fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_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);
 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, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_CACHE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_CACHE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_CACHE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 page_cache_release(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > PAGE_CACHE_SIZE ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_CACHE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_CACHE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it dosen't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_CACHE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_CACHE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         page_cache_release(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 page_cache_release(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 page_cache_release(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827
 828                 /*
 829                  * clear dirty, set writeback and unlock the pages.
 830                  */
 831                 extent_clear_unlock_delalloc(inode, async_extent->start,
 832                                 async_extent->start +
 833                                 async_extent->ram_size - 1,
 834                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 835                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 836                                 PAGE_SET_WRITEBACK);
 837                 ret = btrfs_submit_compressed_write(inode,
 838                                     async_extent->start,
 839                                     async_extent->ram_size,
 840                                     ins.objectid,
 841                                     ins.offset, async_extent->pages,
 842                                     async_extent->nr_pages);
 843                 if (ret) {
 844                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 845                         struct page *p = async_extent->pages[0];
 846                         const u64 start = async_extent->start;
 847                         const u64 end = start + async_extent->ram_size - 1;
 848
 849                         p->mapping = inode->i_mapping;
 850                         tree->ops->writepage_end_io_hook(p, start, end,
 851                                                          NULL, 0);
 852                         p->mapping = NULL;
 853                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 854                                                      PAGE_END_WRITEBACK |
 855                                                      PAGE_SET_ERROR);
 856                         free_async_extent_pages(async_extent);
 857                 }
 858                 alloc_hint = ins.objectid + ins.offset;
 859                 kfree(async_extent);
 860                 cond_resched();
 861         }
 862         return;
 863 out_free_reserve:
 864         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 865 out_free:
 866         extent_clear_unlock_delalloc(inode, async_extent->start,
 867                                      async_extent->start +
 868                                      async_extent->ram_size - 1,
 869                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 870                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 871                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 872                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 873                                      PAGE_SET_ERROR);
 874         free_async_extent_pages(async_extent);
 875         kfree(async_extent);
 876         goto again;
 877 }
 878
 879 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 880                                       u64 num_bytes)
 881 {
 882         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 883         struct extent_map *em;
 884         u64 alloc_hint = 0;
 885
 886         read_lock(&em_tree->lock);
 887         em = search_extent_mapping(em_tree, start, num_bytes);
 888         if (em) {
 889                 /*
 890                  * if block start isn't an actual block number then find the
 891                  * first block in this inode and use that as a hint.  If that
 892                  * block is also bogus then just don't worry about it.
 893                  */
 894                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 895                         free_extent_map(em);
 896                         em = search_extent_mapping(em_tree, 0, 0);
 897                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 898                                 alloc_hint = em->block_start;
 899                         if (em)
 900                                 free_extent_map(em);
 901                 } else {
 902                         alloc_hint = em->block_start;
 903                         free_extent_map(em);
 904                 }
 905         }
 906         read_unlock(&em_tree->lock);
 907
 908         return alloc_hint;
 909 }
 910
 911 /*
 912  * when extent_io.c finds a delayed allocation range in the file,
 913  * the call backs end up in this code.  The basic idea is to
 914  * allocate extents on disk for the range, and create ordered data structs
 915  * in ram to track those extents.
 916  *
 917  * locked_page is the page that writepage had locked already.  We use
 918  * it to make sure we don't do extra locks or unlocks.
 919  *
 920  * *page_started is set to one if we unlock locked_page and do everything
 921  * required to start IO on it.  It may be clean and already done with
 922  * IO when we return.
 923  */
 924 static noinline int cow_file_range(struct inode *inode,
 925                                    struct page *locked_page,
 926                                    u64 start, u64 end, int *page_started,
 927                                    unsigned long *nr_written,
 928                                    int unlock)
 929 {
 930         struct btrfs_root *root = BTRFS_I(inode)->root;
 931         u64 alloc_hint = 0;
 932         u64 num_bytes;
 933         unsigned long ram_size;
 934         u64 disk_num_bytes;
 935         u64 cur_alloc_size;
 936         u64 blocksize = root->sectorsize;
 937         struct btrfs_key ins;
 938         struct extent_map *em;
 939         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 940         int ret = 0;
 941
 942         if (btrfs_is_free_space_inode(inode)) {
 943                 WARN_ON_ONCE(1);
 944                 ret = -EINVAL;
 945                 goto out_unlock;
 946         }
 947
 948         num_bytes = ALIGN(end - start + 1, blocksize);
 949         num_bytes = max(blocksize,  num_bytes);
 950         disk_num_bytes = num_bytes;
 951
 952         /* if this is a small write inside eof, kick off defrag */
 953         if (num_bytes < SZ_64K &&
 954             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 955                 btrfs_add_inode_defrag(NULL, inode);
 956
 957         if (start == 0) {
 958                 /* lets try to make an inline extent */
 959                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 960                                             NULL);
 961                 if (ret == 0) {
 962                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 963                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 964                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 965                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 966                                      PAGE_END_WRITEBACK);
 967
 968                         *nr_written = *nr_written +
 969                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 970                         *page_started = 1;
 971                         goto out;
 972                 } else if (ret < 0) {
 973                         goto out_unlock;
 974                 }
 975         }
 976
 977         BUG_ON(disk_num_bytes >
 978                btrfs_super_total_bytes(root->fs_info->super_copy));
 979
 980         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 981         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 982
 983         while (disk_num_bytes > 0) {
 984                 unsigned long op;
 985
 986                 cur_alloc_size = disk_num_bytes;
 987                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 988                                            root->sectorsize, 0, alloc_hint,
 989                                            &ins, 1, 1);
 990                 if (ret < 0)
 991                         goto out_unlock;
 992
 993                 em = alloc_extent_map();
 994                 if (!em) {
 995                         ret = -ENOMEM;
 996                         goto out_reserve;
 997                 }
 998                 em->start = start;
 999                 em->orig_start = em->start;
1000                 ram_size = ins.offset;
1001                 em->len = ins.offset;
1002                 em->mod_start = em->start;
1003                 em->mod_len = em->len;
1004
1005                 em->block_start = ins.objectid;
1006                 em->block_len = ins.offset;
1007                 em->orig_block_len = ins.offset;
1008                 em->ram_bytes = ram_size;
1009                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1010                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1011                 em->generation = -1;
1012
1013                 while (1) {
1014                         write_lock(&em_tree->lock);
1015                         ret = add_extent_mapping(em_tree, em, 1);
1016                         write_unlock(&em_tree->lock);
1017                         if (ret != -EEXIST) {
1018                                 free_extent_map(em);
1019                                 break;
1020                         }
1021                         btrfs_drop_extent_cache(inode, start,
1022                                                 start + ram_size - 1, 0);
1023                 }
1024                 if (ret)
1025                         goto out_reserve;
1026
1027                 cur_alloc_size = ins.offset;
1028                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1029                                                ram_size, cur_alloc_size, 0);
1030                 if (ret)
1031                         goto out_drop_extent_cache;
1032
1033                 if (root->root_key.objectid ==
1034                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1035                         ret = btrfs_reloc_clone_csums(inode, start,
1036                                                       cur_alloc_size);
1037                         if (ret)
1038                                 goto out_drop_extent_cache;
1039                 }
1040
1041                 if (disk_num_bytes < cur_alloc_size)
1042                         break;
1043
1044                 /* we're not doing compressed IO, don't unlock the first
1045                  * page (which the caller expects to stay locked), don't
1046                  * clear any dirty bits and don't set any writeback bits
1047                  *
1048                  * Do set the Private2 bit so we know this page was properly
1049                  * setup for writepage
1050                  */
1051                 op = unlock ? PAGE_UNLOCK : 0;
1052                 op |= PAGE_SET_PRIVATE2;
1053
1054                 extent_clear_unlock_delalloc(inode, start,
1055                                              start + ram_size - 1, locked_page,
1056                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1057                                              op);
1058                 disk_num_bytes -= cur_alloc_size;
1059                 num_bytes -= cur_alloc_size;
1060                 alloc_hint = ins.objectid + ins.offset;
1061                 start += cur_alloc_size;
1062         }
1063 out:
1064         return ret;
1065
1066 out_drop_extent_cache:
1067         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1068 out_reserve:
1069         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1070 out_unlock:
1071         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1072                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1073                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1074                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1075                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1076         goto out;
1077 }
1078
1079 /*
1080  * work queue call back to started compression on a file and pages
1081  */
1082 static noinline void async_cow_start(struct btrfs_work *work)
1083 {
1084         struct async_cow *async_cow;
1085         int num_added = 0;
1086         async_cow = container_of(work, struct async_cow, work);
1087
1088         compress_file_range(async_cow->inode, async_cow->locked_page,
1089                             async_cow->start, async_cow->end, async_cow,
1090                             &num_added);
1091         if (num_added == 0) {
1092                 btrfs_add_delayed_iput(async_cow->inode);
1093                 async_cow->inode = NULL;
1094         }
1095 }
1096
1097 /*
1098  * work queue call back to submit previously compressed pages
1099  */
1100 static noinline void async_cow_submit(struct btrfs_work *work)
1101 {
1102         struct async_cow *async_cow;
1103         struct btrfs_root *root;
1104         unsigned long nr_pages;
1105
1106         async_cow = container_of(work, struct async_cow, work);
1107
1108         root = async_cow->root;
1109         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1110                 PAGE_CACHE_SHIFT;
1111
1112         /*
1113          * atomic_sub_return implies a barrier for waitqueue_active
1114          */
1115         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1116             5 * SZ_1M &&
1117             waitqueue_active(&root->fs_info->async_submit_wait))
1118                 wake_up(&root->fs_info->async_submit_wait);
1119
1120         if (async_cow->inode)
1121                 submit_compressed_extents(async_cow->inode, async_cow);
1122 }
1123
1124 static noinline void async_cow_free(struct btrfs_work *work)
1125 {
1126         struct async_cow *async_cow;
1127         async_cow = container_of(work, struct async_cow, work);
1128         if (async_cow->inode)
1129                 btrfs_add_delayed_iput(async_cow->inode);
1130         kfree(async_cow);
1131 }
1132
1133 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1134                                 u64 start, u64 end, int *page_started,
1135                                 unsigned long *nr_written)
1136 {
1137         struct async_cow *async_cow;
1138         struct btrfs_root *root = BTRFS_I(inode)->root;
1139         unsigned long nr_pages;
1140         u64 cur_end;
1141         int limit = 10 * SZ_1M;
1142
1143         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1144                          1, 0, NULL, GFP_NOFS);
1145         while (start < end) {
1146                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1147                 BUG_ON(!async_cow); /* -ENOMEM */
1148                 async_cow->inode = igrab(inode);
1149                 async_cow->root = root;
1150                 async_cow->locked_page = locked_page;
1151                 async_cow->start = start;
1152
1153                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1154                     !btrfs_test_opt(root, FORCE_COMPRESS))
1155                         cur_end = end;
1156                 else
1157                         cur_end = min(end, start + SZ_512K - 1);
1158
1159                 async_cow->end = cur_end;
1160                 INIT_LIST_HEAD(&async_cow->extents);
1161
1162                 btrfs_init_work(&async_cow->work,
1163                                 btrfs_delalloc_helper,
1164                                 async_cow_start, async_cow_submit,
1165                                 async_cow_free);
1166
1167                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1168                         PAGE_CACHE_SHIFT;
1169                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1170
1171                 btrfs_queue_work(root->fs_info->delalloc_workers,
1172                                  &async_cow->work);
1173
1174                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1175                         wait_event(root->fs_info->async_submit_wait,
1176                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1177                             limit));
1178                 }
1179
1180                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1181                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1182                         wait_event(root->fs_info->async_submit_wait,
1183                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1184                            0));
1185                 }
1186
1187                 *nr_written += nr_pages;
1188                 start = cur_end + 1;
1189         }
1190         *page_started = 1;
1191         return 0;
1192 }
1193
1194 static noinline int csum_exist_in_range(struct btrfs_root *root,
1195                                         u64 bytenr, u64 num_bytes)
1196 {
1197         int ret;
1198         struct btrfs_ordered_sum *sums;
1199         LIST_HEAD(list);
1200
1201         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1202                                        bytenr + num_bytes - 1, &list, 0);
1203         if (ret == 0 && list_empty(&list))
1204                 return 0;
1205
1206         while (!list_empty(&list)) {
1207                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1208                 list_del(&sums->list);
1209                 kfree(sums);
1210         }
1211         return 1;
1212 }
1213
1214 /*
1215  * when nowcow writeback call back.  This checks for snapshots or COW copies
1216  * of the extents that exist in the file, and COWs the file as required.
1217  *
1218  * If no cow copies or snapshots exist, we write directly to the existing
1219  * blocks on disk
1220  */
1221 static noinline int run_delalloc_nocow(struct inode *inode,
1222                                        struct page *locked_page,
1223                               u64 start, u64 end, int *page_started, int force,
1224                               unsigned long *nr_written)
1225 {
1226         struct btrfs_root *root = BTRFS_I(inode)->root;
1227         struct btrfs_trans_handle *trans;
1228         struct extent_buffer *leaf;
1229         struct btrfs_path *path;
1230         struct btrfs_file_extent_item *fi;
1231         struct btrfs_key found_key;
1232         u64 cow_start;
1233         u64 cur_offset;
1234         u64 extent_end;
1235         u64 extent_offset;
1236         u64 disk_bytenr;
1237         u64 num_bytes;
1238         u64 disk_num_bytes;
1239         u64 ram_bytes;
1240         int extent_type;
1241         int ret, err;
1242         int type;
1243         int nocow;
1244         int check_prev = 1;
1245         bool nolock;
1246         u64 ino = btrfs_ino(inode);
1247
1248         path = btrfs_alloc_path();
1249         if (!path) {
1250                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1251                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1252                                              EXTENT_DO_ACCOUNTING |
1253                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1254                                              PAGE_CLEAR_DIRTY |
1255                                              PAGE_SET_WRITEBACK |
1256                                              PAGE_END_WRITEBACK);
1257                 return -ENOMEM;
1258         }
1259
1260         nolock = btrfs_is_free_space_inode(inode);
1261
1262         if (nolock)
1263                 trans = btrfs_join_transaction_nolock(root);
1264         else
1265                 trans = btrfs_join_transaction(root);
1266
1267         if (IS_ERR(trans)) {
1268                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1269                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1270                                              EXTENT_DO_ACCOUNTING |
1271                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1272                                              PAGE_CLEAR_DIRTY |
1273                                              PAGE_SET_WRITEBACK |
1274                                              PAGE_END_WRITEBACK);
1275                 btrfs_free_path(path);
1276                 return PTR_ERR(trans);
1277         }
1278
1279         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1280
1281         cow_start = (u64)-1;
1282         cur_offset = start;
1283         while (1) {
1284                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1285                                                cur_offset, 0);
1286                 if (ret < 0)
1287                         goto error;
1288                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1289                         leaf = path->nodes[0];
1290                         btrfs_item_key_to_cpu(leaf, &found_key,
1291                                               path->slots[0] - 1);
1292                         if (found_key.objectid == ino &&
1293                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1294                                 path->slots[0]--;
1295                 }
1296                 check_prev = 0;
1297 next_slot:
1298                 leaf = path->nodes[0];
1299                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1300                         ret = btrfs_next_leaf(root, path);
1301                         if (ret < 0)
1302                                 goto error;
1303                         if (ret > 0)
1304                                 break;
1305                         leaf = path->nodes[0];
1306                 }
1307
1308                 nocow = 0;
1309                 disk_bytenr = 0;
1310                 num_bytes = 0;
1311                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1312
1313                 if (found_key.objectid > ino)
1314                         break;
1315                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1316                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1317                         path->slots[0]++;
1318                         goto next_slot;
1319                 }
1320                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1321                     found_key.offset > end)
1322                         break;
1323
1324                 if (found_key.offset > cur_offset) {
1325                         extent_end = found_key.offset;
1326                         extent_type = 0;
1327                         goto out_check;
1328                 }
1329
1330                 fi = btrfs_item_ptr(leaf, path->slots[0],
1331                                     struct btrfs_file_extent_item);
1332                 extent_type = btrfs_file_extent_type(leaf, fi);
1333
1334                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1335                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1336                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1337                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1338                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1339                         extent_end = found_key.offset +
1340                                 btrfs_file_extent_num_bytes(leaf, fi);
1341                         disk_num_bytes =
1342                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1343                         if (extent_end <= start) {
1344                                 path->slots[0]++;
1345                                 goto next_slot;
1346                         }
1347                         if (disk_bytenr == 0)
1348                                 goto out_check;
1349                         if (btrfs_file_extent_compression(leaf, fi) ||
1350                             btrfs_file_extent_encryption(leaf, fi) ||
1351                             btrfs_file_extent_other_encoding(leaf, fi))
1352                                 goto out_check;
1353                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1354                                 goto out_check;
1355                         if (btrfs_extent_readonly(root, disk_bytenr))
1356                                 goto out_check;
1357                         if (btrfs_cross_ref_exist(trans, root, ino,
1358                                                   found_key.offset -
1359                                                   extent_offset, disk_bytenr))
1360                                 goto out_check;
1361                         disk_bytenr += extent_offset;
1362                         disk_bytenr += cur_offset - found_key.offset;
1363                         num_bytes = min(end + 1, extent_end) - cur_offset;
1364                         /*
1365                          * if there are pending snapshots for this root,
1366                          * we fall into common COW way.
1367                          */
1368                         if (!nolock) {
1369                                 err = btrfs_start_write_no_snapshoting(root);
1370                                 if (!err)
1371                                         goto out_check;
1372                         }
1373                         /*
1374                          * force cow if csum exists in the range.
1375                          * this ensure that csum for a given extent are
1376                          * either valid or do not exist.
1377                          */
1378                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1379                                 goto out_check;
1380                         nocow = 1;
1381                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1382                         extent_end = found_key.offset +
1383                                 btrfs_file_extent_inline_len(leaf,
1384                                                      path->slots[0], fi);
1385                         extent_end = ALIGN(extent_end, root->sectorsize);
1386                 } else {
1387                         BUG_ON(1);
1388                 }
1389 out_check:
1390                 if (extent_end <= start) {
1391                         path->slots[0]++;
1392                         if (!nolock && nocow)
1393                                 btrfs_end_write_no_snapshoting(root);
1394                         goto next_slot;
1395                 }
1396                 if (!nocow) {
1397                         if (cow_start == (u64)-1)
1398                                 cow_start = cur_offset;
1399                         cur_offset = extent_end;
1400                         if (cur_offset > end)
1401                                 break;
1402                         path->slots[0]++;
1403                         goto next_slot;
1404                 }
1405
1406                 btrfs_release_path(path);
1407                 if (cow_start != (u64)-1) {
1408                         ret = cow_file_range(inode, locked_page,
1409                                              cow_start, found_key.offset - 1,
1410                                              page_started, nr_written, 1);
1411                         if (ret) {
1412                                 if (!nolock && nocow)
1413                                         btrfs_end_write_no_snapshoting(root);
1414                                 goto error;
1415                         }
1416                         cow_start = (u64)-1;
1417                 }
1418
1419                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1420                         struct extent_map *em;
1421                         struct extent_map_tree *em_tree;
1422                         em_tree = &BTRFS_I(inode)->extent_tree;
1423                         em = alloc_extent_map();
1424                         BUG_ON(!em); /* -ENOMEM */
1425                         em->start = cur_offset;
1426                         em->orig_start = found_key.offset - extent_offset;
1427                         em->len = num_bytes;
1428                         em->block_len = num_bytes;
1429                         em->block_start = disk_bytenr;
1430                         em->orig_block_len = disk_num_bytes;
1431                         em->ram_bytes = ram_bytes;
1432                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1433                         em->mod_start = em->start;
1434                         em->mod_len = em->len;
1435                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1436                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1437                         em->generation = -1;
1438                         while (1) {
1439                                 write_lock(&em_tree->lock);
1440                                 ret = add_extent_mapping(em_tree, em, 1);
1441                                 write_unlock(&em_tree->lock);
1442                                 if (ret != -EEXIST) {
1443                                         free_extent_map(em);
1444                                         break;
1445                                 }
1446                                 btrfs_drop_extent_cache(inode, em->start,
1447                                                 em->start + em->len - 1, 0);
1448                         }
1449                         type = BTRFS_ORDERED_PREALLOC;
1450                 } else {
1451                         type = BTRFS_ORDERED_NOCOW;
1452                 }
1453
1454                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1455                                                num_bytes, num_bytes, type);
1456                 BUG_ON(ret); /* -ENOMEM */
1457
1458                 if (root->root_key.objectid ==
1459                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1460                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1461                                                       num_bytes);
1462                         if (ret) {
1463                                 if (!nolock && nocow)
1464                                         btrfs_end_write_no_snapshoting(root);
1465                                 goto error;
1466                         }
1467                 }
1468
1469                 extent_clear_unlock_delalloc(inode, cur_offset,
1470                                              cur_offset + num_bytes - 1,
1471                                              locked_page, EXTENT_LOCKED |
1472                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1473                                              PAGE_SET_PRIVATE2);
1474                 if (!nolock && nocow)
1475                         btrfs_end_write_no_snapshoting(root);
1476                 cur_offset = extent_end;
1477                 if (cur_offset > end)
1478                         break;
1479         }
1480         btrfs_release_path(path);
1481
1482         if (cur_offset <= end && cow_start == (u64)-1) {
1483                 cow_start = cur_offset;
1484                 cur_offset = end;
1485         }
1486
1487         if (cow_start != (u64)-1) {
1488                 ret = cow_file_range(inode, locked_page, cow_start, end,
1489                                      page_started, nr_written, 1);
1490                 if (ret)
1491                         goto error;
1492         }
1493
1494 error:
1495         err = btrfs_end_transaction(trans, root);
1496         if (!ret)
1497                 ret = err;
1498
1499         if (ret && cur_offset < end)
1500                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1501                                              locked_page, EXTENT_LOCKED |
1502                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1503                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1504                                              PAGE_CLEAR_DIRTY |
1505                                              PAGE_SET_WRITEBACK |
1506                                              PAGE_END_WRITEBACK);
1507         btrfs_free_path(path);
1508         return ret;
1509 }
1510
1511 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1512 {
1513
1514         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1515             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1516                 return 0;
1517
1518         /*
1519          * @defrag_bytes is a hint value, no spinlock held here,
1520          * if is not zero, it means the file is defragging.
1521          * Force cow if given extent needs to be defragged.
1522          */
1523         if (BTRFS_I(inode)->defrag_bytes &&
1524             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1525                            EXTENT_DEFRAG, 0, NULL))
1526                 return 1;
1527
1528         return 0;
1529 }
1530
1531 /*
1532  * extent_io.c call back to do delayed allocation processing
1533  */
1534 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1535                               u64 start, u64 end, int *page_started,
1536                               unsigned long *nr_written)
1537 {
1538         int ret;
1539         int force_cow = need_force_cow(inode, start, end);
1540
1541         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1542                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1543                                          page_started, 1, nr_written);
1544         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1545                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1546                                          page_started, 0, nr_written);
1547         } else if (!inode_need_compress(inode)) {
1548                 ret = cow_file_range(inode, locked_page, start, end,
1549                                       page_started, nr_written, 1);
1550         } else {
1551                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1552                         &BTRFS_I(inode)->runtime_flags);
1553                 ret = cow_file_range_async(inode, locked_page, start, end,
1554                                            page_started, nr_written);
1555         }
1556         return ret;
1557 }
1558
1559 static void btrfs_split_extent_hook(struct inode *inode,
1560                                     struct extent_state *orig, u64 split)
1561 {
1562         u64 size;
1563
1564         /* not delalloc, ignore it */
1565         if (!(orig->state & EXTENT_DELALLOC))
1566                 return;
1567
1568         size = orig->end - orig->start + 1;
1569         if (size > BTRFS_MAX_EXTENT_SIZE) {
1570                 u64 num_extents;
1571                 u64 new_size;
1572
1573                 /*
1574                  * See the explanation in btrfs_merge_extent_hook, the same
1575                  * applies here, just in reverse.
1576                  */
1577                 new_size = orig->end - split + 1;
1578                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1579                                         BTRFS_MAX_EXTENT_SIZE);
1580                 new_size = split - orig->start;
1581                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1582                                         BTRFS_MAX_EXTENT_SIZE);
1583                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1584                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1585                         return;
1586         }
1587
1588         spin_lock(&BTRFS_I(inode)->lock);
1589         BTRFS_I(inode)->outstanding_extents++;
1590         spin_unlock(&BTRFS_I(inode)->lock);
1591 }
1592
1593 /*
1594  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1595  * extents so we can keep track of new extents that are just merged onto old
1596  * extents, such as when we are doing sequential writes, so we can properly
1597  * account for the metadata space we'll need.
1598  */
1599 static void btrfs_merge_extent_hook(struct inode *inode,
1600                                     struct extent_state *new,
1601                                     struct extent_state *other)
1602 {
1603         u64 new_size, old_size;
1604         u64 num_extents;
1605
1606         /* not delalloc, ignore it */
1607         if (!(other->state & EXTENT_DELALLOC))
1608                 return;
1609
1610         if (new->start > other->start)
1611                 new_size = new->end - other->start + 1;
1612         else
1613                 new_size = other->end - new->start + 1;
1614
1615         /* we're not bigger than the max, unreserve the space and go */
1616         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1617                 spin_lock(&BTRFS_I(inode)->lock);
1618                 BTRFS_I(inode)->outstanding_extents--;
1619                 spin_unlock(&BTRFS_I(inode)->lock);
1620                 return;
1621         }
1622
1623         /*
1624          * We have to add up either side to figure out how many extents were
1625          * accounted for before we merged into one big extent.  If the number of
1626          * extents we accounted for is <= the amount we need for the new range
1627          * then we can return, otherwise drop.  Think of it like this
1628          *
1629          * [ 4k][MAX_SIZE]
1630          *
1631          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1632          * need 2 outstanding extents, on one side we have 1 and the other side
1633          * we have 1 so they are == and we can return.  But in this case
1634          *
1635          * [MAX_SIZE+4k][MAX_SIZE+4k]
1636          *
1637          * Each range on their own accounts for 2 extents, but merged together
1638          * they are only 3 extents worth of accounting, so we need to drop in
1639          * this case.
1640          */
1641         old_size = other->end - other->start + 1;
1642         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1643                                 BTRFS_MAX_EXTENT_SIZE);
1644         old_size = new->end - new->start + 1;
1645         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1646                                  BTRFS_MAX_EXTENT_SIZE);
1647
1648         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1649                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1650                 return;
1651
1652         spin_lock(&BTRFS_I(inode)->lock);
1653         BTRFS_I(inode)->outstanding_extents--;
1654         spin_unlock(&BTRFS_I(inode)->lock);
1655 }
1656
1657 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1658                                       struct inode *inode)
1659 {
1660         spin_lock(&root->delalloc_lock);
1661         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1662                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1663                               &root->delalloc_inodes);
1664                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1665                         &BTRFS_I(inode)->runtime_flags);
1666                 root->nr_delalloc_inodes++;
1667                 if (root->nr_delalloc_inodes == 1) {
1668                         spin_lock(&root->fs_info->delalloc_root_lock);
1669                         BUG_ON(!list_empty(&root->delalloc_root));
1670                         list_add_tail(&root->delalloc_root,
1671                                       &root->fs_info->delalloc_roots);
1672                         spin_unlock(&root->fs_info->delalloc_root_lock);
1673                 }
1674         }
1675         spin_unlock(&root->delalloc_lock);
1676 }
1677
1678 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1679                                      struct inode *inode)
1680 {
1681         spin_lock(&root->delalloc_lock);
1682         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1683                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1684                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1685                           &BTRFS_I(inode)->runtime_flags);
1686                 root->nr_delalloc_inodes--;
1687                 if (!root->nr_delalloc_inodes) {
1688                         spin_lock(&root->fs_info->delalloc_root_lock);
1689                         BUG_ON(list_empty(&root->delalloc_root));
1690                         list_del_init(&root->delalloc_root);
1691                         spin_unlock(&root->fs_info->delalloc_root_lock);
1692                 }
1693         }
1694         spin_unlock(&root->delalloc_lock);
1695 }
1696
1697 /*
1698  * extent_io.c set_bit_hook, used to track delayed allocation
1699  * bytes in this file, and to maintain the list of inodes that
1700  * have pending delalloc work to be done.
1701  */
1702 static void btrfs_set_bit_hook(struct inode *inode,
1703                                struct extent_state *state, unsigned *bits)
1704 {
1705
1706         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1707                 WARN_ON(1);
1708         /*
1709          * set_bit and clear bit hooks normally require _irqsave/restore
1710          * but in this case, we are only testing for the DELALLOC
1711          * bit, which is only set or cleared with irqs on
1712          */
1713         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1714                 struct btrfs_root *root = BTRFS_I(inode)->root;
1715                 u64 len = state->end + 1 - state->start;
1716                 bool do_list = !btrfs_is_free_space_inode(inode);
1717
1718                 if (*bits & EXTENT_FIRST_DELALLOC) {
1719                         *bits &= ~EXTENT_FIRST_DELALLOC;
1720                 } else {
1721                         spin_lock(&BTRFS_I(inode)->lock);
1722                         BTRFS_I(inode)->outstanding_extents++;
1723                         spin_unlock(&BTRFS_I(inode)->lock);
1724                 }
1725
1726                 /* For sanity tests */
1727                 if (btrfs_test_is_dummy_root(root))
1728                         return;
1729
1730                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1731                                      root->fs_info->delalloc_batch);
1732                 spin_lock(&BTRFS_I(inode)->lock);
1733                 BTRFS_I(inode)->delalloc_bytes += len;
1734                 if (*bits & EXTENT_DEFRAG)
1735                         BTRFS_I(inode)->defrag_bytes += len;
1736                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1737                                          &BTRFS_I(inode)->runtime_flags))
1738                         btrfs_add_delalloc_inodes(root, inode);
1739                 spin_unlock(&BTRFS_I(inode)->lock);
1740         }
1741 }
1742
1743 /*
1744  * extent_io.c clear_bit_hook, see set_bit_hook for why
1745  */
1746 static void btrfs_clear_bit_hook(struct inode *inode,
1747                                  struct extent_state *state,
1748                                  unsigned *bits)
1749 {
1750         u64 len = state->end + 1 - state->start;
1751         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1752                                     BTRFS_MAX_EXTENT_SIZE);
1753
1754         spin_lock(&BTRFS_I(inode)->lock);
1755         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1756                 BTRFS_I(inode)->defrag_bytes -= len;
1757         spin_unlock(&BTRFS_I(inode)->lock);
1758
1759         /*
1760          * set_bit and clear bit hooks normally require _irqsave/restore
1761          * but in this case, we are only testing for the DELALLOC
1762          * bit, which is only set or cleared with irqs on
1763          */
1764         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1765                 struct btrfs_root *root = BTRFS_I(inode)->root;
1766                 bool do_list = !btrfs_is_free_space_inode(inode);
1767
1768                 if (*bits & EXTENT_FIRST_DELALLOC) {
1769                         *bits &= ~EXTENT_FIRST_DELALLOC;
1770                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1771                         spin_lock(&BTRFS_I(inode)->lock);
1772                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1773                         spin_unlock(&BTRFS_I(inode)->lock);
1774                 }
1775
1776                 /*
1777                  * We don't reserve metadata space for space cache inodes so we
1778                  * don't need to call dellalloc_release_metadata if there is an
1779                  * error.
1780                  */
1781                 if (*bits & EXTENT_DO_ACCOUNTING &&
1782                     root != root->fs_info->tree_root)
1783                         btrfs_delalloc_release_metadata(inode, len);
1784
1785                 /* For sanity tests. */
1786                 if (btrfs_test_is_dummy_root(root))
1787                         return;
1788
1789                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1790                     && do_list && !(state->state & EXTENT_NORESERVE))
1791                         btrfs_free_reserved_data_space_noquota(inode,
1792                                         state->start, len);
1793
1794                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1795                                      root->fs_info->delalloc_batch);
1796                 spin_lock(&BTRFS_I(inode)->lock);
1797                 BTRFS_I(inode)->delalloc_bytes -= len;
1798                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1799                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1800                              &BTRFS_I(inode)->runtime_flags))
1801                         btrfs_del_delalloc_inode(root, inode);
1802                 spin_unlock(&BTRFS_I(inode)->lock);
1803         }
1804 }
1805
1806 /*
1807  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1808  * we don't create bios that span stripes or chunks
1809  */
1810 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1811                          size_t size, struct bio *bio,
1812                          unsigned long bio_flags)
1813 {
1814         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1815         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1816         u64 length = 0;
1817         u64 map_length;
1818         int ret;
1819
1820         if (bio_flags & EXTENT_BIO_COMPRESSED)
1821                 return 0;
1822
1823         length = bio->bi_iter.bi_size;
1824         map_length = length;
1825         ret = btrfs_map_block(root->fs_info, rw, logical,
1826                               &map_length, NULL, 0);
1827         /* Will always return 0 with map_multi == NULL */
1828         BUG_ON(ret < 0);
1829         if (map_length < length + size)
1830                 return 1;
1831         return 0;
1832 }
1833
1834 /*
1835  * in order to insert checksums into the metadata in large chunks,
1836  * we wait until bio submission time.   All the pages in the bio are
1837  * checksummed and sums are attached onto the ordered extent record.
1838  *
1839  * At IO completion time the cums attached on the ordered extent record
1840  * are inserted into the btree
1841  */
1842 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1843                                     struct bio *bio, int mirror_num,
1844                                     unsigned long bio_flags,
1845                                     u64 bio_offset)
1846 {
1847         struct btrfs_root *root = BTRFS_I(inode)->root;
1848         int ret = 0;
1849
1850         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1851         BUG_ON(ret); /* -ENOMEM */
1852         return 0;
1853 }
1854
1855 /*
1856  * in order to insert checksums into the metadata in large chunks,
1857  * we wait until bio submission time.   All the pages in the bio are
1858  * checksummed and sums are attached onto the ordered extent record.
1859  *
1860  * At IO completion time the cums attached on the ordered extent record
1861  * are inserted into the btree
1862  */
1863 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1864                           int mirror_num, unsigned long bio_flags,
1865                           u64 bio_offset)
1866 {
1867         struct btrfs_root *root = BTRFS_I(inode)->root;
1868         int ret;
1869
1870         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1871         if (ret) {
1872                 bio->bi_error = ret;
1873                 bio_endio(bio);
1874         }
1875         return ret;
1876 }
1877
1878 /*
1879  * extent_io.c submission hook. This does the right thing for csum calculation
1880  * on write, or reading the csums from the tree before a read
1881  */
1882 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1883                           int mirror_num, unsigned long bio_flags,
1884                           u64 bio_offset)
1885 {
1886         struct btrfs_root *root = BTRFS_I(inode)->root;
1887         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1888         int ret = 0;
1889         int skip_sum;
1890         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1891
1892         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1893
1894         if (btrfs_is_free_space_inode(inode))
1895                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1896
1897         if (!(rw & REQ_WRITE)) {
1898                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1899                 if (ret)
1900                         goto out;
1901
1902                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1903                         ret = btrfs_submit_compressed_read(inode, bio,
1904                                                            mirror_num,
1905                                                            bio_flags);
1906                         goto out;
1907                 } else if (!skip_sum) {
1908                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1909                         if (ret)
1910                                 goto out;
1911                 }
1912                 goto mapit;
1913         } else if (async && !skip_sum) {
1914                 /* csum items have already been cloned */
1915                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1916                         goto mapit;
1917                 /* we're doing a write, do the async checksumming */
1918                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1919                                    inode, rw, bio, mirror_num,
1920                                    bio_flags, bio_offset,
1921                                    __btrfs_submit_bio_start,
1922                                    __btrfs_submit_bio_done);
1923                 goto out;
1924         } else if (!skip_sum) {
1925                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1926                 if (ret)
1927                         goto out;
1928         }
1929
1930 mapit:
1931         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1932
1933 out:
1934         if (ret < 0) {
1935                 bio->bi_error = ret;
1936                 bio_endio(bio);
1937         }
1938         return ret;
1939 }
1940
1941 /*
1942  * given a list of ordered sums record them in the inode.  This happens
1943  * at IO completion time based on sums calculated at bio submission time.
