fs/btrfs/file.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/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/aio.h>
  28 #include <linux/falloc.h>
  29 #include <linux/swap.h>
  30 #include <linux/writeback.h>
  31 #include <linux/statfs.h>
  32 #include <linux/compat.h>
  33 #include <linux/slab.h>
  34 #include <linux/btrfs.h>
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "tree-log.h"
  41 #include "locking.h"
  42 #include "volumes.h"
  43
  44 static struct kmem_cache *btrfs_inode_defrag_cachep;
  45 /*
  46  * when auto defrag is enabled we
  47  * queue up these defrag structs to remember which
  48  * inodes need defragging passes
  49  */
  50 struct inode_defrag {
  51         struct rb_node rb_node;
  52         /* objectid */
  53         u64 ino;
  54         /*
  55          * transid where the defrag was added, we search for
  56          * extents newer than this
  57          */
  58         u64 transid;
  59
  60         /* root objectid */
  61         u64 root;
  62
  63         /* last offset we were able to defrag */
  64         u64 last_offset;
  65
  66         /* if we've wrapped around back to zero once already */
  67         int cycled;
  68 };
  69
  70 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  71                                   struct inode_defrag *defrag2)
  72 {
  73         if (defrag1->root > defrag2->root)
  74                 return 1;
  75         else if (defrag1->root < defrag2->root)
  76                 return -1;
  77         else if (defrag1->ino > defrag2->ino)
  78                 return 1;
  79         else if (defrag1->ino < defrag2->ino)
  80                 return -1;
  81         else
  82                 return 0;
  83 }
  84
  85 /* pop a record for an inode into the defrag tree.  The lock
  86  * must be held already
  87  *
  88  * If you're inserting a record for an older transid than an
  89  * existing record, the transid already in the tree is lowered
  90  *
  91  * If an existing record is found the defrag item you
  92  * pass in is freed
  93  */
  94 static int __btrfs_add_inode_defrag(struct inode *inode,
  95                                     struct inode_defrag *defrag)
  96 {
  97         struct btrfs_root *root = BTRFS_I(inode)->root;
  98         struct inode_defrag *entry;
  99         struct rb_node **p;
 100         struct rb_node *parent = NULL;
 101         int ret;
 102
 103         p = &root->fs_info->defrag_inodes.rb_node;
 104         while (*p) {
 105                 parent = *p;
 106                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 107
 108                 ret = __compare_inode_defrag(defrag, entry);
 109                 if (ret < 0)
 110                         p = &parent->rb_left;
 111                 else if (ret > 0)
 112                         p = &parent->rb_right;
 113                 else {
 114                         /* if we're reinserting an entry for
 115                          * an old defrag run, make sure to
 116                          * lower the transid of our existing record
 117                          */
 118                         if (defrag->transid < entry->transid)
 119                                 entry->transid = defrag->transid;
 120                         if (defrag->last_offset > entry->last_offset)
 121                                 entry->last_offset = defrag->last_offset;
 122                         return -EEXIST;
 123                 }
 124         }
 125         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 126         rb_link_node(&defrag->rb_node, parent, p);
 127         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 128         return 0;
 129 }
 130
 131 static inline int __need_auto_defrag(struct btrfs_root *root)
 132 {
 133         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 134                 return 0;
 135
 136         if (btrfs_fs_closing(root->fs_info))
 137                 return 0;
 138
 139         return 1;
 140 }
 141
 142 /*
 143  * insert a defrag record for this inode if auto defrag is
 144  * enabled
 145  */
 146 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 147                            struct inode *inode)
 148 {
 149         struct btrfs_root *root = BTRFS_I(inode)->root;
 150         struct inode_defrag *defrag;
 151         u64 transid;
 152         int ret;
 153
 154         if (!__need_auto_defrag(root))
 155                 return 0;
 156
 157         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 158                 return 0;
 159
 160         if (trans)
 161                 transid = trans->transid;
 162         else
 163                 transid = BTRFS_I(inode)->root->last_trans;
 164
 165         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 166         if (!defrag)
 167                 return -ENOMEM;
 168
 169         defrag->ino = btrfs_ino(inode);
 170         defrag->transid = transid;
 171         defrag->root = root->root_key.objectid;
 172
 173         spin_lock(&root->fs_info->defrag_inodes_lock);
 174         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 175                 /*
 176                  * If we set IN_DEFRAG flag and evict the inode from memory,
 177                  * and then re-read this inode, this new inode doesn't have
 178                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 179                  */
 180                 ret = __btrfs_add_inode_defrag(inode, defrag);
 181                 if (ret)
 182                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 183         } else {
 184                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         }
 186         spin_unlock(&root->fs_info->defrag_inodes_lock);
 187         return 0;
 188 }
 189
 190 /*
 191  * Requeue the defrag object. If there is a defrag object that points to
 192  * the same inode in the tree, we will merge them together (by
 193  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 194  */
 195 static void btrfs_requeue_inode_defrag(struct inode *inode,
 196                                        struct inode_defrag *defrag)
 197 {
 198         struct btrfs_root *root = BTRFS_I(inode)->root;
 199         int ret;
 200
 201         if (!__need_auto_defrag(root))
 202                 goto out;
 203
 204         /*
 205          * Here we don't check the IN_DEFRAG flag, because we need merge
 206          * them together.
 207          */
 208         spin_lock(&root->fs_info->defrag_inodes_lock);
 209         ret = __btrfs_add_inode_defrag(inode, defrag);
 210         spin_unlock(&root->fs_info->defrag_inodes_lock);
 211         if (ret)
 212                 goto out;
 213         return;
 214 out:
 215         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 216 }
 217
 218 /*
 219  * pick the defragable inode that we want, if it doesn't exist, we will get
 220  * the next one.
 221  */
 222 static struct inode_defrag *
 223 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 224 {
 225         struct inode_defrag *entry = NULL;
 226         struct inode_defrag tmp;
 227         struct rb_node *p;
 228         struct rb_node *parent = NULL;
 229         int ret;
 230
 231         tmp.ino = ino;
 232         tmp.root = root;
 233
 234         spin_lock(&fs_info->defrag_inodes_lock);
 235         p = fs_info->defrag_inodes.rb_node;
 236         while (p) {
 237                 parent = p;
 238                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 239
 240                 ret = __compare_inode_defrag(&tmp, entry);
 241                 if (ret < 0)
 242                         p = parent->rb_left;
 243                 else if (ret > 0)
 244                         p = parent->rb_right;
 245                 else
 246                         goto out;
 247         }
 248
 249         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 250                 parent = rb_next(parent);
 251                 if (parent)
 252                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 253                 else
 254                         entry = NULL;
 255         }
 256 out:
 257         if (entry)
 258                 rb_erase(parent, &fs_info->defrag_inodes);
 259         spin_unlock(&fs_info->defrag_inodes_lock);
 260         return entry;
 261 }
 262
 263 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 264 {
 265         struct inode_defrag *defrag;
 266         struct rb_node *node;
 267
 268         spin_lock(&fs_info->defrag_inodes_lock);
 269         node = rb_first(&fs_info->defrag_inodes);
 270         while (node) {
 271                 rb_erase(node, &fs_info->defrag_inodes);
 272                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 273                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 274
 275                 if (need_resched()) {
 276                         spin_unlock(&fs_info->defrag_inodes_lock);
 277                         cond_resched();
 278                         spin_lock(&fs_info->defrag_inodes_lock);
 279                 }
 280
 281                 node = rb_first(&fs_info->defrag_inodes);
 282         }
 283         spin_unlock(&fs_info->defrag_inodes_lock);
 284 }
 285
 286 #define BTRFS_DEFRAG_BATCH      1024
 287
 288 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 289                                     struct inode_defrag *defrag)
 290 {
 291         struct btrfs_root *inode_root;
 292         struct inode *inode;
 293         struct btrfs_key key;
 294         struct btrfs_ioctl_defrag_range_args range;
 295         int num_defrag;
 296         int index;
 297         int ret;
 298
 299         /* get the inode */
 300         key.objectid = defrag->root;
 301         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 302         key.offset = (u64)-1;
 303
 304         index = srcu_read_lock(&fs_info->subvol_srcu);
 305
 306         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 307         if (IS_ERR(inode_root)) {
 308                 ret = PTR_ERR(inode_root);
 309                 goto cleanup;
 310         }
 311
 312         key.objectid = defrag->ino;
 313         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 314         key.offset = 0;
 315         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 316         if (IS_ERR(inode)) {
 317                 ret = PTR_ERR(inode);
 318                 goto cleanup;
 319         }
 320         srcu_read_unlock(&fs_info->subvol_srcu, index);
 321
 322         /* do a chunk of defrag */
 323         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 324         memset(&range, 0, sizeof(range));
 325         range.len = (u64)-1;
 326         range.start = defrag->last_offset;
 327
 328         sb_start_write(fs_info->sb);
 329         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 330                                        BTRFS_DEFRAG_BATCH);
 331         sb_end_write(fs_info->sb);
 332         /*
 333          * if we filled the whole defrag batch, there
 334          * must be more work to do.  Queue this defrag
 335          * again
 336          */
 337         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 338                 defrag->last_offset = range.start;
 339                 btrfs_requeue_inode_defrag(inode, defrag);
 340         } else if (defrag->last_offset && !defrag->cycled) {
 341                 /*
 342                  * we didn't fill our defrag batch, but
 343                  * we didn't start at zero.  Make sure we loop
 344                  * around to the start of the file.
 345                  */
 346                 defrag->last_offset = 0;
 347                 defrag->cycled = 1;
 348                 btrfs_requeue_inode_defrag(inode, defrag);
 349         } else {
 350                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 351         }
 352
 353         iput(inode);
 354         return 0;
 355 cleanup:
 356         srcu_read_unlock(&fs_info->subvol_srcu, index);
 357         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 358         return ret;
 359 }
 360
 361 /*
 362  * run through the list of inodes in the FS that need
 363  * defragging
 364  */
 365 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 366 {
 367         struct inode_defrag *defrag;
 368         u64 first_ino = 0;
 369         u64 root_objectid = 0;
 370
 371         atomic_inc(&fs_info->defrag_running);
 372         while (1) {
 373                 /* Pause the auto defragger. */
 374                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 375                              &fs_info->fs_state))
 376                         break;
 377
 378                 if (!__need_auto_defrag(fs_info->tree_root))
 379                         break;
 380
 381                 /* find an inode to defrag */
 382                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 383                                                  first_ino);
 384                 if (!defrag) {
 385                         if (root_objectid || first_ino) {
 386                                 root_objectid = 0;
 387                                 first_ino = 0;
 388                                 continue;
 389                         } else {
 390                                 break;
 391                         }
 392                 }
 393
 394                 first_ino = defrag->ino + 1;
 395                 root_objectid = defrag->root;
 396
 397                 __btrfs_run_defrag_inode(fs_info, defrag);
 398         }
 399         atomic_dec(&fs_info->defrag_running);
 400
 401         /*
 402          * during unmount, we use the transaction_wait queue to
 403          * wait for the defragger to stop
 404          */
 405         wake_up(&fs_info->transaction_wait);
 406         return 0;
 407 }
 408
 409 /* simple helper to fault in pages and copy.  This should go away
 410  * and be replaced with calls into generic code.
