 1 and the other side
1714          * we have 1 so they are == and we can return.  But in this case
1715          *
1716          * [MAX_SIZE+4k][MAX_SIZE+4k]
1717          *
1718          * Each range on their own accounts for 2 extents, but merged together
1719          * they are only 3 extents worth of accounting, so we need to drop in
1720          * this case.
1721          */
1722         old_size = other->end - other->start + 1;
1723         num_extents = count_max_extents(old_size);
1724         old_size = new->end - new->start + 1;
1725         num_extents += count_max_extents(old_size);
1726         if (count_max_extents(new_size) >= num_extents)
1727                 return;
1728
1729         spin_lock(&BTRFS_I(inode)->lock);
1730         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1731         spin_unlock(&BTRFS_I(inode)->lock);
1732 }
1733
1734 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1735                                       struct inode *inode)
1736 {
1737         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1738
1739         spin_lock(&root->delalloc_lock);
1740         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1741                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1742                               &root->delalloc_inodes);
1743                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1744                         &BTRFS_I(inode)->runtime_flags);
1745                 root->nr_delalloc_inodes++;
1746                 if (root->nr_delalloc_inodes == 1) {
1747                         spin_lock(&fs_info->delalloc_root_lock);
1748                         BUG_ON(!list_empty(&root->delalloc_root));
1749                         list_add_tail(&root->delalloc_root,
1750                                       &fs_info->delalloc_roots);
1751                         spin_unlock(&fs_info->delalloc_root_lock);
1752                 }
1753         }
1754         spin_unlock(&root->delalloc_lock);
1755 }
1756
1757 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1758                                      struct btrfs_inode *inode)
1759 {
1760         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1761
1762         spin_lock(&root->delalloc_lock);
1763         if (!list_empty(&inode->delalloc_inodes)) {
1764                 list_del_init(&inode->delalloc_inodes);
1765                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1766                           &inode->runtime_flags);
1767                 root->nr_delalloc_inodes--;
1768                 if (!root->nr_delalloc_inodes) {
1769                         spin_lock(&fs_info->delalloc_root_lock);
1770                         BUG_ON(list_empty(&root->delalloc_root));
1771                         list_del_init(&root->delalloc_root);
1772                         spin_unlock(&fs_info->delalloc_root_lock);
1773                 }
1774         }
1775         spin_unlock(&root->delalloc_lock);
1776 }
1777
1778 /*
1779  * extent_io.c set_bit_hook, used to track delayed allocation
1780  * bytes in this file, and to maintain the list of inodes that
1781  * have pending delalloc work to be done.
1782  */
1783 static void btrfs_set_bit_hook(void *private_data,
1784                                struct extent_state *state, unsigned *bits)
1785 {
1786         struct inode *inode = private_data;
1787
1788         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1789
1790         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1791                 WARN_ON(1);
1792         /*
1793          * set_bit and clear bit hooks normally require _irqsave/restore
1794          * but in this case, we are only testing for the DELALLOC
1795          * bit, which is only set or cleared with irqs on
1796          */
1797         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1798                 struct btrfs_root *root = BTRFS_I(inode)->root;
1799                 u64 len = state->end + 1 - state->start;
1800                 u32 num_extents = count_max_extents(len);
1801                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1802
1803                 spin_lock(&BTRFS_I(inode)->lock);
1804                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1805                 spin_unlock(&BTRFS_I(inode)->lock);
1806
1807                 /* For sanity tests */
1808                 if (btrfs_is_testing(fs_info))
1809                         return;
1810
1811                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1812                                          fs_info->delalloc_batch);
1813                 spin_lock(&BTRFS_I(inode)->lock);
1814                 BTRFS_I(inode)->delalloc_bytes += len;
1815                 if (*bits & EXTENT_DEFRAG)
1816                         BTRFS_I(inode)->defrag_bytes += len;
1817                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1818                                          &BTRFS_I(inode)->runtime_flags))
1819                         btrfs_add_delalloc_inodes(root, inode);
1820                 spin_unlock(&BTRFS_I(inode)->lock);
1821         }
1822
1823         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1824             (*bits & EXTENT_DELALLOC_NEW)) {
1825                 spin_lock(&BTRFS_I(inode)->lock);
1826                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1827                         state->start;
1828                 spin_unlock(&BTRFS_I(inode)->lock);
1829         }
1830 }
1831
1832 /*
1833  * extent_io.c clear_bit_hook, see set_bit_hook for why
1834  */
1835 static void btrfs_clear_bit_hook(void *private_data,
1836                                  struct extent_state *state,
1837                                  unsigned *bits)
1838 {
1839         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1840         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1841         u64 len = state->end + 1 - state->start;
1842         u32 num_extents = count_max_extents(len);
1843
1844         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1845                 spin_lock(&inode->lock);
1846                 inode->defrag_bytes -= len;
1847                 spin_unlock(&inode->lock);
1848         }
1849
1850         /*
1851          * set_bit and clear bit hooks normally require _irqsave/restore
1852          * but in this case, we are only testing for the DELALLOC
1853          * bit, which is only set or cleared with irqs on
1854          */
1855         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1856                 struct btrfs_root *root = inode->root;
1857                 bool do_list = !btrfs_is_free_space_inode(inode);
1858
1859                 spin_lock(&inode->lock);
1860                 btrfs_mod_outstanding_extents(inode, -num_extents);
1861                 spin_unlock(&inode->lock);
1862
1863                 /*
1864                  * We don't reserve metadata space for space cache inodes so we
1865                  * don't need to call dellalloc_release_metadata if there is an
1866                  * error.