1944  */
1945 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1946                              struct inode *inode, u64 file_offset,
1947                              struct list_head *list)
1948 {
1949         struct btrfs_ordered_sum *sum;
1950
1951         list_for_each_entry(sum, list, list) {
1952                 trans->adding_csums = 1;
1953                 btrfs_csum_file_blocks(trans,
1954                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1955                 trans->adding_csums = 0;
1956         }
1957         return 0;
1958 }
1959
1960 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1961                               struct extent_state **cached_state)
1962 {
1963         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1964         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1965                                    cached_state, GFP_NOFS);
1966 }
1967
1968 /* see btrfs_writepage_start_hook for details on why this is required */
1969 struct btrfs_writepage_fixup {
1970         struct page *page;
1971         struct btrfs_work work;
1972 };
1973
1974 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1975 {
1976         struct btrfs_writepage_fixup *fixup;
1977         struct btrfs_ordered_extent *ordered;
1978         struct extent_state *cached_state = NULL;
1979         struct page *page;
1980         struct inode *inode;
1981         u64 page_start;
1982         u64 page_end;
1983         int ret;
1984
1985         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1986         page = fixup->page;
1987 again:
1988         lock_page(page);
1989         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1990                 ClearPageChecked(page);
1991                 goto out_page;
1992         }
1993
1994         inode = page->mapping->host;
1995         page_start = page_offset(page);
1996         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1997
1998         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
1999                          &cached_state);
2000
2001         /* already ordered? We're done */
2002         if (PagePrivate2(page))
2003                 goto out;
2004
2005         ordered = btrfs_lookup_ordered_extent(inode, page_start);
2006         if (ordered) {
2007                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2008                                      page_end, &cached_state, GFP_NOFS);
2009                 unlock_page(page);
2010                 btrfs_start_ordered_extent(inode, ordered, 1);
2011                 btrfs_put_ordered_extent(ordered);
2012                 goto again;
2013         }
2014
2015         ret = btrfs_delalloc_reserve_space(inode, page_start,
2016                                            PAGE_CACHE_SIZE);
2017         if (ret) {
2018                 mapping_set_error(page->mapping, ret);
2019                 end_extent_writepage(page, ret, page_start, page_end);
2020                 ClearPageChecked(page);
2021                 goto out;
2022          }
2023
2024         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2025         ClearPageChecked(page);
2026         set_page_dirty(page);
2027 out:
2028         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2029                              &cached_state, GFP_NOFS);
2030 out_page:
2031         unlock_page(page);
2032         page_cache_release(page);
2033         kfree(fixup);
2034 }
2035
2036 /*
2037  * There are a few paths in the higher layers of the kernel that directly
2038  * set the page dirty bit without asking the filesystem if it is a
2039  * good idea.  This causes problems because we want to make sure COW
2040  * properly happens and the data=ordered rules are followed.
2041  *
2042  * In our case any range that doesn't have the ORDERED bit set
2043  * hasn't been properly setup for IO.  We kick off an async process
2044  * to fix it up.  The async helper will wait for ordered extents, set
2045  * the delalloc bit and make it safe to write the page.
2046  */
2047 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2048 {
2049         struct inode *inode = page->mapping->host;
2050         struct btrfs_writepage_fixup *fixup;
2051         struct btrfs_root *root = BTRFS_I(inode)->root;
2052
2053         /* this page is properly in the ordered list */
2054         if (TestClearPagePrivate2(page))
2055                 return 0;
2056
2057         if (PageChecked(page))
2058                 return -EAGAIN;
2059
2060         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2061         if (!fixup)
2062                 return -EAGAIN;
2063
2064         SetPageChecked(page);
2065         page_cache_get(page);
2066         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2067                         btrfs_writepage_fixup_worker, NULL, NULL);
2068         fixup->page = page;
2069         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2070         return -EBUSY;
2071 }
2072
2073 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2074                                        struct inode *inode, u64 file_pos,
2075                                        u64 disk_bytenr, u64 disk_num_bytes,
2076                                        u64 num_bytes, u64 ram_bytes,
2077                                        u8 compression, u8 encryption,
2078                                        u16 other_encoding, int extent_type)
2079 {
2080         struct btrfs_root *root = BTRFS_I(inode)->root;
2081         struct btrfs_file_extent_item *fi;
2082         struct btrfs_path *path;
2083         struct extent_buffer *leaf;
2084         struct btrfs_key ins;
2085         int extent_inserted = 0;
2086         int ret;
2087
2088         path = btrfs_alloc_path();
2089         if (!path)
2090                 return -ENOMEM;
2091
2092         /*
2093          * we may be replacing one extent in the tree with another.
2094          * The new extent is pinned in the extent map, and we don't want
2095          * to drop it from the cache until it is completely in the btree.
2096          *
2097          * So, tell btrfs_drop_extents to leave this extent in the cache.
2098          * the caller is expected to unpin it and allow it to be merged
2099          * with the others.
2100          */
2101         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2102                                    file_pos + num_bytes, NULL, 0,
2103                                    1, sizeof(*fi), &extent_inserted);
2104         if (ret)
2105                 goto out;
2106
2107         if (!extent_inserted) {
2108                 ins.objectid = btrfs_ino(inode);
2109                 ins.offset = file_pos;
2110                 ins.type = BTRFS_EXTENT_DATA_KEY;
2111
2112                 path->leave_spinning = 1;
2113                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2114                                               sizeof(*fi));
2115                 if (ret)
2116                         goto out;
2117         }
2118         leaf = path->nodes[0];
2119         fi = btrfs_item_ptr(leaf, path->slots[0],
2120                             struct btrfs_file_extent_item);
2121         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2122         btrfs_set_file_extent_type(leaf, fi, extent_type);
2123         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2124         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2125         btrfs_set_file_extent_offset(leaf, fi, 0);
2126         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2127         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2128         btrfs_set_file_extent_compression(leaf, fi, compression);
2129         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2130         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2131
2132         btrfs_mark_buffer_dirty(leaf);
2133         btrfs_release_path(path);
2134
2135         inode_add_bytes(inode, num_bytes);
2136
2137         ins.objectid = disk_bytenr;
2138         ins.offset = disk_num_bytes;
2139         ins.type = BTRFS_EXTENT_ITEM_KEY;
2140         ret = btrfs_alloc_reserved_file_extent(trans, root,
2141                                         root->root_key.objectid,
2142                                         btrfs_ino(inode), file_pos,
2143                                         ram_bytes, &ins);
2144         /*
2145          * Release the reserved range from inode dirty range map, as it is
2146          * already moved into delayed_ref_head
2147          */
2148         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2149 out:
2150         btrfs_free_path(path);
2151
2152         return ret;
2153 }
2154
2155 /* snapshot-aware defrag */
2156 struct sa_defrag_extent_backref {
2157         struct rb_node node;
2158         struct old_sa_defrag_extent *old;
2159         u64 root_id;
2160         u64 inum;
2161         u64 file_pos;
2162         u64 extent_offset;
2163         u64 num_bytes;
2164         u64 generation;
2165 };
2166
2167 struct old_sa_defrag_extent {
2168         struct list_head list;
2169         struct new_sa_defrag_extent *new;
2170
2171         u64 extent_offset;
2172         u64 bytenr;
2173         u64 offset;
2174         u64 len;
2175         int count;
2176 };
2177
2178 struct new_sa_defrag_extent {
2179         struct rb_root root;
2180         struct list_head head;
2181         struct btrfs_path *path;
2182         struct inode *inode;
2183         u64 file_pos;
2184         u64 len;
2185         u64 bytenr;
2186         u64 disk_len;
2187         u8 compress_type;
2188 };
2189
2190 static int backref_comp(struct sa_defrag_extent_backref *b1,
2191                         struct sa_defrag_extent_backref *b2)
2192 {
2193         if (b1->root_id < b2->root_id)
2194                 return -1;
2195         else if (b1->root_id > b2->root_id)
2196                 return 1;
2197
2198         if (b1->inum < b2->inum)
2199                 return -1;
2200         else if (b1->inum > b2->inum)
2201                 return 1;
2202
2203         if (b1->file_pos < b2->file_pos)
2204                 return -1;
2205         else if (b1->file_pos > b2->file_pos)
2206                 return 1;
2207
2208         /*
2209          * [------------------------------] ===> (a range of space)
2210          *     |<--->|   |<---->| =============> (fs/file tree A)
2211          * |<---------------------------->| ===> (fs/file tree B)
2212          *
2213          * A range of space can refer to two file extents in one tree while
2214          * refer to only one file extent in another tree.
2215          *
2216          * So we may process a disk offset more than one time(two extents in A)
2217          * and locate at the same extent(one extent in B), then insert two same
2218          * backrefs(both refer to the extent in B).
2219          */
2220         return 0;
2221 }
2222
2223 static void backref_insert(struct rb_root *root,
2224                            struct sa_defrag_extent_backref *backref)
2225 {
2226         struct rb_node **p = &root->rb_node;
2227         struct rb_node *parent = NULL;
2228         struct sa_defrag_extent_backref *entry;
2229         int ret;
2230
2231         while (*p) {
2232                 parent = *p;
2233                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2234
2235                 ret = backref_comp(backref, entry);
2236                 if (ret < 0)
2237                         p = &(*p)->rb_left;
2238                 else
2239                         p = &(*p)->rb_right;
2240         }
2241
2242         rb_link_node(&backref->node, parent, p);
2243         rb_insert_color(&backref->node, root);
2244 }
2245
2246 /*
2247  * Note the backref might has changed, and in this case we just return 0.
2248  */
2249 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2250                                        void *ctx)
2251 {
2252         struct btrfs_file_extent_item *extent;
2253         struct btrfs_fs_info *fs_info;
2254         struct old_sa_defrag_extent *old = ctx;
2255         struct new_sa_defrag_extent *new = old->new;
2256         struct btrfs_path *path = new->path;
2257         struct btrfs_key key;
2258         struct btrfs_root *root;
2259         struct sa_defrag_extent_backref *backref;
2260         struct extent_buffer *leaf;
2261         struct inode *inode = new->inode;
2262         int slot;
2263         int ret;
2264         u64 extent_offset;
2265         u64 num_bytes;
2266
2267         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2268             inum == btrfs_ino(inode))
2269                 return 0;
2270
2271         key.objectid = root_id;
2272         key.type = BTRFS_ROOT_ITEM_KEY;
2273         key.offset = (u64)-1;
2274
2275         fs_info = BTRFS_I(inode)->root->fs_info;
2276         root = btrfs_read_fs_root_no_name(fs_info, &key);
2277         if (IS_ERR(root)) {
2278                 if (PTR_ERR(root) == -ENOENT)
2279                         return 0;
2280                 WARN_ON(1);
2281                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2282                          inum, offset, root_id);
2283                 return PTR_ERR(root);
2284         }
2285
2286         key.objectid = inum;
2287         key.type = BTRFS_EXTENT_DATA_KEY;
2288         if (offset > (u64)-1 << 32)
2289                 key.offset = 0;
2290         else
2291                 key.offset = offset;
2292
2293         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2294         if (WARN_ON(ret < 0))
2295                 return ret;
2296         ret = 0;
2297
2298         while (1) {
2299                 cond_resched();
2300
2301                 leaf = path->nodes[0];
2302                 slot = path->slots[0];
2303
2304                 if (slot >= btrfs_header_nritems(leaf)) {
2305                         ret = btrfs_next_leaf(root, path);
2306                         if (ret < 0) {
2307                                 goto out;
2308                         } else if (ret > 0) {
2309                                 ret = 0;
2310                                 goto out;
2311                         }
2312                         continue;
2313                 }
2314
2315                 path->slots[0]++;
2316
2317                 btrfs_item_key_to_cpu(leaf, &key, slot);
2318
2319                 if (key.objectid > inum)
2320                         goto out;
2321
2322                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2323                         continue;
2324
2325                 extent = btrfs_item_ptr(leaf, slot,
2326                                         struct btrfs_file_extent_item);
2327
2328                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2329                         continue;
2330
2331                 /*
2332                  * 'offset' refers to the exact key.offset,
2333                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2334                  * (key.offset - extent_offset).
2335                  */
2336                 if (key.offset != offset)
2337                         continue;
2338
2339                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2340                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2341
2342                 if (extent_offset >= old->extent_offset + old->offset +
2343                     old->len || extent_offset + num_bytes <=
2344                     old->extent_offset + old->offset)
2345                         continue;
2346                 break;
2347         }
2348
2349         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2350         if (!backref) {
2351                 ret = -ENOENT;
2352                 goto out;
2353         }
2354
2355         backref->root_id = root_id;
2356         backref->inum = inum;
2357         backref->file_pos = offset;
2358         backref->num_bytes = num_bytes;
2359         backref->extent_offset = extent_offset;
2360         backref->generation = btrfs_file_extent_generation(leaf, extent);
2361         backref->old = old;
2362         backref_insert(&new->root, backref);
2363         old->count++;
2364 out:
2365         btrfs_release_path(path);
2366         WARN_ON(ret);
2367         return ret;
2368 }
2369
2370 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2371                                    struct new_sa_defrag_extent *new)
2372 {
2373         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2374         struct old_sa_defrag_extent *old, *tmp;
2375         int ret;
2376
2377         new->path = path;
2378
2379         list_for_each_entry_safe(old, tmp, &new->head, list) {
2380                 ret = iterate_inodes_from_logical(old->bytenr +
2381                                                   old->extent_offset, fs_info,
2382                                                   path, record_one_backref,
2383                                                   old);
2384                 if (ret < 0 && ret != -ENOENT)
2385                         return false;
2386
2387                 /* no backref to be processed for this extent */
2388                 if (!old->count) {
2389                         list_del(&old->list);
2390                         kfree(old);
2391                 }
2392         }
2393
2394         if (list_empty(&new->head))
2395                 return false;
2396
2397         return true;
2398 }
2399
2400 static int relink_is_mergable(struct extent_buffer *leaf,
2401                               struct btrfs_file_extent_item *fi,
2402                               struct new_sa_defrag_extent *new)
2403 {
2404         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2405                 return 0;
2406
2407         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2408                 return 0;
2409
2410         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2411                 return 0;
2412
2413         if (btrfs_file_extent_encryption(leaf, fi) ||
2414             btrfs_file_extent_other_encoding(leaf, fi))
2415                 return 0;
2416
2417         return 1;
2418 }
2419
2420 /*
2421  * Note the backref might has changed, and in this case we just return 0.
2422  */
2423 static noinline int relink_extent_backref(struct btrfs_path *path,
2424                                  struct sa_defrag_extent_backref *prev,
2425                                  struct sa_defrag_extent_backref *backref)
2426 {
2427         struct btrfs_file_extent_item *extent;
2428         struct btrfs_file_extent_item *item;
2429         struct btrfs_ordered_extent *ordered;
2430         struct btrfs_trans_handle *trans;
2431         struct btrfs_fs_info *fs_info;
2432         struct btrfs_root *root;
2433         struct btrfs_key key;
2434         struct extent_buffer *leaf;
2435         struct old_sa_defrag_extent *old = backref->old;
2436         struct new_sa_defrag_extent *new = old->new;
2437         struct inode *src_inode = new->inode;
2438         struct inode *inode;
2439         struct extent_state *cached = NULL;
2440         int ret = 0;
2441         u64 start;
2442         u64 len;
2443         u64 lock_start;
2444         u64 lock_end;
2445         bool merge = false;
2446         int index;
2447
2448         if (prev && prev->root_id == backref->root_id &&
2449             prev->inum == backref->inum &&
2450             prev->file_pos + prev->num_bytes == backref->file_pos)
2451                 merge = true;
2452
2453         /* step 1: get root */
2454         key.objectid = backref->root_id;
2455         key.type = BTRFS_ROOT_ITEM_KEY;
2456         key.offset = (u64)-1;
2457
2458         fs_info = BTRFS_I(src_inode)->root->fs_info;
2459         index = srcu_read_lock(&fs_info->subvol_srcu);
2460
2461         root = btrfs_read_fs_root_no_name(fs_info, &key);
2462         if (IS_ERR(root)) {
2463                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2464                 if (PTR_ERR(root) == -ENOENT)
2465                         return 0;
2466                 return PTR_ERR(root);
2467         }
2468
2469         if (btrfs_root_readonly(root)) {
2470                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2471                 return 0;
2472         }
2473
2474         /* step 2: get inode */
2475         key.objectid = backref->inum;
2476         key.type = BTRFS_INODE_ITEM_KEY;
2477         key.offset = 0;
2478
2479         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2480         if (IS_ERR(inode)) {
2481                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2482                 return 0;
2483         }
2484
2485         srcu_read_unlock(&fs_info->subvol_srcu, index);
2486
2487         /* step 3: relink backref */
2488         lock_start = backref->file_pos;
2489         lock_end = backref->file_pos + backref->num_bytes - 1;
2490         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2491                          &cached);
2492
2493         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2494         if (ordered) {
2495                 btrfs_put_ordered_extent(ordered);
2496                 goto out_unlock;
2497         }
2498
2499         trans = btrfs_join_transaction(root);
2500         if (IS_ERR(trans)) {
2501                 ret = PTR_ERR(trans);
2502                 goto out_unlock;
2503         }
2504
2505         key.objectid = backref->inum;
2506         key.type = BTRFS_EXTENT_DATA_KEY;
2507         key.offset = backref->file_pos;
2508
2509         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2510         if (ret < 0) {
2511                 goto out_free_path;
2512         } else if (ret > 0) {
2513                 ret = 0;
2514                 goto out_free_path;
2515         }
2516
2517         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2518                                 struct btrfs_file_extent_item);
2519
2520         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2521             backref->generation)
2522                 goto out_free_path;
2523
2524         btrfs_release_path(path);
2525
2526         start = backref->file_pos;
2527         if (backref->extent_offset < old->extent_offset + old->offset)
2528                 start += old->extent_offset + old->offset -
2529                          backref->extent_offset;
2530
2531         len = min(backref->extent_offset + backref->num_bytes,
2532                   old->extent_offset + old->offset + old->len);
2533         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2534
2535         ret = btrfs_drop_extents(trans, root, inode, start,
2536                                  start + len, 1);
2537         if (ret)
2538                 goto out_free_path;
2539 again:
2540         key.objectid = btrfs_ino(inode);
2541         key.type = BTRFS_EXTENT_DATA_KEY;
2542         key.offset = start;
2543
2544         path->leave_spinning = 1;
2545         if (merge) {
2546                 struct btrfs_file_extent_item *fi;
2547                 u64 extent_len;
2548                 struct btrfs_key found_key;
2549
2550                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2551                 if (ret < 0)
2552                         goto out_free_path;
2553
2554                 path->slots[0]--;
2555                 leaf = path->nodes[0];
2556                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2557
2558                 fi = btrfs_item_ptr(leaf, path->slots[0],
2559                                     struct btrfs_file_extent_item);
2560                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2561
2562                 if (extent_len + found_key.offset == start &&
2563                     relink_is_mergable(leaf, fi, new)) {
2564                         btrfs_set_file_extent_num_bytes(leaf, fi,
2565                                                         extent_len + len);
2566                         btrfs_mark_buffer_dirty(leaf);
2567                         inode_add_bytes(inode, len);
2568
2569                         ret = 1;
2570                         goto out_free_path;
2571                 } else {
2572                         merge = false;
2573                         btrfs_release_path(path);
2574                         goto again;
2575                 }
2576         }
2577
2578         ret = btrfs_insert_empty_item(trans, root, path, &key,
2579                                         sizeof(*extent));
2580         if (ret) {
2581                 btrfs_abort_transaction(trans, root, ret);
2582                 goto out_free_path;
2583         }
2584
2585         leaf = path->nodes[0];
2586         item = btrfs_item_ptr(leaf, path->slots[0],
2587                                 struct btrfs_file_extent_item);
2588         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2589         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2590         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2591         btrfs_set_file_extent_num_bytes(leaf, item, len);
2592         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2593         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2594         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2595         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2596         btrfs_set_file_extent_encryption(leaf, item, 0);
2597         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2598
2599         btrfs_mark_buffer_dirty(leaf);
2600         inode_add_bytes(inode, len);
2601         btrfs_release_path(path);
2602
2603         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2604                         new->disk_len, 0,
2605                         backref->root_id, backref->inum,
2606                         new->file_pos); /* start - extent_offset */
2607         if (ret) {
2608                 btrfs_abort_transaction(trans, root, ret);
2609                 goto out_free_path;
2610         }
2611
2612         ret = 1;
2613 out_free_path:
2614         btrfs_release_path(path);
2615         path->leave_spinning = 0;
2616         btrfs_end_transaction(trans, root);
2617 out_unlock:
2618         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2619                              &cached, GFP_NOFS);
2620         iput(inode);
2621         return ret;
2622 }
2623
2624 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2625 {
2626         struct old_sa_defrag_extent *old, *tmp;
2627
2628         if (!new)
2629                 return;
2630
2631         list_for_each_entry_safe(old, tmp, &new->head, list) {
2632                 kfree(old);
2633         }
2634         kfree(new);
2635 }
2636
2637 static void relink_file_extents(struct new_sa_defrag_extent *new)
2638 {
2639         struct btrfs_path *path;
2640         struct sa_defrag_extent_backref *backref;
2641         struct sa_defrag_extent_backref *prev = NULL;
2642         struct inode *inode;
2643         struct btrfs_root *root;
2644         struct rb_node *node;
2645         int ret;
2646
2647         inode = new->inode;
2648         root = BTRFS_I(inode)->root;
2649
2650         path = btrfs_alloc_path();
2651         if (!path)
2652                 return;
2653
2654         if (!record_extent_backrefs(path, new)) {
2655                 btrfs_free_path(path);
2656                 goto out;
2657         }
2658         btrfs_release_path(path);
2659
2660         while (1) {
2661                 node = rb_first(&new->root);
2662                 if (!node)
2663                         break;
2664                 rb_erase(node, &new->root);
2665
2666                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2667
2668                 ret = relink_extent_backref(path, prev, backref);
2669                 WARN_ON(ret < 0);
2670
2671                 kfree(prev);
2672
2673                 if (ret == 1)
2674                         prev = backref;
2675                 else
2676                         prev = NULL;
2677                 cond_resched();
2678         }
2679         kfree(prev);
2680
2681         btrfs_free_path(path);
2682 out:
2683         free_sa_defrag_extent(new);
2684
2685         atomic_dec(&root->fs_info->defrag_running);
2686         wake_up(&root->fs_info->transaction_wait);
2687 }
2688
2689 static struct new_sa_defrag_extent *
2690 record_old_file_extents(struct inode *inode,
2691                         struct btrfs_ordered_extent *ordered)
2692 {
2693         struct btrfs_root *root = BTRFS_I(inode)->root;
2694         struct btrfs_path *path;
2695         struct btrfs_key key;
2696         struct old_sa_defrag_extent *old;
2697         struct new_sa_defrag_extent *new;
2698         int ret;
2699
2700         new = kmalloc(sizeof(*new), GFP_NOFS);
2701         if (!new)
2702                 return NULL;
2703
2704         new->inode = inode;
2705         new->file_pos = ordered->file_offset;
2706         new->len = ordered->len;
2707         new->bytenr = ordered->start;
2708         new->disk_len = ordered->disk_len;
2709         new->compress_type = ordered->compress_type;
2710         new->root = RB_ROOT;
2711         INIT_LIST_HEAD(&new->head);
2712
2713         path = btrfs_alloc_path();
2714         if (!path)
2715                 goto out_kfree;
2716
2717         key.objectid = btrfs_ino(inode);
2718         key.type = BTRFS_EXTENT_DATA_KEY;
2719         key.offset = new->file_pos;
2720
2721         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2722         if (ret < 0)
2723                 goto out_free_path;
2724         if (ret > 0 && path->slots[0] > 0)
2725                 path->slots[0]--;
2726
2727         /* find out all the old extents for the file range */
2728         while (1) {
2729                 struct btrfs_file_extent_item *extent;
2730                 struct extent_buffer *l;
2731                 int slot;
2732                 u64 num_bytes;
2733                 u64 offset;
2734                 u64 end;
2735                 u64 disk_bytenr;
2736                 u64 extent_offset;
2737
2738                 l = path->nodes[0];
2739                 slot = path->slots[0];
2740
2741                 if (slot >= btrfs_header_nritems(l)) {
2742                         ret = btrfs_next_leaf(root, path);
2743                         if (ret < 0)
2744                                 goto out_free_path;
2745                         else if (ret > 0)
2746                                 break;
2747                         continue;
2748                 }
2749
2750                 btrfs_item_key_to_cpu(l, &key, slot);
2751
2752                 if (key.objectid != btrfs_ino(inode))
2753                         break;
2754                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2755                         break;
2756                 if (key.offset >= new->file_pos + new->len)
2757                         break;
2758
2759                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2760
2761                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2762                 if (key.offset + num_bytes < new->file_pos)
2763                         goto next;
2764
2765                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2766                 if (!disk_bytenr)
2767                         goto next;
2768
2769                 extent_offset = btrfs_file_extent_offset(l, extent);
2770
2771                 old = kmalloc(sizeof(*old), GFP_NOFS);
2772                 if (!old)
2773                         goto out_free_path;
2774
2775                 offset = max(new->file_pos, key.offset);
2776                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2777
2778                 old->bytenr = disk_bytenr;
2779                 old->extent_offset = extent_offset;
2780                 old->offset = offset - key.offset;
2781                 old->len = end - offset;
2782                 old->new = new;
2783                 old->count = 0;
2784                 list_add_tail(&old->list, &new->head);
2785 next:
2786                 path->slots[0]++;
2787                 cond_resched();
2788         }
2789
2790         btrfs_free_path(path);
2791         atomic_inc(&root->fs_info->defrag_running);
2792
2793         return new;
2794
2795 out_free_path:
2796         btrfs_free_path(path);
2797 out_kfree:
2798         free_sa_defrag_extent(new);
2799         return NULL;
2800 }
2801
2802 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2803                                          u64 start, u64 len)
2804 {
2805         struct btrfs_block_group_cache *cache;
2806
2807         cache = btrfs_lookup_block_group(root->fs_info, start);
2808         ASSERT(cache);
2809
2810         spin_lock(&cache->lock);
2811         cache->delalloc_bytes -= len;
2812         spin_unlock(&cache->lock);
2813
2814         btrfs_put_block_group(cache);
2815 }
2816
2817 /* as ordered data IO finishes, this gets called so we can finish
2818  * an ordered extent if the range of bytes in the file it covers are
2819  * fully written.
2820  */
2821 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2822 {
2823         struct inode *inode = ordered_extent->inode;
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_trans_handle *trans = NULL;
2826         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2827         struct extent_state *cached_state = NULL;
2828         struct new_sa_defrag_extent *new = NULL;
2829         int compress_type = 0;
2830         int ret = 0;
2831         u64 logical_len = ordered_extent->len;
2832         bool nolock;
2833         bool truncated = false;
2834
2835         nolock = btrfs_is_free_space_inode(inode);
2836
2837         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2838                 ret = -EIO;
2839                 goto out;
2840         }
2841
2842         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2843                                      ordered_extent->file_offset +
2844                                      ordered_extent->len - 1);
2845
2846         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2847                 truncated = true;
2848                 logical_len = ordered_extent->truncated_len;
2849                 /* Truncated the entire extent, don't bother adding */
2850                 if (!logical_len)
2851                         goto out;
2852         }
2853
2854         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2855                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2856
2857                 /*
2858                  * For mwrite(mmap + memset to write) case, we still reserve
2859                  * space for NOCOW range.
2860                  * As NOCOW won't cause a new delayed ref, just free the space
2861                  */
2862                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2863                                        ordered_extent->len);
2864                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2865                 if (nolock)
2866                         trans = btrfs_join_transaction_nolock(root);
2867                 else
2868                         trans = btrfs_join_transaction(root);
2869                 if (IS_ERR(trans)) {
2870                         ret = PTR_ERR(trans);
2871                         trans = NULL;
2872                         goto out;
2873                 }
2874                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2875                 ret = btrfs_update_inode_fallback(trans, root, inode);
2876                 if (ret) /* -ENOMEM or corruption */
2877                         btrfs_abort_transaction(trans, root, ret);
2878                 goto out;
2879         }
2880
2881         lock_extent_bits(io_tree, ordered_extent->file_offset,
2882                          ordered_extent->file_offset + ordered_extent->len - 1,
2883                          &cached_state);
2884
2885         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2886                         ordered_extent->file_offset + ordered_extent->len - 1,
2887                         EXTENT_DEFRAG, 1, cached_state);
2888         if (ret) {
2889                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2890                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2891                         /* the inode is shared */
2892                         new = record_old_file_extents(inode, ordered_extent);
2893
2894                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2895                         ordered_extent->file_offset + ordered_extent->len - 1,
2896                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2897         }
2898
2899         if (nolock)
2900                 trans = btrfs_join_transaction_nolock(root);
2901         else
2902                 trans = btrfs_join_transaction(root);
2903         if (IS_ERR(trans)) {
2904                 ret = PTR_ERR(trans);
2905                 trans = NULL;
2906                 goto out_unlock;
2907         }
2908
2909         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2910
2911         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2912                 compress_type = ordered_extent->compress_type;
2913         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2914                 BUG_ON(compress_type);
2915                 ret = btrfs_mark_extent_written(trans, inode,
2916                                                 ordered_extent->file_offset,
2917                                                 ordered_extent->file_offset +
2918                                                 logical_len);
2919         } else {
2920                 BUG_ON(root == root->fs_info->tree_root);
2921                 ret = insert_reserved_file_extent(trans, inode,
2922                                                 ordered_extent->file_offset,
2923                                                 ordered_extent->start,
2924                                                 ordered_extent->disk_len,
2925                                                 logical_len, logical_len,
2926                                                 compress_type, 0, 0,
2927                                                 BTRFS_FILE_EXTENT_REG);
2928                 if (!ret)
2929                         btrfs_release_delalloc_bytes(root,
2930                                                      ordered_extent->start,
2931                                                      ordered_extent->disk_len);
2932         }
2933         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2934                            ordered_extent->file_offset, ordered_extent->len,
2935                            trans->transid);
2936         if (ret < 0) {
2937                 btrfs_abort_transaction(trans, root, ret);
2938                 goto out_unlock;
2939         }
2940
2941         add_pending_csums(trans, inode, ordered_extent->file_offset,
2942                           &ordered_extent->list);
2943
2944         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2945         ret = btrfs_update_inode_fallback(trans, root, inode);
2946         if (ret) { /* -ENOMEM or corruption */
2947                 btrfs_abort_transaction(trans, root, ret);
2948                 goto out_unlock;
2949         }
2950         ret = 0;
2951 out_unlock:
2952         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2953                              ordered_extent->file_offset +
2954                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2955 out:
2956         if (root != root->fs_info->tree_root)
2957                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2958         if (trans)
2959                 btrfs_end_transaction(trans, root);
2960
2961         if (ret || truncated) {
2962                 u64 start, end;
2963
2964                 if (truncated)
2965                         start = ordered_extent->file_offset + logical_len;
2966                 else
2967                         start = ordered_extent->file_offset;
2968                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2969                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2970
2971                 /* Drop the cache for the part of the extent we didn't write. */
2972                 btrfs_drop_extent_cache(inode, start, end, 0);
2973
2974                 /*
2975                  * If the ordered extent had an IOERR or something else went
2976                  * wrong we need to return the space for this ordered extent
2977                  * back to the allocator.  We only free the extent in the
2978                  * truncated case if we didn't write out the extent at all.
2979                  */
2980                 if ((ret || !logical_len) &&
2981                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2982                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2983                         btrfs_free_reserved_extent(root, ordered_extent->start,
2984                                                    ordered_extent->disk_len, 1);
2985         }
2986
2987
2988         /*
2989          * This needs to be done to make sure anybody waiting knows we are done
2990          * updating everything for this ordered extent.
2991          */
2992         btrfs_remove_ordered_extent(inode, ordered_extent);
2993
2994         /* for snapshot-aware defrag */
2995         if (new) {
2996                 if (ret) {
2997                         free_sa_defrag_extent(new);
2998                         atomic_dec(&root->fs_info->defrag_running);
2999                 } else {
3000                         relink_file_extents(new);
3001                 }
3002         }
3003
3004         /* once for us */
3005         btrfs_put_ordered_extent(ordered_extent);
3006         /* once for the tree */
3007         btrfs_put_ordered_extent(ordered_extent);
3008
3009         return ret;
3010 }
3011
3012 static void finish_ordered_fn(struct btrfs_work *work)
3013 {
3014         struct btrfs_ordered_extent *ordered_extent;
3015         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3016         btrfs_finish_ordered_io(ordered_extent);
3017 }
3018
3019 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3020                                 struct extent_state *state, int uptodate)
3021 {
3022         struct inode *inode = page->mapping->host;
3023         struct btrfs_root *root = BTRFS_I(inode)->root;
3024         struct btrfs_ordered_extent *ordered_extent = NULL;
3025         struct btrfs_workqueue *wq;
3026         btrfs_work_func_t func;
3027
3028         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3029
3030         ClearPagePrivate2(page);
3031         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3032                                             end - start + 1, uptodate))
3033                 return 0;
3034
3035         if (btrfs_is_free_space_inode(inode)) {
3036                 wq = root->fs_info->endio_freespace_worker;
3037                 func = btrfs_freespace_write_helper;
3038         } else {
3039                 wq = root->fs_info->endio_write_workers;
3040                 func = btrfs_endio_write_helper;
3041         }
3042
3043         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3044                         NULL);
3045         btrfs_queue_work(wq, &ordered_extent->work);
3046
3047         return 0;
3048 }
3049
3050 static int __readpage_endio_check(struct inode *inode,
3051                                   struct btrfs_io_bio *io_bio,
3052                                   int icsum, struct page *page,
3053                                   int pgoff, u64 start, size_t len)
3054 {
3055         char *kaddr;
3056         u32 csum_expected;
3057         u32 csum = ~(u32)0;
3058
3059         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3060
3061         kaddr = kmap_atomic(page);
3062         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3063         btrfs_csum_final(csum, (char *)&csum);
3064         if (csum != csum_expected)
3065                 goto zeroit;
3066
3067         kunmap_atomic(kaddr);
3068         return 0;
3069 zeroit:
3070         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3071                 "csum failed ino %llu off %llu csum %u expected csum %u",
3072                            btrfs_ino(inode), start, csum, csum_expected);
3073         memset(kaddr + pgoff, 1, len);
3074         flush_dcache_page(page);
3075         kunmap_atomic(kaddr);
3076         if (csum_expected == 0)
3077                 return 0;
3078         return -EIO;
3079 }
3080
3081 /*
3082  * when reads are done, we need to check csums to verify the data is correct
3083  * if there's a match, we allow the bio to finish.  If not, the code in
3084  * extent_io.c will try to find good copies for us.
3085  */
3086 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3087                                       u64 phy_offset, struct page *page,
3088                                       u64 start, u64 end, int mirror)
3089 {
3090         size_t offset = start - page_offset(page);
3091         struct inode *inode = page->mapping->host;
3092         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3093         struct btrfs_root *root = BTRFS_I(inode)->root;
3094
3095         if (PageChecked(page)) {
3096                 ClearPageChecked(page);
3097                 return 0;
3098         }
3099
3100         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3101                 return 0;
3102
3103         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3104             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3105                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3106                                   GFP_NOFS);
3107                 return 0;
3108         }
3109
3110         phy_offset >>= inode->i_sb->s_blocksize_bits;
3111         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3112                                       start, (size_t)(end - start + 1));
3113 }
3114
3115 void btrfs_add_delayed_iput(struct inode *inode)
3116 {
3117         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3118         struct btrfs_inode *binode = BTRFS_I(inode);
3119
3120         if (atomic_add_unless(&inode->i_count, -1, 1))
3121                 return;
3122
3123         spin_lock(&fs_info->delayed_iput_lock);
3124         if (binode->delayed_iput_count == 0) {
3125                 ASSERT(list_empty(&binode->delayed_iput));
3126                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3127         } else {
3128                 binode->delayed_iput_count++;
3129         }
3130         spin_unlock(&fs_info->delayed_iput_lock);
3131 }
3132
3133 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3134 {
3135         struct btrfs_fs_info *fs_info = root->fs_info;
3136
3137         down_read(&fs_info->delayed_iput_sem);
3138         spin_lock(&fs_info->delayed_iput_lock);
3139         while (!list_empty(&fs_info->delayed_iputs)) {
3140                 struct btrfs_inode *inode;
3141
3142                 inode = list_first_entry(&fs_info->delayed_iputs,
3143                                 struct btrfs_inode, delayed_iput);
3144                 if (inode->delayed_iput_count) {
3145                         inode->delayed_iput_count--;
3146                         list_move_tail(&inode->delayed_iput,
3147                                         &fs_info->delayed_iputs);
3148                 } else {
3149                         list_del_init(&inode->delayed_iput);
3150                 }
3151                 spin_unlock(&fs_info->delayed_iput_lock);
3152                 iput(&inode->vfs_inode);
3153                 spin_lock(&fs_info->delayed_iput_lock);
3154         }
3155         spin_unlock(&fs_info->delayed_iput_lock);
3156         up_read(&root->fs_info->delayed_iput_sem);
3157 }
3158
3159 /*
3160  * This is called in transaction commit time. If there are no orphan
3161  * files in the subvolume, it removes orphan item and frees block_rsv
3162  * structure.
