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 #include "qgroup.h"
  44
  45 static struct kmem_cache *btrfs_inode_defrag_cachep;
  46 /*
  47  * when auto defrag is enabled we
  48  * queue up these defrag structs to remember which
  49  * inodes need defragging passes
  50  */
  51 struct inode_defrag {
  52         struct rb_node rb_node;
  53         /* objectid */
  54         u64 ino;
  55         /*
  56          * transid where the defrag was added, we search for
  57          * extents newer than this
  58          */
  59         u64 transid;
  60
  61         /* root objectid */
  62         u64 root;
  63
  64         /* last offset we were able to defrag */
  65         u64 last_offset;
  66
  67         /* if we've wrapped around back to zero once already */
  68         int cycled;
  69 };
  70
  71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  72                                   struct inode_defrag *defrag2)
  73 {
  74         if (defrag1->root > defrag2->root)
  75                 return 1;
  76         else if (defrag1->root < defrag2->root)
  77                 return -1;
  78         else if (defrag1->ino > defrag2->ino)
  79                 return 1;
  80         else if (defrag1->ino < defrag2->ino)
  81                 return -1;
  82         else
  83                 return 0;
  84 }
  85
  86 /* pop a record for an inode into the defrag tree.  The lock
  87  * must be held already
  88  *
  89  * If you're inserting a record for an older transid than an
  90  * existing record, the transid already in the tree is lowered
  91  *
  92  * If an existing record is found the defrag item you
  93  * pass in is freed
  94  */
  95 static int __btrfs_add_inode_defrag(struct inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_root *root = BTRFS_I(inode)->root;
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &root->fs_info->defrag_inodes.rb_node;
 105         while (*p) {
 106                 parent = *p;
 107                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 108
 109                 ret = __compare_inode_defrag(defrag, entry);
 110                 if (ret < 0)
 111                         p = &parent->rb_left;
 112                 else if (ret > 0)
 113                         p = &parent->rb_right;
 114                 else {
 115                         /* if we're reinserting an entry for
 116                          * an old defrag run, make sure to
 117                          * lower the transid of our existing record
 118                          */
 119                         if (defrag->transid < entry->transid)
 120                                 entry->transid = defrag->transid;
 121                         if (defrag->last_offset > entry->last_offset)
 122                                 entry->last_offset = defrag->last_offset;
 123                         return -EEXIST;
 124                 }
 125         }
 126         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_root *root)
 133 {
 134         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(root->fs_info))
 138                 return 0;
 139
 140         return 1;
 141 }
 142
 143 /*
 144  * insert a defrag record for this inode if auto defrag is
 145  * enabled
 146  */
 147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 148                            struct inode *inode)
 149 {
 150         struct btrfs_root *root = BTRFS_I(inode)->root;
 151         struct inode_defrag *defrag;
 152         u64 transid;
 153         int ret;
 154
 155         if (!__need_auto_defrag(root))
 156                 return 0;
 157
 158         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 159                 return 0;
 160
 161         if (trans)
 162                 transid = trans->transid;
 163         else
 164                 transid = BTRFS_I(inode)->root->last_trans;
 165
 166         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 167         if (!defrag)
 168                 return -ENOMEM;
 169
 170         defrag->ino = btrfs_ino(inode);
 171         defrag->transid = transid;
 172         defrag->root = root->root_key.objectid;
 173
 174         spin_lock(&root->fs_info->defrag_inodes_lock);
 175         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 176                 /*
 177                  * If we set IN_DEFRAG flag and evict the inode from memory,
 178                  * and then re-read this inode, this new inode doesn't have
 179                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 180                  */
 181                 ret = __btrfs_add_inode_defrag(inode, defrag);
 182                 if (ret)
 183                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 184         } else {
 185                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 186         }
 187         spin_unlock(&root->fs_info->defrag_inodes_lock);
 188         return 0;
 189 }
 190
 191 /*
 192  * Requeue the defrag object. If there is a defrag object that points to
 193  * the same inode in the tree, we will merge them together (by
 194  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 195  */
 196 static void btrfs_requeue_inode_defrag(struct inode *inode,
 197                                        struct inode_defrag *defrag)
 198 {
 199         struct btrfs_root *root = BTRFS_I(inode)->root;
 200         int ret;
 201
 202         if (!__need_auto_defrag(root))
 203                 goto out;
 204
 205         /*
 206          * Here we don't check the IN_DEFRAG flag, because we need merge
 207          * them together.
 208          */
 209         spin_lock(&root->fs_info->defrag_inodes_lock);
 210         ret = __btrfs_add_inode_defrag(inode, defrag);
 211         spin_unlock(&root->fs_info->defrag_inodes_lock);
 212         if (ret)
 213                 goto out;
 214         return;
 215 out:
 216         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 217 }
 218
 219 /*
 220  * pick the defragable inode that we want, if it doesn't exist, we will get
 221  * the next one.