1867                  */
1868                 if (*bits & EXTENT_CLEAR_META_RESV &&
1869                     root != fs_info->tree_root)
1870                         btrfs_delalloc_release_metadata(inode, len);
1871
1872                 /* For sanity tests. */
1873                 if (btrfs_is_testing(fs_info))
1874                         return;
1875
1876                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1877                     do_list && !(state->state & EXTENT_NORESERVE) &&
1878                     (*bits & EXTENT_CLEAR_DATA_RESV))
1879                         btrfs_free_reserved_data_space_noquota(
1880                                         &inode->vfs_inode,
1881                                         state->start, len);
1882
1883                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1884                                          fs_info->delalloc_batch);
1885                 spin_lock(&inode->lock);
1886                 inode->delalloc_bytes -= len;
1887                 if (do_list && inode->delalloc_bytes == 0 &&
1888                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1889                                         &inode->runtime_flags))
1890                         btrfs_del_delalloc_inode(root, inode);
1891                 spin_unlock(&inode->lock);
1892         }
1893
1894         if ((state->state & EXTENT_DELALLOC_NEW) &&
1895             (*bits & EXTENT_DELALLOC_NEW)) {
1896                 spin_lock(&inode->lock);
1897                 ASSERT(inode->new_delalloc_bytes >= len);
1898                 inode->new_delalloc_bytes -= len;
1899                 spin_unlock(&inode->lock);
1900         }
1901 }
1902
1903 /*
1904  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1905  * we don't create bios that span stripes or chunks
1906  *
1907  * return 1 if page cannot be merged to bio
1908  * return 0 if page can be merged to bio
1909  * return error otherwise
1910  */
1911 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1912                          size_t size, struct bio *bio,
1913                          unsigned long bio_flags)
1914 {
1915         struct inode *inode = page->mapping->host;
1916         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1917         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1918         u64 length = 0;
1919         u64 map_length;
1920         int ret;
1921
1922         if (bio_flags & EXTENT_BIO_COMPRESSED)
1923                 return 0;
1924
1925         length = bio->bi_iter.bi_size;
1926         map_length = length;
1927         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1928                               NULL, 0);
1929         if (ret < 0)
1930                 return ret;
1931         if (map_length < length + size)
1932                 return 1;
1933         return 0;
1934 }
1935
1936 /*
1937  * in order to insert checksums into the metadata in large chunks,
1938  * we wait until bio submission time.   All the pages in the bio are
1939  * checksummed and sums are attached onto the ordered extent record.
1940  *
1941  * At IO completion time the cums attached on the ordered extent record
1942  * are inserted into the btree
1943  */
1944 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1945                                     u64 bio_offset)
1946 {
1947         struct inode *inode = private_data;
1948         blk_status_t ret = 0;
1949
1950         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1951         BUG_ON(ret); /* -ENOMEM */
1952         return 0;
1953 }
1954
1955 /*
1956  * in order to insert checksums into the metadata in large chunks,
1957  * we wait until bio submission time.   All the pages in the bio are
1958  * checksummed and sums are attached onto the ordered extent record.
1959  *
1960  * At IO completion time the cums attached on the ordered extent record
1961  * are inserted into the btree
1962  */
1963 static blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1964                           int mirror_num)
1965 {
1966         struct inode *inode = private_data;
1967         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1968         blk_status_t ret;
1969
1970         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1971         if (ret) {
1972                 bio->bi_status = ret;
1973                 bio_endio(bio);
1974         }
1975         return ret;
1976 }
1977
1978 /*
1979  * extent_io.c submission hook. This does the right thing for csum calculation
1980  * on write, or reading the csums from the tree before a read.
1981  *
1982  * Rules about async/sync submit,
1983  * a) read:                             sync submit
1984  *
1985  * b) write without checksum:           sync submit
1986  *
1987  * c) write with checksum:
1988  *    c-1) if bio is issued by fsync:   sync submit
1989  *         (sync_writers != 0)
1990  *
1991  *    c-2) if root is reloc root:       sync submit
1992  *         (only in case of buffered IO)
1993  *
1994  *    c-3) otherwise:                   async submit
1995  */
1996 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1997                                  int mirror_num, unsigned long bio_flags,
1998                                  u64 bio_offset)
1999 {
2000         struct inode *inode = private_data;
2001         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2002         struct btrfs_root *root = BTRFS_I(inode)->root;
2003         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2004         blk_status_t ret = 0;
2005         int skip_sum;
2006         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2007
2008         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2009
2010         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2011                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2012
2013         if (bio_op(bio) != REQ_OP_WRITE) {
2014                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2015                 if (ret)
2016                         goto out;
2017
2018                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2019                         ret = btrfs_submit_compressed_read(inode, bio,
2020                                                            mirror_num,
2021                                                            bio_flags);
2022                         goto out;
2023                 } else if (!skip_sum) {
2024                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2025                         if (ret)
2026                                 goto out;
2027                 }
2028                 goto mapit;
2029         } else if (async && !skip_sum) {
2030                 /* csum items have already been cloned */
2031                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2032                         goto mapit;
2033                 /* we're doing a write, do the async checksumming */
2034                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2035                                           bio_offset, inode,
2036                                           btrfs_submit_bio_start,
2037                                           btrfs_submit_bio_done);
2038                 goto out;
2039         } else if (!skip_sum) {
2040                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2041                 if (ret)
2042                         goto out;
2043         }
2044
2045 mapit:
2046         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2047
2048 out:
2049         if (ret) {
2050                 bio->bi_status = ret;
2051                 bio_endio(bio);
2052         }
2053         return ret;
2054 }
2055
2056 /*
2057  * given a list of ordered sums record them in the inode.  This happens
2058  * at IO completion time based on sums calculated at bio submission time.