3163  */
3164 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3165                               struct btrfs_root *root)
3166 {
3167         struct btrfs_block_rsv *block_rsv;
3168         int ret;
3169
3170         if (atomic_read(&root->orphan_inodes) ||
3171             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3172                 return;
3173
3174         spin_lock(&root->orphan_lock);
3175         if (atomic_read(&root->orphan_inodes)) {
3176                 spin_unlock(&root->orphan_lock);
3177                 return;
3178         }
3179
3180         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3181                 spin_unlock(&root->orphan_lock);
3182                 return;
3183         }
3184
3185         block_rsv = root->orphan_block_rsv;
3186         root->orphan_block_rsv = NULL;
3187         spin_unlock(&root->orphan_lock);
3188
3189         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3190             btrfs_root_refs(&root->root_item) > 0) {
3191                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3192                                             root->root_key.objectid);
3193                 if (ret)
3194                         btrfs_abort_transaction(trans, root, ret);
3195                 else
3196                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3197                                   &root->state);
3198         }
3199
3200         if (block_rsv) {
3201                 WARN_ON(block_rsv->size > 0);
3202                 btrfs_free_block_rsv(root, block_rsv);
3203         }
3204 }
3205
3206 /*
3207  * This creates an orphan entry for the given inode in case something goes
3208  * wrong in the middle of an unlink/truncate.
3209  *
3210  * NOTE: caller of this function should reserve 5 units of metadata for
3211  *       this function.
3212  */
3213 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3214 {
3215         struct btrfs_root *root = BTRFS_I(inode)->root;
3216         struct btrfs_block_rsv *block_rsv = NULL;
3217         int reserve = 0;
3218         int insert = 0;
3219         int ret;
3220
3221         if (!root->orphan_block_rsv) {
3222                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3223                 if (!block_rsv)
3224                         return -ENOMEM;
3225         }
3226
3227         spin_lock(&root->orphan_lock);
3228         if (!root->orphan_block_rsv) {
3229                 root->orphan_block_rsv = block_rsv;
3230         } else if (block_rsv) {
3231                 btrfs_free_block_rsv(root, block_rsv);
3232                 block_rsv = NULL;
3233         }
3234
3235         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3236                               &BTRFS_I(inode)->runtime_flags)) {
3237 #if 0
3238                 /*
3239                  * For proper ENOSPC handling, we should do orphan
3240                  * cleanup when mounting. But this introduces backward
3241                  * compatibility issue.
3242                  */
3243                 if (!xchg(&root->orphan_item_inserted, 1))
3244                         insert = 2;
3245                 else
3246                         insert = 1;
3247 #endif
3248                 insert = 1;
3249                 atomic_inc(&root->orphan_inodes);
3250         }
3251
3252         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3253                               &BTRFS_I(inode)->runtime_flags))
3254                 reserve = 1;
3255         spin_unlock(&root->orphan_lock);
3256
3257         /* grab metadata reservation from transaction handle */
3258         if (reserve) {
3259                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3260                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3261         }
3262
3263         /* insert an orphan item to track this unlinked/truncated file */
3264         if (insert >= 1) {
3265                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3266                 if (ret) {
3267                         atomic_dec(&root->orphan_inodes);
3268                         if (reserve) {
3269                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3270                                           &BTRFS_I(inode)->runtime_flags);
3271                                 btrfs_orphan_release_metadata(inode);
3272                         }
3273                         if (ret != -EEXIST) {
3274                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3275                                           &BTRFS_I(inode)->runtime_flags);
3276                                 btrfs_abort_transaction(trans, root, ret);
3277                                 return ret;
3278                         }
3279                 }
3280                 ret = 0;
3281         }
3282
3283         /* insert an orphan item to track subvolume contains orphan files */
3284         if (insert >= 2) {
3285                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3286                                                root->root_key.objectid);
3287                 if (ret && ret != -EEXIST) {
3288                         btrfs_abort_transaction(trans, root, ret);
3289                         return ret;
3290                 }
3291         }
3292         return 0;
3293 }
3294
3295 /*
3296  * We have done the truncate/delete so we can go ahead and remove the orphan
3297  * item for this particular inode.
3298  */
3299 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3300                             struct inode *inode)
3301 {
3302         struct btrfs_root *root = BTRFS_I(inode)->root;
3303         int delete_item = 0;
3304         int release_rsv = 0;
3305         int ret = 0;
3306
3307         spin_lock(&root->orphan_lock);
3308         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3309                                &BTRFS_I(inode)->runtime_flags))
3310                 delete_item = 1;
3311
3312         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3313                                &BTRFS_I(inode)->runtime_flags))
3314                 release_rsv = 1;
3315         spin_unlock(&root->orphan_lock);
3316
3317         if (delete_item) {
3318                 atomic_dec(&root->orphan_inodes);
3319                 if (trans)
3320                         ret = btrfs_del_orphan_item(trans, root,
3321                                                     btrfs_ino(inode));
3322         }
3323
3324         if (release_rsv)
3325                 btrfs_orphan_release_metadata(inode);
3326
3327         return ret;
3328 }
3329
3330 /*
3331  * this cleans up any orphans that may be left on the list from the last use
3332  * of this root.
3333  */
3334 int btrfs_orphan_cleanup(struct btrfs_root *root)
3335 {
3336         struct btrfs_path *path;
3337         struct extent_buffer *leaf;
3338         struct btrfs_key key, found_key;
3339         struct btrfs_trans_handle *trans;
3340         struct inode *inode;
3341         u64 last_objectid = 0;
3342         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3343
3344         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3345                 return 0;
3346
3347         path = btrfs_alloc_path();
3348         if (!path) {
3349                 ret = -ENOMEM;
3350                 goto out;
3351         }
3352         path->reada = READA_BACK;
3353
3354         key.objectid = BTRFS_ORPHAN_OBJECTID;
3355         key.type = BTRFS_ORPHAN_ITEM_KEY;
3356         key.offset = (u64)-1;
3357
3358         while (1) {
3359                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3360                 if (ret < 0)
3361                         goto out;
3362
3363                 /*
3364                  * if ret == 0 means we found what we were searching for, which
3365                  * is weird, but possible, so only screw with path if we didn't
3366                  * find the key and see if we have stuff that matches
3367                  */
3368                 if (ret > 0) {
3369                         ret = 0;
3370                         if (path->slots[0] == 0)
3371                                 break;
3372                         path->slots[0]--;
3373                 }
3374
3375                 /* pull out the item */
3376                 leaf = path->nodes[0];
3377                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3378
3379                 /* make sure the item matches what we want */
3380                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3381                         break;
3382                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3383                         break;
3384
3385                 /* release the path since we're done with it */
3386                 btrfs_release_path(path);
3387
3388                 /*
3389                  * this is where we are basically btrfs_lookup, without the
3390                  * crossing root thing.  we store the inode number in the
3391                  * offset of the orphan item.
3392                  */
3393
3394                 if (found_key.offset == last_objectid) {
3395                         btrfs_err(root->fs_info,
3396                                 "Error removing orphan entry, stopping orphan cleanup");
3397                         ret = -EINVAL;
3398                         goto out;
3399                 }
3400
3401                 last_objectid = found_key.offset;
3402
3403                 found_key.objectid = found_key.offset;
3404                 found_key.type = BTRFS_INODE_ITEM_KEY;
3405                 found_key.offset = 0;
3406                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3407                 ret = PTR_ERR_OR_ZERO(inode);
3408                 if (ret && ret != -ESTALE)
3409                         goto out;
3410
3411                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3412                         struct btrfs_root *dead_root;
3413                         struct btrfs_fs_info *fs_info = root->fs_info;
3414                         int is_dead_root = 0;
3415
3416                         /*
3417                          * this is an orphan in the tree root. Currently these
3418                          * could come from 2 sources:
3419                          *  a) a snapshot deletion in progress
3420                          *  b) a free space cache inode
3421                          * We need to distinguish those two, as the snapshot
3422                          * orphan must not get deleted.
3423                          * find_dead_roots already ran before us, so if this
3424                          * is a snapshot deletion, we should find the root
3425                          * in the dead_roots list
3426                          */
3427                         spin_lock(&fs_info->trans_lock);
3428                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3429                                             root_list) {
3430                                 if (dead_root->root_key.objectid ==
3431                                     found_key.objectid) {
3432                                         is_dead_root = 1;
3433                                         break;
3434                                 }
3435                         }
3436                         spin_unlock(&fs_info->trans_lock);
3437                         if (is_dead_root) {
3438                                 /* prevent this orphan from being found again */
3439                                 key.offset = found_key.objectid - 1;
3440                                 continue;
3441                         }
3442                 }
3443                 /*
3444                  * Inode is already gone but the orphan item is still there,
3445                  * kill the orphan item.
3446                  */
3447                 if (ret == -ESTALE) {
3448                         trans = btrfs_start_transaction(root, 1);
3449                         if (IS_ERR(trans)) {
3450                                 ret = PTR_ERR(trans);
3451                                 goto out;
3452                         }
3453                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3454                                 found_key.objectid);
3455                         ret = btrfs_del_orphan_item(trans, root,
3456                                                     found_key.objectid);
3457                         btrfs_end_transaction(trans, root);
3458                         if (ret)
3459                                 goto out;
3460                         continue;
3461                 }
3462
3463                 /*
3464                  * add this inode to the orphan list so btrfs_orphan_del does
3465                  * the proper thing when we hit it
3466                  */
3467                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3468                         &BTRFS_I(inode)->runtime_flags);
3469                 atomic_inc(&root->orphan_inodes);
3470
3471                 /* if we have links, this was a truncate, lets do that */
3472                 if (inode->i_nlink) {
3473                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3474                                 iput(inode);
3475                                 continue;
3476                         }
3477                         nr_truncate++;
3478
3479                         /* 1 for the orphan item deletion. */
3480                         trans = btrfs_start_transaction(root, 1);
3481                         if (IS_ERR(trans)) {
3482                                 iput(inode);
3483                                 ret = PTR_ERR(trans);
3484                                 goto out;
3485                         }
3486                         ret = btrfs_orphan_add(trans, inode);
3487                         btrfs_end_transaction(trans, root);
3488                         if (ret) {
3489                                 iput(inode);
3490                                 goto out;
3491                         }
3492
3493                         ret = btrfs_truncate(inode);
3494                         if (ret)
3495                                 btrfs_orphan_del(NULL, inode);
3496                 } else {
3497                         nr_unlink++;
3498                 }
3499
3500                 /* this will do delete_inode and everything for us */
3501                 iput(inode);
3502                 if (ret)
3503                         goto out;
3504         }
3505         /* release the path since we're done with it */
3506         btrfs_release_path(path);
3507
3508         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3509
3510         if (root->orphan_block_rsv)
3511                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3512                                         (u64)-1);
3513
3514         if (root->orphan_block_rsv ||
3515             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3516                 trans = btrfs_join_transaction(root);
3517                 if (!IS_ERR(trans))
3518                         btrfs_end_transaction(trans, root);
3519         }
3520
3521         if (nr_unlink)
3522                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3523         if (nr_truncate)
3524                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3525
3526 out:
3527         if (ret)
3528                 btrfs_err(root->fs_info,
3529                         "could not do orphan cleanup %d", ret);
3530         btrfs_free_path(path);
3531         return ret;
3532 }
3533
3534 /*
3535  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3536  * don't find any xattrs, we know there can't be any acls.
3537  *
3538  * slot is the slot the inode is in, objectid is the objectid of the inode
3539  */
3540 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3541                                           int slot, u64 objectid,
3542                                           int *first_xattr_slot)
3543 {
3544         u32 nritems = btrfs_header_nritems(leaf);
3545         struct btrfs_key found_key;
3546         static u64 xattr_access = 0;
3547         static u64 xattr_default = 0;
3548         int scanned = 0;
3549
3550         if (!xattr_access) {
3551                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3552                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3553                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3554                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3555         }
3556
3557         slot++;
3558         *first_xattr_slot = -1;
3559         while (slot < nritems) {
3560                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3561
3562                 /* we found a different objectid, there must not be acls */
3563                 if (found_key.objectid != objectid)
3564                         return 0;
3565
3566                 /* we found an xattr, assume we've got an acl */
3567                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3568                         if (*first_xattr_slot == -1)
3569                                 *first_xattr_slot = slot;
3570                         if (found_key.offset == xattr_access ||
3571                             found_key.offset == xattr_default)
3572                                 return 1;
3573                 }
3574
3575                 /*
3576                  * we found a key greater than an xattr key, there can't
3577                  * be any acls later on
3578                  */
3579                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3580                         return 0;
3581
3582                 slot++;
3583                 scanned++;
3584
3585                 /*
3586                  * it goes inode, inode backrefs, xattrs, extents,
3587                  * so if there are a ton of hard links to an inode there can
3588                  * be a lot of backrefs.  Don't waste time searching too hard,
3589                  * this is just an optimization
3590                  */
3591                 if (scanned >= 8)
3592                         break;
3593         }
3594         /* we hit the end of the leaf before we found an xattr or
3595          * something larger than an xattr.  We have to assume the inode
3596          * has acls
3597          */
3598         if (*first_xattr_slot == -1)
3599                 *first_xattr_slot = slot;
3600         return 1;
3601 }
3602
3603 /*
3604  * read an inode from the btree into the in-memory inode
3605  */
3606 static void btrfs_read_locked_inode(struct inode *inode)
3607 {
3608         struct btrfs_path *path;
3609         struct extent_buffer *leaf;
3610         struct btrfs_inode_item *inode_item;
3611         struct btrfs_root *root = BTRFS_I(inode)->root;
3612         struct btrfs_key location;
3613         unsigned long ptr;
3614         int maybe_acls;
3615         u32 rdev;
3616         int ret;
3617         bool filled = false;
3618         int first_xattr_slot;
3619
3620         ret = btrfs_fill_inode(inode, &rdev);
3621         if (!ret)
3622                 filled = true;
3623
3624         path = btrfs_alloc_path();
3625         if (!path)
3626                 goto make_bad;
3627
3628         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3629
3630         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3631         if (ret)
3632                 goto make_bad;
3633
3634         leaf = path->nodes[0];
3635
3636         if (filled)
3637                 goto cache_index;
3638
3639         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3640                                     struct btrfs_inode_item);
3641         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3642         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3643         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3644         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3645         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3646
3647         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3648         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3649
3650         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3651         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3652
3653         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3654         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3655
3656         BTRFS_I(inode)->i_otime.tv_sec =
3657                 btrfs_timespec_sec(leaf, &inode_item->otime);
3658         BTRFS_I(inode)->i_otime.tv_nsec =
3659                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3660
3661         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3662         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3663         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3664
3665         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3666         inode->i_generation = BTRFS_I(inode)->generation;
3667         inode->i_rdev = 0;
3668         rdev = btrfs_inode_rdev(leaf, inode_item);
3669
3670         BTRFS_I(inode)->index_cnt = (u64)-1;
3671         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3672
3673 cache_index:
3674         /*
3675          * If we were modified in the current generation and evicted from memory
3676          * and then re-read we need to do a full sync since we don't have any
3677          * idea about which extents were modified before we were evicted from
3678          * cache.
3679          *
3680          * This is required for both inode re-read from disk and delayed inode
3681          * in delayed_nodes_tree.
3682          */
3683         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3684                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3685                         &BTRFS_I(inode)->runtime_flags);
3686
3687         /*
3688          * We don't persist the id of the transaction where an unlink operation
3689          * against the inode was last made. So here we assume the inode might
3690          * have been evicted, and therefore the exact value of last_unlink_trans
3691          * lost, and set it to last_trans to avoid metadata inconsistencies
3692          * between the inode and its parent if the inode is fsync'ed and the log
3693          * replayed. For example, in the scenario:
3694          *
3695          * touch mydir/foo
3696          * ln mydir/foo mydir/bar
3697          * sync
3698          * unlink mydir/bar
3699          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3700          * xfs_io -c fsync mydir/foo
3701          * <power failure>
3702          * mount fs, triggers fsync log replay
3703          *
3704          * We must make sure that when we fsync our inode foo we also log its
3705          * parent inode, otherwise after log replay the parent still has the
3706          * dentry with the "bar" name but our inode foo has a link count of 1
3707          * and doesn't have an inode ref with the name "bar" anymore.
3708          *
3709          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3710          * but it guarantees correctness at the expense of ocassional full
3711          * transaction commits on fsync if our inode is a directory, or if our
3712          * inode is not a directory, logging its parent unnecessarily.
3713          */
3714         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3715
3716         path->slots[0]++;
3717         if (inode->i_nlink != 1 ||
3718             path->slots[0] >= btrfs_header_nritems(leaf))
3719                 goto cache_acl;
3720
3721         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3722         if (location.objectid != btrfs_ino(inode))
3723                 goto cache_acl;
3724
3725         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3726         if (location.type == BTRFS_INODE_REF_KEY) {
3727                 struct btrfs_inode_ref *ref;
3728
3729                 ref = (struct btrfs_inode_ref *)ptr;
3730                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3731         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3732                 struct btrfs_inode_extref *extref;
3733
3734                 extref = (struct btrfs_inode_extref *)ptr;
3735                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3736                                                                      extref);
3737         }
3738 cache_acl:
3739         /*
3740          * try to precache a NULL acl entry for files that don't have
3741          * any xattrs or acls
3742          */
3743         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3744                                            btrfs_ino(inode), &first_xattr_slot);
3745         if (first_xattr_slot != -1) {
3746                 path->slots[0] = first_xattr_slot;
3747                 ret = btrfs_load_inode_props(inode, path);
3748                 if (ret)
3749                         btrfs_err(root->fs_info,
3750                                   "error loading props for ino %llu (root %llu): %d",
3751                                   btrfs_ino(inode),
3752                                   root->root_key.objectid, ret);
3753         }
3754         btrfs_free_path(path);
3755
3756         if (!maybe_acls)
3757                 cache_no_acl(inode);
3758
3759         switch (inode->i_mode & S_IFMT) {
3760         case S_IFREG:
3761                 inode->i_mapping->a_ops = &btrfs_aops;
3762                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3763                 inode->i_fop = &btrfs_file_operations;
3764                 inode->i_op = &btrfs_file_inode_operations;
3765                 break;
3766         case S_IFDIR:
3767                 inode->i_fop = &btrfs_dir_file_operations;
3768                 if (root == root->fs_info->tree_root)
3769                         inode->i_op = &btrfs_dir_ro_inode_operations;
3770                 else
3771                         inode->i_op = &btrfs_dir_inode_operations;
3772                 break;
3773         case S_IFLNK:
3774                 inode->i_op = &btrfs_symlink_inode_operations;
3775                 inode_nohighmem(inode);
3776                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3777                 break;
3778         default:
3779                 inode->i_op = &btrfs_special_inode_operations;
3780                 init_special_inode(inode, inode->i_mode, rdev);
3781                 break;
3782         }
3783
3784         btrfs_update_iflags(inode);
3785         return;
3786
3787 make_bad:
3788         btrfs_free_path(path);
3789         make_bad_inode(inode);
3790 }
3791
3792 /*
3793  * given a leaf and an inode, copy the inode fields into the leaf
3794  */
3795 static void fill_inode_item(struct btrfs_trans_handle *trans,
3796                             struct extent_buffer *leaf,
3797                             struct btrfs_inode_item *item,
3798                             struct inode *inode)
3799 {
3800         struct btrfs_map_token token;
3801
3802         btrfs_init_map_token(&token);
3803
3804         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3805         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3806         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3807                                    &token);
3808         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3809         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3810
3811         btrfs_set_token_timespec_sec(leaf, &item->atime,
3812                                      inode->i_atime.tv_sec, &token);
3813         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3814                                       inode->i_atime.tv_nsec, &token);
3815
3816         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3817                                      inode->i_mtime.tv_sec, &token);
3818         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3819                                       inode->i_mtime.tv_nsec, &token);
3820
3821         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3822                                      inode->i_ctime.tv_sec, &token);
3823         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3824                                       inode->i_ctime.tv_nsec, &token);
3825
3826         btrfs_set_token_timespec_sec(leaf, &item->otime,
3827                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3828         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3829                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3830
3831         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3832                                      &token);
3833         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3834                                          &token);
3835         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3836         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3837         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3838         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3839         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3840 }
3841
3842 /*
3843  * copy everything in the in-memory inode into the btree.
3844  */
3845 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3846                                 struct btrfs_root *root, struct inode *inode)
3847 {
3848         struct btrfs_inode_item *inode_item;
3849         struct btrfs_path *path;
3850         struct extent_buffer *leaf;
3851         int ret;
3852
3853         path = btrfs_alloc_path();
3854         if (!path)
3855                 return -ENOMEM;
3856
3857         path->leave_spinning = 1;
3858         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3859                                  1);
3860         if (ret) {
3861                 if (ret > 0)
3862                         ret = -ENOENT;
3863                 goto failed;
3864         }
3865
3866         leaf = path->nodes[0];
3867         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3868                                     struct btrfs_inode_item);
3869
3870         fill_inode_item(trans, leaf, inode_item, inode);
3871         btrfs_mark_buffer_dirty(leaf);
3872         btrfs_set_inode_last_trans(trans, inode);
3873         ret = 0;
3874 failed:
3875         btrfs_free_path(path);
3876         return ret;
3877 }
3878
3879 /*
3880  * copy everything in the in-memory inode into the btree.
3881  */
3882 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3883                                 struct btrfs_root *root, struct inode *inode)
3884 {
3885         int ret;
3886
3887         /*
3888          * If the inode is a free space inode, we can deadlock during commit
3889          * if we put it into the delayed code.
3890          *
3891          * The data relocation inode should also be directly updated
3892          * without delay
3893          */
3894         if (!btrfs_is_free_space_inode(inode)
3895             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3896             && !root->fs_info->log_root_recovering) {
3897                 btrfs_update_root_times(trans, root);
3898
3899                 ret = btrfs_delayed_update_inode(trans, root, inode);
3900                 if (!ret)
3901                         btrfs_set_inode_last_trans(trans, inode);
3902                 return ret;
3903         }
3904
3905         return btrfs_update_inode_item(trans, root, inode);
3906 }
3907
3908 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3909                                          struct btrfs_root *root,
3910                                          struct inode *inode)
3911 {
3912         int ret;
3913
3914         ret = btrfs_update_inode(trans, root, inode);
3915         if (ret == -ENOSPC)
3916                 return btrfs_update_inode_item(trans, root, inode);
3917         return ret;
3918 }
3919
3920 /*
3921  * unlink helper that gets used here in inode.c and in the tree logging
3922  * recovery code.  It remove a link in a directory with a given name, and
3923  * also drops the back refs in the inode to the directory
3924  */
3925 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3926                                 struct btrfs_root *root,
3927                                 struct inode *dir, struct inode *inode,
3928                                 const char *name, int name_len)
3929 {
3930         struct btrfs_path *path;
3931         int ret = 0;
3932         struct extent_buffer *leaf;
3933         struct btrfs_dir_item *di;
3934         struct btrfs_key key;
3935         u64 index;
3936         u64 ino = btrfs_ino(inode);
3937         u64 dir_ino = btrfs_ino(dir);
3938
3939         path = btrfs_alloc_path();
3940         if (!path) {
3941                 ret = -ENOMEM;
3942                 goto out;
3943         }
3944
3945         path->leave_spinning = 1;
3946         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3947                                     name, name_len, -1);
3948         if (IS_ERR(di)) {
3949                 ret = PTR_ERR(di);
3950                 goto err;
3951         }
3952         if (!di) {
3953                 ret = -ENOENT;
3954                 goto err;
3955         }
3956         leaf = path->nodes[0];
3957         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3958         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3959         if (ret)
3960                 goto err;
3961         btrfs_release_path(path);
3962
3963         /*
3964          * If we don't have dir index, we have to get it by looking up
3965          * the inode ref, since we get the inode ref, remove it directly,
3966          * it is unnecessary to do delayed deletion.
3967          *
3968          * But if we have dir index, needn't search inode ref to get it.
3969          * Since the inode ref is close to the inode item, it is better
3970          * that we delay to delete it, and just do this deletion when
3971          * we update the inode item.
3972          */
3973         if (BTRFS_I(inode)->dir_index) {
3974                 ret = btrfs_delayed_delete_inode_ref(inode);
3975                 if (!ret) {
3976                         index = BTRFS_I(inode)->dir_index;
3977                         goto skip_backref;
3978                 }
3979         }
3980
3981         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3982                                   dir_ino, &index);
3983         if (ret) {
3984                 btrfs_info(root->fs_info,
3985                         "failed to delete reference to %.*s, inode %llu parent %llu",
3986                         name_len, name, ino, dir_ino);
3987                 btrfs_abort_transaction(trans, root, ret);
3988                 goto err;
3989         }
3990 skip_backref:
3991         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3992         if (ret) {
3993                 btrfs_abort_transaction(trans, root, ret);
3994                 goto err;
3995         }
3996
3997         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3998                                          inode, dir_ino);
3999         if (ret != 0 && ret != -ENOENT) {
4000                 btrfs_abort_transaction(trans, root, ret);
4001                 goto err;
4002         }
4003
4004         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4005                                            dir, index);
4006         if (ret == -ENOENT)
4007                 ret = 0;
4008         else if (ret)
4009                 btrfs_abort_transaction(trans, root, ret);
4010 err:
4011         btrfs_free_path(path);
4012         if (ret)
4013                 goto out;
4014
4015         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4016         inode_inc_iversion(inode);
4017         inode_inc_iversion(dir);
4018         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4019         ret = btrfs_update_inode(trans, root, dir);
4020 out:
4021         return ret;
4022 }
4023
4024 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4025                        struct btrfs_root *root,
4026                        struct inode *dir, struct inode *inode,
4027                        const char *name, int name_len)
4028 {
4029         int ret;
4030         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4031         if (!ret) {
4032                 drop_nlink(inode);
4033                 ret = btrfs_update_inode(trans, root, inode);
4034         }
4035         return ret;
4036 }
4037
4038 /*
4039  * helper to start transaction for unlink and rmdir.
4040  *
4041  * unlink and rmdir are special in btrfs, they do not always free space, so
4042  * if we cannot make our reservations the normal way try and see if there is
4043  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4044  * allow the unlink to occur.
4045  */
4046 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4047 {
4048         struct btrfs_root *root = BTRFS_I(dir)->root;
4049
4050         /*
4051          * 1 for the possible orphan item
4052          * 1 for the dir item
4053          * 1 for the dir index
4054          * 1 for the inode ref
4055          * 1 for the inode
4056          */
4057         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4058 }
4059
4060 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4061 {
4062         struct btrfs_root *root = BTRFS_I(dir)->root;
4063         struct btrfs_trans_handle *trans;
4064         struct inode *inode = d_inode(dentry);
4065         int ret;
4066
4067         trans = __unlink_start_trans(dir);
4068         if (IS_ERR(trans))
4069                 return PTR_ERR(trans);
4070
4071         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4072
4073         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4074                                  dentry->d_name.name, dentry->d_name.len);
4075         if (ret)
4076                 goto out;
4077
4078         if (inode->i_nlink == 0) {
4079                 ret = btrfs_orphan_add(trans, inode);
4080                 if (ret)
4081                         goto out;
4082         }
4083
4084 out:
4085         btrfs_end_transaction(trans, root);
4086         btrfs_btree_balance_dirty(root);
4087         return ret;
4088 }
4089
4090 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4091                         struct btrfs_root *root,
4092                         struct inode *dir, u64 objectid,
4093                         const char *name, int name_len)
4094 {
4095         struct btrfs_path *path;
4096         struct extent_buffer *leaf;
4097         struct btrfs_dir_item *di;
4098         struct btrfs_key key;
4099         u64 index;
4100         int ret;
4101         u64 dir_ino = btrfs_ino(dir);
4102
4103         path = btrfs_alloc_path();
4104         if (!path)
4105                 return -ENOMEM;
4106
4107         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4108                                    name, name_len, -1);
4109         if (IS_ERR_OR_NULL(di)) {
4110                 if (!di)
4111                         ret = -ENOENT;
4112                 else
4113                         ret = PTR_ERR(di);
4114                 goto out;
4115         }
4116
4117         leaf = path->nodes[0];
4118         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4119         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4120         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4121         if (ret) {
4122                 btrfs_abort_transaction(trans, root, ret);
4123                 goto out;
4124         }
4125         btrfs_release_path(path);
4126
4127         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4128                                  objectid, root->root_key.objectid,
4129                                  dir_ino, &index, name, name_len);
4130         if (ret < 0) {
4131                 if (ret != -ENOENT) {
4132                         btrfs_abort_transaction(trans, root, ret);
4133                         goto out;
4134                 }
4135                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4136                                                  name, name_len);
4137                 if (IS_ERR_OR_NULL(di)) {
4138                         if (!di)
4139                                 ret = -ENOENT;
4140                         else
4141                                 ret = PTR_ERR(di);
4142                         btrfs_abort_transaction(trans, root, ret);
4143                         goto out;
4144                 }
4145
4146                 leaf = path->nodes[0];
4147                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4148                 btrfs_release_path(path);
4149                 index = key.offset;
4150         }
4151         btrfs_release_path(path);
4152
4153         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4154         if (ret) {
4155                 btrfs_abort_transaction(trans, root, ret);
4156                 goto out;
4157         }
4158
4159         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4160         inode_inc_iversion(dir);
4161         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4162         ret = btrfs_update_inode_fallback(trans, root, dir);
4163         if (ret)
4164                 btrfs_abort_transaction(trans, root, ret);
4165 out:
4166         btrfs_free_path(path);
4167         return ret;
4168 }
4169
4170 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4171 {
4172         struct inode *inode = d_inode(dentry);
4173         int err = 0;
4174         struct btrfs_root *root = BTRFS_I(dir)->root;
4175         struct btrfs_trans_handle *trans;
4176
4177         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4178                 return -ENOTEMPTY;
4179         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4180                 return -EPERM;
4181
4182         trans = __unlink_start_trans(dir);
4183         if (IS_ERR(trans))
4184                 return PTR_ERR(trans);
4185
4186         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4187                 err = btrfs_unlink_subvol(trans, root, dir,
4188                                           BTRFS_I(inode)->location.objectid,
4189                                           dentry->d_name.name,
4190                                           dentry->d_name.len);
4191                 goto out;
4192         }
4193
4194         err = btrfs_orphan_add(trans, inode);
4195         if (err)
4196                 goto out;
4197
4198         /* now the directory is empty */
4199         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4200                                  dentry->d_name.name, dentry->d_name.len);
4201         if (!err)
4202                 btrfs_i_size_write(inode, 0);
4203 out:
4204         btrfs_end_transaction(trans, root);
4205         btrfs_btree_balance_dirty(root);
4206
4207         return err;
4208 }
4209
4210 static int truncate_space_check(struct btrfs_trans_handle *trans,
4211                                 struct btrfs_root *root,
4212                                 u64 bytes_deleted)
4213 {
4214         int ret;
4215
4216         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4217         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4218                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4219         if (!ret)
4220                 trans->bytes_reserved += bytes_deleted;
4221         return ret;
4222
4223 }
4224
4225 static int truncate_inline_extent(struct inode *inode,
4226                                   struct btrfs_path *path,
4227                                   struct btrfs_key *found_key,
4228                                   const u64 item_end,
4229                                   const u64 new_size)
4230 {
4231         struct extent_buffer *leaf = path->nodes[0];
4232         int slot = path->slots[0];
4233         struct btrfs_file_extent_item *fi;
4234         u32 size = (u32)(new_size - found_key->offset);
4235         struct btrfs_root *root = BTRFS_I(inode)->root;
4236
4237         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4238
4239         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4240                 loff_t offset = new_size;
4241                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4242
4243                 /*
4244                  * Zero out the remaining of the last page of our inline extent,
4245                  * instead of directly truncating our inline extent here - that
4246                  * would be much more complex (decompressing all the data, then
4247                  * compressing the truncated data, which might be bigger than
4248                  * the size of the inline extent, resize the extent, etc).
4249                  * We release the path because to get the page we might need to
4250                  * read the extent item from disk (data not in the page cache).
4251                  */
4252                 btrfs_release_path(path);
4253                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4254         }
4255
4256         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4257         size = btrfs_file_extent_calc_inline_size(size);
4258         btrfs_truncate_item(root, path, size, 1);
4259
4260         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4261                 inode_sub_bytes(inode, item_end + 1 - new_size);
4262
4263         return 0;
4264 }
4265
4266 /*
4267  * this can truncate away extent items, csum items and directory items.
4268  * It starts at a high offset and removes keys until it can't find
4269  * any higher than new_size
4270  *
4271  * csum items that cross the new i_size are truncated to the new size
4272  * as well.
4273  *
4274  * min_type is the minimum key type to truncate down to.  If set to 0, this
4275  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4276  */
4277 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4278                                struct btrfs_root *root,
4279                                struct inode *inode,
4280                                u64 new_size, u32 min_type)
4281 {
4282         struct btrfs_path *path;
4283         struct extent_buffer *leaf;
4284         struct btrfs_file_extent_item *fi;
4285         struct btrfs_key key;
4286         struct btrfs_key found_key;
4287         u64 extent_start = 0;
4288         u64 extent_num_bytes = 0;
4289         u64 extent_offset = 0;
4290         u64 item_end = 0;
4291         u64 last_size = new_size;
4292         u32 found_type = (u8)-1;
4293         int found_extent;
4294         int del_item;
4295         int pending_del_nr = 0;
4296         int pending_del_slot = 0;
4297         int extent_type = -1;
4298         int ret;
4299         int err = 0;
4300         u64 ino = btrfs_ino(inode);
4301         u64 bytes_deleted = 0;
4302         bool be_nice = 0;
4303         bool should_throttle = 0;
4304         bool should_end = 0;
4305
4306         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4307
4308         /*
4309          * for non-free space inodes and ref cows, we want to back off from
4310          * time to time
4311          */
4312         if (!btrfs_is_free_space_inode(inode) &&
4313             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4314                 be_nice = 1;
4315
4316         path = btrfs_alloc_path();
4317         if (!path)
4318                 return -ENOMEM;
4319         path->reada = READA_BACK;
4320
4321         /*
4322          * We want to drop from the next block forward in case this new size is
4323          * not block aligned since we will be keeping the last block of the
4324          * extent just the way it is.
4325          */
4326         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4327             root == root->fs_info->tree_root)
4328                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4329                                         root->sectorsize), (u64)-1, 0);
4330
4331         /*
4332          * This function is also used to drop the items in the log tree before
4333          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4334          * it is used to drop the loged items. So we shouldn't kill the delayed
4335          * items.