 411  */
 412 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 413                                          size_t write_bytes,
 414                                          struct page **prepared_pages,
 415                                          struct iov_iter *i)
 416 {
 417         size_t copied = 0;
 418         size_t total_copied = 0;
 419         int pg = 0;
 420         int offset = pos & (PAGE_CACHE_SIZE - 1);
 421
 422         while (write_bytes > 0) {
 423                 size_t count = min_t(size_t,
 424                                      PAGE_CACHE_SIZE - offset, write_bytes);
 425                 struct page *page = prepared_pages[pg];
 426                 /*
 427                  * Copy data from userspace to the current page
 428                  */
 429                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 430
 431                 /* Flush processor's dcache for this page */
 432                 flush_dcache_page(page);
 433
 434                 /*
 435                  * if we get a partial write, we can end up with
 436                  * partially up to date pages.  These add
 437                  * a lot of complexity, so make sure they don't
 438                  * happen by forcing this copy to be retried.
 439                  *
 440                  * The rest of the btrfs_file_write code will fall
 441                  * back to page at a time copies after we return 0.
 442                  */
 443                 if (!PageUptodate(page) && copied < count)
 444                         copied = 0;
 445
 446                 iov_iter_advance(i, copied);
 447                 write_bytes -= copied;
 448                 total_copied += copied;
 449
 450                 /* Return to btrfs_file_aio_write to fault page */
 451                 if (unlikely(copied == 0))
 452                         break;
 453
 454                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 455                         offset += copied;
 456                 } else {
 457                         pg++;
 458                         offset = 0;
 459                 }
 460         }
 461         return total_copied;
 462 }
 463
 464 /*
 465  * unlocks pages after btrfs_file_write is done with them
 466  */
 467 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 468 {
 469         size_t i;
 470         for (i = 0; i < num_pages; i++) {
 471                 /* page checked is some magic around finding pages that
 472                  * have been modified without going through btrfs_set_page_dirty
 473                  * clear it here
 474                  */
 475                 ClearPageChecked(pages[i]);
 476                 unlock_page(pages[i]);
 477                 mark_page_accessed(pages[i]);
 478                 page_cache_release(pages[i]);
 479         }
 480 }
 481
 482 /*
 483  * after copy_from_user, pages need to be dirtied and we need to make
 484  * sure holes are created between the current EOF and the start of
 485  * any next extents (if required).
 486  *
 487  * this also makes the decision about creating an inline extent vs
 488  * doing real data extents, marking pages dirty and delalloc as required.
 489  */
 490 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 491                              struct page **pages, size_t num_pages,
 492                              loff_t pos, size_t write_bytes,
 493                              struct extent_state **cached)
 494 {
 495         int err = 0;
 496         int i;
 497         u64 num_bytes;
 498         u64 start_pos;
 499         u64 end_of_last_block;
 500         u64 end_pos = pos + write_bytes;
 501         loff_t isize = i_size_read(inode);
 502
 503         start_pos = pos & ~((u64)root->sectorsize - 1);
 504         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 505
 506         end_of_last_block = start_pos + num_bytes - 1;
 507         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 508                                         cached);
 509         if (err)
 510                 return err;
 511
 512         for (i = 0; i < num_pages; i++) {
 513                 struct page *p = pages[i];
 514                 SetPageUptodate(p);
 515                 ClearPageChecked(p);
 516                 set_page_dirty(p);
 517         }
 518
 519         /*
 520          * we've only changed i_size in ram, and we haven't updated
 521          * the disk i_size.  There is no need to log the inode
 522          * at this time.
 523          */
 524         if (end_pos > isize)
 525                 i_size_write(inode, end_pos);
 526         return 0;
 527 }
 528
 529 /*
 530  * this drops all the extents in the cache that intersect the range
 531  * [start, end].  Existing extents are split as required.
 532  */
 533 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 534                              int skip_pinned)
 535 {
 536         struct extent_map *em;
 537         struct extent_map *split = NULL;
 538         struct extent_map *split2 = NULL;
 539         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 540         u64 len = end - start + 1;
 541         u64 gen;
 542         int ret;
 543         int testend = 1;
 544         unsigned long flags;
 545         int compressed = 0;
 546         bool modified;
 547
 548         WARN_ON(end < start);
 549         if (end == (u64)-1) {
 550                 len = (u64)-1;
 551                 testend = 0;
 552         }
 553         while (1) {
 554                 int no_splits = 0;
 555
 556                 modified = false;
 557                 if (!split)
 558                         split = alloc_extent_map();
 559                 if (!split2)
 560                         split2 = alloc_extent_map();
 561                 if (!split || !split2)
 562                         no_splits = 1;
 563
 564                 write_lock(&em_tree->lock);
 565                 em = lookup_extent_mapping(em_tree, start, len);
 566                 if (!em) {
 567                         write_unlock(&em_tree->lock);
 568                         break;
 569                 }
 570                 flags = em->flags;
 571                 gen = em->generation;
 572                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 573                         if (testend && em->start + em->len >= start + len) {
 574                                 free_extent_map(em);
 575                                 write_unlock(&em_tree->lock);
 576                                 break;
 577                         }
 578                         start = em->start + em->len;
 579                         if (testend)
 580                                 len = start + len - (em->start + em->len);
 581                         free_extent_map(em);
 582                         write_unlock(&em_tree->lock);
 583                         continue;
 584                 }
 585                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 586                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 587                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 588                 modified = !list_empty(&em->list);
 589                 if (no_splits)
 590                         goto next;
 591
 592                 if (em->start < start) {
 593                         split->start = em->start;
 594                         split->len = start - em->start;
 595
 596                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 597                                 split->orig_start = em->orig_start;
 598                                 split->block_start = em->block_start;
 599
 600                                 if (compressed)
 601                                         split->block_len = em->block_len;
 602                                 else
 603                                         split->block_len = split->len;
 604                                 split->orig_block_len = max(split->block_len,
 605                                                 em->orig_block_len);
 606                                 split->ram_bytes = em->ram_bytes;
 607                         } else {
 608                                 split->orig_start = split->start;
 609                                 split->block_len = 0;
 610                                 split->block_start = em->block_start;
 611                                 split->orig_block_len = 0;
 612                                 split->ram_bytes = split->len;
 613                         }
 614
 615                         split->generation = gen;
 616                         split->bdev = em->bdev;
 617                         split->flags = flags;
 618                         split->compress_type = em->compress_type;
 619                         replace_extent_mapping(em_tree, em, split, modified);
 620                         free_extent_map(split);
 621                         split = split2;
 622                         split2 = NULL;
 623                 }
 624                 if (testend && em->start + em->len > start + len) {
 625                         u64 diff = start + len - em->start;
 626
 627                         split->start = start + len;
 628                         split->len = em->start + em->len - (start + len);
 629                         split->bdev = em->bdev;
 630                         split->flags = flags;
 631                         split->compress_type = em->compress_type;
 632                         split->generation = gen;
 633
 634                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 635                                 split->orig_block_len = max(em->block_len,
 636                                                     em->orig_block_len);
 637
 638                                 split->ram_bytes = em->ram_bytes;
 639                                 if (compressed) {
 640                                         split->block_len = em->block_len;
 641                                         split->block_start = em->block_start;
 642                                         split->orig_start = em->orig_start;
 643                                 } else {
 644                                         split->block_len = split->len;
 645                                         split->block_start = em->block_start
 646                                                 + diff;
 647                                         split->orig_start = em->orig_start;
 648                                 }
 649                         } else {
 650                                 split->ram_bytes = split->len;
 651                                 split->orig_start = split->start;
 652                                 split->block_len = 0;
 653                                 split->block_start = em->block_start;
 654                                 split->orig_block_len = 0;
 655                         }
 656
 657                         if (extent_map_in_tree(em)) {
 658                                 replace_extent_mapping(em_tree, em, split,
 659                                                        modified);
 660                         } else {
 661                                 ret = add_extent_mapping(em_tree, split,
 662                                                          modified);
 663                                 ASSERT(ret == 0); /* Logic error */
 664                         }
 665                         free_extent_map(split);
 666                         split = NULL;
 667                 }
 668 next:
 669                 if (extent_map_in_tree(em))
 670                         remove_extent_mapping(em_tree, em);
 671                 write_unlock(&em_tree->lock);
 672
 673                 /* once for us */
 674                 free_extent_map(em);
 675                 /* once for the tree*/
 676                 free_extent_map(em);
 677         }
 678         if (split)
 679                 free_extent_map(split);
 680         if (split2)
 681                 free_extent_map(split2);
 682 }
 683
 684 /*
 685  * this is very complex, but the basic idea is to drop all extents
 686  * in the range start - end.  hint_block is filled in with a block number
 687  * that would be a good hint to the block allocator for this file.
 688  *
 689  * If an extent intersects the range but is not entirely inside the range
 690  * it is either truncated or split.  Anything entirely inside the range
 691  * is deleted from the tree.