2059  */
2060 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2061                              struct inode *inode, struct list_head *list)
2062 {
2063         struct btrfs_ordered_sum *sum;
2064         int ret;
2065
2066         list_for_each_entry(sum, list, list) {
2067                 trans->adding_csums = true;
2068                 ret = btrfs_csum_file_blocks(trans,
2069                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2070                 trans->adding_csums = false;
2071                 if (ret)
2072                         return ret;
2073         }
2074         return 0;
2075 }
2076
2077 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2078                               unsigned int extra_bits,
2079                               struct extent_state **cached_state, int dedupe)
2080 {
2081         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2082         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2083                                    extra_bits, cached_state);
2084 }
2085
2086 /* see btrfs_writepage_start_hook for details on why this is required */
2087 struct btrfs_writepage_fixup {
2088         struct page *page;
2089         struct btrfs_work work;
2090 };
2091
2092 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2093 {
2094         struct btrfs_writepage_fixup *fixup;
2095         struct btrfs_ordered_extent *ordered;
2096         struct extent_state *cached_state = NULL;
2097         struct extent_changeset *data_reserved = NULL;
2098         struct page *page;
2099         struct inode *inode;
2100         u64 page_start;
2101         u64 page_end;
2102         int ret;
2103
2104         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2105         page = fixup->page;
2106 again:
2107         lock_page(page);
2108         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2109                 ClearPageChecked(page);
2110                 goto out_page;
2111         }
2112
2113         inode = page->mapping->host;
2114         page_start = page_offset(page);
2115         page_end = page_offset(page) + PAGE_SIZE - 1;
2116
2117         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2118                          &cached_state);
2119
2120         /* already ordered? We're done */
2121         if (PagePrivate2(page))
2122                 goto out;
2123
2124         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2125                                         PAGE_SIZE);
2126         if (ordered) {
2127                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2128                                      page_end, &cached_state);
2129                 unlock_page(page);
2130                 btrfs_start_ordered_extent(inode, ordered, 1);
2131                 btrfs_put_ordered_extent(ordered);
2132                 goto again;
2133         }
2134
2135         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2136                                            PAGE_SIZE);
2137         if (ret) {
2138                 mapping_set_error(page->mapping, ret);
2139                 end_extent_writepage(page, ret, page_start, page_end);
2140                 ClearPageChecked(page);
2141                 goto out;
2142          }
2143
2144         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2145                                         &cached_state, 0);
2146         if (ret) {
2147                 mapping_set_error(page->mapping, ret);
2148                 end_extent_writepage(page, ret, page_start, page_end);
2149                 ClearPageChecked(page);
2150                 goto out;
2151         }
2152
2153         ClearPageChecked(page);
2154         set_page_dirty(page);
2155         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2156 out:
2157         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2158                              &cached_state);
2159 out_page:
2160         unlock_page(page);
2161         put_page(page);
2162         kfree(fixup);
2163         extent_changeset_free(data_reserved);
2164 }
2165
2166 /*
2167  * There are a few paths in the higher layers of the kernel that directly
2168  * set the page dirty bit without asking the filesystem if it is a
2169  * good idea.  This causes problems because we want to make sure COW
2170  * properly happens and the data=ordered rules are followed.
2171  *
2172  * In our case any range that doesn't have the ORDERED bit set
2173  * hasn't been properly setup for IO.  We kick off an async process
2174  * to fix it up.  The async helper will wait for ordered extents, set
2175  * the delalloc bit and make it safe to write the page.
2176  */
2177 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2178 {
2179         struct inode *inode = page->mapping->host;
2180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2181         struct btrfs_writepage_fixup *fixup;
2182
2183         /* this page is properly in the ordered list */
2184         if (TestClearPagePrivate2(page))
2185                 return 0;
2186
2187         if (PageChecked(page))
2188                 return -EAGAIN;
2189
2190         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2191         if (!fixup)
2192                 return -EAGAIN;
2193
2194         SetPageChecked(page);
2195         get_page(page);
2196         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2197                         btrfs_writepage_fixup_worker, NULL, NULL);
2198         fixup->page = page;
2199         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2200         return -EBUSY;
2201 }
2202
2203 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2204                                        struct inode *inode, u64 file_pos,
2205                                        u64 disk_bytenr, u64 disk_num_bytes,
2206                                        u64 num_bytes, u64 ram_bytes,
2207                                        u8 compression, u8 encryption,
2208                                        u16 other_encoding, int extent_type)
2209 {
2210         struct btrfs_root *root = BTRFS_I(inode)->root;
2211         struct btrfs_file_extent_item *fi;
2212         struct btrfs_path *path;
2213         struct extent_buffer *leaf;
2214         struct btrfs_key ins;
2215         u64 qg_released;
2216         int extent_inserted = 0;
2217         int ret;
2218
2219         path = btrfs_alloc_path();
2220         if (!path)
2221                 return -ENOMEM;
2222
2223         /*
2224          * we may be replacing one extent in the tree with another.