4336          */
4337         if (min_type == 0 && root == BTRFS_I(inode)->root)
4338                 btrfs_kill_delayed_inode_items(inode);
4339
4340         key.objectid = ino;
4341         key.offset = (u64)-1;
4342         key.type = (u8)-1;
4343
4344 search_again:
4345         /*
4346          * with a 16K leaf size and 128MB extents, you can actually queue
4347          * up a huge file in a single leaf.  Most of the time that
4348          * bytes_deleted is > 0, it will be huge by the time we get here
4349          */
4350         if (be_nice && bytes_deleted > SZ_32M) {
4351                 if (btrfs_should_end_transaction(trans, root)) {
4352                         err = -EAGAIN;
4353                         goto error;
4354                 }
4355         }
4356
4357
4358         path->leave_spinning = 1;
4359         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4360         if (ret < 0) {
4361                 err = ret;
4362                 goto out;
4363         }
4364
4365         if (ret > 0) {
4366                 /* there are no items in the tree for us to truncate, we're
4367                  * done
4368                  */
4369                 if (path->slots[0] == 0)
4370                         goto out;
4371                 path->slots[0]--;
4372         }
4373
4374         while (1) {
4375                 fi = NULL;
4376                 leaf = path->nodes[0];
4377                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4378                 found_type = found_key.type;
4379
4380                 if (found_key.objectid != ino)
4381                         break;
4382
4383                 if (found_type < min_type)
4384                         break;
4385
4386                 item_end = found_key.offset;
4387                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4388                         fi = btrfs_item_ptr(leaf, path->slots[0],
4389                                             struct btrfs_file_extent_item);
4390                         extent_type = btrfs_file_extent_type(leaf, fi);
4391                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4392                                 item_end +=
4393                                     btrfs_file_extent_num_bytes(leaf, fi);
4394                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4395                                 item_end += btrfs_file_extent_inline_len(leaf,
4396                                                          path->slots[0], fi);
4397                         }
4398                         item_end--;
4399                 }
4400                 if (found_type > min_type) {
4401                         del_item = 1;
4402                 } else {
4403                         if (item_end < new_size)
4404                                 break;
4405                         if (found_key.offset >= new_size)
4406                                 del_item = 1;
4407                         else
4408                                 del_item = 0;
4409                 }
4410                 found_extent = 0;
4411                 /* FIXME, shrink the extent if the ref count is only 1 */
4412                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4413                         goto delete;
4414
4415                 if (del_item)
4416                         last_size = found_key.offset;
4417                 else
4418                         last_size = new_size;
4419
4420                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4421                         u64 num_dec;
4422                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4423                         if (!del_item) {
4424                                 u64 orig_num_bytes =
4425                                         btrfs_file_extent_num_bytes(leaf, fi);
4426                                 extent_num_bytes = ALIGN(new_size -
4427                                                 found_key.offset,
4428                                                 root->sectorsize);
4429                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4430                                                          extent_num_bytes);
4431                                 num_dec = (orig_num_bytes -
4432                                            extent_num_bytes);
4433                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4434                                              &root->state) &&
4435                                     extent_start != 0)
4436                                         inode_sub_bytes(inode, num_dec);
4437                                 btrfs_mark_buffer_dirty(leaf);
4438                         } else {
4439                                 extent_num_bytes =
4440                                         btrfs_file_extent_disk_num_bytes(leaf,
4441                                                                          fi);
4442                                 extent_offset = found_key.offset -
4443                                         btrfs_file_extent_offset(leaf, fi);
4444
4445                                 /* FIXME blocksize != 4096 */
4446                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4447                                 if (extent_start != 0) {
4448                                         found_extent = 1;
4449                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4450                                                      &root->state))
4451                                                 inode_sub_bytes(inode, num_dec);
4452                                 }
4453                         }
4454                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4455                         /*
4456                          * we can't truncate inline items that have had
4457                          * special encodings
4458                          */
4459                         if (!del_item &&
4460                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4461                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4462
4463                                 /*
4464                                  * Need to release path in order to truncate a
4465                                  * compressed extent. So delete any accumulated
4466                                  * extent items so far.
4467                                  */
4468                                 if (btrfs_file_extent_compression(leaf, fi) !=
4469                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4470                                         err = btrfs_del_items(trans, root, path,
4471                                                               pending_del_slot,
4472                                                               pending_del_nr);
4473                                         if (err) {
4474                                                 btrfs_abort_transaction(trans,
4475                                                                         root,
4476                                                                         err);
4477                                                 goto error;
4478                                         }
4479                                         pending_del_nr = 0;
4480                                 }
4481
4482                                 err = truncate_inline_extent(inode, path,
4483                                                              &found_key,
4484                                                              item_end,
4485                                                              new_size);
4486                                 if (err) {
4487                                         btrfs_abort_transaction(trans,
4488                                                                 root, err);
4489                                         goto error;
4490                                 }
4491                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4492                                             &root->state)) {
4493                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4494                         }
4495                 }
4496 delete:
4497                 if (del_item) {
4498                         if (!pending_del_nr) {
4499                                 /* no pending yet, add ourselves */
4500                                 pending_del_slot = path->slots[0];
4501                                 pending_del_nr = 1;
4502                         } else if (pending_del_nr &&
4503                                    path->slots[0] + 1 == pending_del_slot) {
4504                                 /* hop on the pending chunk */
4505                                 pending_del_nr++;
4506                                 pending_del_slot = path->slots[0];
4507                         } else {
4508                                 BUG();
4509                         }
4510                 } else {
4511                         break;
4512                 }
4513                 should_throttle = 0;
4514
4515                 if (found_extent &&
4516                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4517                      root == root->fs_info->tree_root)) {
4518                         btrfs_set_path_blocking(path);
4519                         bytes_deleted += extent_num_bytes;
4520                         ret = btrfs_free_extent(trans, root, extent_start,
4521                                                 extent_num_bytes, 0,
4522                                                 btrfs_header_owner(leaf),
4523                                                 ino, extent_offset);
4524                         BUG_ON(ret);
4525                         if (btrfs_should_throttle_delayed_refs(trans, root))
4526                                 btrfs_async_run_delayed_refs(root,
4527                                         trans->delayed_ref_updates * 2, 0);
4528                         if (be_nice) {
4529                                 if (truncate_space_check(trans, root,
4530                                                          extent_num_bytes)) {
4531                                         should_end = 1;
4532                                 }
4533                                 if (btrfs_should_throttle_delayed_refs(trans,
4534                                                                        root)) {
4535                                         should_throttle = 1;
4536                                 }
4537                         }
4538                 }
4539
4540                 if (found_type == BTRFS_INODE_ITEM_KEY)
4541                         break;
4542
4543                 if (path->slots[0] == 0 ||
4544                     path->slots[0] != pending_del_slot ||
4545                     should_throttle || should_end) {
4546                         if (pending_del_nr) {
4547                                 ret = btrfs_del_items(trans, root, path,
4548                                                 pending_del_slot,
4549                                                 pending_del_nr);
4550                                 if (ret) {
4551                                         btrfs_abort_transaction(trans,
4552                                                                 root, ret);
4553                                         goto error;
4554                                 }
4555                                 pending_del_nr = 0;
4556                         }
4557                         btrfs_release_path(path);
4558                         if (should_throttle) {
4559                                 unsigned long updates = trans->delayed_ref_updates;
4560                                 if (updates) {
4561                                         trans->delayed_ref_updates = 0;
4562                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4563                                         if (ret && !err)
4564                                                 err = ret;
4565                                 }
4566                         }
4567                         /*
4568                          * if we failed to refill our space rsv, bail out
4569                          * and let the transaction restart
4570                          */
4571                         if (should_end) {
4572                                 err = -EAGAIN;
4573                                 goto error;
4574                         }
4575                         goto search_again;
4576                 } else {
4577                         path->slots[0]--;
4578                 }
4579         }
4580 out:
4581         if (pending_del_nr) {
4582                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4583                                       pending_del_nr);
4584                 if (ret)
4585                         btrfs_abort_transaction(trans, root, ret);
4586         }
4587 error:
4588         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4589                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4590
4591         btrfs_free_path(path);
4592
4593         if (be_nice && bytes_deleted > SZ_32M) {
4594                 unsigned long updates = trans->delayed_ref_updates;
4595                 if (updates) {
4596                         trans->delayed_ref_updates = 0;
4597                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4598                         if (ret && !err)
4599                                 err = ret;
4600                 }
4601         }
4602         return err;
4603 }
4604
4605 /*
4606  * btrfs_truncate_page - read, zero a chunk and write a page
4607  * @inode - inode that we're zeroing
4608  * @from - the offset to start zeroing
4609  * @len - the length to zero, 0 to zero the entire range respective to the
4610  *      offset
4611  * @front - zero up to the offset instead of from the offset on
4612  *
4613  * This will find the page for the "from" offset and cow the page and zero the
4614  * part we want to zero.  This is used with truncate and hole punching.
4615  */
4616 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4617                         int front)
4618 {
4619         struct address_space *mapping = inode->i_mapping;
4620         struct btrfs_root *root = BTRFS_I(inode)->root;
4621         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4622         struct btrfs_ordered_extent *ordered;
4623         struct extent_state *cached_state = NULL;
4624         char *kaddr;
4625         u32 blocksize = root->sectorsize;
4626         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4627         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4628         struct page *page;
4629         gfp_t mask = btrfs_alloc_write_mask(mapping);
4630         int ret = 0;
4631         u64 page_start;
4632         u64 page_end;
4633
4634         if ((offset & (blocksize - 1)) == 0 &&
4635             (!len || ((len & (blocksize - 1)) == 0)))
4636                 goto out;
4637         ret = btrfs_delalloc_reserve_space(inode,
4638                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4639         if (ret)
4640                 goto out;
4641
4642 again:
4643         page = find_or_create_page(mapping, index, mask);
4644         if (!page) {
4645                 btrfs_delalloc_release_space(inode,
4646                                 round_down(from, PAGE_CACHE_SIZE),
4647                                 PAGE_CACHE_SIZE);
4648                 ret = -ENOMEM;
4649                 goto out;
4650         }
4651
4652         page_start = page_offset(page);
4653         page_end = page_start + PAGE_CACHE_SIZE - 1;
4654
4655         if (!PageUptodate(page)) {
4656                 ret = btrfs_readpage(NULL, page);
4657                 lock_page(page);
4658                 if (page->mapping != mapping) {
4659                         unlock_page(page);
4660                         page_cache_release(page);
4661                         goto again;
4662                 }
4663                 if (!PageUptodate(page)) {
4664                         ret = -EIO;
4665                         goto out_unlock;
4666                 }
4667         }
4668         wait_on_page_writeback(page);
4669
4670         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
4671         set_page_extent_mapped(page);
4672
4673         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4674         if (ordered) {
4675                 unlock_extent_cached(io_tree, page_start, page_end,
4676                                      &cached_state, GFP_NOFS);
4677                 unlock_page(page);
4678                 page_cache_release(page);
4679                 btrfs_start_ordered_extent(inode, ordered, 1);
4680                 btrfs_put_ordered_extent(ordered);
4681                 goto again;
4682         }
4683
4684         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4685                           EXTENT_DIRTY | EXTENT_DELALLOC |
4686                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4687                           0, 0, &cached_state, GFP_NOFS);
4688
4689         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4690                                         &cached_state);
4691         if (ret) {
4692                 unlock_extent_cached(io_tree, page_start, page_end,
4693                                      &cached_state, GFP_NOFS);
4694                 goto out_unlock;
4695         }
4696
4697         if (offset != PAGE_CACHE_SIZE) {
4698                 if (!len)
4699                         len = PAGE_CACHE_SIZE - offset;
4700                 kaddr = kmap(page);
4701                 if (front)
4702                         memset(kaddr, 0, offset);
4703                 else
4704                         memset(kaddr + offset, 0, len);
4705                 flush_dcache_page(page);
4706                 kunmap(page);
4707         }
4708         ClearPageChecked(page);
4709         set_page_dirty(page);
4710         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4711                              GFP_NOFS);
4712
4713 out_unlock:
4714         if (ret)
4715                 btrfs_delalloc_release_space(inode, page_start,
4716                                              PAGE_CACHE_SIZE);
4717         unlock_page(page);
4718         page_cache_release(page);
4719 out:
4720         return ret;
4721 }
4722
4723 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4724                              u64 offset, u64 len)
4725 {
4726         struct btrfs_trans_handle *trans;
4727         int ret;
4728
4729         /*
4730          * Still need to make sure the inode looks like it's been updated so
4731          * that any holes get logged if we fsync.
4732          */
4733         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4734                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4735                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4736                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4737                 return 0;
4738         }
4739
4740         /*
4741          * 1 - for the one we're dropping
4742          * 1 - for the one we're adding
4743          * 1 - for updating the inode.
4744          */
4745         trans = btrfs_start_transaction(root, 3);
4746         if (IS_ERR(trans))
4747                 return PTR_ERR(trans);
4748
4749         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4750         if (ret) {
4751                 btrfs_abort_transaction(trans, root, ret);
4752                 btrfs_end_transaction(trans, root);
4753                 return ret;
4754         }
4755
4756         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4757                                        0, 0, len, 0, len, 0, 0, 0);
4758         if (ret)
4759                 btrfs_abort_transaction(trans, root, ret);
4760         else
4761                 btrfs_update_inode(trans, root, inode);
4762         btrfs_end_transaction(trans, root);
4763         return ret;
4764 }
4765
4766 /*
4767  * This function puts in dummy file extents for the area we're creating a hole
4768  * for.  So if we are truncating this file to a larger size we need to insert
4769  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4770  * the range between oldsize and size
4771  */
4772 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4773 {
4774         struct btrfs_root *root = BTRFS_I(inode)->root;
4775         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4776         struct extent_map *em = NULL;
4777         struct extent_state *cached_state = NULL;
4778         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4779         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4780         u64 block_end = ALIGN(size, root->sectorsize);
4781         u64 last_byte;
4782         u64 cur_offset;
4783         u64 hole_size;
4784         int err = 0;
4785
4786         /*
4787          * If our size started in the middle of a page we need to zero out the
4788          * rest of the page before we expand the i_size, otherwise we could
4789          * expose stale data.
4790          */
4791         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4792         if (err)
4793                 return err;
4794
4795         if (size <= hole_start)
4796                 return 0;
4797
4798         while (1) {
4799                 struct btrfs_ordered_extent *ordered;
4800
4801                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4802                                  &cached_state);
4803                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4804                                                      block_end - hole_start);
4805                 if (!ordered)
4806                         break;
4807                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4808                                      &cached_state, GFP_NOFS);
4809                 btrfs_start_ordered_extent(inode, ordered, 1);
4810                 btrfs_put_ordered_extent(ordered);
4811         }
4812
4813         cur_offset = hole_start;
4814         while (1) {
4815                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4816                                 block_end - cur_offset, 0);
4817                 if (IS_ERR(em)) {
4818                         err = PTR_ERR(em);
4819                         em = NULL;
4820                         break;
4821                 }
4822                 last_byte = min(extent_map_end(em), block_end);
4823                 last_byte = ALIGN(last_byte , root->sectorsize);
4824                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4825                         struct extent_map *hole_em;
4826                         hole_size = last_byte - cur_offset;
4827
4828                         err = maybe_insert_hole(root, inode, cur_offset,
4829                                                 hole_size);
4830                         if (err)
4831                                 break;
4832                         btrfs_drop_extent_cache(inode, cur_offset,
4833                                                 cur_offset + hole_size - 1, 0);
4834                         hole_em = alloc_extent_map();
4835                         if (!hole_em) {
4836                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4837                                         &BTRFS_I(inode)->runtime_flags);
4838                                 goto next;
4839                         }
4840                         hole_em->start = cur_offset;
4841                         hole_em->len = hole_size;
4842                         hole_em->orig_start = cur_offset;
4843
4844                         hole_em->block_start = EXTENT_MAP_HOLE;
4845                         hole_em->block_len = 0;
4846                         hole_em->orig_block_len = 0;
4847                         hole_em->ram_bytes = hole_size;
4848                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4849                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4850                         hole_em->generation = root->fs_info->generation;
4851
4852                         while (1) {
4853                                 write_lock(&em_tree->lock);
4854                                 err = add_extent_mapping(em_tree, hole_em, 1);
4855                                 write_unlock(&em_tree->lock);
4856                                 if (err != -EEXIST)
4857                                         break;
4858                                 btrfs_drop_extent_cache(inode, cur_offset,
4859                                                         cur_offset +
4860                                                         hole_size - 1, 0);
4861                         }
4862                         free_extent_map(hole_em);
4863                 }
4864 next:
4865                 free_extent_map(em);
4866                 em = NULL;
4867                 cur_offset = last_byte;
4868                 if (cur_offset >= block_end)
4869                         break;
4870         }
4871         free_extent_map(em);
4872         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4873                              GFP_NOFS);
4874         return err;
4875 }
4876
4877 static int wait_snapshoting_atomic_t(atomic_t *a)
4878 {
4879         schedule();
4880         return 0;
4881 }
4882
4883 static void wait_for_snapshot_creation(struct btrfs_root *root)
4884 {
4885         while (true) {
4886                 int ret;
4887
4888                 ret = btrfs_start_write_no_snapshoting(root);
4889                 if (ret)
4890                         break;
4891                 wait_on_atomic_t(&root->will_be_snapshoted,
4892                                  wait_snapshoting_atomic_t,
4893                                  TASK_UNINTERRUPTIBLE);
4894         }
4895 }
4896
4897 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4898 {
4899         struct btrfs_root *root = BTRFS_I(inode)->root;
4900         struct btrfs_trans_handle *trans;
4901         loff_t oldsize = i_size_read(inode);
4902         loff_t newsize = attr->ia_size;
4903         int mask = attr->ia_valid;
4904         int ret;
4905
4906         /*
4907          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4908          * special case where we need to update the times despite not having
4909          * these flags set.  For all other operations the VFS set these flags
4910          * explicitly if it wants a timestamp update.
4911          */
4912         if (newsize != oldsize) {
4913                 inode_inc_iversion(inode);
4914                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4915                         inode->i_ctime = inode->i_mtime =
4916                                 current_fs_time(inode->i_sb);
4917         }
4918
4919         if (newsize > oldsize) {
4920                 truncate_pagecache(inode, newsize);
4921                 /*
4922                  * Don't do an expanding truncate while snapshoting is ongoing.
4923                  * This is to ensure the snapshot captures a fully consistent
4924                  * state of this file - if the snapshot captures this expanding
4925                  * truncation, it must capture all writes that happened before
4926                  * this truncation.
4927                  */
4928                 wait_for_snapshot_creation(root);
4929                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4930                 if (ret) {
4931                         btrfs_end_write_no_snapshoting(root);
4932                         return ret;
4933                 }
4934
4935                 trans = btrfs_start_transaction(root, 1);
4936                 if (IS_ERR(trans)) {
4937                         btrfs_end_write_no_snapshoting(root);
4938                         return PTR_ERR(trans);
4939                 }
4940
4941                 i_size_write(inode, newsize);
4942                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4943                 ret = btrfs_update_inode(trans, root, inode);
4944                 btrfs_end_write_no_snapshoting(root);
4945                 btrfs_end_transaction(trans, root);
4946         } else {
4947
4948                 /*
4949                  * We're truncating a file that used to have good data down to
4950                  * zero. Make sure it gets into the ordered flush list so that
4951                  * any new writes get down to disk quickly.
4952                  */
4953                 if (newsize == 0)
4954                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4955                                 &BTRFS_I(inode)->runtime_flags);
4956
4957                 /*
4958                  * 1 for the orphan item we're going to add
4959                  * 1 for the orphan item deletion.
4960                  */
4961                 trans = btrfs_start_transaction(root, 2);
4962                 if (IS_ERR(trans))
4963                         return PTR_ERR(trans);
4964
4965                 /*
4966                  * We need to do this in case we fail at _any_ point during the
4967                  * actual truncate.  Once we do the truncate_setsize we could
4968                  * invalidate pages which forces any outstanding ordered io to
4969                  * be instantly completed which will give us extents that need
4970                  * to be truncated.  If we fail to get an orphan inode down we
4971                  * could have left over extents that were never meant to live,
4972                  * so we need to garuntee from this point on that everything
4973                  * will be consistent.
4974                  */
4975                 ret = btrfs_orphan_add(trans, inode);
4976                 btrfs_end_transaction(trans, root);
4977                 if (ret)
4978                         return ret;
4979
4980                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4981                 truncate_setsize(inode, newsize);
4982
4983                 /* Disable nonlocked read DIO to avoid the end less truncate */
4984                 btrfs_inode_block_unlocked_dio(inode);
4985                 inode_dio_wait(inode);
4986                 btrfs_inode_resume_unlocked_dio(inode);
4987
4988                 ret = btrfs_truncate(inode);
4989                 if (ret && inode->i_nlink) {
4990                         int err;
4991
4992                         /*
4993                          * failed to truncate, disk_i_size is only adjusted down
4994                          * as we remove extents, so it should represent the true
4995                          * size of the inode, so reset the in memory size and
4996                          * delete our orphan entry.
4997                          */
4998                         trans = btrfs_join_transaction(root);
4999                         if (IS_ERR(trans)) {
5000                                 btrfs_orphan_del(NULL, inode);
5001                                 return ret;
5002                         }
5003                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5004                         err = btrfs_orphan_del(trans, inode);
5005                         if (err)
5006                                 btrfs_abort_transaction(trans, root, err);
5007                         btrfs_end_transaction(trans, root);
5008                 }
5009         }
5010
5011         return ret;
5012 }
5013
5014 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5015 {
5016         struct inode *inode = d_inode(dentry);
5017         struct btrfs_root *root = BTRFS_I(inode)->root;
5018         int err;
5019
5020         if (btrfs_root_readonly(root))
5021                 return -EROFS;
5022
5023         err = inode_change_ok(inode, attr);
5024         if (err)
5025                 return err;
5026
5027         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5028                 err = btrfs_setsize(inode, attr);
5029                 if (err)
5030                         return err;
5031         }
5032
5033         if (attr->ia_valid) {
5034                 setattr_copy(inode, attr);
5035                 inode_inc_iversion(inode);
5036                 err = btrfs_dirty_inode(inode);
5037
5038                 if (!err && attr->ia_valid & ATTR_MODE)
5039                         err = posix_acl_chmod(inode, inode->i_mode);
5040         }
5041
5042         return err;
5043 }
5044
5045 /*
5046  * While truncating the inode pages during eviction, we get the VFS calling
5047  * btrfs_invalidatepage() against each page of the inode. This is slow because
5048  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5049  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5050  * extent_state structures over and over, wasting lots of time.
5051  *
5052  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5053  * those expensive operations on a per page basis and do only the ordered io
5054  * finishing, while we release here the extent_map and extent_state structures,
5055  * without the excessive merging and splitting.
5056  */
5057 static void evict_inode_truncate_pages(struct inode *inode)
5058 {
5059         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5060         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5061         struct rb_node *node;
5062
5063         ASSERT(inode->i_state & I_FREEING);
5064         truncate_inode_pages_final(&inode->i_data);
5065
5066         write_lock(&map_tree->lock);
5067         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5068                 struct extent_map *em;
5069
5070                 node = rb_first(&map_tree->map);
5071                 em = rb_entry(node, struct extent_map, rb_node);
5072                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5073                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5074                 remove_extent_mapping(map_tree, em);
5075                 free_extent_map(em);
5076                 if (need_resched()) {
5077                         write_unlock(&map_tree->lock);
5078                         cond_resched();
5079                         write_lock(&map_tree->lock);
5080                 }
5081         }
5082         write_unlock(&map_tree->lock);
5083
5084         /*
5085          * Keep looping until we have no more ranges in the io tree.
5086          * We can have ongoing bios started by readpages (called from readahead)
5087          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5088          * still in progress (unlocked the pages in the bio but did not yet
5089          * unlocked the ranges in the io tree). Therefore this means some
5090          * ranges can still be locked and eviction started because before
5091          * submitting those bios, which are executed by a separate task (work
5092          * queue kthread), inode references (inode->i_count) were not taken
5093          * (which would be dropped in the end io callback of each bio).
5094          * Therefore here we effectively end up waiting for those bios and
5095          * anyone else holding locked ranges without having bumped the inode's
5096          * reference count - if we don't do it, when they access the inode's
5097          * io_tree to unlock a range it may be too late, leading to an
5098          * use-after-free issue.
5099          */
5100         spin_lock(&io_tree->lock);
5101         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5102                 struct extent_state *state;
5103                 struct extent_state *cached_state = NULL;
5104                 u64 start;
5105                 u64 end;
5106
5107                 node = rb_first(&io_tree->state);
5108                 state = rb_entry(node, struct extent_state, rb_node);
5109                 start = state->start;
5110                 end = state->end;
5111                 spin_unlock(&io_tree->lock);
5112
5113                 lock_extent_bits(io_tree, start, end, &cached_state);
5114
5115                 /*
5116                  * If still has DELALLOC flag, the extent didn't reach disk,
5117                  * and its reserved space won't be freed by delayed_ref.
5118                  * So we need to free its reserved space here.
5119                  * (Refer to comment in btrfs_invalidatepage, case 2)
5120                  *
5121                  * Note, end is the bytenr of last byte, so we need + 1 here.
5122                  */
5123                 if (state->state & EXTENT_DELALLOC)
5124                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5125
5126                 clear_extent_bit(io_tree, start, end,
5127                                  EXTENT_LOCKED | EXTENT_DIRTY |
5128                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5129                                  EXTENT_DEFRAG, 1, 1,
5130                                  &cached_state, GFP_NOFS);
5131
5132                 cond_resched();
5133                 spin_lock(&io_tree->lock);
5134         }
5135         spin_unlock(&io_tree->lock);
5136 }
5137
5138 void btrfs_evict_inode(struct inode *inode)
5139 {
5140         struct btrfs_trans_handle *trans;
5141         struct btrfs_root *root = BTRFS_I(inode)->root;
5142         struct btrfs_block_rsv *rsv, *global_rsv;
5143         int steal_from_global = 0;
5144         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5145         int ret;
5146
5147         trace_btrfs_inode_evict(inode);
5148
5149         evict_inode_truncate_pages(inode);
5150
5151         if (inode->i_nlink &&
5152             ((btrfs_root_refs(&root->root_item) != 0 &&
5153               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5154              btrfs_is_free_space_inode(inode)))
5155                 goto no_delete;
5156
5157         if (is_bad_inode(inode)) {
5158                 btrfs_orphan_del(NULL, inode);
5159                 goto no_delete;
5160         }
5161         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5162         if (!special_file(inode->i_mode))
5163                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5164
5165         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5166
5167         if (root->fs_info->log_root_recovering) {
5168                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5169                                  &BTRFS_I(inode)->runtime_flags));
5170                 goto no_delete;
5171         }
5172
5173         if (inode->i_nlink > 0) {
5174                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5175                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5176                 goto no_delete;
5177         }
5178
5179         ret = btrfs_commit_inode_delayed_inode(inode);
5180         if (ret) {
5181                 btrfs_orphan_del(NULL, inode);
5182                 goto no_delete;
5183         }
5184
5185         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5186         if (!rsv) {
5187                 btrfs_orphan_del(NULL, inode);
5188                 goto no_delete;
5189         }
5190         rsv->size = min_size;
5191         rsv->failfast = 1;
5192         global_rsv = &root->fs_info->global_block_rsv;
5193
5194         btrfs_i_size_write(inode, 0);
5195
5196         /*
5197          * This is a bit simpler than btrfs_truncate since we've already
5198          * reserved our space for our orphan item in the unlink, so we just
5199          * need to reserve some slack space in case we add bytes and update
5200          * inode item when doing the truncate.
5201          */
5202         while (1) {
5203                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5204                                              BTRFS_RESERVE_FLUSH_LIMIT);
5205
5206                 /*
5207                  * Try and steal from the global reserve since we will
5208                  * likely not use this space anyway, we want to try as
5209                  * hard as possible to get this to work.
5210                  */
5211                 if (ret)
5212                         steal_from_global++;
5213                 else
5214                         steal_from_global = 0;
5215                 ret = 0;
5216
5217                 /*
5218                  * steal_from_global == 0: we reserved stuff, hooray!
5219                  * steal_from_global == 1: we didn't reserve stuff, boo!
5220                  * steal_from_global == 2: we've committed, still not a lot of
5221                  * room but maybe we'll have room in the global reserve this
5222                  * time.
5223                  * steal_from_global == 3: abandon all hope!
5224                  */
5225                 if (steal_from_global > 2) {
5226                         btrfs_warn(root->fs_info,
5227                                 "Could not get space for a delete, will truncate on mount %d",
5228                                 ret);
5229                         btrfs_orphan_del(NULL, inode);
5230                         btrfs_free_block_rsv(root, rsv);
5231                         goto no_delete;
5232                 }
5233
5234                 trans = btrfs_join_transaction(root);
5235                 if (IS_ERR(trans)) {
5236                         btrfs_orphan_del(NULL, inode);
5237                         btrfs_free_block_rsv(root, rsv);
5238                         goto no_delete;
5239                 }
5240
5241                 /*
5242                  * We can't just steal from the global reserve, we need tomake
5243                  * sure there is room to do it, if not we need to commit and try
5244                  * again.
5245                  */
5246                 if (steal_from_global) {
5247                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5248                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5249                                                               min_size);
5250                         else
5251                                 ret = -ENOSPC;
5252                 }
5253
5254                 /*
5255                  * Couldn't steal from the global reserve, we have too much
5256                  * pending stuff built up, commit the transaction and try it
5257                  * again.
5258                  */
5259                 if (ret) {
5260                         ret = btrfs_commit_transaction(trans, root);
5261                         if (ret) {
5262                                 btrfs_orphan_del(NULL, inode);
5263                                 btrfs_free_block_rsv(root, rsv);
5264                                 goto no_delete;
5265                         }
5266                         continue;
5267                 } else {
5268                         steal_from_global = 0;
5269                 }
5270
5271                 trans->block_rsv = rsv;
5272
5273                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5274                 if (ret != -ENOSPC && ret != -EAGAIN)
5275                         break;
5276
5277                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5278                 btrfs_end_transaction(trans, root);
5279                 trans = NULL;
5280                 btrfs_btree_balance_dirty(root);
5281         }
5282
5283         btrfs_free_block_rsv(root, rsv);
5284
5285         /*
5286          * Errors here aren't a big deal, it just means we leave orphan items
5287          * in the tree.  They will be cleaned up on the next mount.
5288          */
5289         if (ret == 0) {
5290                 trans->block_rsv = root->orphan_block_rsv;
5291                 btrfs_orphan_del(trans, inode);
5292         } else {
5293                 btrfs_orphan_del(NULL, inode);
5294         }
5295
5296         trans->block_rsv = &root->fs_info->trans_block_rsv;
5297         if (!(root == root->fs_info->tree_root ||
5298               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5299                 btrfs_return_ino(root, btrfs_ino(inode));
5300
5301         btrfs_end_transaction(trans, root);
5302         btrfs_btree_balance_dirty(root);
5303 no_delete:
5304         btrfs_remove_delayed_node(inode);
5305         clear_inode(inode);
5306 }
5307
5308 /*
5309  * this returns the key found in the dir entry in the location pointer.
5310  * If no dir entries were found, location->objectid is 0.
5311  */
5312 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5313                                struct btrfs_key *location)
5314 {
5315         const char *name = dentry->d_name.name;
5316         int namelen = dentry->d_name.len;
5317         struct btrfs_dir_item *di;
5318         struct btrfs_path *path;
5319         struct btrfs_root *root = BTRFS_I(dir)->root;
5320         int ret = 0;
5321
5322         path = btrfs_alloc_path();
5323         if (!path)
5324                 return -ENOMEM;
5325
5326         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5327                                     namelen, 0);
5328         if (IS_ERR(di))
5329                 ret = PTR_ERR(di);
5330
5331         if (IS_ERR_OR_NULL(di))
5332                 goto out_err;
5333
5334         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5335 out:
5336         btrfs_free_path(path);
5337         return ret;
5338 out_err:
5339         location->objectid = 0;
5340         goto out;
5341 }
5342
5343 /*
5344  * when we hit a tree root in a directory, the btrfs part of the inode
5345  * needs to be changed to reflect the root directory of the tree root.  This
5346  * is kind of like crossing a mount point.
5347  */
5348 static int fixup_tree_root_location(struct btrfs_root *root,
5349                                     struct inode *dir,
5350                                     struct dentry *dentry,
5351                                     struct btrfs_key *location,
5352                                     struct btrfs_root **sub_root)
5353 {
5354         struct btrfs_path *path;
5355         struct btrfs_root *new_root;
5356         struct btrfs_root_ref *ref;
5357         struct extent_buffer *leaf;
5358         struct btrfs_key key;
5359         int ret;
5360         int err = 0;
5361
5362         path = btrfs_alloc_path();
5363         if (!path) {
5364                 err = -ENOMEM;
5365                 goto out;
5366         }
5367
5368         err = -ENOENT;
5369         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5370         key.type = BTRFS_ROOT_REF_KEY;
5371         key.offset = location->objectid;
5372
5373         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5374                                 0, 0);
5375         if (ret) {
5376                 if (ret < 0)
5377                         err = ret;
5378                 goto out;
5379         }
5380
5381         leaf = path->nodes[0];
5382         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5383         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5384             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5385                 goto out;
5386
5387         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5388                                    (unsigned long)(ref + 1),
5389                                    dentry->d_name.len);
5390         if (ret)
5391                 goto out;
5392
5393         btrfs_release_path(path);
5394
5395         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5396         if (IS_ERR(new_root)) {
5397                 err = PTR_ERR(new_root);
5398                 goto out;
5399         }
5400
5401         *sub_root = new_root;
5402         location->objectid = btrfs_root_dirid(&new_root->root_item);
5403         location->type = BTRFS_INODE_ITEM_KEY;
5404         location->offset = 0;
5405         err = 0;
5406 out:
5407         btrfs_free_path(path);
5408         return err;
5409 }
5410
5411 static void inode_tree_add(struct inode *inode)
5412 {
5413         struct btrfs_root *root = BTRFS_I(inode)->root;
5414         struct btrfs_inode *entry;
5415         struct rb_node **p;
5416         struct rb_node *parent;
5417         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5418         u64 ino = btrfs_ino(inode);
5419
5420         if (inode_unhashed(inode))
5421                 return;
5422         parent = NULL;
5423         spin_lock(&root->inode_lock);
5424         p = &root->inode_tree.rb_node;
5425         while (*p) {
5426                 parent = *p;
5427                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5428
5429                 if (ino < btrfs_ino(&entry->vfs_inode))
5430                         p = &parent->rb_left;
5431                 else if (ino > btrfs_ino(&entry->vfs_inode))
5432                         p = &parent->rb_right;
5433                 else {
5434                         WARN_ON(!(entry->vfs_inode.i_state &
5435                                   (I_WILL_FREE | I_FREEING)));
5436                         rb_replace_node(parent, new, &root->inode_tree);
5437                         RB_CLEAR_NODE(parent);
5438                         spin_unlock(&root->inode_lock);
5439                         return;
5440                 }
5441         }
5442         rb_link_node(new, parent, p);
5443         rb_insert_color(new, &root->inode_tree);
5444         spin_unlock(&root->inode_lock);
5445 }
5446
5447 static void inode_tree_del(struct inode *inode)
5448 {
5449         struct btrfs_root *root = BTRFS_I(inode)->root;
5450         int empty = 0;
5451
5452         spin_lock(&root->inode_lock);
5453         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5454                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5455                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5456                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5457         }
5458         spin_unlock(&root->inode_lock);
5459
5460         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5461                 synchronize_srcu(&root->fs_info->subvol_srcu);
5462                 spin_lock(&root->inode_lock);
5463                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5464                 spin_unlock(&root->inode_lock);
5465                 if (empty)
5466                         btrfs_add_dead_root(root);
5467         }
5468 }
5469
5470 void btrfs_invalidate_inodes(struct btrfs_root *root)
5471 {
5472         struct rb_node *node;
5473         struct rb_node *prev;
5474         struct btrfs_inode *entry;
5475         struct inode *inode;
5476         u64 objectid = 0;
5477
5478         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5479                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5480
5481         spin_lock(&root->inode_lock);
5482 again:
5483         node = root->inode_tree.rb_node;
5484         prev = NULL;
5485         while (node) {
5486                 prev = node;
5487                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5488
5489                 if (objectid < btrfs_ino(&entry->vfs_inode))
5490                         node = node->rb_left;
5491                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5492                         node = node->rb_right;
5493                 else
5494                         break;
5495         }
5496         if (!node) {
5497                 while (prev) {
5498                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5499                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5500                                 node = prev;
5501                                 break;
5502                         }
5503                         prev = rb_next(prev);
5504                 }
5505         }
5506         while (node) {
5507                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5508                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5509                 inode = igrab(&entry->vfs_inode);
5510                 if (inode) {
5511                         spin_unlock(&root->inode_lock);
5512                         if (atomic_read(&inode->i_count) > 1)
5513                                 d_prune_aliases(inode);
5514                         /*
5515                          * btrfs_drop_inode will have it removed from
5516                          * the inode cache when its usage count
5517                          * hits zero.