 222  */
 223 static struct inode_defrag *
 224 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 225 {
 226         struct inode_defrag *entry = NULL;
 227         struct inode_defrag tmp;
 228         struct rb_node *p;
 229         struct rb_node *parent = NULL;
 230         int ret;
 231
 232         tmp.ino = ino;
 233         tmp.root = root;
 234
 235         spin_lock(&fs_info->defrag_inodes_lock);
 236         p = fs_info->defrag_inodes.rb_node;
 237         while (p) {
 238                 parent = p;
 239                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 240
 241                 ret = __compare_inode_defrag(&tmp, entry);
 242                 if (ret < 0)
 243                         p = parent->rb_left;
 244                 else if (ret > 0)
 245                         p = parent->rb_right;
 246                 else
 247                         goto out;
 248         }
 249
 250         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 251                 parent = rb_next(parent);
 252                 if (parent)
 253                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 254                 else
 255                         entry = NULL;
 256         }
 257 out:
 258         if (entry)
 259                 rb_erase(parent, &fs_info->defrag_inodes);
 260         spin_unlock(&fs_info->defrag_inodes_lock);
 261         return entry;
 262 }
 263
 264 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 265 {
 266         struct inode_defrag *defrag;
 267         struct rb_node *node;
 268
 269         spin_lock(&fs_info->defrag_inodes_lock);
 270         node = rb_first(&fs_info->defrag_inodes);
 271         while (node) {
 272                 rb_erase(node, &fs_info->defrag_inodes);
 273                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 274                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 275
 276                 if (need_resched()) {
 277                         spin_unlock(&fs_info->defrag_inodes_lock);
 278                         cond_resched();
 279                         spin_lock(&fs_info->defrag_inodes_lock);
 280                 }
 281
 282                 node = rb_first(&fs_info->defrag_inodes);
 283         }
 284         spin_unlock(&fs_info->defrag_inodes_lock);
 285 }
 286
 287 #define BTRFS_DEFRAG_BATCH      1024
 288
 289 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 290                                     struct inode_defrag *defrag)
 291 {
 292         struct btrfs_root *inode_root;
 293         struct inode *inode;
 294         struct btrfs_key key;
 295         struct btrfs_ioctl_defrag_range_args range;
 296         int num_defrag;
 297         int index;
 298         int ret;
 299
 300         /* get the inode */
 301         key.objectid = defrag->root;
 302         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 303         key.offset = (u64)-1;
 304
 305         index = srcu_read_lock(&fs_info->subvol_srcu);
 306
 307         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 308         if (IS_ERR(inode_root)) {
 309                 ret = PTR_ERR(inode_root);
 310                 goto cleanup;
 311         }
 312
 313         key.objectid = defrag->ino;
 314         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 315         key.offset = 0;
 316         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 317         if (IS_ERR(inode)) {
 318                 ret = PTR_ERR(inode);
 319                 goto cleanup;
 320         }
 321         srcu_read_unlock(&fs_info->subvol_srcu, index);
 322
 323         /* do a chunk of defrag */
 324         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 325         memset(&range, 0, sizeof(range));
 326         range.len = (u64)-1;
 327         range.start = defrag->last_offset;
 328
 329         sb_start_write(fs_info->sb);
 330         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 331                                        BTRFS_DEFRAG_BATCH);
 332         sb_end_write(fs_info->sb);
 333         /*
 334          * if we filled the whole defrag batch, there
 335          * must be more work to do.  Queue this defrag
 336          * again
 337          */
 338         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 339                 defrag->last_offset = range.start;
 340                 btrfs_requeue_inode_defrag(inode, defrag);
 341         } else if (defrag->last_offset && !defrag->cycled) {
 342                 /*
 343                  * we didn't fill our defrag batch, but
 344                  * we didn't start at zero.  Make sure we loop
 345                  * around to the start of the file.
 346                  */
 347                 defrag->last_offset = 0;
 348                 defrag->cycled = 1;
 349                 btrfs_requeue_inode_defrag(inode, defrag);
 350         } else {
 351                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 352         }
 353
 354         iput(inode);
 355         return 0;
 356 cleanup:
 357         srcu_read_unlock(&fs_info->subvol_srcu, index);
 358         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 359         return ret;
 360 }
 361
 362 /*
 363  * run through the list of inodes in the FS that need
 364  * defragging
 365  */
 366 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 367 {
 368         struct inode_defrag *defrag;
 369         u64 first_ino = 0;
 370         u64 root_objectid = 0;
 371
 372         atomic_inc(&fs_info->defrag_running);
 373         while (1) {
 374                 /* Pause the auto defragger. */
 375                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 376                              &fs_info->fs_state))
 377                         break;
 378
 379                 if (!__need_auto_defrag(fs_info->tree_root))
 380                         break;
 381
 382                 /* find an inode to defrag */
 383                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 384                                                  first_ino);
 385                 if (!defrag) {
 386                         if (root_objectid || first_ino) {
 387                                 root_objectid = 0;
 388                                 first_ino = 0;
 389                                 continue;
 390                         } else {
 391                                 break;
 392                         }
 393                 }
 394
 395                 first_ino = defrag->ino + 1;
 396                 root_objectid = defrag->root;
 397
 398                 __btrfs_run_defrag_inode(fs_info, defrag);
 399         }
 400         atomic_dec(&fs_info->defrag_running);
 401
 402         /*
 403          * during unmount, we use the transaction_wait queue to
 404          * wait for the defragger to stop
 405          */
 406         wake_up(&fs_info->transaction_wait);
 407         return 0;
 408 }
 409
 410 /* simple helper to fault in pages and copy.  This should go away
 411  * and be replaced with calls into generic code.