2225          * The new extent is pinned in the extent map, and we don't want
2226          * to drop it from the cache until it is completely in the btree.
2227          *
2228          * So, tell btrfs_drop_extents to leave this extent in the cache.
2229          * the caller is expected to unpin it and allow it to be merged
2230          * with the others.
2231          */
2232         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2233                                    file_pos + num_bytes, NULL, 0,
2234                                    1, sizeof(*fi), &extent_inserted);
2235         if (ret)
2236                 goto out;
2237
2238         if (!extent_inserted) {
2239                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2240                 ins.offset = file_pos;
2241                 ins.type = BTRFS_EXTENT_DATA_KEY;
2242
2243                 path->leave_spinning = 1;
2244                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2245                                               sizeof(*fi));
2246                 if (ret)
2247                         goto out;
2248         }
2249         leaf = path->nodes[0];
2250         fi = btrfs_item_ptr(leaf, path->slots[0],
2251                             struct btrfs_file_extent_item);
2252         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2253         btrfs_set_file_extent_type(leaf, fi, extent_type);
2254         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2255         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2256         btrfs_set_file_extent_offset(leaf, fi, 0);
2257         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2258         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2259         btrfs_set_file_extent_compression(leaf, fi, compression);
2260         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2261         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2262
2263         btrfs_mark_buffer_dirty(leaf);
2264         btrfs_release_path(path);
2265
2266         inode_add_bytes(inode, num_bytes);
2267
2268         ins.objectid = disk_bytenr;
2269         ins.offset = disk_num_bytes;
2270         ins.type = BTRFS_EXTENT_ITEM_KEY;
2271
2272         /*
2273          * Release the reserved range from inode dirty range map, as it is
2274          * already moved into delayed_ref_head
2275          */
2276         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2277         if (ret < 0)
2278                 goto out;
2279         qg_released = ret;
2280         ret = btrfs_alloc_reserved_file_extent(trans, root,
2281                                                btrfs_ino(BTRFS_I(inode)),
2282                                                file_pos, qg_released, &ins);
2283 out:
2284         btrfs_free_path(path);
2285
2286         return ret;
2287 }
2288
2289 /* snapshot-aware defrag */
2290 struct sa_defrag_extent_backref {
2291         struct rb_node node;
2292         struct old_sa_defrag_extent *old;
2293         u64 root_id;
2294         u64 inum;
2295         u64 file_pos;
2296         u64 extent_offset;
2297         u64 num_bytes;
2298         u64 generation;
2299 };
2300
2301 struct old_sa_defrag_extent {
2302         struct list_head list;
2303         struct new_sa_defrag_extent *new;
2304
2305         u64 extent_offset;
2306         u64 bytenr;
2307         u64 offset;
2308         u64 len;
2309         int count;
2310 };
2311
2312 struct new_sa_defrag_extent {
2313         struct rb_root root;
2314         struct list_head head;
2315         struct btrfs_path *path;
2316         struct inode *inode;
2317         u64 file_pos;
2318         u64 len;
2319         u64 bytenr;
2320         u64 disk_len;
2321         u8 compress_type;
2322 };
2323
2324 static int backref_comp(struct sa_defrag_extent_backref *b1,
2325                         struct sa_defrag_extent_backref *b2)
2326 {
2327         if (b1->root_id < b2->root_id)
2328                 return -1;
2329         else if (b1->root_id > b2->root_id)
2330                 return 1;
2331
2332         if (b1->inum < b2->inum)
2333                 return -1;
2334         else if (b1->inum > b2->inum)
2335                 return 1;
2336
2337         if (b1->file_pos < b2->file_pos)
2338                 return -1;
2339         else if (b1->file_pos > b2->file_pos)
2340                 return 1;
2341
2342         /*
2343          * [------------------------------] ===> (a range of space)
2344          *     |<--->|   |<---->| =============> (fs/file tree A)
2345          * |<---------------------------->| ===> (fs/file tree B)
2346          *
2347          * A range of space can refer to two file extents in one tree while
2348          * refer to only one file extent in another tree.
2349          *
2350          * So we may process a disk offset more than one time(two extents in A)
2351          * and locate at the same extent(one extent in B), then insert two same
2352          * backrefs(both refer to the extent in B).
2353          */
2354         return 0;
2355 }
2356
2357 static void backref_insert(struct rb_root *root,
2358                            struct sa_defrag_extent_backref *backref)
2359 {
2360         struct rb_node **p = &root->rb_node;
2361         struct rb_node *parent = NULL;
2362         struct sa_defrag_extent_backref *entry;
2363         int ret;
2364
2365         while (*p) {
2366                 parent = *p;
2367                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2368
2369                 ret = backref_comp(backref, entry);
2370                 if (ret < 0)
2371                         p = &(*p)->rb_left;
2372                 else
2373                         p = &(*p)->rb_right;
2374         }
2375
2376         rb_link_node(&backref->node, parent, p);
2377         rb_insert_color(&backref->node, root);
2378 }
2379
2380 /*
2381  * Note the backref might has changed, and in this case we just return 0.