5518                          */
5519                         iput(inode);
5520                         cond_resched();
5521                         spin_lock(&root->inode_lock);
5522                         goto again;
5523                 }
5524
5525                 if (cond_resched_lock(&root->inode_lock))
5526                         goto again;
5527
5528                 node = rb_next(node);
5529         }
5530         spin_unlock(&root->inode_lock);
5531 }
5532
5533 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5534 {
5535         struct btrfs_iget_args *args = p;
5536         inode->i_ino = args->location->objectid;
5537         memcpy(&BTRFS_I(inode)->location, args->location,
5538                sizeof(*args->location));
5539         BTRFS_I(inode)->root = args->root;
5540         return 0;
5541 }
5542
5543 static int btrfs_find_actor(struct inode *inode, void *opaque)
5544 {
5545         struct btrfs_iget_args *args = opaque;
5546         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5547                 args->root == BTRFS_I(inode)->root;
5548 }
5549
5550 static struct inode *btrfs_iget_locked(struct super_block *s,
5551                                        struct btrfs_key *location,
5552                                        struct btrfs_root *root)
5553 {
5554         struct inode *inode;
5555         struct btrfs_iget_args args;
5556         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5557
5558         args.location = location;
5559         args.root = root;
5560
5561         inode = iget5_locked(s, hashval, btrfs_find_actor,
5562                              btrfs_init_locked_inode,
5563                              (void *)&args);
5564         return inode;
5565 }
5566
5567 /* Get an inode object given its location and corresponding root.
5568  * Returns in *is_new if the inode was read from disk
5569  */
5570 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5571                          struct btrfs_root *root, int *new)
5572 {
5573         struct inode *inode;
5574
5575         inode = btrfs_iget_locked(s, location, root);
5576         if (!inode)
5577                 return ERR_PTR(-ENOMEM);
5578
5579         if (inode->i_state & I_NEW) {
5580                 btrfs_read_locked_inode(inode);
5581                 if (!is_bad_inode(inode)) {
5582                         inode_tree_add(inode);
5583                         unlock_new_inode(inode);
5584                         if (new)
5585                                 *new = 1;
5586                 } else {
5587                         unlock_new_inode(inode);
5588                         iput(inode);
5589                         inode = ERR_PTR(-ESTALE);
5590                 }
5591         }
5592
5593         return inode;
5594 }
5595
5596 static struct inode *new_simple_dir(struct super_block *s,
5597                                     struct btrfs_key *key,
5598                                     struct btrfs_root *root)
5599 {
5600         struct inode *inode = new_inode(s);
5601
5602         if (!inode)
5603                 return ERR_PTR(-ENOMEM);
5604
5605         BTRFS_I(inode)->root = root;
5606         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5607         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5608
5609         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5610         inode->i_op = &btrfs_dir_ro_inode_operations;
5611         inode->i_fop = &simple_dir_operations;
5612         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5613         inode->i_mtime = CURRENT_TIME;
5614         inode->i_atime = inode->i_mtime;
5615         inode->i_ctime = inode->i_mtime;
5616         BTRFS_I(inode)->i_otime = inode->i_mtime;
5617
5618         return inode;
5619 }
5620
5621 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5622 {
5623         struct inode *inode;
5624         struct btrfs_root *root = BTRFS_I(dir)->root;
5625         struct btrfs_root *sub_root = root;
5626         struct btrfs_key location;
5627         int index;
5628         int ret = 0;
5629
5630         if (dentry->d_name.len > BTRFS_NAME_LEN)
5631                 return ERR_PTR(-ENAMETOOLONG);
5632
5633         ret = btrfs_inode_by_name(dir, dentry, &location);
5634         if (ret < 0)
5635                 return ERR_PTR(ret);
5636
5637         if (location.objectid == 0)
5638                 return ERR_PTR(-ENOENT);
5639
5640         if (location.type == BTRFS_INODE_ITEM_KEY) {
5641                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5642                 return inode;
5643         }
5644
5645         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5646
5647         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5648         ret = fixup_tree_root_location(root, dir, dentry,
5649                                        &location, &sub_root);
5650         if (ret < 0) {
5651                 if (ret != -ENOENT)
5652                         inode = ERR_PTR(ret);
5653                 else
5654                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5655         } else {
5656                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5657         }
5658         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5659
5660         if (!IS_ERR(inode) && root != sub_root) {
5661                 down_read(&root->fs_info->cleanup_work_sem);
5662                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5663                         ret = btrfs_orphan_cleanup(sub_root);
5664                 up_read(&root->fs_info->cleanup_work_sem);
5665                 if (ret) {
5666                         iput(inode);
5667                         inode = ERR_PTR(ret);
5668                 }
5669         }
5670
5671         return inode;
5672 }
5673
5674 static int btrfs_dentry_delete(const struct dentry *dentry)
5675 {
5676         struct btrfs_root *root;
5677         struct inode *inode = d_inode(dentry);
5678
5679         if (!inode && !IS_ROOT(dentry))
5680                 inode = d_inode(dentry->d_parent);
5681
5682         if (inode) {
5683                 root = BTRFS_I(inode)->root;
5684                 if (btrfs_root_refs(&root->root_item) == 0)
5685                         return 1;
5686
5687                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5688                         return 1;
5689         }
5690         return 0;
5691 }
5692
5693 static void btrfs_dentry_release(struct dentry *dentry)
5694 {
5695         kfree(dentry->d_fsdata);
5696 }
5697
5698 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5699                                    unsigned int flags)
5700 {
5701         struct inode *inode;
5702
5703         inode = btrfs_lookup_dentry(dir, dentry);
5704         if (IS_ERR(inode)) {
5705                 if (PTR_ERR(inode) == -ENOENT)
5706                         inode = NULL;
5707                 else
5708                         return ERR_CAST(inode);
5709         }
5710
5711         return d_splice_alias(inode, dentry);
5712 }
5713
5714 unsigned char btrfs_filetype_table[] = {
5715         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5716 };
5717
5718 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5719 {
5720         struct inode *inode = file_inode(file);
5721         struct btrfs_root *root = BTRFS_I(inode)->root;
5722         struct btrfs_item *item;
5723         struct btrfs_dir_item *di;
5724         struct btrfs_key key;
5725         struct btrfs_key found_key;
5726         struct btrfs_path *path;
5727         struct list_head ins_list;
5728         struct list_head del_list;
5729         int ret;
5730         struct extent_buffer *leaf;
5731         int slot;
5732         unsigned char d_type;
5733         int over = 0;
5734         u32 di_cur;
5735         u32 di_total;
5736         u32 di_len;
5737         int key_type = BTRFS_DIR_INDEX_KEY;
5738         char tmp_name[32];
5739         char *name_ptr;
5740         int name_len;
5741         int is_curr = 0;        /* ctx->pos points to the current index? */
5742
5743         /* FIXME, use a real flag for deciding about the key type */
5744         if (root->fs_info->tree_root == root)
5745                 key_type = BTRFS_DIR_ITEM_KEY;
5746
5747         if (!dir_emit_dots(file, ctx))
5748                 return 0;
5749
5750         path = btrfs_alloc_path();
5751         if (!path)
5752                 return -ENOMEM;
5753
5754         path->reada = READA_FORWARD;
5755
5756         if (key_type == BTRFS_DIR_INDEX_KEY) {
5757                 INIT_LIST_HEAD(&ins_list);
5758                 INIT_LIST_HEAD(&del_list);
5759                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5760         }
5761
5762         key.type = key_type;
5763         key.offset = ctx->pos;
5764         key.objectid = btrfs_ino(inode);
5765
5766         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5767         if (ret < 0)
5768                 goto err;
5769
5770         while (1) {
5771                 leaf = path->nodes[0];
5772                 slot = path->slots[0];
5773                 if (slot >= btrfs_header_nritems(leaf)) {
5774                         ret = btrfs_next_leaf(root, path);
5775                         if (ret < 0)
5776                                 goto err;
5777                         else if (ret > 0)
5778                                 break;
5779                         continue;
5780                 }
5781
5782                 item = btrfs_item_nr(slot);
5783                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5784
5785                 if (found_key.objectid != key.objectid)
5786                         break;
5787                 if (found_key.type != key_type)
5788                         break;
5789                 if (found_key.offset < ctx->pos)
5790                         goto next;
5791                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5792                     btrfs_should_delete_dir_index(&del_list,
5793                                                   found_key.offset))
5794                         goto next;
5795
5796                 ctx->pos = found_key.offset;
5797                 is_curr = 1;
5798
5799                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5800                 di_cur = 0;
5801                 di_total = btrfs_item_size(leaf, item);
5802
5803                 while (di_cur < di_total) {
5804                         struct btrfs_key location;
5805
5806                         if (verify_dir_item(root, leaf, di))
5807                                 break;
5808
5809                         name_len = btrfs_dir_name_len(leaf, di);
5810                         if (name_len <= sizeof(tmp_name)) {
5811                                 name_ptr = tmp_name;
5812                         } else {
5813                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5814                                 if (!name_ptr) {
5815                                         ret = -ENOMEM;
5816                                         goto err;
5817                                 }
5818                         }
5819                         read_extent_buffer(leaf, name_ptr,
5820                                            (unsigned long)(di + 1), name_len);
5821
5822                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5823                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5824
5825
5826                         /* is this a reference to our own snapshot? If so
5827                          * skip it.
5828                          *
5829                          * In contrast to old kernels, we insert the snapshot's
5830                          * dir item and dir index after it has been created, so
5831                          * we won't find a reference to our own snapshot. We
5832                          * still keep the following code for backward
5833                          * compatibility.
5834                          */
5835                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5836                             location.objectid == root->root_key.objectid) {
5837                                 over = 0;
5838                                 goto skip;
5839                         }
5840                         over = !dir_emit(ctx, name_ptr, name_len,
5841                                        location.objectid, d_type);
5842
5843 skip:
5844                         if (name_ptr != tmp_name)
5845                                 kfree(name_ptr);
5846
5847                         if (over)
5848                                 goto nopos;
5849                         di_len = btrfs_dir_name_len(leaf, di) +
5850                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5851                         di_cur += di_len;
5852                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5853                 }
5854 next:
5855                 path->slots[0]++;
5856         }
5857
5858         if (key_type == BTRFS_DIR_INDEX_KEY) {
5859                 if (is_curr)
5860                         ctx->pos++;
5861                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5862                 if (ret)
5863                         goto nopos;
5864         }
5865
5866         /* Reached end of directory/root. Bump pos past the last item. */
5867         ctx->pos++;
5868
5869         /*
5870          * Stop new entries from being returned after we return the last
5871          * entry.
5872          *
5873          * New directory entries are assigned a strictly increasing
5874          * offset.  This means that new entries created during readdir
5875          * are *guaranteed* to be seen in the future by that readdir.
5876          * This has broken buggy programs which operate on names as
5877          * they're returned by readdir.  Until we re-use freed offsets
5878          * we have this hack to stop new entries from being returned
5879          * under the assumption that they'll never reach this huge
5880          * offset.
5881          *
5882          * This is being careful not to overflow 32bit loff_t unless the
5883          * last entry requires it because doing so has broken 32bit apps
5884          * in the past.
5885          */
5886         if (key_type == BTRFS_DIR_INDEX_KEY) {
5887                 if (ctx->pos >= INT_MAX)
5888                         ctx->pos = LLONG_MAX;
5889                 else
5890                         ctx->pos = INT_MAX;
5891         }
5892 nopos:
5893         ret = 0;
5894 err:
5895         if (key_type == BTRFS_DIR_INDEX_KEY)
5896                 btrfs_put_delayed_items(&ins_list, &del_list);
5897         btrfs_free_path(path);
5898         return ret;
5899 }
5900
5901 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5902 {
5903         struct btrfs_root *root = BTRFS_I(inode)->root;
5904         struct btrfs_trans_handle *trans;
5905         int ret = 0;
5906         bool nolock = false;
5907
5908         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5909                 return 0;
5910
5911         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5912                 nolock = true;
5913
5914         if (wbc->sync_mode == WB_SYNC_ALL) {
5915                 if (nolock)
5916                         trans = btrfs_join_transaction_nolock(root);
5917                 else
5918                         trans = btrfs_join_transaction(root);
5919                 if (IS_ERR(trans))
5920                         return PTR_ERR(trans);
5921                 ret = btrfs_commit_transaction(trans, root);
5922         }
5923         return ret;
5924 }
5925
5926 /*
5927  * This is somewhat expensive, updating the tree every time the
5928  * inode changes.  But, it is most likely to find the inode in cache.
5929  * FIXME, needs more benchmarking...there are no reasons other than performance
5930  * to keep or drop this code.
5931  */
5932 static int btrfs_dirty_inode(struct inode *inode)
5933 {
5934         struct btrfs_root *root = BTRFS_I(inode)->root;
5935         struct btrfs_trans_handle *trans;
5936         int ret;
5937
5938         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5939                 return 0;
5940
5941         trans = btrfs_join_transaction(root);
5942         if (IS_ERR(trans))
5943                 return PTR_ERR(trans);
5944
5945         ret = btrfs_update_inode(trans, root, inode);
5946         if (ret && ret == -ENOSPC) {
5947                 /* whoops, lets try again with the full transaction */
5948                 btrfs_end_transaction(trans, root);
5949                 trans = btrfs_start_transaction(root, 1);
5950                 if (IS_ERR(trans))
5951                         return PTR_ERR(trans);
5952
5953                 ret = btrfs_update_inode(trans, root, inode);
5954         }
5955         btrfs_end_transaction(trans, root);
5956         if (BTRFS_I(inode)->delayed_node)
5957                 btrfs_balance_delayed_items(root);
5958
5959         return ret;
5960 }
5961
5962 /*
5963  * This is a copy of file_update_time.  We need this so we can return error on
5964  * ENOSPC for updating the inode in the case of file write and mmap writes.
5965  */
5966 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5967                              int flags)
5968 {
5969         struct btrfs_root *root = BTRFS_I(inode)->root;
5970
5971         if (btrfs_root_readonly(root))
5972                 return -EROFS;
5973
5974         if (flags & S_VERSION)
5975                 inode_inc_iversion(inode);
5976         if (flags & S_CTIME)
5977                 inode->i_ctime = *now;
5978         if (flags & S_MTIME)
5979                 inode->i_mtime = *now;
5980         if (flags & S_ATIME)
5981                 inode->i_atime = *now;
5982         return btrfs_dirty_inode(inode);
5983 }
5984
5985 /*
5986  * find the highest existing sequence number in a directory
5987  * and then set the in-memory index_cnt variable to reflect
5988  * free sequence numbers
5989  */
5990 static int btrfs_set_inode_index_count(struct inode *inode)
5991 {
5992         struct btrfs_root *root = BTRFS_I(inode)->root;
5993         struct btrfs_key key, found_key;
5994         struct btrfs_path *path;
5995         struct extent_buffer *leaf;
5996         int ret;
5997
5998         key.objectid = btrfs_ino(inode);
5999         key.type = BTRFS_DIR_INDEX_KEY;
6000         key.offset = (u64)-1;
6001
6002         path = btrfs_alloc_path();
6003         if (!path)
6004                 return -ENOMEM;
6005
6006         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6007         if (ret < 0)
6008                 goto out;
6009         /* FIXME: we should be able to handle this */
6010         if (ret == 0)
6011                 goto out;
6012         ret = 0;
6013
6014         /*
6015          * MAGIC NUMBER EXPLANATION:
6016          * since we search a directory based on f_pos we have to start at 2
6017          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6018          * else has to start at 2
6019          */
6020         if (path->slots[0] == 0) {
6021                 BTRFS_I(inode)->index_cnt = 2;
6022                 goto out;
6023         }
6024
6025         path->slots[0]--;
6026
6027         leaf = path->nodes[0];
6028         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6029
6030         if (found_key.objectid != btrfs_ino(inode) ||
6031             found_key.type != BTRFS_DIR_INDEX_KEY) {
6032                 BTRFS_I(inode)->index_cnt = 2;
6033                 goto out;
6034         }
6035
6036         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6037 out:
6038         btrfs_free_path(path);
6039         return ret;
6040 }
6041
6042 /*
6043  * helper to find a free sequence number in a given directory.  This current
6044  * code is very simple, later versions will do smarter things in the btree
6045  */
6046 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6047 {
6048         int ret = 0;
6049
6050         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6051                 ret = btrfs_inode_delayed_dir_index_count(dir);
6052                 if (ret) {
6053                         ret = btrfs_set_inode_index_count(dir);
6054                         if (ret)
6055                                 return ret;
6056                 }
6057         }
6058
6059         *index = BTRFS_I(dir)->index_cnt;
6060         BTRFS_I(dir)->index_cnt++;
6061
6062         return ret;
6063 }
6064
6065 static int btrfs_insert_inode_locked(struct inode *inode)
6066 {
6067         struct btrfs_iget_args args;
6068         args.location = &BTRFS_I(inode)->location;
6069         args.root = BTRFS_I(inode)->root;
6070
6071         return insert_inode_locked4(inode,
6072                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6073                    btrfs_find_actor, &args);
6074 }
6075
6076 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6077                                      struct btrfs_root *root,
6078                                      struct inode *dir,
6079                                      const char *name, int name_len,
6080                                      u64 ref_objectid, u64 objectid,
6081                                      umode_t mode, u64 *index)
6082 {
6083         struct inode *inode;
6084         struct btrfs_inode_item *inode_item;
6085         struct btrfs_key *location;
6086         struct btrfs_path *path;
6087         struct btrfs_inode_ref *ref;
6088         struct btrfs_key key[2];
6089         u32 sizes[2];
6090         int nitems = name ? 2 : 1;
6091         unsigned long ptr;
6092         int ret;
6093
6094         path = btrfs_alloc_path();
6095         if (!path)
6096                 return ERR_PTR(-ENOMEM);
6097
6098         inode = new_inode(root->fs_info->sb);
6099         if (!inode) {
6100                 btrfs_free_path(path);
6101                 return ERR_PTR(-ENOMEM);
6102         }
6103
6104         /*
6105          * O_TMPFILE, set link count to 0, so that after this point,
6106          * we fill in an inode item with the correct link count.
6107          */
6108         if (!name)
6109                 set_nlink(inode, 0);
6110
6111         /*
6112          * we have to initialize this early, so we can reclaim the inode
6113          * number if we fail afterwards in this function.
6114          */
6115         inode->i_ino = objectid;
6116
6117         if (dir && name) {
6118                 trace_btrfs_inode_request(dir);
6119
6120                 ret = btrfs_set_inode_index(dir, index);
6121                 if (ret) {
6122                         btrfs_free_path(path);
6123                         iput(inode);
6124                         return ERR_PTR(ret);
6125                 }
6126         } else if (dir) {
6127                 *index = 0;
6128         }
6129         /*
6130          * index_cnt is ignored for everything but a dir,
6131          * btrfs_get_inode_index_count has an explanation for the magic
6132          * number
6133          */
6134         BTRFS_I(inode)->index_cnt = 2;
6135         BTRFS_I(inode)->dir_index = *index;
6136         BTRFS_I(inode)->root = root;
6137         BTRFS_I(inode)->generation = trans->transid;
6138         inode->i_generation = BTRFS_I(inode)->generation;
6139
6140         /*
6141          * We could have gotten an inode number from somebody who was fsynced
6142          * and then removed in this same transaction, so let's just set full
6143          * sync since it will be a full sync anyway and this will blow away the
6144          * old info in the log.
6145          */
6146         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6147
6148         key[0].objectid = objectid;
6149         key[0].type = BTRFS_INODE_ITEM_KEY;
6150         key[0].offset = 0;
6151
6152         sizes[0] = sizeof(struct btrfs_inode_item);
6153
6154         if (name) {
6155                 /*
6156                  * Start new inodes with an inode_ref. This is slightly more
6157                  * efficient for small numbers of hard links since they will
6158                  * be packed into one item. Extended refs will kick in if we
6159                  * add more hard links than can fit in the ref item.
6160                  */
6161                 key[1].objectid = objectid;
6162                 key[1].type = BTRFS_INODE_REF_KEY;
6163                 key[1].offset = ref_objectid;
6164
6165                 sizes[1] = name_len + sizeof(*ref);
6166         }
6167
6168         location = &BTRFS_I(inode)->location;
6169         location->objectid = objectid;
6170         location->offset = 0;
6171         location->type = BTRFS_INODE_ITEM_KEY;
6172
6173         ret = btrfs_insert_inode_locked(inode);
6174         if (ret < 0)
6175                 goto fail;
6176
6177         path->leave_spinning = 1;
6178         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6179         if (ret != 0)
6180                 goto fail_unlock;
6181
6182         inode_init_owner(inode, dir, mode);
6183         inode_set_bytes(inode, 0);
6184
6185         inode->i_mtime = CURRENT_TIME;
6186         inode->i_atime = inode->i_mtime;
6187         inode->i_ctime = inode->i_mtime;
6188         BTRFS_I(inode)->i_otime = inode->i_mtime;
6189
6190         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6191                                   struct btrfs_inode_item);
6192         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6193                              sizeof(*inode_item));
6194         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6195
6196         if (name) {
6197                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6198                                      struct btrfs_inode_ref);
6199                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6200                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6201                 ptr = (unsigned long)(ref + 1);
6202                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6203         }
6204
6205         btrfs_mark_buffer_dirty(path->nodes[0]);
6206         btrfs_free_path(path);
6207
6208         btrfs_inherit_iflags(inode, dir);
6209
6210         if (S_ISREG(mode)) {
6211                 if (btrfs_test_opt(root, NODATASUM))
6212                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6213                 if (btrfs_test_opt(root, NODATACOW))
6214                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6215                                 BTRFS_INODE_NODATASUM;
6216         }
6217
6218         inode_tree_add(inode);
6219
6220         trace_btrfs_inode_new(inode);
6221         btrfs_set_inode_last_trans(trans, inode);
6222
6223         btrfs_update_root_times(trans, root);
6224
6225         ret = btrfs_inode_inherit_props(trans, inode, dir);
6226         if (ret)
6227                 btrfs_err(root->fs_info,
6228                           "error inheriting props for ino %llu (root %llu): %d",
6229                           btrfs_ino(inode), root->root_key.objectid, ret);
6230
6231         return inode;
6232
6233 fail_unlock:
6234         unlock_new_inode(inode);
6235 fail:
6236         if (dir && name)
6237                 BTRFS_I(dir)->index_cnt--;
6238         btrfs_free_path(path);
6239         iput(inode);
6240         return ERR_PTR(ret);
6241 }
6242
6243 static inline u8 btrfs_inode_type(struct inode *inode)
6244 {
6245         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6246 }
6247
6248 /*
6249  * utility function to add 'inode' into 'parent_inode' with
6250  * a give name and a given sequence number.
6251  * if 'add_backref' is true, also insert a backref from the
6252  * inode to the parent directory.
6253  */
6254 int btrfs_add_link(struct btrfs_trans_handle *trans,
6255                    struct inode *parent_inode, struct inode *inode,
6256                    const char *name, int name_len, int add_backref, u64 index)
6257 {
6258         int ret = 0;
6259         struct btrfs_key key;
6260         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6261         u64 ino = btrfs_ino(inode);
6262         u64 parent_ino = btrfs_ino(parent_inode);
6263
6264         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6265                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6266         } else {
6267                 key.objectid = ino;
6268                 key.type = BTRFS_INODE_ITEM_KEY;
6269                 key.offset = 0;
6270         }
6271
6272         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6273                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6274                                          key.objectid, root->root_key.objectid,
6275                                          parent_ino, index, name, name_len);
6276         } else if (add_backref) {
6277                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6278                                              parent_ino, index);
6279         }
6280
6281         /* Nothing to clean up yet */
6282         if (ret)
6283                 return ret;
6284
6285         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6286                                     parent_inode, &key,
6287                                     btrfs_inode_type(inode), index);
6288         if (ret == -EEXIST || ret == -EOVERFLOW)
6289                 goto fail_dir_item;
6290         else if (ret) {
6291                 btrfs_abort_transaction(trans, root, ret);
6292                 return ret;
6293         }
6294
6295         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6296                            name_len * 2);
6297         inode_inc_iversion(parent_inode);
6298         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6299         ret = btrfs_update_inode(trans, root, parent_inode);
6300         if (ret)
6301                 btrfs_abort_transaction(trans, root, ret);
6302         return ret;
6303
6304 fail_dir_item:
6305         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6306                 u64 local_index;
6307                 int err;
6308                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6309                                  key.objectid, root->root_key.objectid,
6310                                  parent_ino, &local_index, name, name_len);
6311
6312         } else if (add_backref) {
6313                 u64 local_index;
6314                 int err;
6315
6316                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6317                                           ino, parent_ino, &local_index);
6318         }
6319         return ret;
6320 }
6321
6322 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6323                             struct inode *dir, struct dentry *dentry,
6324                             struct inode *inode, int backref, u64 index)
6325 {
6326         int err = btrfs_add_link(trans, dir, inode,
6327                                  dentry->d_name.name, dentry->d_name.len,
6328                                  backref, index);
6329         if (err > 0)
6330                 err = -EEXIST;
6331         return err;
6332 }
6333
6334 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6335                         umode_t mode, dev_t rdev)
6336 {
6337         struct btrfs_trans_handle *trans;
6338         struct btrfs_root *root = BTRFS_I(dir)->root;
6339         struct inode *inode = NULL;
6340         int err;
6341         int drop_inode = 0;
6342         u64 objectid;
6343         u64 index = 0;
6344
6345         /*
6346          * 2 for inode item and ref
6347          * 2 for dir items
6348          * 1 for xattr if selinux is on
6349          */
6350         trans = btrfs_start_transaction(root, 5);
6351         if (IS_ERR(trans))
6352                 return PTR_ERR(trans);
6353
6354         err = btrfs_find_free_ino(root, &objectid);
6355         if (err)
6356                 goto out_unlock;
6357
6358         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6359                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6360                                 mode, &index);
6361         if (IS_ERR(inode)) {
6362                 err = PTR_ERR(inode);
6363                 goto out_unlock;
6364         }
6365
6366         /*
6367         * If the active LSM wants to access the inode during
6368         * d_instantiate it needs these. Smack checks to see
6369         * if the filesystem supports xattrs by looking at the
6370         * ops vector.
6371         */
6372         inode->i_op = &btrfs_special_inode_operations;
6373         init_special_inode(inode, inode->i_mode, rdev);
6374
6375         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6376         if (err)
6377                 goto out_unlock_inode;
6378
6379         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6380         if (err) {
6381                 goto out_unlock_inode;
6382         } else {
6383                 btrfs_update_inode(trans, root, inode);
6384                 unlock_new_inode(inode);
6385                 d_instantiate(dentry, inode);
6386         }
6387
6388 out_unlock:
6389         btrfs_end_transaction(trans, root);
6390         btrfs_balance_delayed_items(root);
6391         btrfs_btree_balance_dirty(root);
6392         if (drop_inode) {
6393                 inode_dec_link_count(inode);
6394                 iput(inode);
6395         }
6396         return err;
6397
6398 out_unlock_inode:
6399         drop_inode = 1;
6400         unlock_new_inode(inode);
6401         goto out_unlock;
6402
6403 }
6404
6405 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6406                         umode_t mode, bool excl)
6407 {
6408         struct btrfs_trans_handle *trans;
6409         struct btrfs_root *root = BTRFS_I(dir)->root;
6410         struct inode *inode = NULL;
6411         int drop_inode_on_err = 0;
6412         int err;
6413         u64 objectid;
6414         u64 index = 0;
6415
6416         /*
6417          * 2 for inode item and ref
6418          * 2 for dir items
6419          * 1 for xattr if selinux is on
6420          */
6421         trans = btrfs_start_transaction(root, 5);
6422         if (IS_ERR(trans))
6423                 return PTR_ERR(trans);
6424
6425         err = btrfs_find_free_ino(root, &objectid);
6426         if (err)
6427                 goto out_unlock;
6428
6429         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6430                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6431                                 mode, &index);
6432         if (IS_ERR(inode)) {
6433                 err = PTR_ERR(inode);
6434                 goto out_unlock;
6435         }
6436         drop_inode_on_err = 1;
6437         /*
6438         * If the active LSM wants to access the inode during
6439         * d_instantiate it needs these. Smack checks to see
6440         * if the filesystem supports xattrs by looking at the
6441         * ops vector.
6442         */
6443         inode->i_fop = &btrfs_file_operations;
6444         inode->i_op = &btrfs_file_inode_operations;
6445         inode->i_mapping->a_ops = &btrfs_aops;
6446
6447         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6448         if (err)
6449                 goto out_unlock_inode;
6450
6451         err = btrfs_update_inode(trans, root, inode);
6452         if (err)
6453                 goto out_unlock_inode;
6454
6455         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6456         if (err)
6457                 goto out_unlock_inode;
6458
6459         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6460         unlock_new_inode(inode);
6461         d_instantiate(dentry, inode);
6462
6463 out_unlock:
6464         btrfs_end_transaction(trans, root);
6465         if (err && drop_inode_on_err) {
6466                 inode_dec_link_count(inode);
6467                 iput(inode);
6468         }
6469         btrfs_balance_delayed_items(root);
6470         btrfs_btree_balance_dirty(root);
6471         return err;
6472
6473 out_unlock_inode:
6474         unlock_new_inode(inode);
6475         goto out_unlock;
6476
6477 }
6478
6479 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6480                       struct dentry *dentry)
6481 {
6482         struct btrfs_trans_handle *trans = NULL;
6483         struct btrfs_root *root = BTRFS_I(dir)->root;
6484         struct inode *inode = d_inode(old_dentry);
6485         u64 index;
6486         int err;
6487         int drop_inode = 0;
6488
6489         /* do not allow sys_link's with other subvols of the same device */
6490         if (root->objectid != BTRFS_I(inode)->root->objectid)
6491                 return -EXDEV;
6492
6493         if (inode->i_nlink >= BTRFS_LINK_MAX)
6494                 return -EMLINK;
6495
6496         err = btrfs_set_inode_index(dir, &index);
6497         if (err)
6498                 goto fail;
6499
6500         /*
6501          * 2 items for inode and inode ref
6502          * 2 items for dir items
6503          * 1 item for parent inode
6504          */
6505         trans = btrfs_start_transaction(root, 5);
6506         if (IS_ERR(trans)) {
6507                 err = PTR_ERR(trans);
6508                 trans = NULL;
6509                 goto fail;
6510         }
6511
6512         /* There are several dir indexes for this inode, clear the cache. */
6513         BTRFS_I(inode)->dir_index = 0ULL;
6514         inc_nlink(inode);
6515         inode_inc_iversion(inode);
6516         inode->i_ctime = CURRENT_TIME;
6517         ihold(inode);
6518         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6519
6520         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6521
6522         if (err) {
6523                 drop_inode = 1;
6524         } else {
6525                 struct dentry *parent = dentry->d_parent;
6526                 err = btrfs_update_inode(trans, root, inode);
6527                 if (err)
6528                         goto fail;
6529                 if (inode->i_nlink == 1) {
6530                         /*
6531                          * If new hard link count is 1, it's a file created
6532                          * with open(2) O_TMPFILE flag.
6533                          */
6534                         err = btrfs_orphan_del(trans, inode);
6535                         if (err)
6536                                 goto fail;
6537                 }
6538                 d_instantiate(dentry, inode);
6539                 btrfs_log_new_name(trans, inode, NULL, parent);
6540         }
6541
6542         btrfs_balance_delayed_items(root);
6543 fail:
6544         if (trans)
6545                 btrfs_end_transaction(trans, root);
6546         if (drop_inode) {
6547                 inode_dec_link_count(inode);
6548                 iput(inode);
6549         }
6550         btrfs_btree_balance_dirty(root);
6551         return err;
6552 }
6553
6554 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6555 {
6556         struct inode *inode = NULL;
6557         struct btrfs_trans_handle *trans;
6558         struct btrfs_root *root = BTRFS_I(dir)->root;
6559         int err = 0;
6560         int drop_on_err = 0;
6561         u64 objectid = 0;
6562         u64 index = 0;
6563
6564         /*
6565          * 2 items for inode and ref
6566          * 2 items for dir items
6567          * 1 for xattr if selinux is on
6568          */
6569         trans = btrfs_start_transaction(root, 5);
6570         if (IS_ERR(trans))
6571                 return PTR_ERR(trans);
6572
6573         err = btrfs_find_free_ino(root, &objectid);
6574         if (err)
6575                 goto out_fail;
6576
6577         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6578                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6579                                 S_IFDIR | mode, &index);
6580         if (IS_ERR(inode)) {
6581                 err = PTR_ERR(inode);
6582                 goto out_fail;
6583         }
6584
6585         drop_on_err = 1;
6586         /* these must be set before we unlock the inode */
6587         inode->i_op = &btrfs_dir_inode_operations;
6588         inode->i_fop = &btrfs_dir_file_operations;
6589
6590         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6591         if (err)
6592                 goto out_fail_inode;
6593
6594         btrfs_i_size_write(inode, 0);
6595         err = btrfs_update_inode(trans, root, inode);
6596         if (err)
6597                 goto out_fail_inode;
6598
6599         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6600                              dentry->d_name.len, 0, index);
6601         if (err)
6602                 goto out_fail_inode;
6603
6604         d_instantiate(dentry, inode);
6605         /*
6606          * mkdir is special.  We're unlocking after we call d_instantiate
6607          * to avoid a race with nfsd calling d_instantiate.
6608          */
6609         unlock_new_inode(inode);
6610         drop_on_err = 0;
6611
6612 out_fail:
6613         btrfs_end_transaction(trans, root);
6614         if (drop_on_err) {
6615                 inode_dec_link_count(inode);
6616                 iput(inode);
6617         }
6618         btrfs_balance_delayed_items(root);
6619         btrfs_btree_balance_dirty(root);
6620         return err;
6621
6622 out_fail_inode:
6623         unlock_new_inode(inode);
6624         goto out_fail;
6625 }
6626
6627 /* Find next extent map of a given extent map, caller needs to ensure locks */
6628 static struct extent_map *next_extent_map(struct extent_map *em)
6629 {
6630         struct rb_node *next;
6631
6632         next = rb_next(&em->rb_node);
6633         if (!next)
6634                 return NULL;
6635         return container_of(next, struct extent_map, rb_node);
6636 }
6637
6638 static struct extent_map *prev_extent_map(struct extent_map *em)
6639 {
6640         struct rb_node *prev;
6641
6642         prev = rb_prev(&em->rb_node);
6643         if (!prev)
6644                 return NULL;
6645         return container_of(prev, struct extent_map, rb_node);
6646 }
6647
6648 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6649  * the existing extent is the nearest extent to map_start,
6650  * and an extent that you want to insert, deal with overlap and insert
6651  * the best fitted new extent into the tree.