 692  */
 693 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 694                          struct btrfs_root *root, struct inode *inode,
 695                          struct btrfs_path *path, u64 start, u64 end,
 696                          u64 *drop_end, int drop_cache,
 697                          int replace_extent,
 698                          u32 extent_item_size,
 699                          int *key_inserted)
 700 {
 701         struct extent_buffer *leaf;
 702         struct btrfs_file_extent_item *fi;
 703         struct btrfs_key key;
 704         struct btrfs_key new_key;
 705         u64 ino = btrfs_ino(inode);
 706         u64 search_start = start;
 707         u64 disk_bytenr = 0;
 708         u64 num_bytes = 0;
 709         u64 extent_offset = 0;
 710         u64 extent_end = 0;
 711         int del_nr = 0;
 712         int del_slot = 0;
 713         int extent_type;
 714         int recow;
 715         int ret;
 716         int modify_tree = -1;
 717         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 718         int found = 0;
 719         int leafs_visited = 0;
 720
 721         if (drop_cache)
 722                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 723
 724         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 725                 modify_tree = 0;
 726
 727         while (1) {
 728                 recow = 0;
 729                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 730                                                search_start, modify_tree);
 731                 if (ret < 0)
 732                         break;
 733                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 734                         leaf = path->nodes[0];
 735                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 736                         if (key.objectid == ino &&
 737                             key.type == BTRFS_EXTENT_DATA_KEY)
 738                                 path->slots[0]--;
 739                 }
 740                 ret = 0;
 741                 leafs_visited++;
 742 next_slot:
 743                 leaf = path->nodes[0];
 744                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 745                         BUG_ON(del_nr > 0);
 746                         ret = btrfs_next_leaf(root, path);
 747                         if (ret < 0)
 748                                 break;
 749                         if (ret > 0) {
 750                                 ret = 0;
 751                                 break;
 752                         }
 753                         leafs_visited++;
 754                         leaf = path->nodes[0];
 755                         recow = 1;
 756                 }
 757
 758                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 759                 if (key.objectid > ino ||
 760                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 761                         break;
 762
 763                 fi = btrfs_item_ptr(leaf, path->slots[0],
 764                                     struct btrfs_file_extent_item);
 765                 extent_type = btrfs_file_extent_type(leaf, fi);
 766
 767                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 768                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 769                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 770                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 771                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 772                         extent_end = key.offset +
 773                                 btrfs_file_extent_num_bytes(leaf, fi);
 774                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 775                         extent_end = key.offset +
 776                                 btrfs_file_extent_inline_len(leaf,
 777                                                      path->slots[0], fi);
 778                 } else {
 779                         WARN_ON(1);
 780                         extent_end = search_start;
 781                 }
 782
 783                 /*
 784                  * Don't skip extent items representing 0 byte lengths. They
 785                  * used to be created (bug) if while punching holes we hit
 786                  * -ENOSPC condition. So if we find one here, just ensure we
 787                  * delete it, otherwise we would insert a new file extent item
 788                  * with the same key (offset) as that 0 bytes length file
 789                  * extent item in the call to setup_items_for_insert() later
 790                  * in this function.
 791                  */
 792                 if (extent_end == key.offset && extent_end >= search_start)
 793                         goto delete_extent_item;
 794
 795                 if (extent_end <= search_start) {
 796                         path->slots[0]++;
 797                         goto next_slot;
 798                 }
 799
 800                 found = 1;
 801                 search_start = max(key.offset, start);
 802                 if (recow || !modify_tree) {
 803                         modify_tree = -1;
 804                         btrfs_release_path(path);
 805                         continue;
 806                 }
 807
 808                 /*
 809                  *     | - range to drop - |
 810                  *  | -------- extent -------- |
 811                  */
 812                 if (start > key.offset && end < extent_end) {
 813                         BUG_ON(del_nr > 0);
 814                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 815                                 ret = -EOPNOTSUPP;
 816                                 break;
 817                         }
 818
 819                         memcpy(&new_key, &key, sizeof(new_key));
 820                         new_key.offset = start;
 821                         ret = btrfs_duplicate_item(trans, root, path,
 822                                                    &new_key);
 823                         if (ret == -EAGAIN) {
 824                                 btrfs_release_path(path);
 825                                 continue;
 826                         }
 827                         if (ret < 0)
 828                                 break;
 829
 830                         leaf = path->nodes[0];
 831                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 832                                             struct btrfs_file_extent_item);
 833                         btrfs_set_file_extent_num_bytes(leaf, fi,
 834                                                         start - key.offset);
 835
 836                         fi = btrfs_item_ptr(leaf, path->slots[0],
 837                                             struct btrfs_file_extent_item);
 838
 839                         extent_offset += start - key.offset;
 840                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 841                         btrfs_set_file_extent_num_bytes(leaf, fi,
 842                                                         extent_end - start);
 843                         btrfs_mark_buffer_dirty(leaf);
 844
 845                         if (update_refs && disk_bytenr > 0) {
 846                                 ret = btrfs_inc_extent_ref(trans, root,
 847                                                 disk_bytenr, num_bytes, 0,
 848                                                 root->root_key.objectid,
 849                                                 new_key.objectid,
 850                                                 start - extent_offset, 0);
 851                                 BUG_ON(ret); /* -ENOMEM */
 852                         }
 853                         key.offset = start;
 854                 }
 855                 /*
 856                  *  | ---- range to drop ----- |
 857                  *      | -------- extent -------- |
 858                  */
 859                 if (start <= key.offset && end < extent_end) {
 860                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 861                                 ret = -EOPNOTSUPP;
 862                                 break;
 863                         }
 864
 865                         memcpy(&new_key, &key, sizeof(new_key));
 866                         new_key.offset = end;
 867                         btrfs_set_item_key_safe(root, path, &new_key);
 868
 869                         extent_offset += end - key.offset;
 870                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 871                         btrfs_set_file_extent_num_bytes(leaf, fi,
 872                                                         extent_end - end);
 873                         btrfs_mark_buffer_dirty(leaf);
 874                         if (update_refs && disk_bytenr > 0)
 875                                 inode_sub_bytes(inode, end - key.offset);
 876                         break;
 877                 }
 878
 879                 search_start = extent_end;
 880                 /*
 881                  *       | ---- range to drop ----- |
 882                  *  | -------- extent -------- |
 883                  */
 884                 if (start > key.offset && end >= extent_end) {
 885                         BUG_ON(del_nr > 0);
 886                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 887                                 ret = -EOPNOTSUPP;
 888                                 break;
 889                         }
 890
 891                         btrfs_set_file_extent_num_bytes(leaf, fi,
 892                                                         start - key.offset);
 893                         btrfs_mark_buffer_dirty(leaf);
 894                         if (update_refs && disk_bytenr > 0)
 895                                 inode_sub_bytes(inode, extent_end - start);
 896                         if (end == extent_end)
 897                                 break;
 898
 899                         path->slots[0]++;
 900                         goto next_slot;
 901                 }
 902
 903                 /*
 904                  *  | ---- range to drop ----- |
 905                  *    | ------ extent ------ |
 906                  */
 907                 if (start <= key.offset && end >= extent_end) {
 908 delete_extent_item:
 909                         if (del_nr == 0) {
 910                                 del_slot = path->slots[0];
 911                                 del_nr = 1;
 912                         } else {
 913                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 914                                 del_nr++;
 915                         }
 916
 917                         if (update_refs &&
 918                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 919                                 inode_sub_bytes(inode,
 920                                                 extent_end - key.offset);
 921                                 extent_end = ALIGN(extent_end,
 922                                                    root->sectorsize);
 923                         } else if (update_refs && disk_bytenr > 0) {
 924                                 ret = btrfs_free_extent(trans, root,
 925                                                 disk_bytenr, num_bytes, 0,
 926                                                 root->root_key.objectid,
 927                                                 key.objectid, key.offset -
 928                                                 extent_offset, 0);
 929                                 BUG_ON(ret); /* -ENOMEM */
 930                                 inode_sub_bytes(inode,
 931                                                 extent_end - key.offset);
 932                         }
 933
 934                         if (end == extent_end)
 935                                 break;
 936
 937                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 938                                 path->slots[0]++;
 939                                 goto next_slot;
 940                         }
 941
 942                         ret = btrfs_del_items(trans, root, path, del_slot,
 943                                               del_nr);
 944                         if (ret) {
 945                                 btrfs_abort_transaction(trans, root, ret);
 946                                 break;
 947                         }
 948
 949                         del_nr = 0;
 950                         del_slot = 0;
 951
 952                         btrfs_release_path(path);
 953                         continue;
 954                 }
 955
 956                 BUG_ON(1);
 957         }
 958
 959         if (!ret && del_nr > 0) {
 960                 /*
 961                  * Set path->slots[0] to first slot, so that after the delete
 962                  * if items are move off from our leaf to its immediate left or
 963                  * right neighbor leafs, we end up with a correct and adjusted
 964                  * path->slots[0] for our insertion (if replace_extent != 0).
 965                  */
 966                 path->slots[0] = del_slot;
 967                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 968                 if (ret)
 969                         btrfs_abort_transaction(trans, root, ret);
 970         }
 971
 972         leaf = path->nodes[0];
 973         /*
 974          * If btrfs_del_items() was called, it might have deleted a leaf, in
 975          * which case it unlocked our path, so check path->locks[0] matches a
 976          * write lock.
 977          */
 978         if (!ret && replace_extent && leafs_visited == 1 &&
 979             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 980              path->locks[0] == BTRFS_WRITE_LOCK) &&
 981             btrfs_leaf_free_space(root, leaf) >=
 982             sizeof(struct btrfs_item) + extent_item_size) {
 983
 984                 key.objectid = ino;
 985                 key.type = BTRFS_EXTENT_DATA_KEY;
 986                 key.offset = start;
 987                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 988                         struct btrfs_key slot_key;
 989
 990                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 991                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
 992                                 path->slots[0]++;
 993                 }
 994                 setup_items_for_insert(root, path, &key,
 995                                        &extent_item_size,
 996                                        extent_item_size,
 997                                        sizeof(struct btrfs_item) +
 998                                        extent_item_size, 1);
 999                 *key_inserted = 1;
1000         }
1001
1002         if (!replace_extent || !(*key_inserted))
1003                 btrfs_release_path(path);
1004         if (drop_end)
1005                 *drop_end = found ? min(end, extent_end) : end;
1006         return ret;
1007 }
1008
1009 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1010                        struct btrfs_root *root, struct inode *inode, u64 start,
1011                        u64 end, int drop_cache)
1012 {
1013         struct btrfs_path *path;
1014         int ret;
1015
1016         path = btrfs_alloc_path();
1017         if (!path)
1018                 return -ENOMEM;
1019         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1020                                    drop_cache, 0, 0, NULL);
1021         btrfs_free_path(path);
1022         return ret;
1023 }
1024
1025 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1026                             u64 objectid, u64 bytenr, u64 orig_offset,
1027                             u64 *start, u64 *end)
1028 {
1029         struct btrfs_file_extent_item *fi;
1030         struct btrfs_key key;
1031         u64 extent_end;
1032
1033         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1034                 return 0;
1035
1036         btrfs_item_key_to_cpu(leaf, &key, slot);
1037         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1038                 return 0;
1039
1040         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1041         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1042             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1043             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1044             btrfs_file_extent_compression(leaf, fi) ||
1045             btrfs_file_extent_encryption(leaf, fi) ||
1046             btrfs_file_extent_other_encoding(leaf, fi))
1047                 return 0;
1048
1049         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1050         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1051                 return 0;
1052
1053         *start = key.offset;
1054         *end = extent_end;
1055         return 1;
1056 }
1057
1058 /*
1059  * Mark extent in the range start - end as written.