2382  */
2383 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2384                                        void *ctx)
2385 {
2386         struct btrfs_file_extent_item *extent;
2387         struct old_sa_defrag_extent *old = ctx;
2388         struct new_sa_defrag_extent *new = old->new;
2389         struct btrfs_path *path = new->path;
2390         struct btrfs_key key;
2391         struct btrfs_root *root;
2392         struct sa_defrag_extent_backref *backref;
2393         struct extent_buffer *leaf;
2394         struct inode *inode = new->inode;
2395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2396         int slot;
2397         int ret;
2398         u64 extent_offset;
2399         u64 num_bytes;
2400
2401         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2402             inum == btrfs_ino(BTRFS_I(inode)))
2403                 return 0;
2404
2405         key.objectid = root_id;
2406         key.type = BTRFS_ROOT_ITEM_KEY;
2407         key.offset = (u64)-1;
2408
2409         root = btrfs_read_fs_root_no_name(fs_info, &key);
2410         if (IS_ERR(root)) {
2411                 if (PTR_ERR(root) == -ENOENT)
2412                         return 0;
2413                 WARN_ON(1);
2414                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2415                          inum, offset, root_id);
2416                 return PTR_ERR(root);
2417         }
2418
2419         key.objectid = inum;
2420         key.type = BTRFS_EXTENT_DATA_KEY;
2421         if (offset > (u64)-1 << 32)
2422                 key.offset = 0;
2423         else
2424                 key.offset = offset;
2425
2426         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2427         if (WARN_ON(ret < 0))
2428                 return ret;
2429         ret = 0;
2430
2431         while (1) {
2432                 cond_resched();
2433
2434                 leaf = path->nodes[0];
2435                 slot = path->slots[0];
2436
2437                 if (slot >= btrfs_header_nritems(leaf)) {
2438                         ret = btrfs_next_leaf(root, path);
2439                         if (ret < 0) {
2440                                 goto out;
2441                         } else if (ret > 0) {
2442                                 ret = 0;
2443                                 goto out;
2444                         }
2445                         continue;
2446                 }
2447
2448                 path->slots[0]++;
2449
2450                 btrfs_item_key_to_cpu(leaf, &key, slot);
2451
2452                 if (key.objectid > inum)
2453                         goto out;
2454
2455                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2456                         continue;
2457
2458                 extent = btrfs_item_ptr(leaf, slot,
2459                                         struct btrfs_file_extent_item);
2460
2461                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2462                         continue;
2463
2464                 /*
2465                  * 'offset' refers to the exact key.offset,
2466                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2467                  * (key.offset - extent_offset).
2468                  */
2469                 if (key.offset != offset)
2470                         continue;
2471
2472                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2473                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2474
2475                 if (extent_offset >= old->extent_offset + old->offset +
2476                     old->len || extent_offset + num_bytes <=
2477                     old->extent_offset + old->offset)
2478                         continue;
2479                 break;
2480         }
2481
2482         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2483         if (!backref) {
2484                 ret = -ENOENT;
2485                 goto out;
2486         }
2487
2488         backref->root_id = root_id;
2489         backref->inum = inum;
2490         backref->file_pos = offset;
2491         backref->num_bytes = num_bytes;
2492         backref->extent_offset = extent_offset;
2493         backref->generation = btrfs_file_extent_generation(leaf, extent);
2494         backref->old = old;
2495         backref_insert(&new->root, backref);
2496         old->count++;
2497 out:
2498         btrfs_release_path(path);
2499         WARN_ON(ret);
2500         return ret;
2501 }
2502
2503 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2504                                    struct new_sa_defrag_extent *new)
2505 {
2506         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2507         struct old_sa_defrag_extent *old, *tmp;
2508         int ret;
2509
2510         new->path = path;
2511
2512         list_for_each_entry_safe(old, tmp, &new->head, list) {
2513                 ret = iterate_inodes_from_logical(old->bytenr +
2514                                                   old->extent_offset, fs_info,
2515                                                   path, record_one_backref,
2516                                                   old, false);
2517                 if (ret < 0 && ret != -ENOENT)
2518                         return false;
2519
2520                 /* no backref to be processed for this extent */
2521                 if (!old->count) {
2522                         list_del(&old->list);
2523                         kfree(old);
2524                 }
2525         }
2526
2527         if (list_empty(&new->head))
2528                 return false;
2529
2530         return true;
2531 }
2532
2533 static int relink_is_mergable(struct extent_buffer *leaf,
2534                               struct btrfs_file_extent_item *fi,
2535                               struct new_sa_defrag_extent *new)
2536 {
2537         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2538                 return 0;
2539
2540         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2541                 return 0;
2542
2543         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2544                 return 0;
2545
2546         if (btrfs_file_extent_encryption(leaf, fi) ||
2547             btrfs_file_extent_other_encoding(leaf, fi))
2548                 return 0;
2549
2550         return 1;
2551 }
2552
2553 /*
2554  * Note the backref might has changed, and in this case we just return 0.