6652  */
6653 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6654                                 struct extent_map *existing,
6655                                 struct extent_map *em,
6656                                 u64 map_start)
6657 {
6658         struct extent_map *prev;
6659         struct extent_map *next;
6660         u64 start;
6661         u64 end;
6662         u64 start_diff;
6663
6664         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6665
6666         if (existing->start > map_start) {
6667                 next = existing;
6668                 prev = prev_extent_map(next);
6669         } else {
6670                 prev = existing;
6671                 next = next_extent_map(prev);
6672         }
6673
6674         start = prev ? extent_map_end(prev) : em->start;
6675         start = max_t(u64, start, em->start);
6676         end = next ? next->start : extent_map_end(em);
6677         end = min_t(u64, end, extent_map_end(em));
6678         start_diff = start - em->start;
6679         em->start = start;
6680         em->len = end - start;
6681         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6682             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6683                 em->block_start += start_diff;
6684                 em->block_len -= start_diff;
6685         }
6686         return add_extent_mapping(em_tree, em, 0);
6687 }
6688
6689 static noinline int uncompress_inline(struct btrfs_path *path,
6690                                       struct page *page,
6691                                       size_t pg_offset, u64 extent_offset,
6692                                       struct btrfs_file_extent_item *item)
6693 {
6694         int ret;
6695         struct extent_buffer *leaf = path->nodes[0];
6696         char *tmp;
6697         size_t max_size;
6698         unsigned long inline_size;
6699         unsigned long ptr;
6700         int compress_type;
6701
6702         WARN_ON(pg_offset != 0);
6703         compress_type = btrfs_file_extent_compression(leaf, item);
6704         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6705         inline_size = btrfs_file_extent_inline_item_len(leaf,
6706                                         btrfs_item_nr(path->slots[0]));
6707         tmp = kmalloc(inline_size, GFP_NOFS);
6708         if (!tmp)
6709                 return -ENOMEM;
6710         ptr = btrfs_file_extent_inline_start(item);
6711
6712         read_extent_buffer(leaf, tmp, ptr, inline_size);
6713
6714         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6715         ret = btrfs_decompress(compress_type, tmp, page,
6716                                extent_offset, inline_size, max_size);
6717         kfree(tmp);
6718         return ret;
6719 }
6720
6721 /*
6722  * a bit scary, this does extent mapping from logical file offset to the disk.
6723  * the ugly parts come from merging extents from the disk with the in-ram
6724  * representation.  This gets more complex because of the data=ordered code,
6725  * where the in-ram extents might be locked pending data=ordered completion.
6726  *
6727  * This also copies inline extents directly into the page.
6728  */
6729
6730 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6731                                     size_t pg_offset, u64 start, u64 len,
6732                                     int create)
6733 {
6734         int ret;
6735         int err = 0;
6736         u64 extent_start = 0;
6737         u64 extent_end = 0;
6738         u64 objectid = btrfs_ino(inode);
6739         u32 found_type;
6740         struct btrfs_path *path = NULL;
6741         struct btrfs_root *root = BTRFS_I(inode)->root;
6742         struct btrfs_file_extent_item *item;
6743         struct extent_buffer *leaf;
6744         struct btrfs_key found_key;
6745         struct extent_map *em = NULL;
6746         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6747         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6748         struct btrfs_trans_handle *trans = NULL;
6749         const bool new_inline = !page || create;
6750
6751 again:
6752         read_lock(&em_tree->lock);
6753         em = lookup_extent_mapping(em_tree, start, len);
6754         if (em)
6755                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6756         read_unlock(&em_tree->lock);
6757
6758         if (em) {
6759                 if (em->start > start || em->start + em->len <= start)
6760                         free_extent_map(em);
6761                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6762                         free_extent_map(em);
6763                 else
6764                         goto out;
6765         }
6766         em = alloc_extent_map();
6767         if (!em) {
6768                 err = -ENOMEM;
6769                 goto out;
6770         }
6771         em->bdev = root->fs_info->fs_devices->latest_bdev;
6772         em->start = EXTENT_MAP_HOLE;
6773         em->orig_start = EXTENT_MAP_HOLE;
6774         em->len = (u64)-1;
6775         em->block_len = (u64)-1;
6776
6777         if (!path) {
6778                 path = btrfs_alloc_path();
6779                 if (!path) {
6780                         err = -ENOMEM;
6781                         goto out;
6782                 }
6783                 /*
6784                  * Chances are we'll be called again, so go ahead and do
6785                  * readahead
6786                  */
6787                 path->reada = READA_FORWARD;
6788         }
6789
6790         ret = btrfs_lookup_file_extent(trans, root, path,
6791                                        objectid, start, trans != NULL);
6792         if (ret < 0) {
6793                 err = ret;
6794                 goto out;
6795         }
6796
6797         if (ret != 0) {
6798                 if (path->slots[0] == 0)
6799                         goto not_found;
6800                 path->slots[0]--;
6801         }
6802
6803         leaf = path->nodes[0];
6804         item = btrfs_item_ptr(leaf, path->slots[0],
6805                               struct btrfs_file_extent_item);
6806         /* are we inside the extent that was found? */
6807         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6808         found_type = found_key.type;
6809         if (found_key.objectid != objectid ||
6810             found_type != BTRFS_EXTENT_DATA_KEY) {
6811                 /*
6812                  * If we backup past the first extent we want to move forward
6813                  * and see if there is an extent in front of us, otherwise we'll
6814                  * say there is a hole for our whole search range which can
6815                  * cause problems.
6816                  */
6817                 extent_end = start;
6818                 goto next;
6819         }
6820
6821         found_type = btrfs_file_extent_type(leaf, item);
6822         extent_start = found_key.offset;
6823         if (found_type == BTRFS_FILE_EXTENT_REG ||
6824             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6825                 extent_end = extent_start +
6826                        btrfs_file_extent_num_bytes(leaf, item);
6827         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6828                 size_t size;
6829                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6830                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6831         }
6832 next:
6833         if (start >= extent_end) {
6834                 path->slots[0]++;
6835                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6836                         ret = btrfs_next_leaf(root, path);
6837                         if (ret < 0) {
6838                                 err = ret;
6839                                 goto out;
6840                         }
6841                         if (ret > 0)
6842                                 goto not_found;
6843                         leaf = path->nodes[0];
6844                 }
6845                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6846                 if (found_key.objectid != objectid ||
6847                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6848                         goto not_found;
6849                 if (start + len <= found_key.offset)
6850                         goto not_found;
6851                 if (start > found_key.offset)
6852                         goto next;
6853                 em->start = start;
6854                 em->orig_start = start;
6855                 em->len = found_key.offset - start;
6856                 goto not_found_em;
6857         }
6858
6859         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6860
6861         if (found_type == BTRFS_FILE_EXTENT_REG ||
6862             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6863                 goto insert;
6864         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6865                 unsigned long ptr;
6866                 char *map;
6867                 size_t size;
6868                 size_t extent_offset;
6869                 size_t copy_size;
6870
6871                 if (new_inline)
6872                         goto out;
6873
6874                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6875                 extent_offset = page_offset(page) + pg_offset - extent_start;
6876                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6877                                 size - extent_offset);
6878                 em->start = extent_start + extent_offset;
6879                 em->len = ALIGN(copy_size, root->sectorsize);
6880                 em->orig_block_len = em->len;
6881                 em->orig_start = em->start;
6882                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6883                 if (create == 0 && !PageUptodate(page)) {
6884                         if (btrfs_file_extent_compression(leaf, item) !=
6885                             BTRFS_COMPRESS_NONE) {
6886                                 ret = uncompress_inline(path, page, pg_offset,
6887                                                         extent_offset, item);
6888                                 if (ret) {
6889                                         err = ret;
6890                                         goto out;
6891                                 }
6892                         } else {
6893                                 map = kmap(page);
6894                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6895                                                    copy_size);
6896                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6897                                         memset(map + pg_offset + copy_size, 0,
6898                                                PAGE_CACHE_SIZE - pg_offset -
6899                                                copy_size);
6900                                 }
6901                                 kunmap(page);
6902                         }
6903                         flush_dcache_page(page);
6904                 } else if (create && PageUptodate(page)) {
6905                         BUG();
6906                         if (!trans) {
6907                                 kunmap(page);
6908                                 free_extent_map(em);
6909                                 em = NULL;
6910
6911                                 btrfs_release_path(path);
6912                                 trans = btrfs_join_transaction(root);
6913
6914                                 if (IS_ERR(trans))
6915                                         return ERR_CAST(trans);
6916                                 goto again;
6917                         }
6918                         map = kmap(page);
6919                         write_extent_buffer(leaf, map + pg_offset, ptr,
6920                                             copy_size);
6921                         kunmap(page);
6922                         btrfs_mark_buffer_dirty(leaf);
6923                 }
6924                 set_extent_uptodate(io_tree, em->start,
6925                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6926                 goto insert;
6927         }
6928 not_found:
6929         em->start = start;
6930         em->orig_start = start;
6931         em->len = len;
6932 not_found_em:
6933         em->block_start = EXTENT_MAP_HOLE;
6934         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6935 insert:
6936         btrfs_release_path(path);
6937         if (em->start > start || extent_map_end(em) <= start) {
6938                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6939                         em->start, em->len, start, len);
6940                 err = -EIO;
6941                 goto out;
6942         }
6943
6944         err = 0;
6945         write_lock(&em_tree->lock);
6946         ret = add_extent_mapping(em_tree, em, 0);
6947         /* it is possible that someone inserted the extent into the tree
6948          * while we had the lock dropped.  It is also possible that
6949          * an overlapping map exists in the tree
6950          */
6951         if (ret == -EEXIST) {
6952                 struct extent_map *existing;
6953
6954                 ret = 0;
6955
6956                 existing = search_extent_mapping(em_tree, start, len);
6957                 /*
6958                  * existing will always be non-NULL, since there must be
6959                  * extent causing the -EEXIST.
6960                  */
6961                 if (start >= extent_map_end(existing) ||
6962                     start <= existing->start) {
6963                         /*
6964                          * The existing extent map is the one nearest to
6965                          * the [start, start + len) range which overlaps
6966                          */
6967                         err = merge_extent_mapping(em_tree, existing,
6968                                                    em, start);
6969                         free_extent_map(existing);
6970                         if (err) {
6971                                 free_extent_map(em);
6972                                 em = NULL;
6973                         }
6974                 } else {
6975                         free_extent_map(em);
6976                         em = existing;
6977                         err = 0;
6978                 }
6979         }
6980         write_unlock(&em_tree->lock);
6981 out:
6982
6983         trace_btrfs_get_extent(root, em);
6984
6985         btrfs_free_path(path);
6986         if (trans) {
6987                 ret = btrfs_end_transaction(trans, root);
6988                 if (!err)
6989                         err = ret;
6990         }
6991         if (err) {
6992                 free_extent_map(em);
6993                 return ERR_PTR(err);
6994         }
6995         BUG_ON(!em); /* Error is always set */
6996         return em;
6997 }
6998
6999 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7000                                            size_t pg_offset, u64 start, u64 len,
7001                                            int create)
7002 {
7003         struct extent_map *em;
7004         struct extent_map *hole_em = NULL;
7005         u64 range_start = start;
7006         u64 end;
7007         u64 found;
7008         u64 found_end;
7009         int err = 0;
7010
7011         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7012         if (IS_ERR(em))
7013                 return em;
7014         if (em) {
7015                 /*
7016                  * if our em maps to
7017                  * -  a hole or
7018                  * -  a pre-alloc extent,
7019                  * there might actually be delalloc bytes behind it.
7020                  */
7021                 if (em->block_start != EXTENT_MAP_HOLE &&
7022                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7023                         return em;
7024                 else
7025                         hole_em = em;
7026         }
7027
7028         /* check to see if we've wrapped (len == -1 or similar) */
7029         end = start + len;
7030         if (end < start)
7031                 end = (u64)-1;
7032         else
7033                 end -= 1;
7034
7035         em = NULL;
7036
7037         /* ok, we didn't find anything, lets look for delalloc */
7038         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7039                                  end, len, EXTENT_DELALLOC, 1);
7040         found_end = range_start + found;
7041         if (found_end < range_start)
7042                 found_end = (u64)-1;
7043
7044         /*
7045          * we didn't find anything useful, return
7046          * the original results from get_extent()
7047          */
7048         if (range_start > end || found_end <= start) {
7049                 em = hole_em;
7050                 hole_em = NULL;
7051                 goto out;
7052         }
7053
7054         /* adjust the range_start to make sure it doesn't
7055          * go backwards from the start they passed in
7056          */
7057         range_start = max(start, range_start);
7058         found = found_end - range_start;
7059
7060         if (found > 0) {
7061                 u64 hole_start = start;
7062                 u64 hole_len = len;
7063
7064                 em = alloc_extent_map();
7065                 if (!em) {
7066                         err = -ENOMEM;
7067                         goto out;
7068                 }
7069                 /*
7070                  * when btrfs_get_extent can't find anything it
7071                  * returns one huge hole
7072                  *
7073                  * make sure what it found really fits our range, and
7074                  * adjust to make sure it is based on the start from
7075                  * the caller
7076                  */
7077                 if (hole_em) {
7078                         u64 calc_end = extent_map_end(hole_em);
7079
7080                         if (calc_end <= start || (hole_em->start > end)) {
7081                                 free_extent_map(hole_em);
7082                                 hole_em = NULL;
7083                         } else {
7084                                 hole_start = max(hole_em->start, start);
7085                                 hole_len = calc_end - hole_start;
7086                         }
7087                 }
7088                 em->bdev = NULL;
7089                 if (hole_em && range_start > hole_start) {
7090                         /* our hole starts before our delalloc, so we
7091                          * have to return just the parts of the hole
7092                          * that go until  the delalloc starts
7093                          */
7094                         em->len = min(hole_len,
7095                                       range_start - hole_start);
7096                         em->start = hole_start;
7097                         em->orig_start = hole_start;
7098                         /*
7099                          * don't adjust block start at all,
7100                          * it is fixed at EXTENT_MAP_HOLE
7101                          */
7102                         em->block_start = hole_em->block_start;
7103                         em->block_len = hole_len;
7104                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7105                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7106                 } else {
7107                         em->start = range_start;
7108                         em->len = found;
7109                         em->orig_start = range_start;
7110                         em->block_start = EXTENT_MAP_DELALLOC;
7111                         em->block_len = found;
7112                 }
7113         } else if (hole_em) {
7114                 return hole_em;
7115         }
7116 out:
7117
7118         free_extent_map(hole_em);
7119         if (err) {
7120                 free_extent_map(em);
7121                 return ERR_PTR(err);
7122         }
7123         return em;
7124 }
7125
7126 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7127                                                   u64 start, u64 len)
7128 {
7129         struct btrfs_root *root = BTRFS_I(inode)->root;
7130         struct extent_map *em;
7131         struct btrfs_key ins;
7132         u64 alloc_hint;
7133         int ret;
7134
7135         alloc_hint = get_extent_allocation_hint(inode, start, len);
7136         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7137                                    alloc_hint, &ins, 1, 1);
7138         if (ret)
7139                 return ERR_PTR(ret);
7140
7141         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7142                               ins.offset, ins.offset, ins.offset, 0);
7143         if (IS_ERR(em)) {
7144                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7145                 return em;
7146         }
7147
7148         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7149                                            ins.offset, ins.offset, 0);
7150         if (ret) {
7151                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7152                 free_extent_map(em);
7153                 return ERR_PTR(ret);
7154         }
7155
7156         return em;
7157 }
7158
7159 /*
7160  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7161  * block must be cow'd
7162  */
7163 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7164                               u64 *orig_start, u64 *orig_block_len,
7165                               u64 *ram_bytes)
7166 {
7167         struct btrfs_trans_handle *trans;
7168         struct btrfs_path *path;
7169         int ret;
7170         struct extent_buffer *leaf;
7171         struct btrfs_root *root = BTRFS_I(inode)->root;
7172         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7173         struct btrfs_file_extent_item *fi;
7174         struct btrfs_key key;
7175         u64 disk_bytenr;
7176         u64 backref_offset;
7177         u64 extent_end;
7178         u64 num_bytes;
7179         int slot;
7180         int found_type;
7181         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7182
7183         path = btrfs_alloc_path();
7184         if (!path)
7185                 return -ENOMEM;
7186
7187         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7188                                        offset, 0);
7189         if (ret < 0)
7190                 goto out;
7191
7192         slot = path->slots[0];
7193         if (ret == 1) {
7194                 if (slot == 0) {
7195                         /* can't find the item, must cow */
7196                         ret = 0;
7197                         goto out;
7198                 }
7199                 slot--;
7200         }
7201         ret = 0;
7202         leaf = path->nodes[0];
7203         btrfs_item_key_to_cpu(leaf, &key, slot);
7204         if (key.objectid != btrfs_ino(inode) ||
7205             key.type != BTRFS_EXTENT_DATA_KEY) {
7206                 /* not our file or wrong item type, must cow */
7207                 goto out;
7208         }
7209
7210         if (key.offset > offset) {
7211                 /* Wrong offset, must cow */
7212                 goto out;
7213         }
7214
7215         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7216         found_type = btrfs_file_extent_type(leaf, fi);
7217         if (found_type != BTRFS_FILE_EXTENT_REG &&
7218             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7219                 /* not a regular extent, must cow */
7220                 goto out;
7221         }
7222
7223         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7224                 goto out;
7225
7226         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7227         if (extent_end <= offset)
7228                 goto out;
7229
7230         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7231         if (disk_bytenr == 0)
7232                 goto out;
7233
7234         if (btrfs_file_extent_compression(leaf, fi) ||
7235             btrfs_file_extent_encryption(leaf, fi) ||
7236             btrfs_file_extent_other_encoding(leaf, fi))
7237                 goto out;
7238
7239         backref_offset = btrfs_file_extent_offset(leaf, fi);
7240
7241         if (orig_start) {
7242                 *orig_start = key.offset - backref_offset;
7243                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7244                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7245         }
7246
7247         if (btrfs_extent_readonly(root, disk_bytenr))
7248                 goto out;
7249
7250         num_bytes = min(offset + *len, extent_end) - offset;
7251         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7252                 u64 range_end;
7253
7254                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7255                 ret = test_range_bit(io_tree, offset, range_end,
7256                                      EXTENT_DELALLOC, 0, NULL);
7257                 if (ret) {
7258                         ret = -EAGAIN;
7259                         goto out;
7260                 }
7261         }
7262
7263         btrfs_release_path(path);
7264
7265         /*
7266          * look for other files referencing this extent, if we
7267          * find any we must cow
7268          */
7269         trans = btrfs_join_transaction(root);
7270         if (IS_ERR(trans)) {
7271                 ret = 0;
7272                 goto out;
7273         }
7274
7275         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7276                                     key.offset - backref_offset, disk_bytenr);
7277         btrfs_end_transaction(trans, root);
7278         if (ret) {
7279                 ret = 0;
7280                 goto out;
7281         }
7282
7283         /*
7284          * adjust disk_bytenr and num_bytes to cover just the bytes
7285          * in this extent we are about to write.  If there
7286          * are any csums in that range we have to cow in order
7287          * to keep the csums correct
7288          */
7289         disk_bytenr += backref_offset;
7290         disk_bytenr += offset - key.offset;
7291         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7292                                 goto out;
7293         /*
7294          * all of the above have passed, it is safe to overwrite this extent
7295          * without cow
7296          */
7297         *len = num_bytes;
7298         ret = 1;
7299 out:
7300         btrfs_free_path(path);
7301         return ret;
7302 }
7303
7304 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7305 {
7306         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7307         int found = false;
7308         void **pagep = NULL;
7309         struct page *page = NULL;
7310         int start_idx;
7311         int end_idx;
7312
7313         start_idx = start >> PAGE_CACHE_SHIFT;
7314
7315         /*
7316          * end is the last byte in the last page.  end == start is legal
7317          */
7318         end_idx = end >> PAGE_CACHE_SHIFT;
7319
7320         rcu_read_lock();
7321
7322         /* Most of the code in this while loop is lifted from
7323          * find_get_page.  It's been modified to begin searching from a
7324          * page and return just the first page found in that range.  If the
7325          * found idx is less than or equal to the end idx then we know that
7326          * a page exists.  If no pages are found or if those pages are
7327          * outside of the range then we're fine (yay!) */
7328         while (page == NULL &&
7329                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7330                 page = radix_tree_deref_slot(pagep);
7331                 if (unlikely(!page))
7332                         break;
7333
7334                 if (radix_tree_exception(page)) {
7335                         if (radix_tree_deref_retry(page)) {
7336                                 page = NULL;
7337                                 continue;
7338                         }
7339                         /*
7340                          * Otherwise, shmem/tmpfs must be storing a swap entry
7341                          * here as an exceptional entry: so return it without
7342                          * attempting to raise page count.
7343                          */
7344                         page = NULL;
7345                         break; /* TODO: Is this relevant for this use case? */
7346                 }
7347
7348                 if (!page_cache_get_speculative(page)) {
7349                         page = NULL;
7350                         continue;
7351                 }
7352
7353                 /*
7354                  * Has the page moved?
7355                  * This is part of the lockless pagecache protocol. See
7356                  * include/linux/pagemap.h for details.
7357                  */
7358                 if (unlikely(page != *pagep)) {
7359                         page_cache_release(page);
7360                         page = NULL;
7361                 }
7362         }
7363
7364         if (page) {
7365                 if (page->index <= end_idx)
7366                         found = true;
7367                 page_cache_release(page);
7368         }
7369
7370         rcu_read_unlock();
7371         return found;
7372 }
7373
7374 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7375                               struct extent_state **cached_state, int writing)
7376 {
7377         struct btrfs_ordered_extent *ordered;
7378         int ret = 0;
7379
7380         while (1) {
7381                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7382                                  cached_state);
7383                 /*
7384                  * We're concerned with the entire range that we're going to be
7385                  * doing DIO to, so we need to make sure theres no ordered
7386                  * extents in this range.
7387                  */
7388                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7389                                                      lockend - lockstart + 1);
7390
7391                 /*
7392                  * We need to make sure there are no buffered pages in this
7393                  * range either, we could have raced between the invalidate in
7394                  * generic_file_direct_write and locking the extent.  The
7395                  * invalidate needs to happen so that reads after a write do not
7396                  * get stale data.
7397                  */
7398                 if (!ordered &&
7399                     (!writing ||
7400                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7401                         break;
7402
7403                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7404                                      cached_state, GFP_NOFS);
7405
7406                 if (ordered) {
7407                         btrfs_start_ordered_extent(inode, ordered, 1);
7408                         btrfs_put_ordered_extent(ordered);
7409                 } else {
7410                         /*
7411                          * We could trigger writeback for this range (and wait
7412                          * for it to complete) and then invalidate the pages for
7413                          * this range (through invalidate_inode_pages2_range()),
7414                          * but that can lead us to a deadlock with a concurrent
7415                          * call to readpages() (a buffered read or a defrag call
7416                          * triggered a readahead) on a page lock due to an
7417                          * ordered dio extent we created before but did not have
7418                          * yet a corresponding bio submitted (whence it can not
7419                          * complete), which makes readpages() wait for that
7420                          * ordered extent to complete while holding a lock on
7421                          * that page.
7422                          */
7423                         ret = -ENOTBLK;
7424                         break;
7425                 }
7426
7427                 cond_resched();
7428         }
7429
7430         return ret;
7431 }
7432
7433 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7434                                            u64 len, u64 orig_start,
7435                                            u64 block_start, u64 block_len,
7436                                            u64 orig_block_len, u64 ram_bytes,
7437                                            int type)
7438 {
7439         struct extent_map_tree *em_tree;
7440         struct extent_map *em;
7441         struct btrfs_root *root = BTRFS_I(inode)->root;
7442         int ret;
7443
7444         em_tree = &BTRFS_I(inode)->extent_tree;
7445         em = alloc_extent_map();
7446         if (!em)
7447                 return ERR_PTR(-ENOMEM);
7448
7449         em->start = start;
7450         em->orig_start = orig_start;
7451         em->mod_start = start;
7452         em->mod_len = len;
7453         em->len = len;
7454         em->block_len = block_len;
7455         em->block_start = block_start;
7456         em->bdev = root->fs_info->fs_devices->latest_bdev;
7457         em->orig_block_len = orig_block_len;
7458         em->ram_bytes = ram_bytes;
7459         em->generation = -1;
7460         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7461         if (type == BTRFS_ORDERED_PREALLOC)
7462                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7463
7464         do {
7465                 btrfs_drop_extent_cache(inode, em->start,
7466                                 em->start + em->len - 1, 0);
7467                 write_lock(&em_tree->lock);
7468                 ret = add_extent_mapping(em_tree, em, 1);
7469                 write_unlock(&em_tree->lock);
7470         } while (ret == -EEXIST);
7471
7472         if (ret) {
7473                 free_extent_map(em);
7474                 return ERR_PTR(ret);
7475         }
7476
7477         return em;
7478 }
7479
7480 static void adjust_dio_outstanding_extents(struct inode *inode,
7481                                            struct btrfs_dio_data *dio_data,
7482                                            const u64 len)
7483 {
7484         unsigned num_extents;
7485
7486         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7487                                            BTRFS_MAX_EXTENT_SIZE);
7488         /*
7489          * If we have an outstanding_extents count still set then we're
7490          * within our reservation, otherwise we need to adjust our inode
7491          * counter appropriately.
7492          */
7493         if (dio_data->outstanding_extents) {
7494                 dio_data->outstanding_extents -= num_extents;
7495         } else {
7496                 spin_lock(&BTRFS_I(inode)->lock);
7497                 BTRFS_I(inode)->outstanding_extents += num_extents;
7498                 spin_unlock(&BTRFS_I(inode)->lock);
7499         }
7500 }
7501
7502 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7503                                    struct buffer_head *bh_result, int create)
7504 {
7505         struct extent_map *em;
7506         struct btrfs_root *root = BTRFS_I(inode)->root;
7507         struct extent_state *cached_state = NULL;
7508         struct btrfs_dio_data *dio_data = NULL;
7509         u64 start = iblock << inode->i_blkbits;
7510         u64 lockstart, lockend;
7511         u64 len = bh_result->b_size;
7512         int unlock_bits = EXTENT_LOCKED;
7513         int ret = 0;
7514
7515         if (create)
7516                 unlock_bits |= EXTENT_DIRTY;
7517         else
7518                 len = min_t(u64, len, root->sectorsize);
7519
7520         lockstart = start;
7521         lockend = start + len - 1;
7522
7523         if (current->journal_info) {
7524                 /*
7525                  * Need to pull our outstanding extents and set journal_info to NULL so
7526                  * that anything that needs to check if there's a transction doesn't get
7527                  * confused.
7528                  */
7529                 dio_data = current->journal_info;
7530                 current->journal_info = NULL;
7531         }
7532
7533         /*
7534          * If this errors out it's because we couldn't invalidate pagecache for
7535          * this range and we need to fallback to buffered.
7536          */
7537         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7538                                create)) {
7539                 ret = -ENOTBLK;
7540                 goto err;
7541         }
7542
7543         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7544         if (IS_ERR(em)) {
7545                 ret = PTR_ERR(em);
7546                 goto unlock_err;
7547         }
7548
7549         /*
7550          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7551          * io.  INLINE is special, and we could probably kludge it in here, but
7552          * it's still buffered so for safety lets just fall back to the generic
7553          * buffered path.
7554          *
7555          * For COMPRESSED we _have_ to read the entire extent in so we can
7556          * decompress it, so there will be buffering required no matter what we
7557          * do, so go ahead and fallback to buffered.
7558          *
7559          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7560          * to buffered IO.  Don't blame me, this is the price we pay for using
7561          * the generic code.
7562          */
7563         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7564             em->block_start == EXTENT_MAP_INLINE) {
7565                 free_extent_map(em);
7566                 ret = -ENOTBLK;
7567                 goto unlock_err;
7568         }
7569
7570         /* Just a good old fashioned hole, return */
7571         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7572                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7573                 free_extent_map(em);
7574                 goto unlock_err;
7575         }
7576
7577         /*
7578          * We don't allocate a new extent in the following cases
7579          *
7580          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7581          * existing extent.
7582          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7583          * just use the extent.
7584          *
7585          */
7586         if (!create) {
7587                 len = min(len, em->len - (start - em->start));
7588                 lockstart = start + len;
7589                 goto unlock;
7590         }
7591
7592         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7593             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7594              em->block_start != EXTENT_MAP_HOLE)) {
7595                 int type;
7596                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7597
7598                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7599                         type = BTRFS_ORDERED_PREALLOC;
7600                 else
7601                         type = BTRFS_ORDERED_NOCOW;
7602                 len = min(len, em->len - (start - em->start));
7603                 block_start = em->block_start + (start - em->start);
7604
7605                 if (can_nocow_extent(inode, start, &len, &orig_start,
7606                                      &orig_block_len, &ram_bytes) == 1) {
7607                         if (type == BTRFS_ORDERED_PREALLOC) {
7608                                 free_extent_map(em);
7609                                 em = create_pinned_em(inode, start, len,
7610                                                        orig_start,
7611                                                        block_start, len,
7612                                                        orig_block_len,
7613                                                        ram_bytes, type);
7614                                 if (IS_ERR(em)) {
7615                                         ret = PTR_ERR(em);
7616                                         goto unlock_err;
7617                                 }
7618                         }
7619
7620                         ret = btrfs_add_ordered_extent_dio(inode, start,
7621                                            block_start, len, len, type);
7622                         if (ret) {
7623                                 free_extent_map(em);
7624                                 goto unlock_err;
7625                         }
7626                         goto unlock;
7627                 }
7628         }
7629
7630         /*
7631          * this will cow the extent, reset the len in case we changed
7632          * it above
7633          */
7634         len = bh_result->b_size;
7635         free_extent_map(em);
7636         em = btrfs_new_extent_direct(inode, start, len);
7637         if (IS_ERR(em)) {
7638                 ret = PTR_ERR(em);
7639                 goto unlock_err;
7640         }
7641         len = min(len, em->len - (start - em->start));
7642 unlock:
7643         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7644                 inode->i_blkbits;
7645         bh_result->b_size = len;
7646         bh_result->b_bdev = em->bdev;
7647         set_buffer_mapped(bh_result);
7648         if (create) {
7649                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7650                         set_buffer_new(bh_result);
7651
7652                 /*
7653                  * Need to update the i_size under the extent lock so buffered
7654                  * readers will get the updated i_size when we unlock.
7655                  */
7656                 if (start + len > i_size_read(inode))
7657                         i_size_write(inode, start + len);
7658
7659                 adjust_dio_outstanding_extents(inode, dio_data, len);
7660                 btrfs_free_reserved_data_space(inode, start, len);
7661                 WARN_ON(dio_data->reserve < len);
7662                 dio_data->reserve -= len;
7663                 dio_data->unsubmitted_oe_range_end = start + len;
7664                 current->journal_info = dio_data;
7665         }
7666
7667         /*
7668          * In the case of write we need to clear and unlock the entire range,
7669          * in the case of read we need to unlock only the end area that we
7670          * aren't using if there is any left over space.
7671          */
7672         if (lockstart < lockend) {
7673                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7674                                  lockend, unlock_bits, 1, 0,
7675                                  &cached_state, GFP_NOFS);
7676         } else {
7677                 free_extent_state(cached_state);
7678         }
7679
7680         free_extent_map(em);
7681
7682         return 0;
7683
7684 unlock_err:
7685         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7686                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7687 err:
7688         if (dio_data)
7689                 current->journal_info = dio_data;
7690         /*
7691          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7692          * write less data then expected, so that we don't underflow our inode's
7693          * outstanding extents counter.
7694          */
7695         if (create && dio_data)
7696                 adjust_dio_outstanding_extents(inode, dio_data, len);
7697
7698         return ret;
7699 }
7700
7701 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7702                                         int rw, int mirror_num)
7703 {
7704         struct btrfs_root *root = BTRFS_I(inode)->root;
7705         int ret;
7706
7707         BUG_ON(rw & REQ_WRITE);
7708
7709         bio_get(bio);
7710
7711         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7712                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7713         if (ret)
7714                 goto err;
7715
7716         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7717 err:
7718         bio_put(bio);
7719         return ret;
7720 }
7721
7722 static int btrfs_check_dio_repairable(struct inode *inode,
7723                                       struct bio *failed_bio,
7724                                       struct io_failure_record *failrec,
7725                                       int failed_mirror)
7726 {
7727         int num_copies;
7728
7729         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7730                                       failrec->logical, failrec->len);
7731         if (num_copies == 1) {
7732                 /*
7733                  * we only have a single copy of the data, so don't bother with
7734                  * all the retry and error correction code that follows. no
7735                  * matter what the error is, it is very likely to persist.