1060  *
1061  * This changes extent type from 'pre-allocated' to 'regular'. If only
1062  * part of extent is marked as written, the extent will be split into
1063  * two or three.
1064  */
1065 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1066                               struct inode *inode, u64 start, u64 end)
1067 {
1068         struct btrfs_root *root = BTRFS_I(inode)->root;
1069         struct extent_buffer *leaf;
1070         struct btrfs_path *path;
1071         struct btrfs_file_extent_item *fi;
1072         struct btrfs_key key;
1073         struct btrfs_key new_key;
1074         u64 bytenr;
1075         u64 num_bytes;
1076         u64 extent_end;
1077         u64 orig_offset;
1078         u64 other_start;
1079         u64 other_end;
1080         u64 split;
1081         int del_nr = 0;
1082         int del_slot = 0;
1083         int recow;
1084         int ret;
1085         u64 ino = btrfs_ino(inode);
1086
1087         path = btrfs_alloc_path();
1088         if (!path)
1089                 return -ENOMEM;
1090 again:
1091         recow = 0;
1092         split = start;
1093         key.objectid = ino;
1094         key.type = BTRFS_EXTENT_DATA_KEY;
1095         key.offset = split;
1096
1097         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1098         if (ret < 0)
1099                 goto out;
1100         if (ret > 0 && path->slots[0] > 0)
1101                 path->slots[0]--;
1102
1103         leaf = path->nodes[0];
1104         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1105         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1106         fi = btrfs_item_ptr(leaf, path->slots[0],
1107                             struct btrfs_file_extent_item);
1108         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1109                BTRFS_FILE_EXTENT_PREALLOC);
1110         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1111         BUG_ON(key.offset > start || extent_end < end);
1112
1113         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1114         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1115         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1116         memcpy(&new_key, &key, sizeof(new_key));
1117
1118         if (start == key.offset && end < extent_end) {
1119                 other_start = 0;
1120                 other_end = start;
1121                 if (extent_mergeable(leaf, path->slots[0] - 1,
1122                                      ino, bytenr, orig_offset,
1123                                      &other_start, &other_end)) {
1124                         new_key.offset = end;
1125                         btrfs_set_item_key_safe(root, path, &new_key);
1126                         fi = btrfs_item_ptr(leaf, path->slots[0],
1127                                             struct btrfs_file_extent_item);
1128                         btrfs_set_file_extent_generation(leaf, fi,
1129                                                          trans->transid);
1130                         btrfs_set_file_extent_num_bytes(leaf, fi,
1131                                                         extent_end - end);
1132                         btrfs_set_file_extent_offset(leaf, fi,
1133                                                      end - orig_offset);
1134                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1135                                             struct btrfs_file_extent_item);
1136                         btrfs_set_file_extent_generation(leaf, fi,
1137                                                          trans->transid);
1138                         btrfs_set_file_extent_num_bytes(leaf, fi,
1139                                                         end - other_start);
1140                         btrfs_mark_buffer_dirty(leaf);
1141                         goto out;
1142                 }
1143         }
1144
1145         if (start > key.offset && end == extent_end) {
1146                 other_start = end;
1147                 other_end = 0;
1148                 if (extent_mergeable(leaf, path->slots[0] + 1,
1149                                      ino, bytenr, orig_offset,
1150                                      &other_start, &other_end)) {
1151                         fi = btrfs_item_ptr(leaf, path->slots[0],
1152                                             struct btrfs_file_extent_item);
1153                         btrfs_set_file_extent_num_bytes(leaf, fi,
1154                                                         start - key.offset);
1155                         btrfs_set_file_extent_generation(leaf, fi,
1156                                                          trans->transid);
1157                         path->slots[0]++;
1158                         new_key.offset = start;
1159                         btrfs_set_item_key_safe(root, path, &new_key);
1160
1161                         fi = btrfs_item_ptr(leaf, path->slots[0],
1162                                             struct btrfs_file_extent_item);
1163                         btrfs_set_file_extent_generation(leaf, fi,
1164                                                          trans->transid);
1165                         btrfs_set_file_extent_num_bytes(leaf, fi,
1166                                                         other_end - start);
1167                         btrfs_set_file_extent_offset(leaf, fi,
1168                                                      start - orig_offset);
1169                         btrfs_mark_buffer_dirty(leaf);
1170                         goto out;
1171                 }
1172         }
1173
1174         while (start > key.offset || end < extent_end) {
1175                 if (key.offset == start)
1176                         split = end;
1177
1178                 new_key.offset = split;
1179                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1180                 if (ret == -EAGAIN) {
1181                         btrfs_release_path(path);
1182                         goto again;
1183                 }
1184                 if (ret < 0) {
1185                         btrfs_abort_transaction(trans, root, ret);
1186                         goto out;
1187                 }
1188
1189                 leaf = path->nodes[0];
1190                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1191                                     struct btrfs_file_extent_item);
1192                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1193                 btrfs_set_file_extent_num_bytes(leaf, fi,
1194                                                 split - key.offset);
1195
1196                 fi = btrfs_item_ptr(leaf, path->slots[0],
1197                                     struct btrfs_file_extent_item);
1198
1199                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1200                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1201                 btrfs_set_file_extent_num_bytes(leaf, fi,
1202                                                 extent_end - split);
1203                 btrfs_mark_buffer_dirty(leaf);
1204
1205                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1206                                            root->root_key.objectid,
1207                                            ino, orig_offset, 0);
1208                 BUG_ON(ret); /* -ENOMEM */
1209
1210                 if (split == start) {
1211                         key.offset = start;
1212                 } else {
1213                         BUG_ON(start != key.offset);
1214                         path->slots[0]--;
1215                         extent_end = end;
1216                 }
1217                 recow = 1;
1218         }
1219
1220         other_start = end;
1221         other_end = 0;
1222         if (extent_mergeable(leaf, path->slots[0] + 1,
1223                              ino, bytenr, orig_offset,
1224                              &other_start, &other_end)) {
1225                 if (recow) {
1226                         btrfs_release_path(path);
1227                         goto again;
1228                 }
1229                 extent_end = other_end;
1230                 del_slot = path->slots[0] + 1;
1231                 del_nr++;
1232                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1233                                         0, root->root_key.objectid,
1234                                         ino, orig_offset, 0);
1235                 BUG_ON(ret); /* -ENOMEM */
1236         }
1237         other_start = 0;
1238         other_end = start;
1239         if (extent_mergeable(leaf, path->slots[0] - 1,
1240                              ino, bytenr, orig_offset,
1241                              &other_start, &other_end)) {
1242                 if (recow) {
1243                         btrfs_release_path(path);
1244                         goto again;
1245                 }
1246                 key.offset = other_start;
1247                 del_slot = path->slots[0];
1248                 del_nr++;
1249                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1250                                         0, root->root_key.objectid,
1251                                         ino, orig_offset, 0);
1252                 BUG_ON(ret); /* -ENOMEM */
1253         }
1254         if (del_nr == 0) {
1255                 fi = btrfs_item_ptr(leaf, path->slots[0],
1256                            struct btrfs_file_extent_item);
1257                 btrfs_set_file_extent_type(leaf, fi,
1258                                            BTRFS_FILE_EXTENT_REG);
1259                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1260                 btrfs_mark_buffer_dirty(leaf);
1261         } else {
1262                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1263                            struct btrfs_file_extent_item);
1264                 btrfs_set_file_extent_type(leaf, fi,
1265                                            BTRFS_FILE_EXTENT_REG);
1266                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1267                 btrfs_set_file_extent_num_bytes(leaf, fi,
1268                                                 extent_end - key.offset);
1269                 btrfs_mark_buffer_dirty(leaf);
1270
1271                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1272                 if (ret < 0) {
1273                         btrfs_abort_transaction(trans, root, ret);
1274                         goto out;
1275                 }
1276         }
1277 out:
1278         btrfs_free_path(path);
1279         return 0;
1280 }
1281
1282 /*
1283  * on error we return an unlocked page and the error value
1284  * on success we return a locked page and 0
1285  */
1286 static int prepare_uptodate_page(struct page *page, u64 pos,
1287                                  bool force_uptodate)
1288 {
1289         int ret = 0;
1290
1291         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1292             !PageUptodate(page)) {
1293                 ret = btrfs_readpage(NULL, page);
1294                 if (ret)
1295                         return ret;
1296                 lock_page(page);
1297                 if (!PageUptodate(page)) {
1298                         unlock_page(page);
1299                         return -EIO;
1300                 }
1301         }
1302         return 0;
1303 }
1304
1305 /*
1306  * this just gets pages into the page cache and locks them down.
1307  */
1308 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1309                                   size_t num_pages, loff_t pos,
1310                                   size_t write_bytes, bool force_uptodate)
1311 {
1312         int i;
1313         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1314         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1315         int err = 0;
1316         int faili;
1317
1318         for (i = 0; i < num_pages; i++) {
1319                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1320                                                mask | __GFP_WRITE);
1321                 if (!pages[i]) {
1322                         faili = i - 1;
1323                         err = -ENOMEM;
1324                         goto fail;
1325                 }
1326
1327                 if (i == 0)
1328                         err = prepare_uptodate_page(pages[i], pos,
1329                                                     force_uptodate);
1330                 if (i == num_pages - 1)
1331                         err = prepare_uptodate_page(pages[i],
1332                                                     pos + write_bytes, false);
1333                 if (err) {
1334                         page_cache_release(pages[i]);
1335                         faili = i - 1;
1336                         goto fail;
1337                 }
1338                 wait_on_page_writeback(pages[i]);
1339         }
1340
1341         return 0;
1342 fail:
1343         while (faili >= 0) {
1344                 unlock_page(pages[faili]);
1345                 page_cache_release(pages[faili]);
1346                 faili--;
1347         }
1348         return err;
1349
1350 }
1351
1352 /*
1353  * This function locks the extent and properly waits for data=ordered extents
1354  * to finish before allowing the pages to be modified if need.