2555  */
2556 static noinline int relink_extent_backref(struct btrfs_path *path,
2557                                  struct sa_defrag_extent_backref *prev,
2558                                  struct sa_defrag_extent_backref *backref)
2559 {
2560         struct btrfs_file_extent_item *extent;
2561         struct btrfs_file_extent_item *item;
2562         struct btrfs_ordered_extent *ordered;
2563         struct btrfs_trans_handle *trans;
2564         struct btrfs_root *root;
2565         struct btrfs_key key;
2566         struct extent_buffer *leaf;
2567         struct old_sa_defrag_extent *old = backref->old;
2568         struct new_sa_defrag_extent *new = old->new;
2569         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2570         struct inode *inode;
2571         struct extent_state *cached = NULL;
2572         int ret = 0;
2573         u64 start;
2574         u64 len;
2575         u64 lock_start;
2576         u64 lock_end;
2577         bool merge = false;
2578         int index;
2579
2580         if (prev && prev->root_id == backref->root_id &&
2581             prev->inum == backref->inum &&
2582             prev->file_pos + prev->num_bytes == backref->file_pos)
2583                 merge = true;
2584
2585         /* step 1: get root */
2586         key.objectid = backref->root_id;
2587         key.type = BTRFS_ROOT_ITEM_KEY;
2588         key.offset = (u64)-1;
2589
2590         index = srcu_read_lock(&fs_info->subvol_srcu);
2591
2592         root = btrfs_read_fs_root_no_name(fs_info, &key);
2593         if (IS_ERR(root)) {
2594                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2595                 if (PTR_ERR(root) == -ENOENT)
2596                         return 0;
2597                 return PTR_ERR(root);
2598         }
2599
2600         if (btrfs_root_readonly(root)) {
2601                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2602                 return 0;
2603         }
2604
2605         /* step 2: get inode */
2606         key.objectid = backref->inum;
2607         key.type = BTRFS_INODE_ITEM_KEY;
2608         key.offset = 0;
2609
2610         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2611         if (IS_ERR(inode)) {
2612                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2613                 return 0;
2614         }
2615
2616         srcu_read_unlock(&fs_info->subvol_srcu, index);
2617
2618         /* step 3: relink backref */
2619         lock_start = backref->file_pos;
2620         lock_end = backref->file_pos + backref->num_bytes - 1;
2621         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2622                          &cached);
2623
2624         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2625         if (ordered) {
2626                 btrfs_put_ordered_extent(ordered);
2627                 goto out_unlock;
2628         }
2629
2630         trans = btrfs_join_transaction(root);
2631         if (IS_ERR(trans)) {
2632                 ret = PTR_ERR(trans);
2633                 goto out_unlock;
2634         }
2635
2636         key.objectid = backref->inum;
2637         key.type = BTRFS_EXTENT_DATA_KEY;
2638         key.offset = backref->file_pos;
2639
2640         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2641         if (ret < 0) {
2642                 goto out_free_path;
2643         } else if (ret > 0) {
2644                 ret = 0;
2645                 goto out_free_path;
2646         }
2647
2648         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2649                                 struct btrfs_file_extent_item);
2650
2651         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2652             backref->generation)
2653                 goto out_free_path;
2654
2655         btrfs_release_path(path);
2656
2657         start = backref->file_pos;
2658         if (backref->extent_offset < old->extent_offset + old->offset)
2659                 start += old->extent_offset + old->offset -
2660                          backref->extent_offset;
2661
2662         len = min(backref->extent_offset + backref->num_bytes,
2663                   old->extent_offset + old->offset + old->len);
2664         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2665
2666         ret = btrfs_drop_extents(trans, root, inode, start,
2667                                  start + len, 1);
2668         if (ret)
2669                 goto out_free_path;
2670 again:
2671         key.objectid = btrfs_ino(BTRFS_I(inode));
2672         key.type = BTRFS_EXTENT_DATA_KEY;
2673         key.offset = start;
2674
2675         path->leave_spinning = 1;
2676         if (merge) {
2677                 struct btrfs_file_extent_item *fi;
2678                 u64 extent_len;
2679                 struct btrfs_key found_key;
2680
2681                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2682                 if (ret < 0)
2683                         goto out_free_path;
2684
2685                 path->slots[0]--;
2686                 leaf = path->nodes[0];
2687                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2688
2689                 fi = btrfs_item_ptr(leaf, path->slots[0],
2690                                     struct btrfs_file_extent_item);
2691                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2692
2693                 if (extent_len + found_key.offset == start &&
2694                     relink_is_mergable(leaf, fi, new)) {
2695                         btrfs_set_file_extent_num_bytes(leaf, fi,
2696                                                         extent_len + len);
2697                         btrfs_mark_buffer_dirty(leaf);
2698                         inode_add_bytes(inode, len);
2699
2700                         ret = 1;
2701                         goto out_free_path;
2702                 } else {
2703                         merge = false;
2704                         btrfs_release_path(path);
2705                         goto again;
2706                 }
2707         }
2708
2709         ret = btrfs_insert_empty_item(trans, root, path, &key,
2710                                         sizeof(*extent));
2711         if (ret) {
2712                 btrfs_abort_transaction(trans, ret);
2713                 goto out_free_path;
2714         }
2715
2716         leaf = path->nodes[0];
2717         item = btrfs_item_ptr(leaf, path->slots[0],
2718                                 struct btrfs_file_extent_item);
2719         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2720         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2721         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2722         btrfs_set_file_extent_num_bytes(leaf, item, len);
2723         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2724         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2725         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2726         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2727         btrfs_set_file_extent_encryption(leaf, item, 0);
2728         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2729
2730         btrfs_mark_buffer_dirty(leaf);
2731         inode_add_bytes(inode, len);
2732         btrfs_release_path(path);
2733
2734         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2735                         new->disk_len, 0,