7736                  */
7737                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7738                          num_copies, failrec->this_mirror, failed_mirror);
7739                 return 0;
7740         }
7741
7742         failrec->failed_mirror = failed_mirror;
7743         failrec->this_mirror++;
7744         if (failrec->this_mirror == failed_mirror)
7745                 failrec->this_mirror++;
7746
7747         if (failrec->this_mirror > num_copies) {
7748                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7749                          num_copies, failrec->this_mirror, failed_mirror);
7750                 return 0;
7751         }
7752
7753         return 1;
7754 }
7755
7756 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7757                           struct page *page, u64 start, u64 end,
7758                           int failed_mirror, bio_end_io_t *repair_endio,
7759                           void *repair_arg)
7760 {
7761         struct io_failure_record *failrec;
7762         struct bio *bio;
7763         int isector;
7764         int read_mode;
7765         int ret;
7766
7767         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7768
7769         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7770         if (ret)
7771                 return ret;
7772
7773         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7774                                          failed_mirror);
7775         if (!ret) {
7776                 free_io_failure(inode, failrec);
7777                 return -EIO;
7778         }
7779
7780         if (failed_bio->bi_vcnt > 1)
7781                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7782         else
7783                 read_mode = READ_SYNC;
7784
7785         isector = start - btrfs_io_bio(failed_bio)->logical;
7786         isector >>= inode->i_sb->s_blocksize_bits;
7787         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7788                                       0, isector, repair_endio, repair_arg);
7789         if (!bio) {
7790                 free_io_failure(inode, failrec);
7791                 return -EIO;
7792         }
7793
7794         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7795                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7796                     read_mode, failrec->this_mirror, failrec->in_validation);
7797
7798         ret = submit_dio_repair_bio(inode, bio, read_mode,
7799                                     failrec->this_mirror);
7800         if (ret) {
7801                 free_io_failure(inode, failrec);
7802                 bio_put(bio);
7803         }
7804
7805         return ret;
7806 }
7807
7808 struct btrfs_retry_complete {
7809         struct completion done;
7810         struct inode *inode;
7811         u64 start;
7812         int uptodate;
7813 };
7814
7815 static void btrfs_retry_endio_nocsum(struct bio *bio)
7816 {
7817         struct btrfs_retry_complete *done = bio->bi_private;
7818         struct bio_vec *bvec;
7819         int i;
7820
7821         if (bio->bi_error)
7822                 goto end;
7823
7824         done->uptodate = 1;
7825         bio_for_each_segment_all(bvec, bio, i)
7826                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7827 end:
7828         complete(&done->done);
7829         bio_put(bio);
7830 }
7831
7832 static int __btrfs_correct_data_nocsum(struct inode *inode,
7833                                        struct btrfs_io_bio *io_bio)
7834 {
7835         struct bio_vec *bvec;
7836         struct btrfs_retry_complete done;
7837         u64 start;
7838         int i;
7839         int ret;
7840
7841         start = io_bio->logical;
7842         done.inode = inode;
7843
7844         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7845 try_again:
7846                 done.uptodate = 0;
7847                 done.start = start;
7848                 init_completion(&done.done);
7849
7850                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7851                                      start + bvec->bv_len - 1,
7852                                      io_bio->mirror_num,
7853                                      btrfs_retry_endio_nocsum, &done);
7854                 if (ret)
7855                         return ret;
7856
7857                 wait_for_completion(&done.done);
7858
7859                 if (!done.uptodate) {
7860                         /* We might have another mirror, so try again */
7861                         goto try_again;
7862                 }
7863
7864                 start += bvec->bv_len;
7865         }
7866
7867         return 0;
7868 }
7869
7870 static void btrfs_retry_endio(struct bio *bio)
7871 {
7872         struct btrfs_retry_complete *done = bio->bi_private;
7873         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7874         struct bio_vec *bvec;
7875         int uptodate;
7876         int ret;
7877         int i;
7878
7879         if (bio->bi_error)
7880                 goto end;
7881
7882         uptodate = 1;
7883         bio_for_each_segment_all(bvec, bio, i) {
7884                 ret = __readpage_endio_check(done->inode, io_bio, i,
7885                                              bvec->bv_page, 0,
7886                                              done->start, bvec->bv_len);
7887                 if (!ret)
7888                         clean_io_failure(done->inode, done->start,
7889                                          bvec->bv_page, 0);
7890                 else
7891                         uptodate = 0;
7892         }
7893
7894         done->uptodate = uptodate;
7895 end:
7896         complete(&done->done);
7897         bio_put(bio);
7898 }
7899
7900 static int __btrfs_subio_endio_read(struct inode *inode,
7901                                     struct btrfs_io_bio *io_bio, int err)
7902 {
7903         struct bio_vec *bvec;
7904         struct btrfs_retry_complete done;
7905         u64 start;
7906         u64 offset = 0;
7907         int i;
7908         int ret;
7909
7910         err = 0;
7911         start = io_bio->logical;
7912         done.inode = inode;
7913
7914         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7915                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7916                                              0, start, bvec->bv_len);
7917                 if (likely(!ret))
7918                         goto next;
7919 try_again:
7920                 done.uptodate = 0;
7921                 done.start = start;
7922                 init_completion(&done.done);
7923
7924                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7925                                      start + bvec->bv_len - 1,
7926                                      io_bio->mirror_num,
7927                                      btrfs_retry_endio, &done);
7928                 if (ret) {
7929                         err = ret;
7930                         goto next;
7931                 }
7932
7933                 wait_for_completion(&done.done);
7934
7935                 if (!done.uptodate) {
7936                         /* We might have another mirror, so try again */
7937                         goto try_again;
7938                 }
7939 next:
7940                 offset += bvec->bv_len;
7941                 start += bvec->bv_len;
7942         }
7943
7944         return err;
7945 }
7946
7947 static int btrfs_subio_endio_read(struct inode *inode,
7948                                   struct btrfs_io_bio *io_bio, int err)
7949 {
7950         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7951
7952         if (skip_csum) {
7953                 if (unlikely(err))
7954                         return __btrfs_correct_data_nocsum(inode, io_bio);
7955                 else
7956                         return 0;
7957         } else {
7958                 return __btrfs_subio_endio_read(inode, io_bio, err);
7959         }
7960 }
7961
7962 static void btrfs_endio_direct_read(struct bio *bio)
7963 {
7964         struct btrfs_dio_private *dip = bio->bi_private;
7965         struct inode *inode = dip->inode;
7966         struct bio *dio_bio;
7967         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7968         int err = bio->bi_error;
7969
7970         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7971                 err = btrfs_subio_endio_read(inode, io_bio, err);
7972
7973         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7974                       dip->logical_offset + dip->bytes - 1);
7975         dio_bio = dip->dio_bio;
7976
7977         kfree(dip);
7978
7979         dio_end_io(dio_bio, bio->bi_error);
7980
7981         if (io_bio->end_io)
7982                 io_bio->end_io(io_bio, err);
7983         bio_put(bio);
7984 }
7985
7986 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
7987                                                     const u64 offset,
7988                                                     const u64 bytes,
7989                                                     const int uptodate)
7990 {
7991         struct btrfs_root *root = BTRFS_I(inode)->root;
7992         struct btrfs_ordered_extent *ordered = NULL;
7993         u64 ordered_offset = offset;
7994         u64 ordered_bytes = bytes;
7995         int ret;
7996
7997 again:
7998         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
7999                                                    &ordered_offset,
8000                                                    ordered_bytes,
8001                                                    uptodate);
8002         if (!ret)
8003                 goto out_test;
8004
8005         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8006                         finish_ordered_fn, NULL, NULL);
8007         btrfs_queue_work(root->fs_info->endio_write_workers,
8008                          &ordered->work);
8009 out_test:
8010         /*
8011          * our bio might span multiple ordered extents.  If we haven't
8012          * completed the accounting for the whole dio, go back and try again
8013          */
8014         if (ordered_offset < offset + bytes) {
8015                 ordered_bytes = offset + bytes - ordered_offset;
8016                 ordered = NULL;
8017                 goto again;
8018         }
8019 }
8020
8021 static void btrfs_endio_direct_write(struct bio *bio)
8022 {
8023         struct btrfs_dio_private *dip = bio->bi_private;
8024         struct bio *dio_bio = dip->dio_bio;
8025
8026         btrfs_endio_direct_write_update_ordered(dip->inode,
8027                                                 dip->logical_offset,
8028                                                 dip->bytes,
8029                                                 !bio->bi_error);
8030
8031         kfree(dip);
8032
8033         dio_end_io(dio_bio, bio->bi_error);
8034         bio_put(bio);
8035 }
8036
8037 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8038                                     struct bio *bio, int mirror_num,
8039                                     unsigned long bio_flags, u64 offset)
8040 {
8041         int ret;
8042         struct btrfs_root *root = BTRFS_I(inode)->root;
8043         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8044         BUG_ON(ret); /* -ENOMEM */
8045         return 0;
8046 }
8047
8048 static void btrfs_end_dio_bio(struct bio *bio)
8049 {
8050         struct btrfs_dio_private *dip = bio->bi_private;
8051         int err = bio->bi_error;
8052
8053         if (err)
8054                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8055                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8056                            btrfs_ino(dip->inode), bio->bi_rw,
8057                            (unsigned long long)bio->bi_iter.bi_sector,
8058                            bio->bi_iter.bi_size, err);
8059
8060         if (dip->subio_endio)
8061                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8062
8063         if (err) {
8064                 dip->errors = 1;
8065
8066                 /*
8067                  * before atomic variable goto zero, we must make sure
8068                  * dip->errors is perceived to be set.
8069                  */
8070                 smp_mb__before_atomic();
8071         }
8072
8073         /* if there are more bios still pending for this dio, just exit */
8074         if (!atomic_dec_and_test(&dip->pending_bios))
8075                 goto out;
8076
8077         if (dip->errors) {
8078                 bio_io_error(dip->orig_bio);
8079         } else {
8080                 dip->dio_bio->bi_error = 0;
8081                 bio_endio(dip->orig_bio);
8082         }
8083 out:
8084         bio_put(bio);
8085 }
8086
8087 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8088                                        u64 first_sector, gfp_t gfp_flags)
8089 {
8090         struct bio *bio;
8091         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8092         if (bio)
8093                 bio_associate_current(bio);
8094         return bio;
8095 }
8096
8097 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8098                                                  struct inode *inode,
8099                                                  struct btrfs_dio_private *dip,
8100                                                  struct bio *bio,
8101                                                  u64 file_offset)
8102 {
8103         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8104         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8105         int ret;
8106
8107         /*
8108          * We load all the csum data we need when we submit
8109          * the first bio to reduce the csum tree search and
8110          * contention.
8111          */
8112         if (dip->logical_offset == file_offset) {
8113                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8114                                                 file_offset);
8115                 if (ret)
8116                         return ret;
8117         }
8118
8119         if (bio == dip->orig_bio)
8120                 return 0;
8121
8122         file_offset -= dip->logical_offset;
8123         file_offset >>= inode->i_sb->s_blocksize_bits;
8124         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8125
8126         return 0;
8127 }
8128
8129 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8130                                          int rw, u64 file_offset, int skip_sum,
8131                                          int async_submit)
8132 {
8133         struct btrfs_dio_private *dip = bio->bi_private;
8134         int write = rw & REQ_WRITE;
8135         struct btrfs_root *root = BTRFS_I(inode)->root;
8136         int ret;
8137
8138         if (async_submit)
8139                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8140
8141         bio_get(bio);
8142
8143         if (!write) {
8144                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8145                                 BTRFS_WQ_ENDIO_DATA);
8146                 if (ret)
8147                         goto err;
8148         }
8149
8150         if (skip_sum)
8151                 goto map;
8152
8153         if (write && async_submit) {
8154                 ret = btrfs_wq_submit_bio(root->fs_info,
8155                                    inode, rw, bio, 0, 0,
8156                                    file_offset,
8157                                    __btrfs_submit_bio_start_direct_io,
8158                                    __btrfs_submit_bio_done);
8159                 goto err;
8160         } else if (write) {
8161                 /*
8162                  * If we aren't doing async submit, calculate the csum of the
8163                  * bio now.
8164                  */
8165                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8166                 if (ret)
8167                         goto err;
8168         } else {
8169                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8170                                                      file_offset);
8171                 if (ret)
8172                         goto err;
8173         }
8174 map:
8175         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8176 err:
8177         bio_put(bio);
8178         return ret;
8179 }
8180
8181 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8182                                     int skip_sum)
8183 {
8184         struct inode *inode = dip->inode;
8185         struct btrfs_root *root = BTRFS_I(inode)->root;
8186         struct bio *bio;
8187         struct bio *orig_bio = dip->orig_bio;
8188         struct bio_vec *bvec = orig_bio->bi_io_vec;
8189         u64 start_sector = orig_bio->bi_iter.bi_sector;
8190         u64 file_offset = dip->logical_offset;
8191         u64 submit_len = 0;
8192         u64 map_length;
8193         int nr_pages = 0;
8194         int ret;
8195         int async_submit = 0;
8196
8197         map_length = orig_bio->bi_iter.bi_size;
8198         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8199                               &map_length, NULL, 0);
8200         if (ret)
8201                 return -EIO;
8202
8203         if (map_length >= orig_bio->bi_iter.bi_size) {
8204                 bio = orig_bio;
8205                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8206                 goto submit;
8207         }
8208
8209         /* async crcs make it difficult to collect full stripe writes. */
8210         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8211                 async_submit = 0;
8212         else
8213                 async_submit = 1;
8214
8215         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8216         if (!bio)
8217                 return -ENOMEM;
8218
8219         bio->bi_private = dip;
8220         bio->bi_end_io = btrfs_end_dio_bio;
8221         btrfs_io_bio(bio)->logical = file_offset;
8222         atomic_inc(&dip->pending_bios);
8223
8224         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8225                 if (map_length < submit_len + bvec->bv_len ||
8226                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8227                                  bvec->bv_offset) < bvec->bv_len) {
8228                         /*
8229                          * inc the count before we submit the bio so
8230                          * we know the end IO handler won't happen before
8231                          * we inc the count. Otherwise, the dip might get freed
8232                          * before we're done setting it up
8233                          */
8234                         atomic_inc(&dip->pending_bios);
8235                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8236                                                      file_offset, skip_sum,
8237                                                      async_submit);
8238                         if (ret) {
8239                                 bio_put(bio);
8240                                 atomic_dec(&dip->pending_bios);
8241                                 goto out_err;
8242                         }
8243
8244                         start_sector += submit_len >> 9;
8245                         file_offset += submit_len;
8246
8247                         submit_len = 0;
8248                         nr_pages = 0;
8249
8250                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8251                                                   start_sector, GFP_NOFS);
8252                         if (!bio)
8253                                 goto out_err;
8254                         bio->bi_private = dip;
8255                         bio->bi_end_io = btrfs_end_dio_bio;
8256                         btrfs_io_bio(bio)->logical = file_offset;
8257
8258                         map_length = orig_bio->bi_iter.bi_size;
8259                         ret = btrfs_map_block(root->fs_info, rw,
8260                                               start_sector << 9,
8261                                               &map_length, NULL, 0);
8262                         if (ret) {
8263                                 bio_put(bio);
8264                                 goto out_err;
8265                         }
8266                 } else {
8267                         submit_len += bvec->bv_len;
8268                         nr_pages++;
8269                         bvec++;
8270                 }
8271         }
8272
8273 submit:
8274         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8275                                      async_submit);
8276         if (!ret)
8277                 return 0;
8278
8279         bio_put(bio);
8280 out_err:
8281         dip->errors = 1;
8282         /*
8283          * before atomic variable goto zero, we must
8284          * make sure dip->errors is perceived to be set.
8285          */
8286         smp_mb__before_atomic();
8287         if (atomic_dec_and_test(&dip->pending_bios))
8288                 bio_io_error(dip->orig_bio);
8289
8290         /* bio_end_io() will handle error, so we needn't return it */
8291         return 0;
8292 }
8293
8294 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8295                                 struct inode *inode, loff_t file_offset)
8296 {
8297         struct btrfs_dio_private *dip = NULL;
8298         struct bio *io_bio = NULL;
8299         struct btrfs_io_bio *btrfs_bio;
8300         int skip_sum;
8301         int write = rw & REQ_WRITE;
8302         int ret = 0;
8303
8304         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8305
8306         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8307         if (!io_bio) {
8308                 ret = -ENOMEM;
8309                 goto free_ordered;
8310         }
8311
8312         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8313         if (!dip) {
8314                 ret = -ENOMEM;
8315                 goto free_ordered;
8316         }
8317
8318         dip->private = dio_bio->bi_private;
8319         dip->inode = inode;
8320         dip->logical_offset = file_offset;
8321         dip->bytes = dio_bio->bi_iter.bi_size;
8322         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8323         io_bio->bi_private = dip;
8324         dip->orig_bio = io_bio;
8325         dip->dio_bio = dio_bio;
8326         atomic_set(&dip->pending_bios, 0);
8327         btrfs_bio = btrfs_io_bio(io_bio);
8328         btrfs_bio->logical = file_offset;
8329
8330         if (write) {
8331                 io_bio->bi_end_io = btrfs_endio_direct_write;
8332         } else {
8333                 io_bio->bi_end_io = btrfs_endio_direct_read;
8334                 dip->subio_endio = btrfs_subio_endio_read;
8335         }
8336
8337         /*
8338          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8339          * even if we fail to submit a bio, because in such case we do the
8340          * corresponding error handling below and it must not be done a second
8341          * time by btrfs_direct_IO().
8342          */
8343         if (write) {
8344                 struct btrfs_dio_data *dio_data = current->journal_info;
8345
8346                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8347                         dip->bytes;
8348                 dio_data->unsubmitted_oe_range_start =
8349                         dio_data->unsubmitted_oe_range_end;
8350         }
8351
8352         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8353         if (!ret)
8354                 return;
8355
8356         if (btrfs_bio->end_io)
8357                 btrfs_bio->end_io(btrfs_bio, ret);
8358
8359 free_ordered:
8360         /*
8361          * If we arrived here it means either we failed to submit the dip
8362          * or we either failed to clone the dio_bio or failed to allocate the
8363          * dip. If we cloned the dio_bio and allocated the dip, we can just
8364          * call bio_endio against our io_bio so that we get proper resource
8365          * cleanup if we fail to submit the dip, otherwise, we must do the
8366          * same as btrfs_endio_direct_[write|read] because we can't call these
8367          * callbacks - they require an allocated dip and a clone of dio_bio.
8368          */
8369         if (io_bio && dip) {
8370                 io_bio->bi_error = -EIO;
8371                 bio_endio(io_bio);
8372                 /*
8373                  * The end io callbacks free our dip, do the final put on io_bio
8374                  * and all the cleanup and final put for dio_bio (through
8375                  * dio_end_io()).
8376                  */
8377                 dip = NULL;
8378                 io_bio = NULL;
8379         } else {
8380                 if (write)
8381                         btrfs_endio_direct_write_update_ordered(inode,
8382                                                 file_offset,
8383                                                 dio_bio->bi_iter.bi_size,
8384                                                 0);
8385                 else
8386                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8387                               file_offset + dio_bio->bi_iter.bi_size - 1);
8388
8389                 dio_bio->bi_error = -EIO;
8390                 /*
8391                  * Releases and cleans up our dio_bio, no need to bio_put()
8392                  * nor bio_endio()/bio_io_error() against dio_bio.
8393                  */
8394                 dio_end_io(dio_bio, ret);
8395         }
8396         if (io_bio)
8397                 bio_put(io_bio);
8398         kfree(dip);
8399 }
8400
8401 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8402                         const struct iov_iter *iter, loff_t offset)
8403 {
8404         int seg;
8405         int i;
8406         unsigned blocksize_mask = root->sectorsize - 1;
8407         ssize_t retval = -EINVAL;
8408
8409         if (offset & blocksize_mask)
8410                 goto out;
8411
8412         if (iov_iter_alignment(iter) & blocksize_mask)
8413                 goto out;
8414
8415         /* If this is a write we don't need to check anymore */
8416         if (iov_iter_rw(iter) == WRITE)
8417                 return 0;
8418         /*
8419          * Check to make sure we don't have duplicate iov_base's in this
8420          * iovec, if so return EINVAL, otherwise we'll get csum errors
8421          * when reading back.
8422          */
8423         for (seg = 0; seg < iter->nr_segs; seg++) {
8424                 for (i = seg + 1; i < iter->nr_segs; i++) {
8425                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8426                                 goto out;
8427                 }
8428         }
8429         retval = 0;
8430 out:
8431         return retval;
8432 }
8433
8434 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8435                                loff_t offset)
8436 {
8437         struct file *file = iocb->ki_filp;
8438         struct inode *inode = file->f_mapping->host;
8439         struct btrfs_root *root = BTRFS_I(inode)->root;
8440         struct btrfs_dio_data dio_data = { 0 };
8441         size_t count = 0;
8442         int flags = 0;
8443         bool wakeup = true;
8444         bool relock = false;
8445         ssize_t ret;
8446
8447         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8448                 return 0;
8449
8450         inode_dio_begin(inode);
8451         smp_mb__after_atomic();
8452
8453         /*
8454          * The generic stuff only does filemap_write_and_wait_range, which
8455          * isn't enough if we've written compressed pages to this area, so
8456          * we need to flush the dirty pages again to make absolutely sure
8457          * that any outstanding dirty pages are on disk.
8458          */
8459         count = iov_iter_count(iter);
8460         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8461                      &BTRFS_I(inode)->runtime_flags))
8462                 filemap_fdatawrite_range(inode->i_mapping, offset,
8463                                          offset + count - 1);
8464
8465         if (iov_iter_rw(iter) == WRITE) {
8466                 /*
8467                  * If the write DIO is beyond the EOF, we need update
8468                  * the isize, but it is protected by i_mutex. So we can
8469                  * not unlock the i_mutex at this case.
8470                  */
8471                 if (offset + count <= inode->i_size) {
8472                         mutex_unlock(&inode->i_mutex);
8473                         relock = true;
8474                 }
8475                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8476                 if (ret)
8477                         goto out;
8478                 dio_data.outstanding_extents = div64_u64(count +
8479                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8480                                                 BTRFS_MAX_EXTENT_SIZE);
8481
8482                 /*
8483                  * We need to know how many extents we reserved so that we can
8484                  * do the accounting properly if we go over the number we
8485                  * originally calculated.  Abuse current->journal_info for this.
8486                  */
8487                 dio_data.reserve = round_up(count, root->sectorsize);
8488                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8489                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8490                 current->journal_info = &dio_data;
8491         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8492                                      &BTRFS_I(inode)->runtime_flags)) {
8493                 inode_dio_end(inode);
8494                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8495                 wakeup = false;
8496         }
8497
8498         ret = __blockdev_direct_IO(iocb, inode,
8499                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8500                                    iter, offset, btrfs_get_blocks_direct, NULL,
8501                                    btrfs_submit_direct, flags);
8502         if (iov_iter_rw(iter) == WRITE) {
8503                 current->journal_info = NULL;
8504                 if (ret < 0 && ret != -EIOCBQUEUED) {
8505                         if (dio_data.reserve)
8506                                 btrfs_delalloc_release_space(inode, offset,
8507                                                              dio_data.reserve);
8508                         /*
8509                          * On error we might have left some ordered extents
8510                          * without submitting corresponding bios for them, so
8511                          * cleanup them up to avoid other tasks getting them
8512                          * and waiting for them to complete forever.
8513                          */
8514                         if (dio_data.unsubmitted_oe_range_start <
8515                             dio_data.unsubmitted_oe_range_end)
8516                                 btrfs_endio_direct_write_update_ordered(inode,
8517                                         dio_data.unsubmitted_oe_range_start,
8518                                         dio_data.unsubmitted_oe_range_end -
8519                                         dio_data.unsubmitted_oe_range_start,
8520                                         0);
8521                 } else if (ret >= 0 && (size_t)ret < count)
8522                         btrfs_delalloc_release_space(inode, offset,
8523                                                      count - (size_t)ret);
8524         }
8525 out:
8526         if (wakeup)
8527                 inode_dio_end(inode);
8528         if (relock)
8529                 mutex_lock(&inode->i_mutex);
8530
8531         return ret;
8532 }
8533
8534 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8535
8536 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8537                 __u64 start, __u64 len)
8538 {
8539         int     ret;
8540
8541         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8542         if (ret)
8543                 return ret;
8544
8545         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8546 }
8547
8548 int btrfs_readpage(struct file *file, struct page *page)
8549 {
8550         struct extent_io_tree *tree;
8551         tree = &BTRFS_I(page->mapping->host)->io_tree;
8552         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8553 }
8554
8555 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8556 {
8557         struct extent_io_tree *tree;
8558         struct inode *inode = page->mapping->host;
8559         int ret;
8560
8561         if (current->flags & PF_MEMALLOC) {
8562                 redirty_page_for_writepage(wbc, page);
8563                 unlock_page(page);
8564                 return 0;
8565         }
8566
8567         /*
8568          * If we are under memory pressure we will call this directly from the
8569          * VM, we need to make sure we have the inode referenced for the ordered
8570          * extent.  If not just return like we didn't do anything.
8571          */
8572         if (!igrab(inode)) {
8573                 redirty_page_for_writepage(wbc, page);
8574                 return AOP_WRITEPAGE_ACTIVATE;
8575         }
8576         tree = &BTRFS_I(page->mapping->host)->io_tree;
8577         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8578         btrfs_add_delayed_iput(inode);
8579         return ret;
8580 }
8581
8582 static int btrfs_writepages(struct address_space *mapping,
8583                             struct writeback_control *wbc)
8584 {
8585         struct extent_io_tree *tree;
8586
8587         tree = &BTRFS_I(mapping->host)->io_tree;
8588         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8589 }
8590
8591 static int
8592 btrfs_readpages(struct file *file, struct address_space *mapping,
8593                 struct list_head *pages, unsigned nr_pages)
8594 {
8595         struct extent_io_tree *tree;
8596         tree = &BTRFS_I(mapping->host)->io_tree;
8597         return extent_readpages(tree, mapping, pages, nr_pages,
8598                                 btrfs_get_extent);
8599 }
8600 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8601 {
8602         struct extent_io_tree *tree;
8603         struct extent_map_tree *map;
8604         int ret;
8605
8606         tree = &BTRFS_I(page->mapping->host)->io_tree;
8607         map = &BTRFS_I(page->mapping->host)->extent_tree;
8608         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8609         if (ret == 1) {
8610                 ClearPagePrivate(page);
8611                 set_page_private(page, 0);
8612                 page_cache_release(page);
8613         }
8614         return ret;
8615 }
8616
8617 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8618 {
8619         if (PageWriteback(page) || PageDirty(page))
8620                 return 0;
8621         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8622 }
8623
8624 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8625                                  unsigned int length)
8626 {
8627         struct inode *inode = page->mapping->host;
8628         struct extent_io_tree *tree;
8629         struct btrfs_ordered_extent *ordered;
8630         struct extent_state *cached_state = NULL;
8631         u64 page_start = page_offset(page);
8632         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8633         int inode_evicting = inode->i_state & I_FREEING;
8634
8635         /*
8636          * we have the page locked, so new writeback can't start,
8637          * and the dirty bit won't be cleared while we are here.
8638          *
8639          * Wait for IO on this page so that we can safely clear
8640          * the PagePrivate2 bit and do ordered accounting
8641          */
8642         wait_on_page_writeback(page);
8643
8644         tree = &BTRFS_I(inode)->io_tree;
8645         if (offset) {
8646                 btrfs_releasepage(page, GFP_NOFS);
8647                 return;
8648         }
8649
8650         if (!inode_evicting)
8651                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8652         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8653         if (ordered) {
8654                 /*
8655                  * IO on this page will never be started, so we need
8656                  * to account for any ordered extents now
8657                  */
8658                 if (!inode_evicting)
8659                         clear_extent_bit(tree, page_start, page_end,
8660                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8661                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8662                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8663                                          GFP_NOFS);
8664                 /*
8665                  * whoever cleared the private bit is responsible
8666                  * for the finish_ordered_io
8667                  */
8668                 if (TestClearPagePrivate2(page)) {
8669                         struct btrfs_ordered_inode_tree *tree;
8670                         u64 new_len;
8671
8672                         tree = &BTRFS_I(inode)->ordered_tree;
8673
8674                         spin_lock_irq(&tree->lock);
8675                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8676                         new_len = page_start - ordered->file_offset;
8677                         if (new_len < ordered->truncated_len)
8678                                 ordered->truncated_len = new_len;
8679                         spin_unlock_irq(&tree->lock);
8680
8681                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8682                                                            page_start,
8683                                                            PAGE_CACHE_SIZE, 1))
8684                                 btrfs_finish_ordered_io(ordered);
8685                 }
8686                 btrfs_put_ordered_extent(ordered);
8687                 if (!inode_evicting) {
8688                         cached_state = NULL;
8689                         lock_extent_bits(tree, page_start, page_end,
8690                                          &cached_state);
8691                 }
8692         }
8693
8694         /*
8695          * Qgroup reserved space handler
8696          * Page here will be either
8697          * 1) Already written to disk
8698          *    In this case, its reserved space is released from data rsv map
8699          *    and will be freed by delayed_ref handler finally.
8700          *    So even we call qgroup_free_data(), it won't decrease reserved
8701          *    space.
8702          * 2) Not written to disk
8703          *    This means the reserved space should be freed here.
8704          */
8705         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8706         if (!inode_evicting) {
8707                 clear_extent_bit(tree, page_start, page_end,
8708                                  EXTENT_LOCKED | EXTENT_DIRTY |
8709                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8710                                  EXTENT_DEFRAG, 1, 1,
8711                                  &cached_state, GFP_NOFS);
8712
8713                 __btrfs_releasepage(page, GFP_NOFS);
8714         }
8715
8716         ClearPageChecked(page);
8717         if (PagePrivate(page)) {
8718                 ClearPagePrivate(page);
8719                 set_page_private(page, 0);
8720                 page_cache_release(page);
8721         }
8722 }
8723
8724 /*
8725  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8726  * called from a page fault handler when a page is first dirtied. Hence we must
8727  * be careful to check for EOF conditions here. We set the page up correctly
8728  * for a written page which means we get ENOSPC checking when writing into
8729  * holes and correct delalloc and unwritten extent mapping on filesystems that
8730  * support these features.
8731  *
8732  * We are not allowed to take the i_mutex here so we have to play games to
8733  * protect against truncate races as the page could now be beyond EOF.  Because
8734  * vmtruncate() writes the inode size before removing pages, once we have the
8735  * page lock we can determine safely if the page is beyond EOF. If it is not
8736  * beyond EOF, then the page is guaranteed safe against truncation until we
8737  * unlock the page.
8738  */
8739 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8740 {
8741         struct page *page = vmf->page;
8742         struct inode *inode = file_inode(vma->vm_file);
8743         struct btrfs_root *root = BTRFS_I(inode)->root;
8744         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8745         struct btrfs_ordered_extent *ordered;
8746         struct extent_state *cached_state = NULL;
8747         char *kaddr;
8748         unsigned long zero_start;
8749         loff_t size;
8750         int ret;
8751         int reserved = 0;
8752         u64 page_start;
8753         u64 page_end;
8754
8755         sb_start_pagefault(inode->i_sb);
8756         page_start = page_offset(page);
8757         page_end = page_start + PAGE_CACHE_SIZE - 1;
8758
8759         ret = btrfs_delalloc_reserve_space(inode, page_start,
8760                                            PAGE_CACHE_SIZE);
8761         if (!ret) {
8762                 ret = file_update_time(vma->vm_file);
8763                 reserved = 1;
8764         }
8765         if (ret) {
8766                 if (ret == -ENOMEM)
8767                         ret = VM_FAULT_OOM;
8768                 else /* -ENOSPC, -EIO, etc */
8769                         ret = VM_FAULT_SIGBUS;
8770                 if (reserved)
8771                         goto out;
8772                 goto out_noreserve;
8773         }
8774
8775         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8776 again:
8777         lock_page(page);
8778         size = i_size_read(inode);
8779
8780         if ((page->mapping != inode->i_mapping) ||
8781             (page_start >= size)) {
8782                 /* page got truncated out from underneath us */
8783                 goto out_unlock;
8784         }
8785         wait_on_page_writeback(page);
8786
8787         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8788         set_page_extent_mapped(page);
8789
8790         /*
8791          * we can't set the delalloc bits if there are pending ordered
8792          * extents.  Drop our locks and wait for them to finish
8793          */
8794         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8795         if (ordered) {
8796                 unlock_extent_cached(io_tree, page_start, page_end,
8797                                      &cached_state, GFP_NOFS);
8798                 unlock_page(page);
8799                 btrfs_start_ordered_extent(inode, ordered, 1);
8800                 btrfs_put_ordered_extent(ordered);
8801                 goto again;
8802         }
8803
8804         /*
8805          * XXX - page_mkwrite gets called every time the page is dirtied, even
8806          * if it was already dirty, so for space accounting reasons we need to
8807          * clear any delalloc bits for the range we are fixing to save.  There
8808          * is probably a better way to do this, but for now keep consistent with
8809          * prepare_pages in the normal write path.
8810          */
8811         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8812                           EXTENT_DIRTY | EXTENT_DELALLOC |
8813                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8814                           0, 0, &cached_state, GFP_NOFS);
8815
8816         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8817                                         &cached_state);
8818         if (ret) {
8819                 unlock_extent_cached(io_tree, page_start, page_end,
8820                                      &cached_state, GFP_NOFS);
8821                 ret = VM_FAULT_SIGBUS;
8822                 goto out_unlock;
8823         }
8824         ret = 0;
8825
8826         /* page is wholly or partially inside EOF */
8827         if (page_start + PAGE_CACHE_SIZE > size)
8828                 zero_start = size & ~PAGE_CACHE_MASK;
8829         else
8830                 zero_start = PAGE_CACHE_SIZE;
8831
8832         if (zero_start != PAGE_CACHE_SIZE) {
8833                 kaddr = kmap(page);
8834                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8835                 flush_dcache_page(page);
8836                 kunmap(page);
8837         }
8838         ClearPageChecked(page);
8839         set_page_dirty(page);
8840         SetPageUptodate(page);
8841
8842         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8843         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8844         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8845
8846         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8847
8848 out_unlock:
8849         if (!ret) {
8850                 sb_end_pagefault(inode->i_sb);
8851                 return VM_FAULT_LOCKED;
8852         }
8853         unlock_page(page);
8854 out:
8855         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8856 out_noreserve:
8857         sb_end_pagefault(inode->i_sb);
8858         return ret;
8859 }
8860
8861 static int btrfs_truncate(struct inode *inode)
8862 {
8863         struct btrfs_root *root = BTRFS_I(inode)->root;
8864         struct btrfs_block_rsv *rsv;
8865         int ret = 0;
8866         int err = 0;
8867         struct btrfs_trans_handle *trans;
8868         u64 mask = root->sectorsize - 1;
8869         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8870
8871         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8872                                        (u64)-1);
8873         if (ret)
8874                 return ret;
8875
8876         /*
8877          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8878          * 3 things going on here
8879          *
8880          * 1) We need to reserve space for our orphan item and the space to
8881          * delete our orphan item.  Lord knows we don't want to have a dangling
8882          * orphan item because we didn't reserve space to remove it.
8883          *
8884          * 2) We need to reserve space to update our inode.
8885          *
8886          * 3) We need to have something to cache all the space that is going to
8887          * be free'd up by the truncate operation, but also have some slack
8888          * space reserved in case it uses space during the truncate (thank you
8889          * very much snapshotting).
8890          *
8891          * And we need these to all be seperate.  The fact is we can use alot of
8892          * space doing the truncate, and we have no earthly idea how much space
8893          * we will use, so we need the truncate reservation to be seperate so it
8894          * doesn't end up using space reserved for updating the inode or
8895          * removing the orphan item.  We also need to be able to stop the
8896          * transaction and start a new one, which means we need to be able to
8897          * update the inode several times, and we have no idea of knowing how
8898          * many times that will be, so we can't just reserve 1 item for the
8899          * entirety of the opration, so that has to be done seperately as well.
8900          * Then there is the orphan item, which does indeed need to be held on
8901          * to for the whole operation, and we need nobody to touch this reserved
8902          * space except the orphan code.
8903          *
8904          * So that leaves us with
8905          *
8906          * 1) root->orphan_block_rsv - for the orphan deletion.
8907          * 2) rsv - for the truncate reservation, which we will steal from the
8908          * transaction reservation.
8909          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8910          * updating the inode.
8911          */
8912         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8913         if (!rsv)
8914                 return -ENOMEM;
8915         rsv->size = min_size;
8916         rsv->failfast = 1;
8917
8918         /*
8919          * 1 for the truncate slack space
8920          * 1 for updating the inode.
8921          */
8922         trans = btrfs_start_transaction(root, 2);
8923         if (IS_ERR(trans)) {
8924                 err = PTR_ERR(trans);
8925                 goto out;
8926         }
8927
8928         /* Migrate the slack space for the truncate to our reserve */
8929         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8930                                       min_size);
8931         BUG_ON(ret);
8932
8933         /*
8934          * So if we truncate and then write and fsync we normally would just
8935          * write the extents that changed, which is a problem if we need to
8936          * first truncate that entire inode.  So set this flag so we write out
8937          * all of the extents in the inode to the sync log so we're completely
8938          * safe.
8939          */
8940         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8941         trans->block_rsv = rsv;
8942
8943         while (1) {
8944                 ret = btrfs_truncate_inode_items(trans, root, inode,
8945                                                  inode->i_size,
8946                                                  BTRFS_EXTENT_DATA_KEY);
8947                 if (ret != -ENOSPC && ret != -EAGAIN) {
8948                         err = ret;
8949                         break;
8950                 }
8951
8952                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8953                 ret = btrfs_update_inode(trans, root, inode);
8954                 if (ret) {
8955                         err = ret;
8956                         break;
8957                 }
8958
8959                 btrfs_end_transaction(trans, root);
8960                 btrfs_btree_balance_dirty(root);
8961
8962                 trans = btrfs_start_transaction(root, 2);
8963                 if (IS_ERR(trans)) {
8964                         ret = err = PTR_ERR(trans);
8965                         trans = NULL;
8966                         break;
8967                 }
8968
8969                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8970                                               rsv, min_size);
8971                 BUG_ON(ret);    /* shouldn't happen */
8972                 trans->block_rsv = rsv;
8973         }
8974
8975         if (ret == 0 && inode->i_nlink > 0) {
8976                 trans->block_rsv = root->orphan_block_rsv;
8977                 ret = btrfs_orphan_del(trans, inode);
8978                 if (ret)
8979                         err = ret;
8980         }
8981
8982         if (trans) {
8983                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8984                 ret = btrfs_update_inode(trans, root, inode);
8985                 if (ret && !err)
8986                         err = ret;
8987
8988                 ret = btrfs_end_transaction(trans, root);
8989                 btrfs_btree_balance_dirty(root);
8990         }
8991
8992 out:
8993         btrfs_free_block_rsv(root, rsv);
8994
8995         if (ret && !err)
8996                 err = ret;
8997
8998         return err;
8999 }
9000
9001 /*
9002  * create a new subvolume directory/inode (helper for the ioctl).