1355  *
1356  * The return value:
1357  * 1 - the extent is locked
1358  * 0 - the extent is not locked, and everything is OK
1359  * -EAGAIN - need re-prepare the pages
1360  * the other < 0 number - Something wrong happens
1361  */
1362 static noinline int
1363 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1364                                 size_t num_pages, loff_t pos,
1365                                 u64 *lockstart, u64 *lockend,
1366                                 struct extent_state **cached_state)
1367 {
1368         u64 start_pos;
1369         u64 last_pos;
1370         int i;
1371         int ret = 0;
1372
1373         start_pos = pos & ~((u64)PAGE_CACHE_SIZE - 1);
1374         last_pos = start_pos + ((u64)num_pages << PAGE_CACHE_SHIFT) - 1;
1375
1376         if (start_pos < inode->i_size) {
1377                 struct btrfs_ordered_extent *ordered;
1378                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1379                                  start_pos, last_pos, 0, cached_state);
1380                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1381                                                      last_pos - start_pos + 1);
1382                 if (ordered &&
1383                     ordered->file_offset + ordered->len > start_pos &&
1384                     ordered->file_offset <= last_pos) {
1385                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1386                                              start_pos, last_pos,
1387                                              cached_state, GFP_NOFS);
1388                         for (i = 0; i < num_pages; i++) {
1389                                 unlock_page(pages[i]);
1390                                 page_cache_release(pages[i]);
1391                         }
1392                         btrfs_start_ordered_extent(inode, ordered, 1);
1393                         btrfs_put_ordered_extent(ordered);
1394                         return -EAGAIN;
1395                 }
1396                 if (ordered)
1397                         btrfs_put_ordered_extent(ordered);
1398
1399                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1400                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1401                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1402                                   0, 0, cached_state, GFP_NOFS);
1403                 *lockstart = start_pos;
1404                 *lockend = last_pos;
1405                 ret = 1;
1406         }
1407
1408         for (i = 0; i < num_pages; i++) {
1409                 if (clear_page_dirty_for_io(pages[i]))
1410                         account_page_redirty(pages[i]);
1411                 set_page_extent_mapped(pages[i]);
1412                 WARN_ON(!PageLocked(pages[i]));
1413         }
1414
1415         return ret;
1416 }
1417
1418 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1419                                     size_t *write_bytes)
1420 {
1421         struct btrfs_root *root = BTRFS_I(inode)->root;
1422         struct btrfs_ordered_extent *ordered;
1423         u64 lockstart, lockend;
1424         u64 num_bytes;
1425         int ret;
1426
1427         ret = btrfs_start_nocow_write(root);
1428         if (!ret)
1429                 return -ENOSPC;
1430
1431         lockstart = round_down(pos, root->sectorsize);
1432         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1433
1434         while (1) {
1435                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1436                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1437                                                      lockend - lockstart + 1);
1438                 if (!ordered) {
1439                         break;
1440                 }
1441                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1442                 btrfs_start_ordered_extent(inode, ordered, 1);
1443                 btrfs_put_ordered_extent(ordered);
1444         }
1445
1446         num_bytes = lockend - lockstart + 1;
1447         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1448         if (ret <= 0) {
1449                 ret = 0;
1450                 btrfs_end_nocow_write(root);
1451         } else {
1452                 *write_bytes = min_t(size_t, *write_bytes ,
1453                                      num_bytes - pos + lockstart);
1454         }
1455
1456         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1457
1458         return ret;
1459 }
1460
1461 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1462                                                struct iov_iter *i,
1463                                                loff_t pos)
1464 {
1465         struct inode *inode = file_inode(file);
1466         struct btrfs_root *root = BTRFS_I(inode)->root;
1467         struct page **pages = NULL;
1468         struct extent_state *cached_state = NULL;
1469         u64 release_bytes = 0;
1470         u64 lockstart;
1471         u64 lockend;
1472         unsigned long first_index;
1473         size_t num_written = 0;
1474         int nrptrs;
1475         int ret = 0;
1476         bool only_release_metadata = false;
1477         bool force_page_uptodate = false;
1478         bool need_unlock;
1479
1480         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1481                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1482                      (sizeof(struct page *)));
1483         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1484         nrptrs = max(nrptrs, 8);
1485         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1486         if (!pages)
1487                 return -ENOMEM;
1488
1489         first_index = pos >> PAGE_CACHE_SHIFT;
1490
1491         while (iov_iter_count(i) > 0) {
1492                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1493                 size_t write_bytes = min(iov_iter_count(i),
1494                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1495                                          offset);
1496                 size_t num_pages = (write_bytes + offset +
1497                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1498                 size_t reserve_bytes;
1499                 size_t dirty_pages;
1500                 size_t copied;
1501
1502                 WARN_ON(num_pages > nrptrs);
1503
1504                 /*
1505                  * Fault pages before locking them in prepare_pages
1506                  * to avoid recursive lock
1507                  */
1508                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1509                         ret = -EFAULT;
1510                         break;
1511                 }
1512
1513                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1514                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1515                 if (ret == -ENOSPC &&
1516                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1517                                               BTRFS_INODE_PREALLOC))) {
1518                         ret = check_can_nocow(inode, pos, &write_bytes);
1519                         if (ret > 0) {
1520                                 only_release_metadata = true;
1521                                 /*
1522                                  * our prealloc extent may be smaller than
1523                                  * write_bytes, so scale down.
1524                                  */
1525                                 num_pages = (write_bytes + offset +
1526                                              PAGE_CACHE_SIZE - 1) >>
1527                                         PAGE_CACHE_SHIFT;
1528                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1529                                 ret = 0;
1530                         } else {
1531                                 ret = -ENOSPC;
1532                         }
1533                 }
1534
1535                 if (ret)
1536                         break;
1537
1538                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1539                 if (ret) {
1540                         if (!only_release_metadata)
1541                                 btrfs_free_reserved_data_space(inode,
1542                                                                reserve_bytes);
1543                         else
1544                                 btrfs_end_nocow_write(root);
1545                         break;
1546                 }
1547
1548                 release_bytes = reserve_bytes;
1549                 need_unlock = false;
1550 again:
1551                 /*
1552                  * This is going to setup the pages array with the number of
1553                  * pages we want, so we don't really need to worry about the
1554                  * contents of pages from loop to loop
1555                  */
1556                 ret = prepare_pages(inode, pages, num_pages,
1557                                     pos, write_bytes,
1558                                     force_page_uptodate);
1559                 if (ret)
1560                         break;
1561
1562                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1563                                                       pos, &lockstart, &lockend,
1564                                                       &cached_state);
1565                 if (ret < 0) {
1566                         if (ret == -EAGAIN)
1567                                 goto again;
1568                         break;
1569                 } else if (ret > 0) {
1570                         need_unlock = true;
1571                         ret = 0;
1572                 }
1573
1574                 copied = btrfs_copy_from_user(pos, num_pages,
1575                                            write_bytes, pages, i);
1576
1577                 /*
1578                  * if we have trouble faulting in the pages, fall
1579                  * back to one page at a time
1580                  */
1581                 if (copied < write_bytes)
1582                         nrptrs = 1;
1583
1584                 if (copied == 0) {
1585                         force_page_uptodate = true;
1586                         dirty_pages = 0;
1587                 } else {
1588                         force_page_uptodate = false;
1589                         dirty_pages = (copied + offset +
1590                                        PAGE_CACHE_SIZE - 1) >>
1591                                        PAGE_CACHE_SHIFT;
1592                 }
1593
1594                 /*
1595                  * If we had a short copy we need to release the excess delaloc
1596                  * bytes we reserved.  We need to increment outstanding_extents
1597                  * because btrfs_delalloc_release_space will decrement it, but
1598                  * we still have an outstanding extent for the chunk we actually
1599                  * managed to copy.
1600                  */
1601                 if (num_pages > dirty_pages) {
1602                         release_bytes = (num_pages - dirty_pages) <<
1603                                 PAGE_CACHE_SHIFT;
1604                         if (copied > 0) {
1605                                 spin_lock(&BTRFS_I(inode)->lock);
1606                                 BTRFS_I(inode)->outstanding_extents++;
1607                                 spin_unlock(&BTRFS_I(inode)->lock);
1608                         }
1609                         if (only_release_metadata)
1610                                 btrfs_delalloc_release_metadata(inode,
1611                                                                 release_bytes);
1612                         else
1613                                 btrfs_delalloc_release_space(inode,
1614                                                              release_bytes);
1615                 }
1616
1617                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1618
1619                 if (copied > 0)
1620                         ret = btrfs_dirty_pages(root, inode, pages,
1621                                                 dirty_pages, pos, copied,
1622                                                 NULL);
1623                 if (need_unlock)
1624                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1625                                              lockstart, lockend, &cached_state,
1626                                              GFP_NOFS);
1627                 if (ret) {
1628                         btrfs_drop_pages(pages, num_pages);
1629                         break;
1630                 }
1631
1632                 release_bytes = 0;
1633                 if (only_release_metadata)
1634                         btrfs_end_nocow_write(root);
1635
1636                 if (only_release_metadata && copied > 0) {
1637                         u64 lockstart = round_down(pos, root->sectorsize);
1638                         u64 lockend = lockstart +
1639                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1640
1641                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1642                                        lockend, EXTENT_NORESERVE, NULL,
1643                                        NULL, GFP_NOFS);
1644                         only_release_metadata = false;
1645                 }
1646
1647                 btrfs_drop_pages(pages, num_pages);
1648
1649                 cond_resched();
1650
1651                 balance_dirty_pages_ratelimited(inode->i_mapping);
1652                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1653                         btrfs_btree_balance_dirty(root);
1654
1655                 pos += copied;
1656                 num_written += copied;
1657         }
1658
1659         kfree(pages);
1660
1661         if (release_bytes) {
1662                 if (only_release_metadata) {
1663                         btrfs_end_nocow_write(root);
1664                         btrfs_delalloc_release_metadata(inode, release_bytes);
1665                 } else {
1666                         btrfs_delalloc_release_space(inode, release_bytes);
1667                 }
1668         }
1669
1670         return num_written ? num_written : ret;
1671 }
1672
1673 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1674                                     const struct iovec *iov,
1675                                     unsigned long nr_segs, loff_t pos,
1676                                     size_t count, size_t ocount)
1677 {
1678         struct file *file = iocb->ki_filp;
1679         struct iov_iter i;
1680         ssize_t written;
1681         ssize_t written_buffered;
1682         loff_t endbyte;
1683         int err;
1684
1685         written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1686                                             count, ocount);
1687
1688         if (written < 0 || written == count)
1689                 return written;
1690
1691         pos += written;
1692         count -= written;
1693         iov_iter_init(&i, iov, nr_segs, count, written);
1694         written_buffered = __btrfs_buffered_write(file, &i, pos);
1695         if (written_buffered < 0) {
1696                 err = written_buffered;
1697                 goto out;
1698         }
1699         endbyte = pos + written_buffered - 1;
1700         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1701         if (err)
1702                 goto out;
1703         written += written_buffered;
1704         iocb->ki_pos = pos + written_buffered;
1705         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1706                                  endbyte >> PAGE_CACHE_SHIFT);
1707 out:
1708         return written ? written : err;
1709 }
1710
1711 static void update_time_for_write(struct inode *inode)
1712 {
1713         struct timespec now;
1714
1715         if (IS_NOCMTIME(inode))
1716                 return;
1717
1718         now = current_fs_time(inode->i_sb);
1719         if (!timespec_equal(&inode->i_mtime, &now))
1720                 inode->i_mtime = now;
1721
1722         if (!timespec_equal(&inode->i_ctime, &now))
1723                 inode->i_ctime = now;
1724
1725         if (IS_I_VERSION(inode))
1726                 inode_inc_iversion(inode);
1727 }
1728
1729 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1730                                     const struct iovec *iov,
1731                                     unsigned long nr_segs, loff_t pos)
1732 {
1733         struct file *file = iocb->ki_filp;
1734         struct inode *inode = file_inode(file);
1735         struct btrfs_root *root = BTRFS_I(inode)->root;
1736         u64 start_pos;
1737         u64 end_pos;
1738         ssize_t num_written = 0;
1739         ssize_t err = 0;
1740         size_t count, ocount;
1741         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1742
1743         mutex_lock(&inode->i_mutex);
1744
1745         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1746         if (err) {
1747                 mutex_unlock(&inode->i_mutex);
1748                 goto out;
1749         }
1750         count = ocount;
1751
1752         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1753         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1754         if (err) {
1755                 mutex_unlock(&inode->i_mutex);
1756                 goto out;
1757         }
1758
1759         if (count == 0) {
1760                 mutex_unlock(&inode->i_mutex);
1761                 goto out;
1762         }
1763
1764         err = file_remove_suid(file);
1765         if (err) {
1766                 mutex_unlock(&inode->i_mutex);
1767                 goto out;
1768         }
1769
1770         /*
1771          * If BTRFS flips readonly due to some impossible error
1772          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1773          * although we have opened a file as writable, we have
1774          * to stop this write operation to ensure FS consistency.