2736                         backref->root_id, backref->inum,
2737                         new->file_pos); /* start - extent_offset */
2738         if (ret) {
2739                 btrfs_abort_transaction(trans, ret);
2740                 goto out_free_path;
2741         }
2742
2743         ret = 1;
2744 out_free_path:
2745         btrfs_release_path(path);
2746         path->leave_spinning = 0;
2747         btrfs_end_transaction(trans);
2748 out_unlock:
2749         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2750                              &cached);
2751         iput(inode);
2752         return ret;
2753 }
2754
2755 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2756 {
2757         struct old_sa_defrag_extent *old, *tmp;
2758
2759         if (!new)
2760                 return;
2761
2762         list_for_each_entry_safe(old, tmp, &new->head, list) {
2763                 kfree(old);
2764         }
2765         kfree(new);
2766 }
2767
2768 static void relink_file_extents(struct new_sa_defrag_extent *new)
2769 {
2770         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2771         struct btrfs_path *path;
2772         struct sa_defrag_extent_backref *backref;
2773         struct sa_defrag_extent_backref *prev = NULL;
2774         struct inode *inode;
2775         struct rb_node *node;
2776         int ret;
2777
2778         inode = new->inode;
2779
2780         path = btrfs_alloc_path();
2781         if (!path)
2782                 return;
2783
2784         if (!record_extent_backrefs(path, new)) {
2785                 btrfs_free_path(path);
2786                 goto out;
2787         }
2788         btrfs_release_path(path);
2789
2790         while (1) {
2791                 node = rb_first(&new->root);
2792                 if (!node)
2793                         break;
2794                 rb_erase(node, &new->root);
2795
2796                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2797
2798                 ret = relink_extent_backref(path, prev, backref);
2799                 WARN_ON(ret < 0);
2800
2801                 kfree(prev);
2802
2803                 if (ret == 1)
2804                         prev = backref;
2805                 else
2806                         prev = NULL;
2807                 cond_resched();
2808         }
2809         kfree(prev);
2810
2811         btrfs_free_path(path);
2812 out:
2813         free_sa_defrag_extent(new);
2814
2815         atomic_dec(&fs_info->defrag_running);
2816         wake_up(&fs_info->transaction_wait);
2817 }
2818
2819 static struct new_sa_defrag_extent *
2820 record_old_file_extents(struct inode *inode,
2821                         struct btrfs_ordered_extent *ordered)
2822 {
2823         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_path *path;
2826         struct btrfs_key key;
2827         struct old_sa_defrag_extent *old;
2828         struct new_sa_defrag_extent *new;
2829         int ret;
2830
2831         new = kmalloc(sizeof(*new), GFP_NOFS);
2832         if (!new)
2833                 return NULL;
2834
2835         new->inode = inode;
2836         new->file_pos = ordered->file_offset;
2837         new->len = ordered->len;
2838         new->bytenr = ordered->start;
2839         new->disk_len = ordered->disk_len;
2840         new->compress_type = ordered->compress_type;
2841         new->root = RB_ROOT;
2842         INIT_LIST_HEAD(&new->head);
2843
2844         path = btrfs_alloc_path();
2845         if (!path)
2846                 goto out_kfree;
2847
2848         key.objectid = btrfs_ino(BTRFS_I(inode));
2849         key.type = BTRFS_EXTENT_DATA_KEY;
2850         key.offset = new->file_pos;
2851
2852         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2853         if (ret < 0)
2854                 goto out_free_path;
2855         if (ret > 0 && path->slots[0] > 0)
2856                 path->slots[0]--;
2857
2858         /* find out all the old extents for the file range */
2859         while (1) {
2860                 struct btrfs_file_extent_item *extent;
2861                 struct extent_buffer *l;
2862                 int slot;
2863                 u64 num_bytes;
2864                 u64 offset;
2865                 u64 end;
2866                 u64 disk_bytenr;
2867                 u64 extent_offset;
2868
2869                 l = path->nodes[0];
2870                 slot = path->slots[0];
2871
2872                 if (slot >= btrfs_header_nritems(l)) {
2873                         ret = btrfs_next_leaf(root, path);
2874                         if (ret < 0)
2875                                 goto out_free_path;
2876                         else if (ret > 0)
2877                                 break;
2878                         continue;
2879                 }
2880
2881                 btrfs_item_key_to_cpu(l, &key, slot);
2882
2883                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2884                         break;
2885                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2886                         break;
2887                 if (key.offset >= new->file_pos + new->len)
2888                         break;
2889
2890                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2891
2892                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2893                 if (key.offset + num_bytes < new->file_pos)
2894                         goto next;
2895
2896                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2897                 if (!disk_bytenr)
2898                         goto next;
2899
2900                 extent_offset = btrfs_file_extent_offset(l, extent);
2901
2902                 old = kmalloc(sizeof(*old), GFP_NOFS);
2903                 if (!old)
2904                         goto out_free_path;
2905
2906                 offset = max(new->file_pos, key.offset);
2907                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2908
2909                 old->bytenr = disk_bytenr;
2910                 old->extent_offset = extent_offset;
2911                 old->offset = offset - key.offset;
2912                 old->len = end - offset;
2913                 old->new = new;
2914                 old->count = 0;
2915                 list_add_tail(&old->list, &new->head);
2916 next:
2917                 path->slots[0]++;
2918                 cond_resched();
2919         }
2920
2921         btrfs_free_path(path);
2922         atomic_inc(&fs_info->defrag_running);
2923
2924         return new;
2925
2926 out_free_path:
2927         btrfs_free_path(path);
2928 out_kfree:
2929         free_sa_defrag_extent(new);
2930         return NULL;
2931 }
2932
2933 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2934                                          u64 start, u64 len)
2935 {
2936         struct btrfs_block_group_cache *cache;
2937
2938         cache = btrfs_lookup_block_group(fs_info, start);
2939         ASSERT(cache);
2940
2941         spin_lock(&cache->lock);
2942         cache->delalloc_bytes -= len;
2943         spin_unlock(&cache->lock);
2944
2945         btrfs_put_block_group(cache);
2946 }
2947
2948 /* as ordered data IO finishes, this gets called so we can finish
2949  * an ordered extent if the range of bytes in the file it covers are
2950  * fully written.