9003  */
9004 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9005                              struct btrfs_root *new_root,
9006                              struct btrfs_root *parent_root,
9007                              u64 new_dirid)
9008 {
9009         struct inode *inode;
9010         int err;
9011         u64 index = 0;
9012
9013         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9014                                 new_dirid, new_dirid,
9015                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9016                                 &index);
9017         if (IS_ERR(inode))
9018                 return PTR_ERR(inode);
9019         inode->i_op = &btrfs_dir_inode_operations;
9020         inode->i_fop = &btrfs_dir_file_operations;
9021
9022         set_nlink(inode, 1);
9023         btrfs_i_size_write(inode, 0);
9024         unlock_new_inode(inode);
9025
9026         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9027         if (err)
9028                 btrfs_err(new_root->fs_info,
9029                           "error inheriting subvolume %llu properties: %d",
9030                           new_root->root_key.objectid, err);
9031
9032         err = btrfs_update_inode(trans, new_root, inode);
9033
9034         iput(inode);
9035         return err;
9036 }
9037
9038 struct inode *btrfs_alloc_inode(struct super_block *sb)
9039 {
9040         struct btrfs_inode *ei;
9041         struct inode *inode;
9042
9043         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9044         if (!ei)
9045                 return NULL;
9046
9047         ei->root = NULL;
9048         ei->generation = 0;
9049         ei->last_trans = 0;
9050         ei->last_sub_trans = 0;
9051         ei->logged_trans = 0;
9052         ei->delalloc_bytes = 0;
9053         ei->defrag_bytes = 0;
9054         ei->disk_i_size = 0;
9055         ei->flags = 0;
9056         ei->csum_bytes = 0;
9057         ei->index_cnt = (u64)-1;
9058         ei->dir_index = 0;
9059         ei->last_unlink_trans = 0;
9060         ei->last_log_commit = 0;
9061         ei->delayed_iput_count = 0;
9062
9063         spin_lock_init(&ei->lock);
9064         ei->outstanding_extents = 0;
9065         ei->reserved_extents = 0;
9066
9067         ei->runtime_flags = 0;
9068         ei->force_compress = BTRFS_COMPRESS_NONE;
9069
9070         ei->delayed_node = NULL;
9071
9072         ei->i_otime.tv_sec = 0;
9073         ei->i_otime.tv_nsec = 0;
9074
9075         inode = &ei->vfs_inode;
9076         extent_map_tree_init(&ei->extent_tree);
9077         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9078         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9079         ei->io_tree.track_uptodate = 1;
9080         ei->io_failure_tree.track_uptodate = 1;
9081         atomic_set(&ei->sync_writers, 0);
9082         mutex_init(&ei->log_mutex);
9083         mutex_init(&ei->delalloc_mutex);
9084         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9085         INIT_LIST_HEAD(&ei->delalloc_inodes);
9086         INIT_LIST_HEAD(&ei->delayed_iput);
9087         RB_CLEAR_NODE(&ei->rb_node);
9088
9089         return inode;
9090 }
9091
9092 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9093 void btrfs_test_destroy_inode(struct inode *inode)
9094 {
9095         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9096         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9097 }
9098 #endif
9099
9100 static void btrfs_i_callback(struct rcu_head *head)
9101 {
9102         struct inode *inode = container_of(head, struct inode, i_rcu);
9103         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9104 }
9105
9106 void btrfs_destroy_inode(struct inode *inode)
9107 {
9108         struct btrfs_ordered_extent *ordered;
9109         struct btrfs_root *root = BTRFS_I(inode)->root;
9110
9111         WARN_ON(!hlist_empty(&inode->i_dentry));
9112         WARN_ON(inode->i_data.nrpages);
9113         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9114         WARN_ON(BTRFS_I(inode)->reserved_extents);
9115         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9116         WARN_ON(BTRFS_I(inode)->csum_bytes);
9117         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9118
9119         /*
9120          * This can happen where we create an inode, but somebody else also
9121          * created the same inode and we need to destroy the one we already
9122          * created.
9123          */
9124         if (!root)
9125                 goto free;
9126
9127         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9128                      &BTRFS_I(inode)->runtime_flags)) {
9129                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9130                         btrfs_ino(inode));
9131                 atomic_dec(&root->orphan_inodes);
9132         }
9133
9134         while (1) {
9135                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9136                 if (!ordered)
9137                         break;
9138                 else {
9139                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9140                                 ordered->file_offset, ordered->len);
9141                         btrfs_remove_ordered_extent(inode, ordered);
9142                         btrfs_put_ordered_extent(ordered);
9143                         btrfs_put_ordered_extent(ordered);
9144                 }
9145         }
9146         btrfs_qgroup_check_reserved_leak(inode);
9147         inode_tree_del(inode);
9148         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9149 free:
9150         call_rcu(&inode->i_rcu, btrfs_i_callback);
9151 }
9152
9153 int btrfs_drop_inode(struct inode *inode)
9154 {
9155         struct btrfs_root *root = BTRFS_I(inode)->root;
9156
9157         if (root == NULL)
9158                 return 1;
9159
9160         /* the snap/subvol tree is on deleting */
9161         if (btrfs_root_refs(&root->root_item) == 0)
9162                 return 1;
9163         else
9164                 return generic_drop_inode(inode);
9165 }
9166
9167 static void init_once(void *foo)
9168 {
9169         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9170
9171         inode_init_once(&ei->vfs_inode);
9172 }
9173
9174 void btrfs_destroy_cachep(void)
9175 {
9176         /*
9177          * Make sure all delayed rcu free inodes are flushed before we
9178          * destroy cache.
9179          */
9180         rcu_barrier();
9181         if (btrfs_inode_cachep)
9182                 kmem_cache_destroy(btrfs_inode_cachep);
9183         if (btrfs_trans_handle_cachep)
9184                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9185         if (btrfs_transaction_cachep)
9186                 kmem_cache_destroy(btrfs_transaction_cachep);
9187         if (btrfs_path_cachep)
9188                 kmem_cache_destroy(btrfs_path_cachep);
9189         if (btrfs_free_space_cachep)
9190                 kmem_cache_destroy(btrfs_free_space_cachep);
9191 }
9192
9193 int btrfs_init_cachep(void)
9194 {
9195         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9196                         sizeof(struct btrfs_inode), 0,
9197                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9198                         init_once);
9199         if (!btrfs_inode_cachep)
9200                 goto fail;
9201
9202         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9203                         sizeof(struct btrfs_trans_handle), 0,
9204                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9205         if (!btrfs_trans_handle_cachep)
9206                 goto fail;
9207
9208         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9209                         sizeof(struct btrfs_transaction), 0,
9210                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9211         if (!btrfs_transaction_cachep)
9212                 goto fail;
9213
9214         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9215                         sizeof(struct btrfs_path), 0,
9216                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9217         if (!btrfs_path_cachep)
9218                 goto fail;
9219
9220         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9221                         sizeof(struct btrfs_free_space), 0,
9222                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9223         if (!btrfs_free_space_cachep)
9224                 goto fail;
9225
9226         return 0;
9227 fail:
9228         btrfs_destroy_cachep();
9229         return -ENOMEM;
9230 }
9231
9232 static int btrfs_getattr(struct vfsmount *mnt,
9233                          struct dentry *dentry, struct kstat *stat)
9234 {
9235         u64 delalloc_bytes;
9236         struct inode *inode = d_inode(dentry);
9237         u32 blocksize = inode->i_sb->s_blocksize;
9238
9239         generic_fillattr(inode, stat);
9240         stat->dev = BTRFS_I(inode)->root->anon_dev;
9241         stat->blksize = PAGE_CACHE_SIZE;
9242
9243         spin_lock(&BTRFS_I(inode)->lock);
9244         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9245         spin_unlock(&BTRFS_I(inode)->lock);
9246         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9247                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9248         return 0;
9249 }
9250
9251 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9252                            struct inode *new_dir, struct dentry *new_dentry)
9253 {
9254         struct btrfs_trans_handle *trans;
9255         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9256         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9257         struct inode *new_inode = d_inode(new_dentry);
9258         struct inode *old_inode = d_inode(old_dentry);
9259         struct timespec ctime = CURRENT_TIME;
9260         u64 index = 0;
9261         u64 root_objectid;
9262         int ret;
9263         u64 old_ino = btrfs_ino(old_inode);
9264
9265         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9266                 return -EPERM;
9267
9268         /* we only allow rename subvolume link between subvolumes */
9269         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9270                 return -EXDEV;
9271
9272         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9273             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9274                 return -ENOTEMPTY;
9275
9276         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9277             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9278                 return -ENOTEMPTY;
9279
9280
9281         /* check for collisions, even if the  name isn't there */
9282         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9283                              new_dentry->d_name.name,
9284                              new_dentry->d_name.len);
9285
9286         if (ret) {
9287                 if (ret == -EEXIST) {
9288                         /* we shouldn't get
9289                          * eexist without a new_inode */
9290                         if (WARN_ON(!new_inode)) {
9291                                 return ret;
9292                         }
9293                 } else {
9294                         /* maybe -EOVERFLOW */
9295                         return ret;
9296                 }
9297         }
9298         ret = 0;
9299
9300         /*
9301          * we're using rename to replace one file with another.  Start IO on it
9302          * now so  we don't add too much work to the end of the transaction
9303          */
9304         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9305                 filemap_flush(old_inode->i_mapping);
9306
9307         /* close the racy window with snapshot create/destroy ioctl */
9308         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9309                 down_read(&root->fs_info->subvol_sem);
9310         /*
9311          * We want to reserve the absolute worst case amount of items.  So if
9312          * both inodes are subvols and we need to unlink them then that would
9313          * require 4 item modifications, but if they are both normal inodes it
9314          * would require 5 item modifications, so we'll assume their normal
9315          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9316          * should cover the worst case number of items we'll modify.
9317          */
9318         trans = btrfs_start_transaction(root, 11);
9319         if (IS_ERR(trans)) {
9320                 ret = PTR_ERR(trans);
9321                 goto out_notrans;
9322         }
9323
9324         if (dest != root)
9325                 btrfs_record_root_in_trans(trans, dest);
9326
9327         ret = btrfs_set_inode_index(new_dir, &index);
9328         if (ret)
9329                 goto out_fail;
9330
9331         BTRFS_I(old_inode)->dir_index = 0ULL;
9332         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9333                 /* force full log commit if subvolume involved. */
9334                 btrfs_set_log_full_commit(root->fs_info, trans);
9335         } else {
9336                 ret = btrfs_insert_inode_ref(trans, dest,
9337                                              new_dentry->d_name.name,
9338                                              new_dentry->d_name.len,
9339                                              old_ino,
9340                                              btrfs_ino(new_dir), index);
9341                 if (ret)
9342                         goto out_fail;
9343                 /*
9344                  * this is an ugly little race, but the rename is required
9345                  * to make sure that if we crash, the inode is either at the
9346                  * old name or the new one.  pinning the log transaction lets
9347                  * us make sure we don't allow a log commit to come in after
9348                  * we unlink the name but before we add the new name back in.
9349                  */
9350                 btrfs_pin_log_trans(root);
9351         }
9352
9353         inode_inc_iversion(old_dir);
9354         inode_inc_iversion(new_dir);
9355         inode_inc_iversion(old_inode);
9356         old_dir->i_ctime = old_dir->i_mtime = ctime;
9357         new_dir->i_ctime = new_dir->i_mtime = ctime;
9358         old_inode->i_ctime = ctime;
9359
9360         if (old_dentry->d_parent != new_dentry->d_parent)
9361                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9362
9363         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9364                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9365                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9366                                         old_dentry->d_name.name,
9367                                         old_dentry->d_name.len);
9368         } else {
9369                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9370                                         d_inode(old_dentry),
9371                                         old_dentry->d_name.name,
9372                                         old_dentry->d_name.len);
9373                 if (!ret)
9374                         ret = btrfs_update_inode(trans, root, old_inode);
9375         }
9376         if (ret) {
9377                 btrfs_abort_transaction(trans, root, ret);
9378                 goto out_fail;
9379         }
9380
9381         if (new_inode) {
9382                 inode_inc_iversion(new_inode);
9383                 new_inode->i_ctime = CURRENT_TIME;
9384                 if (unlikely(btrfs_ino(new_inode) ==
9385                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9386                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9387                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9388                                                 root_objectid,
9389                                                 new_dentry->d_name.name,
9390                                                 new_dentry->d_name.len);
9391                         BUG_ON(new_inode->i_nlink == 0);
9392                 } else {
9393                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9394                                                  d_inode(new_dentry),
9395                                                  new_dentry->d_name.name,
9396                                                  new_dentry->d_name.len);
9397                 }
9398                 if (!ret && new_inode->i_nlink == 0)
9399                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9400                 if (ret) {
9401                         btrfs_abort_transaction(trans, root, ret);
9402                         goto out_fail;
9403                 }
9404         }
9405
9406         ret = btrfs_add_link(trans, new_dir, old_inode,
9407                              new_dentry->d_name.name,
9408                              new_dentry->d_name.len, 0, index);
9409         if (ret) {
9410                 btrfs_abort_transaction(trans, root, ret);
9411                 goto out_fail;
9412         }
9413
9414         if (old_inode->i_nlink == 1)
9415                 BTRFS_I(old_inode)->dir_index = index;
9416
9417         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9418                 struct dentry *parent = new_dentry->d_parent;
9419                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9420                 btrfs_end_log_trans(root);
9421         }
9422 out_fail:
9423         btrfs_end_transaction(trans, root);
9424 out_notrans:
9425         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9426                 up_read(&root->fs_info->subvol_sem);
9427
9428         return ret;
9429 }
9430
9431 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9432                          struct inode *new_dir, struct dentry *new_dentry,
9433                          unsigned int flags)
9434 {
9435         if (flags & ~RENAME_NOREPLACE)
9436                 return -EINVAL;
9437
9438         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9439 }
9440
9441 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9442 {
9443         struct btrfs_delalloc_work *delalloc_work;
9444         struct inode *inode;
9445
9446         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9447                                      work);
9448         inode = delalloc_work->inode;
9449         filemap_flush(inode->i_mapping);
9450         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9451                                 &BTRFS_I(inode)->runtime_flags))
9452                 filemap_flush(inode->i_mapping);
9453
9454         if (delalloc_work->delay_iput)
9455                 btrfs_add_delayed_iput(inode);
9456         else
9457                 iput(inode);
9458         complete(&delalloc_work->completion);
9459 }
9460
9461 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9462                                                     int delay_iput)
9463 {
9464         struct btrfs_delalloc_work *work;
9465
9466         work = kmalloc(sizeof(*work), GFP_NOFS);
9467         if (!work)
9468                 return NULL;
9469
9470         init_completion(&work->completion);
9471         INIT_LIST_HEAD(&work->list);
9472         work->inode = inode;
9473         work->delay_iput = delay_iput;
9474         WARN_ON_ONCE(!inode);
9475         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9476                         btrfs_run_delalloc_work, NULL, NULL);
9477
9478         return work;
9479 }
9480
9481 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9482 {
9483         wait_for_completion(&work->completion);
9484         kfree(work);
9485 }
9486
9487 /*
9488  * some fairly slow code that needs optimization. This walks the list
9489  * of all the inodes with pending delalloc and forces them to disk.
9490  */
9491 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9492                                    int nr)
9493 {
9494         struct btrfs_inode *binode;
9495         struct inode *inode;
9496         struct btrfs_delalloc_work *work, *next;
9497         struct list_head works;
9498         struct list_head splice;
9499         int ret = 0;
9500
9501         INIT_LIST_HEAD(&works);
9502         INIT_LIST_HEAD(&splice);
9503
9504         mutex_lock(&root->delalloc_mutex);
9505         spin_lock(&root->delalloc_lock);
9506         list_splice_init(&root->delalloc_inodes, &splice);
9507         while (!list_empty(&splice)) {
9508                 binode = list_entry(splice.next, struct btrfs_inode,
9509                                     delalloc_inodes);
9510
9511                 list_move_tail(&binode->delalloc_inodes,
9512                                &root->delalloc_inodes);
9513                 inode = igrab(&binode->vfs_inode);
9514                 if (!inode) {
9515                         cond_resched_lock(&root->delalloc_lock);
9516                         continue;
9517                 }
9518                 spin_unlock(&root->delalloc_lock);
9519
9520                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
9521                 if (!work) {
9522                         if (delay_iput)
9523                                 btrfs_add_delayed_iput(inode);
9524                         else
9525                                 iput(inode);
9526                         ret = -ENOMEM;
9527                         goto out;
9528                 }
9529                 list_add_tail(&work->list, &works);
9530                 btrfs_queue_work(root->fs_info->flush_workers,
9531                                  &work->work);
9532                 ret++;
9533                 if (nr != -1 && ret >= nr)
9534                         goto out;
9535                 cond_resched();
9536                 spin_lock(&root->delalloc_lock);
9537         }
9538         spin_unlock(&root->delalloc_lock);
9539
9540 out:
9541         list_for_each_entry_safe(work, next, &works, list) {
9542                 list_del_init(&work->list);
9543                 btrfs_wait_and_free_delalloc_work(work);
9544         }
9545
9546         if (!list_empty_careful(&splice)) {
9547                 spin_lock(&root->delalloc_lock);
9548                 list_splice_tail(&splice, &root->delalloc_inodes);
9549                 spin_unlock(&root->delalloc_lock);
9550         }
9551         mutex_unlock(&root->delalloc_mutex);
9552         return ret;
9553 }
9554
9555 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9556 {
9557         int ret;
9558
9559         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9560                 return -EROFS;
9561
9562         ret = __start_delalloc_inodes(root, delay_iput, -1);
9563         if (ret > 0)
9564                 ret = 0;
9565         /*
9566          * the filemap_flush will queue IO into the worker threads, but
9567          * we have to make sure the IO is actually started and that
9568          * ordered extents get created before we return
9569          */
9570         atomic_inc(&root->fs_info->async_submit_draining);
9571         while (atomic_read(&root->fs_info->nr_async_submits) ||
9572               atomic_read(&root->fs_info->async_delalloc_pages)) {
9573                 wait_event(root->fs_info->async_submit_wait,
9574                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9575                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9576         }
9577         atomic_dec(&root->fs_info->async_submit_draining);
9578         return ret;
9579 }
9580
9581 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9582                                int nr)
9583 {
9584         struct btrfs_root *root;
9585         struct list_head splice;
9586         int ret;
9587
9588         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9589                 return -EROFS;
9590
9591         INIT_LIST_HEAD(&splice);
9592
9593         mutex_lock(&fs_info->delalloc_root_mutex);
9594         spin_lock(&fs_info->delalloc_root_lock);
9595         list_splice_init(&fs_info->delalloc_roots, &splice);
9596         while (!list_empty(&splice) && nr) {
9597                 root = list_first_entry(&splice, struct btrfs_root,
9598                                         delalloc_root);
9599                 root = btrfs_grab_fs_root(root);
9600                 BUG_ON(!root);
9601                 list_move_tail(&root->delalloc_root,
9602                                &fs_info->delalloc_roots);
9603                 spin_unlock(&fs_info->delalloc_root_lock);
9604
9605                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9606                 btrfs_put_fs_root(root);
9607                 if (ret < 0)
9608                         goto out;
9609
9610                 if (nr != -1) {
9611                         nr -= ret;
9612                         WARN_ON(nr < 0);
9613                 }
9614                 spin_lock(&fs_info->delalloc_root_lock);
9615         }
9616         spin_unlock(&fs_info->delalloc_root_lock);
9617
9618         ret = 0;
9619         atomic_inc(&fs_info->async_submit_draining);
9620         while (atomic_read(&fs_info->nr_async_submits) ||
9621               atomic_read(&fs_info->async_delalloc_pages)) {
9622                 wait_event(fs_info->async_submit_wait,
9623                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9624                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9625         }
9626         atomic_dec(&fs_info->async_submit_draining);
9627 out:
9628         if (!list_empty_careful(&splice)) {
9629                 spin_lock(&fs_info->delalloc_root_lock);
9630                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9631                 spin_unlock(&fs_info->delalloc_root_lock);
9632         }
9633         mutex_unlock(&fs_info->delalloc_root_mutex);
9634         return ret;
9635 }
9636
9637 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9638                          const char *symname)
9639 {
9640         struct btrfs_trans_handle *trans;
9641         struct btrfs_root *root = BTRFS_I(dir)->root;
9642         struct btrfs_path *path;
9643         struct btrfs_key key;
9644         struct inode *inode = NULL;
9645         int err;
9646         int drop_inode = 0;
9647         u64 objectid;
9648         u64 index = 0;
9649         int name_len;
9650         int datasize;
9651         unsigned long ptr;
9652         struct btrfs_file_extent_item *ei;
9653         struct extent_buffer *leaf;
9654
9655         name_len = strlen(symname);
9656         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9657                 return -ENAMETOOLONG;
9658
9659         /*
9660          * 2 items for inode item and ref
9661          * 2 items for dir items
9662          * 1 item for updating parent inode item
9663          * 1 item for the inline extent item
9664          * 1 item for xattr if selinux is on
9665          */
9666         trans = btrfs_start_transaction(root, 7);
9667         if (IS_ERR(trans))
9668                 return PTR_ERR(trans);
9669
9670         err = btrfs_find_free_ino(root, &objectid);
9671         if (err)
9672                 goto out_unlock;
9673
9674         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9675                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9676                                 S_IFLNK|S_IRWXUGO, &index);
9677         if (IS_ERR(inode)) {
9678                 err = PTR_ERR(inode);
9679                 goto out_unlock;
9680         }
9681
9682         /*
9683         * If the active LSM wants to access the inode during
9684         * d_instantiate it needs these. Smack checks to see
9685         * if the filesystem supports xattrs by looking at the
9686         * ops vector.
9687         */
9688         inode->i_fop = &btrfs_file_operations;
9689         inode->i_op = &btrfs_file_inode_operations;
9690         inode->i_mapping->a_ops = &btrfs_aops;
9691         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9692
9693         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9694         if (err)
9695                 goto out_unlock_inode;
9696
9697         path = btrfs_alloc_path();
9698         if (!path) {
9699                 err = -ENOMEM;
9700                 goto out_unlock_inode;
9701         }
9702         key.objectid = btrfs_ino(inode);
9703         key.offset = 0;
9704         key.type = BTRFS_EXTENT_DATA_KEY;
9705         datasize = btrfs_file_extent_calc_inline_size(name_len);
9706         err = btrfs_insert_empty_item(trans, root, path, &key,
9707                                       datasize);
9708         if (err) {
9709                 btrfs_free_path(path);
9710                 goto out_unlock_inode;
9711         }
9712         leaf = path->nodes[0];
9713         ei = btrfs_item_ptr(leaf, path->slots[0],
9714                             struct btrfs_file_extent_item);
9715         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9716         btrfs_set_file_extent_type(leaf, ei,
9717                                    BTRFS_FILE_EXTENT_INLINE);
9718         btrfs_set_file_extent_encryption(leaf, ei, 0);
9719         btrfs_set_file_extent_compression(leaf, ei, 0);
9720         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9721         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9722
9723         ptr = btrfs_file_extent_inline_start(ei);
9724         write_extent_buffer(leaf, symname, ptr, name_len);
9725         btrfs_mark_buffer_dirty(leaf);
9726         btrfs_free_path(path);
9727
9728         inode->i_op = &btrfs_symlink_inode_operations;
9729         inode_nohighmem(inode);
9730         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9731         inode_set_bytes(inode, name_len);
9732         btrfs_i_size_write(inode, name_len);
9733         err = btrfs_update_inode(trans, root, inode);
9734         /*
9735          * Last step, add directory indexes for our symlink inode. This is the
9736          * last step to avoid extra cleanup of these indexes if an error happens
9737          * elsewhere above.
9738          */
9739         if (!err)
9740                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9741         if (err) {
9742                 drop_inode = 1;
9743                 goto out_unlock_inode;
9744         }
9745
9746         unlock_new_inode(inode);
9747         d_instantiate(dentry, inode);
9748
9749 out_unlock:
9750         btrfs_end_transaction(trans, root);
9751         if (drop_inode) {
9752                 inode_dec_link_count(inode);
9753                 iput(inode);
9754         }
9755         btrfs_btree_balance_dirty(root);
9756         return err;
9757
9758 out_unlock_inode:
9759         drop_inode = 1;
9760         unlock_new_inode(inode);
9761         goto out_unlock;
9762 }
9763
9764 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9765                                        u64 start, u64 num_bytes, u64 min_size,
9766                                        loff_t actual_len, u64 *alloc_hint,
9767                                        struct btrfs_trans_handle *trans)
9768 {
9769         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9770         struct extent_map *em;
9771         struct btrfs_root *root = BTRFS_I(inode)->root;
9772         struct btrfs_key ins;
9773         u64 cur_offset = start;
9774         u64 i_size;
9775         u64 cur_bytes;
9776         u64 last_alloc = (u64)-1;
9777         int ret = 0;
9778         bool own_trans = true;
9779
9780         if (trans)
9781                 own_trans = false;
9782         while (num_bytes > 0) {
9783                 if (own_trans) {
9784                         trans = btrfs_start_transaction(root, 3);
9785                         if (IS_ERR(trans)) {
9786                                 ret = PTR_ERR(trans);
9787                                 break;
9788                         }
9789                 }
9790
9791                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9792                 cur_bytes = max(cur_bytes, min_size);
9793                 /*
9794                  * If we are severely fragmented we could end up with really
9795                  * small allocations, so if the allocator is returning small
9796                  * chunks lets make its job easier by only searching for those
9797                  * sized chunks.
9798                  */
9799                 cur_bytes = min(cur_bytes, last_alloc);
9800                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9801                                            *alloc_hint, &ins, 1, 0);
9802                 if (ret) {
9803                         if (own_trans)
9804                                 btrfs_end_transaction(trans, root);
9805                         break;
9806                 }
9807
9808                 last_alloc = ins.offset;
9809                 ret = insert_reserved_file_extent(trans, inode,
9810                                                   cur_offset, ins.objectid,
9811                                                   ins.offset, ins.offset,
9812                                                   ins.offset, 0, 0, 0,
9813                                                   BTRFS_FILE_EXTENT_PREALLOC);
9814                 if (ret) {
9815                         btrfs_free_reserved_extent(root, ins.objectid,
9816                                                    ins.offset, 0);
9817                         btrfs_abort_transaction(trans, root, ret);
9818                         if (own_trans)
9819                                 btrfs_end_transaction(trans, root);
9820                         break;
9821                 }
9822
9823                 btrfs_drop_extent_cache(inode, cur_offset,
9824                                         cur_offset + ins.offset -1, 0);
9825
9826                 em = alloc_extent_map();
9827                 if (!em) {
9828                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9829                                 &BTRFS_I(inode)->runtime_flags);
9830                         goto next;
9831                 }
9832
9833                 em->start = cur_offset;
9834                 em->orig_start = cur_offset;
9835                 em->len = ins.offset;
9836                 em->block_start = ins.objectid;
9837                 em->block_len = ins.offset;
9838                 em->orig_block_len = ins.offset;
9839                 em->ram_bytes = ins.offset;
9840                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9841                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9842                 em->generation = trans->transid;
9843
9844                 while (1) {
9845                         write_lock(&em_tree->lock);
9846                         ret = add_extent_mapping(em_tree, em, 1);
9847                         write_unlock(&em_tree->lock);
9848                         if (ret != -EEXIST)
9849                                 break;
9850                         btrfs_drop_extent_cache(inode, cur_offset,
9851                                                 cur_offset + ins.offset - 1,
9852                                                 0);
9853                 }
9854                 free_extent_map(em);
9855 next:
9856                 num_bytes -= ins.offset;
9857                 cur_offset += ins.offset;
9858                 *alloc_hint = ins.objectid + ins.offset;
9859
9860                 inode_inc_iversion(inode);
9861                 inode->i_ctime = CURRENT_TIME;
9862                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9863                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9864                     (actual_len > inode->i_size) &&
9865                     (cur_offset > inode->i_size)) {
9866                         if (cur_offset > actual_len)
9867                                 i_size = actual_len;
9868                         else
9869                                 i_size = cur_offset;
9870                         i_size_write(inode, i_size);
9871                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9872                 }
9873
9874                 ret = btrfs_update_inode(trans, root, inode);
9875
9876                 if (ret) {
9877                         btrfs_abort_transaction(trans, root, ret);
9878                         if (own_trans)
9879                                 btrfs_end_transaction(trans, root);
9880                         break;
9881                 }
9882
9883                 if (own_trans)
9884                         btrfs_end_transaction(trans, root);
9885         }
9886         return ret;
9887 }
9888
9889 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9890                               u64 start, u64 num_bytes, u64 min_size,
9891                               loff_t actual_len, u64 *alloc_hint)
9892 {
9893         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9894                                            min_size, actual_len, alloc_hint,
9895                                            NULL);
9896 }
9897
9898 int btrfs_prealloc_file_range_trans(struct inode *inode,
9899                                     struct btrfs_trans_handle *trans, int mode,
9900                                     u64 start, u64 num_bytes, u64 min_size,
9901                                     loff_t actual_len, u64 *alloc_hint)
9902 {
9903         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9904                                            min_size, actual_len, alloc_hint, trans);
9905 }
9906
9907 static int btrfs_set_page_dirty(struct page *page)
9908 {
9909         return __set_page_dirty_nobuffers(page);
9910 }
9911
9912 static int btrfs_permission(struct inode *inode, int mask)
9913 {
9914         struct btrfs_root *root = BTRFS_I(inode)->root;
9915         umode_t mode = inode->i_mode;
9916
9917         if (mask & MAY_WRITE &&
9918             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9919                 if (btrfs_root_readonly(root))
9920                         return -EROFS;
9921                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9922                         return -EACCES;
9923         }
9924         return generic_permission(inode, mask);
9925 }
9926
9927 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9928 {
9929         struct btrfs_trans_handle *trans;
9930         struct btrfs_root *root = BTRFS_I(dir)->root;
9931         struct inode *inode = NULL;
9932         u64 objectid;
9933         u64 index;
9934         int ret = 0;
9935
9936         /*
9937          * 5 units required for adding orphan entry
9938          */
9939         trans = btrfs_start_transaction(root, 5);
9940         if (IS_ERR(trans))
9941                 return PTR_ERR(trans);
9942
9943         ret = btrfs_find_free_ino(root, &objectid);
9944         if (ret)
9945                 goto out;
9946
9947         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9948                                 btrfs_ino(dir), objectid, mode, &index);
9949         if (IS_ERR(inode)) {
9950                 ret = PTR_ERR(inode);
9951                 inode = NULL;
9952                 goto out;
9953         }
9954
9955         inode->i_fop = &btrfs_file_operations;
9956         inode->i_op = &btrfs_file_inode_operations;
9957
9958         inode->i_mapping->a_ops = &btrfs_aops;
9959         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9960
9961         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9962         if (ret)
9963                 goto out_inode;
9964
9965         ret = btrfs_update_inode(trans, root, inode);
9966         if (ret)
9967                 goto out_inode;
9968         ret = btrfs_orphan_add(trans, inode);
9969         if (ret)
9970                 goto out_inode;
9971
9972         /*
9973          * We set number of links to 0 in btrfs_new_inode(), and here we set
9974          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9975          * through:
9976          *
9977          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9978          */
9979         set_nlink(inode, 1);
9980         unlock_new_inode(inode);
9981         d_tmpfile(dentry, inode);
9982         mark_inode_dirty(inode);
9983
9984 out:
9985         btrfs_end_transaction(trans, root);
9986         if (ret)
9987                 iput(inode);
9988         btrfs_balance_delayed_items(root);
9989         btrfs_btree_balance_dirty(root);
9990         return ret;
9991
9992 out_inode:
9993         unlock_new_inode(inode);
9994         goto out;
9995
9996 }
9997
9998 /* Inspired by filemap_check_errors() */
9999 int btrfs_inode_check_errors(struct inode *inode)
10000 {
10001         int ret = 0;
10002
10003         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10004             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10005                 ret = -ENOSPC;
10006         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10007             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10008                 ret = -EIO;
10009
10010         return ret;
10011 }
10012
10013 static const struct inode_operations btrfs_dir_inode_operations = {
10014         .getattr        = btrfs_getattr,
10015         .lookup         = btrfs_lookup,
10016         .create         = btrfs_create,
10017         .unlink         = btrfs_unlink,
10018         .link           = btrfs_link,
10019         .mkdir          = btrfs_mkdir,
10020         .rmdir          = btrfs_rmdir,
10021         .rename2        = btrfs_rename2,
10022         .symlink        = btrfs_symlink,
10023         .setattr        = btrfs_setattr,
10024         .mknod          = btrfs_mknod,
10025         .setxattr       = btrfs_setxattr,
10026         .getxattr       = generic_getxattr,
10027         .listxattr      = btrfs_listxattr,
10028         .removexattr    = btrfs_removexattr,
10029         .permission     = btrfs_permission,
10030         .get_acl        = btrfs_get_acl,
10031         .set_acl        = btrfs_set_acl,
10032         .update_time    = btrfs_update_time,
10033         .tmpfile        = btrfs_tmpfile,
10034 };
10035 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10036         .lookup         = btrfs_lookup,
10037         .permission     = btrfs_permission,
10038         .get_acl        = btrfs_get_acl,
10039         .set_acl        = btrfs_set_acl,
10040         .update_time    = btrfs_update_time,
10041 };
10042
10043 static const struct file_operations btrfs_dir_file_operations = {
10044         .llseek         = generic_file_llseek,
10045         .read           = generic_read_dir,
10046         .iterate        = btrfs_real_readdir,
10047         .unlocked_ioctl = btrfs_ioctl,
10048 #ifdef CONFIG_COMPAT
10049         .compat_ioctl   = btrfs_ioctl,
10050 #endif
10051         .release        = btrfs_release_file,
10052         .fsync          = btrfs_sync_file,
10053 };
10054
10055 static const struct extent_io_ops btrfs_extent_io_ops = {
10056         .fill_delalloc = run_delalloc_range,
10057         .submit_bio_hook = btrfs_submit_bio_hook,
10058         .merge_bio_hook = btrfs_merge_bio_hook,
10059         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10060         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10061         .writepage_start_hook = btrfs_writepage_start_hook,
10062         .set_bit_hook = btrfs_set_bit_hook,
10063         .clear_bit_hook = btrfs_clear_bit_hook,
10064         .merge_extent_hook = btrfs_merge_extent_hook,
10065         .split_extent_hook = btrfs_split_extent_hook,
10066 };
10067
10068 /*
10069  * btrfs doesn't support the bmap operation because swapfiles
10070  * use bmap to make a mapping of extents in the file.  They assume
10071  * these extents won't change over the life of the file and they
10072  * use the bmap result to do IO directly to the drive.
10073  *
10074  * the btrfs bmap call would return logical addresses that aren't
10075  * suitable for IO and they also will change frequently as COW
10076  * operations happen.  So, swapfile + btrfs == corruption.
10077  *
10078  * For now we're avoiding this by dropping bmap.
10079  */
10080 static const struct address_space_operations btrfs_aops = {
10081         .readpage       = btrfs_readpage,
10082         .writepage      = btrfs_writepage,
10083         .writepages     = btrfs_writepages,
10084         .readpages      = btrfs_readpages,
10085         .direct_IO      = btrfs_direct_IO,
10086         .invalidatepage = btrfs_invalidatepage,
10087         .releasepage    = btrfs_releasepage,
10088         .set_page_dirty = btrfs_set_page_dirty,
10089         .error_remove_page = generic_error_remove_page,
10090 };
10091
10092 static const struct address_space_operations btrfs_symlink_aops = {
10093         .readpage       = btrfs_readpage,
10094         .writepage      = btrfs_writepage,
10095         .invalidatepage = btrfs_invalidatepage,
10096         .releasepage    = btrfs_releasepage,
10097 };
10098
10099 static const struct inode_operations btrfs_file_inode_operations = {
10100         .getattr        = btrfs_getattr,
10101         .setattr        = btrfs_setattr,
10102         .setxattr       = btrfs_setxattr,
10103         .getxattr       = generic_getxattr,
10104         .listxattr      = btrfs_listxattr,
10105         .removexattr    = btrfs_removexattr,
10106         .permission     = btrfs_permission,
10107         .fiemap         = btrfs_fiemap,
10108         .get_acl        = btrfs_get_acl,
10109         .set_acl        = btrfs_set_acl,
10110         .update_time    = btrfs_update_time,
10111 };
10112 static const struct inode_operations btrfs_special_inode_operations = {
10113         .getattr        = btrfs_getattr,
10114         .setattr        = btrfs_setattr,
10115         .permission     = btrfs_permission,
10116         .setxattr       = btrfs_setxattr,
10117         .getxattr       = generic_getxattr,
10118         .listxattr      = btrfs_listxattr,
10119         .removexattr    = btrfs_removexattr,
10120         .get_acl        = btrfs_get_acl,
10121         .set_acl        = btrfs_set_acl,
10122         .update_time    = btrfs_update_time,
10123 };
10124 static const struct inode_operations btrfs_symlink_inode_operations = {
10125         .readlink       = generic_readlink,
10126         .get_link       = page_get_link,
10127         .getattr        = btrfs_getattr,
10128         .setattr        = btrfs_setattr,
10129         .permission     = btrfs_permission,
10130         .setxattr       = btrfs_setxattr,
10131         .getxattr       = generic_getxattr,
10132         .listxattr      = btrfs_listxattr,
10133         .removexattr    = btrfs_removexattr,
10134         .update_time    = btrfs_update_time,
10135 };
10136
10137 const struct dentry_operations btrfs_dentry_operations = {
10138         .d_delete       = btrfs_dentry_delete,
10139         .d_release      = btrfs_dentry_release,
10140 };