1775          */
1776         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1777                 mutex_unlock(&inode->i_mutex);
1778                 err = -EROFS;
1779                 goto out;
1780         }
1781
1782         /*
1783          * We reserve space for updating the inode when we reserve space for the
1784          * extent we are going to write, so we will enospc out there.  We don't
1785          * need to start yet another transaction to update the inode as we will
1786          * update the inode when we finish writing whatever data we write.
1787          */
1788         update_time_for_write(inode);
1789
1790         start_pos = round_down(pos, root->sectorsize);
1791         if (start_pos > i_size_read(inode)) {
1792                 /* Expand hole size to cover write data, preventing empty gap */
1793                 end_pos = round_up(pos + count, root->sectorsize);
1794                 err = btrfs_cont_expand(inode, i_size_read(inode), end_pos);
1795                 if (err) {
1796                         mutex_unlock(&inode->i_mutex);
1797                         goto out;
1798                 }
1799         }
1800
1801         if (sync)
1802                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1803
1804         if (unlikely(file->f_flags & O_DIRECT)) {
1805                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1806                                                    pos, count, ocount);
1807         } else {
1808                 struct iov_iter i;
1809
1810                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1811
1812                 num_written = __btrfs_buffered_write(file, &i, pos);
1813                 if (num_written > 0)
1814                         iocb->ki_pos = pos + num_written;
1815         }
1816
1817         mutex_unlock(&inode->i_mutex);
1818
1819         /*
1820          * we want to make sure fsync finds this change
1821          * but we haven't joined a transaction running right now.
1822          *
1823          * Later on, someone is sure to update the inode and get the
1824          * real transid recorded.
1825          *
1826          * We set last_trans now to the fs_info generation + 1,
1827          * this will either be one more than the running transaction
1828          * or the generation used for the next transaction if there isn't
1829          * one running right now.
1830          *
1831          * We also have to set last_sub_trans to the current log transid,
1832          * otherwise subsequent syncs to a file that's been synced in this
1833          * transaction will appear to have already occured.
1834          */
1835         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1836         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1837         if (num_written > 0) {
1838                 err = generic_write_sync(file, pos, num_written);
1839                 if (err < 0)
1840                         num_written = err;
1841         }
1842
1843         if (sync)
1844                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1845 out:
1846         current->backing_dev_info = NULL;
1847         return num_written ? num_written : err;
1848 }
1849
1850 int btrfs_release_file(struct inode *inode, struct file *filp)
1851 {
1852         /*
1853          * ordered_data_close is set by settattr when we are about to truncate
1854          * a file from a non-zero size to a zero size.  This tries to
1855          * flush down new bytes that may have been written if the
1856          * application were using truncate to replace a file in place.
1857          */
1858         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1859                                &BTRFS_I(inode)->runtime_flags)) {
1860                 struct btrfs_trans_handle *trans;
1861                 struct btrfs_root *root = BTRFS_I(inode)->root;
1862
1863                 /*
1864                  * We need to block on a committing transaction to keep us from
1865                  * throwing a ordered operation on to the list and causing
1866                  * something like sync to deadlock trying to flush out this
1867                  * inode.
1868                  */
1869                 trans = btrfs_start_transaction(root, 0);
1870                 if (IS_ERR(trans))
1871                         return PTR_ERR(trans);
1872                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1873                 btrfs_end_transaction(trans, root);
1874                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1875                         filemap_flush(inode->i_mapping);
1876         }
1877         if (filp->private_data)
1878                 btrfs_ioctl_trans_end(filp);
1879         return 0;
1880 }
1881
1882 /*
1883  * fsync call for both files and directories.  This logs the inode into
1884  * the tree log instead of forcing full commits whenever possible.
1885  *
1886  * It needs to call filemap_fdatawait so that all ordered extent updates are
1887  * in the metadata btree are up to date for copying to the log.
1888  *
1889  * It drops the inode mutex before doing the tree log commit.  This is an
1890  * important optimization for directories because holding the mutex prevents
1891  * new operations on the dir while we write to disk.
1892  */
1893 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1894 {
1895         struct dentry *dentry = file->f_path.dentry;
1896         struct inode *inode = dentry->d_inode;
1897         struct btrfs_root *root = BTRFS_I(inode)->root;
1898         struct btrfs_trans_handle *trans;
1899         struct btrfs_log_ctx ctx;
1900         int ret = 0;
1901         bool full_sync = 0;
1902
1903         trace_btrfs_sync_file(file, datasync);
1904
1905         /*
1906          * We write the dirty pages in the range and wait until they complete
1907          * out of the ->i_mutex. If so, we can flush the dirty pages by
1908          * multi-task, and make the performance up.  See
1909          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1910          */
1911         atomic_inc(&BTRFS_I(inode)->sync_writers);
1912         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1913         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1914                              &BTRFS_I(inode)->runtime_flags))
1915                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1916         atomic_dec(&BTRFS_I(inode)->sync_writers);
1917         if (ret)
1918                 return ret;
1919
1920         mutex_lock(&inode->i_mutex);
1921
1922         /*
1923          * We flush the dirty pages again to avoid some dirty pages in the
1924          * range being left.
1925          */
1926         atomic_inc(&root->log_batch);
1927         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1928                              &BTRFS_I(inode)->runtime_flags);
1929         if (full_sync) {
1930                 ret = btrfs_wait_ordered_range(inode, start, end - start + 1);
1931                 if (ret) {
1932                         mutex_unlock(&inode->i_mutex);
1933                         goto out;
1934                 }
1935         }
1936         atomic_inc(&root->log_batch);
1937
1938         /*
1939          * check the transaction that last modified this inode
1940          * and see if its already been committed
1941          */
1942         if (!BTRFS_I(inode)->last_trans) {
1943                 mutex_unlock(&inode->i_mutex);
1944                 goto out;
1945         }
1946
1947         /*
1948          * if the last transaction that changed this file was before
1949          * the current transaction, we can bail out now without any
1950          * syncing
1951          */
1952         smp_mb();
1953         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1954             BTRFS_I(inode)->last_trans <=
1955             root->fs_info->last_trans_committed) {
1956                 BTRFS_I(inode)->last_trans = 0;
1957
1958                 /*
1959                  * We'v had everything committed since the last time we were
1960                  * modified so clear this flag in case it was set for whatever
1961                  * reason, it's no longer relevant.
1962                  */
1963                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1964                           &BTRFS_I(inode)->runtime_flags);
1965                 mutex_unlock(&inode->i_mutex);
1966                 goto out;
1967         }
1968
1969         /*
1970          * ok we haven't committed the transaction yet, lets do a commit
1971          */
1972         if (file->private_data)
1973                 btrfs_ioctl_trans_end(file);
1974
1975         /*
1976          * We use start here because we will need to wait on the IO to complete
1977          * in btrfs_sync_log, which could require joining a transaction (for
1978          * example checking cross references in the nocow path).  If we use join
1979          * here we could get into a situation where we're waiting on IO to
1980          * happen that is blocked on a transaction trying to commit.  With start
1981          * we inc the extwriter counter, so we wait for all extwriters to exit
1982          * before we start blocking join'ers.  This comment is to keep somebody
1983          * from thinking they are super smart and changing this to
1984          * btrfs_join_transaction *cough*Josef*cough*.
1985          */
1986         trans = btrfs_start_transaction(root, 0);
1987         if (IS_ERR(trans)) {
1988                 ret = PTR_ERR(trans);
1989                 mutex_unlock(&inode->i_mutex);
1990                 goto out;
1991         }
1992         trans->sync = true;
1993
1994         btrfs_init_log_ctx(&ctx);
1995
1996         ret = btrfs_log_dentry_safe(trans, root, dentry, &ctx);
1997         if (ret < 0) {
1998                 /* Fallthrough and commit/free transaction. */
1999                 ret = 1;
2000         }
2001
2002         /* we've logged all the items and now have a consistent
2003          * version of the file in the log.  It is possible that
2004          * someone will come in and modify the file, but that's
2005          * fine because the log is consistent on disk, and we
2006          * have references to all of the file's extents
2007          *
2008          * It is possible that someone will come in and log the
2009          * file again, but that will end up using the synchronization
2010          * inside btrfs_sync_log to keep things safe.