2951  */
2952 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2953 {
2954         struct inode *inode = ordered_extent->inode;
2955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2956         struct btrfs_root *root = BTRFS_I(inode)->root;
2957         struct btrfs_trans_handle *trans = NULL;
2958         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2959         struct extent_state *cached_state = NULL;
2960         struct new_sa_defrag_extent *new = NULL;
2961         int compress_type = 0;
2962         int ret = 0;
2963         u64 logical_len = ordered_extent->len;
2964         bool nolock;
2965         bool truncated = false;
2966         bool range_locked = false;
2967         bool clear_new_delalloc_bytes = false;
2968
2969         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2970             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2971             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2972                 clear_new_delalloc_bytes = true;
2973
2974         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2975
2976         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2977                 ret = -EIO;
2978                 goto out;
2979         }
2980
2981         btrfs_free_io_failure_record(BTRFS_I(inode),
2982                         ordered_extent->file_offset,
2983                         ordered_extent->file_offset +
2984                         ordered_extent->len - 1);
2985
2986         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2987                 truncated = true;
2988                 logical_len = ordered_extent->truncated_len;
2989                 /* Truncated the entire extent, don't bother adding */
2990                 if (!logical_len)
2991                         goto out;
2992         }
2993
2994         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2995                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2996
2997                 /*
2998                  * For mwrite(mmap + memset to write) case, we still reserve
2999                  * space for NOCOW range.
3000                  * As NOCOW won't cause a new delayed ref, just free the space
3001                  */
3002                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3003                                        ordered_extent->len);
3004                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3005                 if (nolock)
3006                         trans = btrfs_join_transaction_nolock(root);
3007                 else
3008                         trans = btrfs_join_transaction(root);
3009                 if (IS_ERR(trans)) {
3010                         ret = PTR_ERR(trans);
3011                         trans = NULL;
3012                         goto out;
3013                 }
3014                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3015                 ret = btrfs_update_inode_fallback(trans, root, inode);
3016                 if (ret) /* -ENOMEM or corruption */
3017                         btrfs_abort_transaction(trans, ret);
3018                 goto out;
3019         }
3020
3021         range_locked = true;
3022         lock_extent_bits(io_tree, ordered_extent->file_offset,
3023                          ordered_extent->file_offset + ordered_extent->len - 1,
3024                          &cached_state);
3025
3026         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3027                         ordered_extent->file_offset + ordered_extent->len - 1,
3028                         EXTENT_DEFRAG, 0, cached_state);
3029         if (ret) {
3030                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3031                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3032                         /* the inode is shared */
3033                         new = record_old_file_extents(inode, ordered_extent);
3034
3035                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3036                         ordered_extent->file_offset + ordered_extent->len - 1,
3037                         EXTENT_DEFRAG, 0, 0, &cached_state);
3038         }
3039
3040         if (nolock)
3041                 trans = btrfs_join_transaction_nolock(root);
3042         else
3043                 trans = btrfs_join_transaction(root);
3044         if (IS_ERR(trans)) {
3045                 ret = PTR_ERR(trans);
3046                 trans = NULL;
3047                 goto out;
3048         }
3049
3050         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3051
3052         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3053                 compress_type = ordered_extent->compress_type;
3054         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3055                 BUG_ON(compress_type);
3056                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3057                                        ordered_extent->len);
3058                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3059                                                 ordered_extent->file_offset,
3060                                                 ordered_extent->file_offset +
3061                                                 logical_len);
3062         } else {
3063                 BUG_ON(root == fs_info->tree_root);
3064                 ret = insert_reserved_file_extent(trans, inode,
3065                                                 ordered_extent->file_offset,
3066                                                 ordered_extent->start,
3067                                                 ordered_extent->disk_len,
3068                                                 logical_len, logical_len,
3069                      &n