2011          */
2012         mutex_unlock(&inode->i_mutex);
2013
2014         if (ret != BTRFS_NO_LOG_SYNC) {
2015                 if (!ret) {
2016                         ret = btrfs_sync_log(trans, root, &ctx);
2017                         if (!ret) {
2018                                 ret = btrfs_end_transaction(trans, root);
2019                                 goto out;
2020                         }
2021                 }
2022                 if (!full_sync) {
2023                         ret = btrfs_wait_ordered_range(inode, start,
2024                                                        end - start + 1);
2025                         if (ret)
2026                                 goto out;
2027                 }
2028                 ret = btrfs_commit_transaction(trans, root);
2029         } else {
2030                 ret = btrfs_end_transaction(trans, root);
2031         }
2032 out:
2033         return ret > 0 ? -EIO : ret;
2034 }
2035
2036 static const struct vm_operations_struct btrfs_file_vm_ops = {
2037         .fault          = filemap_fault,
2038         .map_pages      = filemap_map_pages,
2039         .page_mkwrite   = btrfs_page_mkwrite,
2040         .remap_pages    = generic_file_remap_pages,
2041 };
2042
2043 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2044 {
2045         struct address_space *mapping = filp->f_mapping;
2046
2047         if (!mapping->a_ops->readpage)
2048                 return -ENOEXEC;
2049
2050         file_accessed(filp);
2051         vma->vm_ops = &btrfs_file_vm_ops;
2052
2053         return 0;
2054 }
2055
2056 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2057                           int slot, u64 start, u64 end)
2058 {
2059         struct btrfs_file_extent_item *fi;
2060         struct btrfs_key key;
2061
2062         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2063                 return 0;
2064
2065         btrfs_item_key_to_cpu(leaf, &key, slot);
2066         if (key.objectid != btrfs_ino(inode) ||
2067             key.type != BTRFS_EXTENT_DATA_KEY)
2068                 return 0;
2069
2070         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2071
2072         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2073                 return 0;
2074
2075         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2076                 return 0;
2077
2078         if (key.offset == end)
2079                 return 1;
2080         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2081                 return 1;
2082         return 0;
2083 }
2084
2085 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2086                       struct btrfs_path *path, u64 offset, u64 end)
2087 {
2088         struct btrfs_root *root = BTRFS_I(inode)->root;
2089         struct extent_buffer *leaf;
2090         struct btrfs_file_extent_item *fi;
2091         struct extent_map *hole_em;
2092         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2093         struct btrfs_key key;
2094         int ret;
2095
2096         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2097                 goto out;
2098
2099         key.objectid = btrfs_ino(inode);
2100         key.type = BTRFS_EXTENT_DATA_KEY;
2101         key.offset = offset;
2102
2103         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2104         if (ret < 0)
2105                 return ret;
2106         BUG_ON(!ret);
2107
2108         leaf = path->nodes[0];
2109         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2110                 u64 num_bytes;
2111
2112                 path->slots[0]--;
2113                 fi = btrfs_item_ptr(leaf, path->slots[0],
2114                                     struct btrfs_file_extent_item);
2115                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2116                         end - offset;
2117                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2118                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2119                 btrfs_set_file_extent_offset(leaf, fi, 0);
2120                 btrfs_mark_buffer_dirty(leaf);
2121                 goto out;
2122         }
2123
2124         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
2125                 u64 num_bytes;
2126
2127                 path->slots[0]++;
2128                 key.offset = offset;
2129                 btrfs_set_item_key_safe(root, path, &key);
2130                 fi = btrfs_item_ptr(leaf, path->slots[0],
2131                                     struct btrfs_file_extent_item);
2132                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2133                         offset;
2134                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2135                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2136                 btrfs_set_file_extent_offset(leaf, fi, 0);
2137                 btrfs_mark_buffer_dirty(leaf);
2138                 goto out;
2139         }
2140         btrfs_release_path(path);
2141
2142         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2143                                        0, 0, end - offset, 0, end - offset,
2144                                        0, 0, 0);
2145         if (ret)
2146                 return ret;
2147
2148 out:
2149         btrfs_release_path(path);
2150
2151         hole_em = alloc_extent_map();
2152         if (!hole_em) {
2153                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2154                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2155                         &BTRFS_I(inode)->runtime_flags);
2156         } else {
2157                 hole_em->start = offset;
2158                 hole_em->len = end - offset;
2159                 hole_em->ram_bytes = hole_em->len;
2160                 hole_em->orig_start = offset;
2161
2162                 hole_em->block_start = EXTENT_MAP_HOLE;
2163                 hole_em->block_len = 0;
2164                 hole_em->orig_block_len = 0;
2165                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2166                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2167                 hole_em->generation = trans->transid;
2168
2169                 do {
2170                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2171                         write_lock(&em_tree->lock);
2172                         ret = add_extent_mapping(em_tree, hole_em, 1);
2173                         write_unlock(&em_tree->lock);
2174                 } while (ret == -EEXIST);
2175                 free_extent_map(hole_em);
2176                 if (ret)
2177                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2178                                 &BTRFS_I(inode)->runtime_flags);
2179         }
2180
2181         return 0;
2182 }
2183
2184 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2185 {
2186         struct btrfs_root *root = BTRFS_I(inode)->root;
2187         struct extent_state *cached_state = NULL;
2188         struct btrfs_path *path;
2189         struct btrfs_block_rsv *rsv;
2190         struct btrfs_trans_handle *trans;
2191         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2192         u64 lockend = round_down(offset + len,
2193                                  BTRFS_I(inode)->root->sectorsize) - 1;
2194         u64 cur_offset = lockstart;
2195         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2196         u64 drop_end;
2197         int ret = 0;
2198         int err = 0;
2199         int rsv_count;
2200         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2201                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2202         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2203         u64 ino_size;
2204
2205         ret = btrfs_wait_ordered_range(inode, offset, len);
2206         if (ret)
2207                 return ret;
2208
2209         mutex_lock(&inode->i_mutex);
2210         ino_size = round_up(inode->i_size, PAGE_CACHE_SIZE);
2211         /*
2212          * We needn't truncate any page which is beyond the end of the file
2213          * because we are sure there is no data there.
2214          */
2215         /*
2216          * Only do this if we are in the same page and we aren't doing the
2217          * entire page.
2218          */
2219         if (same_page && len < PAGE_CACHE_SIZE) {
2220                 if (offset < ino_size)
2221                         ret = btrfs_truncate_page(inode, offset, len, 0);
2222                 mutex_unlock(&inode->i_mutex);
2223                 return ret;
2224         }
2225
2226         /* zero back part of the first page */
2227         if (offset < ino_size) {
2228                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2229                 if (ret) {
2230                         mutex_unlock(&inode->i_mutex);
2231                         return ret;
2232                 }
2233         }
2234
2235         /* zero the front end of the last page */
2236         if (offset + len < ino_size) {
2237                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
2238                 if (ret) {
2239                         mutex_unlock(&inode->i_mutex);
2240                         return ret;
2241                 }
2242         }
2243
2244         if (lockend < lockstart) {
2245                 mutex_unlock(&inode->i_mutex);
2246                 return 0;
2247         }
2248
2249         while (1) {
2250                 struct btrfs_ordered_extent *ordered;
2251
2252                 truncate_pagecache_range(inode, lockstart, lockend);
2253
2254                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2255                                  0, &cached_state);
2256                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2257
2258                 /*
2259                  * We need to make sure we have no ordered extents in this range
2260                  * and nobody raced in and read a page in this range, if we did
2261                  * we need to try again.
2262                  */
2263                 if ((!ordered ||
2264                     (ordered->file_offset + ordered->len <= lockstart ||
2265                      ordered->file_offset > lockend)) &&
2266                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2267                                      lockend, EXTENT_UPTODATE, 0,
2268                                      cached_state)) {
2269                         if (ordered)
2270                                 btrfs_put_ordered_extent(ordered);
2271                         break;
2272                 }
2273                 if (ordered)
2274                         btrfs_put_ordered_extent(ordered);
2275                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2276                                      lockend, &cached_state, GFP_NOFS);
2277                 ret = btrfs_wait_ordered_range(inode, lockstart,
2278                                                lockend - lockstart + 1);
2279                 if (ret) {
2280                         mutex_unlock(&inode->i_mutex);
2281                         return ret;
2282                 }
2283         }
2284
2285         path = btrfs_alloc_path();
2286         if (!path) {
2287                 ret = -ENOMEM;
2288                 goto out;
2289         }
2290
2291         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2292         if (!rsv) {
2293                 ret = -ENOMEM;
2294                 goto out_free;
2295         }
2296         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2297         rsv->failfast = 1;
2298
2299         /*
2300          * 1 - update the inode
2301          * 1 - removing the extents in the range
2302          * 1 - adding the hole extent if no_holes isn't set
2303          */
2304         rsv_count = no_holes ? 2 : 3;
2305         trans = btrfs_start_transaction(root, rsv_count);
2306         if (IS_ERR(trans)) {
2307                 err = PTR_ERR(trans);
2308                 goto out_free;
2309         }
2310
2311         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2312                                       min_size);
2313         BUG_ON(ret);
2314         trans->block_rsv = rsv;
2315
2316         while (cur_offset < lockend) {
2317                 ret = __btrfs_drop_extents(trans, root, inode, path,
2318                                            cur_offset, lockend + 1,
2319                                            &drop_end, 1, 0, 0, NULL);
2320                 if (ret != -ENOSPC)
2321                         break;
2322
2323                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2324
2325                 if (cur_offset < ino_size) {
2326                         ret = fill_holes(trans, inode, path, cur_offset,
2327                                          drop_end);
2328                         if (ret) {
2329                                 err = ret;
2330                                 break;
2331                         }
2332                 }
2333
2334                 cur_offset = drop_end;
2335
2336                 ret = btrfs_update_inode(trans, root, inode);
2337                 if (ret) {
2338                         err = ret;
2339                         break;
2340                 }
2341
2342                 btrfs_end_transaction(trans, root);
2343                 btrfs_btree_balance_dirty(root);
2344
2345                 trans = btrfs_start_transaction(root, rsv_count);
2346                 if (IS_ERR(trans)) {
2347                         ret = PTR_ERR(trans);
2348                         trans = NULL;
2349                         break;
2350                 }
2351
2352                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2353                                               rsv, min_size);
2354                 BUG_ON(ret);    /* shouldn't happen */
2355                 trans->block_rsv = rsv;
2356         }
2357
2358         if (ret) {
2359                 err = ret;
2360                 goto out_trans;
2361         }
2362
2363         trans->block_rsv = &root->fs_info->trans_block_rsv;
2364         /*
2365          * Don't insert file hole extent item if it's for a range beyond eof
2366          * (because it's useless) or if it represents a 0 bytes range (when
2367          * cur_offset == drop_end).
2368          */
2369         if (cur_offset < ino_size && cur_offset < drop_end) {
2370                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2371                 if (ret) {
2372                         err = ret;
2373                         goto out_trans;
2374                 }
2375         }
2376
2377 out_trans:
2378         if (!trans)
2379                 goto out_free;
2380
2381         inode_inc_iversion(inode);
2382         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2383
2384         trans->block_rsv = &root->fs_info->trans_block_rsv;
2385         ret = btrfs_update_inode(trans, root, inode);
2386         btrfs_end_transaction(trans, root);
2387         btrfs_btree_balance_dirty(root);
2388 out_free:
2389         btrfs_free_path(path);
2390         btrfs_free_block_rsv(root, rsv);
2391 out:
2392         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2393                              &cached_state, GFP_NOFS);
2394         mutex_unlock(&inode->i_mutex);
2395         if (ret && !err)
2396                 err = ret;
2397         return err;
2398 }
2399
2400 static long btrfs_fallocate(struct file *file, int mode,
2401                             loff_t offset, loff_t len)
2402 {
2403         struct inode *inode = file_inode(file);
2404         struct extent_state *cached_state = NULL;
2405         struct btrfs_root *root = BTRFS_I(inode)->root;
2406         u64 cur_offset;
2407         u64 last_byte;
2408         u64 alloc_start;
2409         u64 alloc_end;
2410         u64 alloc_hint = 0;
2411         u64 locked_end;
2412         struct extent_map *em;
2413         int blocksize = BTRFS_I(inode)->root->sectorsize;
2414         int ret;
2415
2416         alloc_start = round_down(offset, blocksize);
2417         alloc_end = round_up(offset + len, blocksize);
2418
2419         /* Make sure we aren't being give some crap mode */
2420         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2421                 return -EOPNOTSUPP;
2422
2423         if (mode & FALLOC_FL_PUNCH_HOLE)
2424                 return btrfs_punch_hole(inode, offset, len);
2425
2426         /*
2427          * Make sure we have enough space before we do the
2428          * allocation.
2429          */
2430         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2431         if (ret)
2432                 return ret;
2433         if (root->fs_info->quota_enabled) {
2434                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2435                 if (ret)
2436                         goto out_reserve_fail;
2437         }
2438
2439         mutex_lock(&inode->i_mutex);
2440         ret = inode_newsize_ok(inode, alloc_end);
2441         if (ret)
2442                 goto out;
2443
2444         if (alloc_start > inode->i_size) {
2445                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2446                                         alloc_start);
2447                 if (ret)
2448                         goto out;
2449         } else {
2450                 /*
2451                  * If we are fallocating from the end of the file onward we
2452                  * need to zero out the end of the page if i_size lands in the
2453                  * middle of a page.
2454                  */
2455                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2456                 if (ret)
2457                         goto out;
2458         }
2459
2460         /*
2461          * wait for ordered IO before we have any locks.  We'll loop again
2462          * below with the locks held.
2463          */
2464         ret = btrfs_wait_ordered_range(inode, alloc_start,
2465                                        alloc_end - alloc_start);
2466         if (ret)
2467                 goto out;
2468
2469         locked_end = alloc_end - 1;
2470         while (1) {
2471                 struct btrfs_ordered_extent *ordered;
2472
2473                 /* the extent lock is ordered inside the running
2474                  * transaction
2475                  */
2476                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2477                                  locked_end, 0, &cached_state);
2478                 ordered = btrfs_lookup_first_ordered_extent(inode,
2479                                                             alloc_end - 1);
2480                 if (ordered &&
2481                     ordered->file_offset + ordered->len > alloc_start &&
2482                     ordered->file_offset < alloc_end) {
2483                         btrfs_put_ordered_extent(ordered);
2484                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2485                                              alloc_start, locked_end,
2486                                              &cached_state, GFP_NOFS);
2487                         /*
2488                          * we can't wait on the range with the transaction
2489                          * running or with the extent lock held
2490                          */
2491                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2492                                                        alloc_end - alloc_start);
2493                         if (ret)
2494                                 goto out;
2495                 } else {
2496                         if (ordered)
2497                                 btrfs_put_ordered_extent(ordered);
2498                         break;
2499                 }
2500         }
2501
2502         cur_offset = alloc_start;
2503         while (1) {
2504                 u64 actual_end;
2505
2506                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2507                                       alloc_end - cur_offset, 0);
2508                 if (IS_ERR_OR_NULL(em)) {
2509                         if (!em)
2510                                 ret = -ENOMEM;
2511                         else
2512                                 ret = PTR_ERR(em);
2513                         break;
2514                 }
2515                 last_byte = min(extent_map_end(em), alloc_end);
2516                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2517                 last_byte = ALIGN(last_byte, blocksize);
2518
2519                 if (em->block_start == EXTENT_MAP_HOLE ||
2520                     (cur_offset >= inode->i_size &&
2521                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2522                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2523                                                         last_byte - cur_offset,
2524                                                         1 << inode->i_blkbits,
2525                                                         offset + len,
2526                                                         &alloc_hint);
2527
2528                         if (ret < 0) {
2529                                 free_extent_map(em);
2530                                 break;
2531                         }
2532                 } else if (actual_end > inode->i_size &&
2533                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2534                         /*
2535                          * We didn't need to allocate any more space, but we
2536                          * still extended the size of the file so we need to
2537                          * update i_size.
2538                          */
2539                         inode->i_ctime = CURRENT_TIME;
2540                         i_size_write(inode, actual_end);
2541                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2542                 }
2543                 free_extent_map(em);
2544
2545                 cur_offset = last_byte;
2546                 if (cur_offset >= alloc_end) {
2547                         ret = 0;
2548                         break;
2549                 }
2550         }
2551         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2552                              &cached_state, GFP_NOFS);
2553 out:
2554         mutex_unlock(&inode->i_mutex);
2555         if (root->fs_info->quota_enabled)
2556                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2557 out_reserve_fail:
2558         /* Let go of our reservation. */
2559         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2560         return ret;
2561 }
2562
2563 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2564 {
2565         struct btrfs_root *root = BTRFS_I(inode)->root;
2566         struct extent_map *em = NULL;
2567         struct extent_state *cached_state = NULL;
2568         u64 lockstart = *offset;
2569         u64 lockend = i_size_read(inode);
2570         u64 start = *offset;
2571         u64 len = i_size_read(inode);
2572         int ret = 0;
2573
2574         lockend = max_t(u64, root->sectorsize, lockend);
2575         if (lockend <= lockstart)
2576                 lockend = lockstart + root->sectorsize;
2577
2578         lockend--;
2579         len = lockend - lockstart + 1;
2580
2581         len = max_t(u64, len, root->sectorsize);
2582         if (inode->i_size == 0)
2583                 return -ENXIO;
2584
2585         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2586                          &cached_state);
2587
2588         while (start < inode->i_size) {
2589                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2590                 if (IS_ERR(em)) {
2591                         ret = PTR_ERR(em);
2592                         em = NULL;
2593                         break;
2594                 }
2595
2596                 if (whence == SEEK_HOLE &&
2597                     (em->block_start == EXTENT_MAP_HOLE ||
2598                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2599                         break;
2600                 else if (whence == SEEK_DATA &&
2601                            (em->block_start != EXTENT_MAP_HOLE &&
2602                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2603                         break;
2604
2605                 start = em->start + em->len;
2606                 free_extent_map(em);
2607                 em = NULL;
2608                 cond_resched();
2609         }
2610         free_extent_map(em);
2611         if (!ret) {
2612                 if (whence == SEEK_DATA && start >= inode->i_size)
2613                         ret = -ENXIO;
2614                 else
2615                         *offset = min_t(loff_t, start, inode->i_size);
2616         }
2617         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2618                              &cached_state, GFP_NOFS);
2619         return ret;
2620 }
2621
2622 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2623 {
2624         struct inode *inode = file->f_mapping->host;
2625         int ret;
2626
2627         mutex_lock(&inode->i_mutex);
2628         switch (whence) {
2629         case SEEK_END:
2630         case SEEK_CUR:
2631                 offset = generic_file_llseek(file, offset, whence);
2632                 goto out;
2633         case SEEK_DATA:
2634         case SEEK_HOLE:
2635                 if (offset >= i_size_read(inode)) {
2636                         mutex_unlock(&inode->i_mutex);
2637                         return -ENXIO;
2638                 }
2639
2640                 ret = find_desired_extent(inode, &offset, whence);
2641                 if (ret) {
2642                         mutex_unlock(&inode->i_mutex);
2643                         return ret;
2644                 }
2645         }
2646
2647         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2648 out:
2649         mutex_unlock(&inode->i_mutex);
2650         return offset;
2651 }
2652
2653 const struct file_operations btrfs_file_operations = {
2654         .llseek         = btrfs_file_llseek,
2655         .read           = do_sync_read,
2656         .write          = do_sync_write,
2657         .aio_read       = generic_file_aio_read,
2658         .splice_read    = generic_file_splice_read,
2659         .aio_write      = btrfs_file_aio_write,
2660         .mmap           = btrfs_file_mmap,
2661         .open           = generic_file_open,
2662         .release        = btrfs_release_file,
2663         .fsync          = btrfs_sync_file,
2664         .fallocate      = btrfs_fallocate,
2665         .unlocked_ioctl = btrfs_ioctl,
2666 #ifdef CONFIG_COMPAT
2667         .compat_ioctl   = btrfs_ioctl,
2668 #endif
2669 };
2670
2671 void btrfs_auto_defrag_exit(void)
2672 {
2673         if (btrfs_inode_defrag_cachep)
2674                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2675 }
2676
2677 int btrfs_auto_defrag_init(void)
2678 {
2679         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2680                                         sizeof(struct inode_defrag), 0,
2681                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2682                                         NULL);
2683         if (!btrfs_inode_defrag_cachep)
2684                 return -ENOMEM;
2685
2686         return 0;
2687 }