 412  */
 413 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 414                                          size_t write_bytes,
 415                                          struct page **prepared_pages,
 416                                          struct iov_iter *i)
 417 {
 418         size_t copied = 0;
 419         size_t total_copied = 0;
 420         int pg = 0;
 421         int offset = pos & (PAGE_CACHE_SIZE - 1);
 422
 423         while (write_bytes > 0) {
 424                 size_t count = min_t(size_t,
 425                                      PAGE_CACHE_SIZE - offset, write_bytes);
 426                 struct page *page = prepared_pages[pg];
 427                 /*
 428                  * Copy data from userspace to the current page
 429                  */
 430                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 431
 432                 /* Flush processor's dcache for this page */
 433                 flush_dcache_page(page);
 434
 435                 /*
 436                  * if we get a partial write, we can end up with
 437                  * partially up to date pages.  These add
 438                  * a lot of complexity, so make sure they don't
 439                  * happen by forcing this copy to be retried.
 440                  *
 441                  * The rest of the btrfs_file_write code will fall
 442                  * back to page at a time copies after we return 0.
 443                  */
 444                 if (!PageUptodate(page) && copied < count)
 445                         copied = 0;
 446
 447                 iov_iter_advance(i, copied);
 448                 write_bytes -= copied;
 449                 total_copied += copied;
 450
 451                 /* Return to btrfs_file_aio_write to fault page */
 452                 if (unlikely(copied == 0))
 453                         break;
 454
 455                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 456                         offset += copied;
 457                 } else {
 458                         pg++;
 459                         offset = 0;
 460                 }
 461         }
 462         return total_copied;
 463 }
 464
 465 /*
 466  * unlocks pages after btrfs_file_write is done with them
 467  */
 468 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 469 {
 470         size_t i;
 471         for (i = 0; i < num_pages; i++) {
 472                 /* page checked is some magic around finding pages that
 473                  * have been modified without going through btrfs_set_page_dirty
 474                  * clear it here. There should be no need to mark the pages
 475                  * accessed as prepare_pages should have marked them accessed
 476                  * in prepare_pages via find_or_create_page()
 477                  */
 478                 ClearPageChecked(pages[i]);
 479                 unlock_page(pages[i]);
 480                 page_cache_release(pages[i]);
 481         }
 482 }
 483
 484 /*
 485  * after copy_from_user, pages need to be dirtied and we need to make
 486  * sure holes are created between the current EOF and the start of
 487  * any next extents (if required).
 488  *
 489  * this also makes the decision about creating an inline extent vs
 490  * doing real data extents, marking pages dirty and delalloc as required.
 491  */
 492 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 493                              struct page **pages, size_t num_pages,
 494                              loff_t pos, size_t write_bytes,
 495                              struct extent_state **cached)
 496 {
 497         int err = 0;
 498         int i;
 499         u64 num_bytes;
 500         u64 start_pos;
 501         u64 end_of_last_block;
 502         u64 end_pos = pos + write_bytes;
 503         loff_t isize = i_size_read(inode);
 504
 505         start_pos = pos & ~((u64)root->sectorsize - 1);
 506         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 507
 508         end_of_last_block = start_pos + num_bytes - 1;
 509         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 510                                         cached);
 511         if (err)
 512                 return err;
 513
 514         for (i = 0; i < num_pages; i++) {
 515                 struct page *p = pages[i];
 516                 SetPageUptodate(p);
 517                 ClearPageChecked(p);
 518                 set_page_dirty(p);
 519         }
 520
 521         /*
 522          * we've only changed i_size in ram, and we haven't updated
 523          * the disk i_size.  There is no need to log the inode
 524          * at this time.
 525          */
 526         if (end_pos > isize)
 527                 i_size_write(inode, end_pos);
 528         return 0;
 529 }
 530
 531 /*
 532  * this drops all the extents in the cache that intersect the range
 533  * [start, end].  Existing extents are split as required.
 534  */
 535 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 536                              int skip_pinned)
 537 {
 538         struct extent_map *em;
 539         struct extent_map *split = NULL;
 540         struct extent_map *split2 = NULL;
 541         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 542         u64 len = end - start + 1;
 543         u64 gen;
 544         int ret;
 545         int testend = 1;
 546         unsigned long flags;
 547         int compressed = 0;
 548         bool modified;
 549
 550         WARN_ON(end < start);
 551         if (end == (u64)-1) {
 552                 len = (u64)-1;
 553                 testend = 0;
 554         }
 555         while (1) {
 556                 int no_splits = 0;
 557
 558                 modified = false;
 559                 if (!split)
 560                         split = alloc_extent_map();
 561                 if (!split2)
 562                         split2 = alloc_extent_map();
 563                 if (!split || !split2)
 564                         no_splits = 1;
 565
 566                 write_lock(&em_tree->lock);
 567                 em = lookup_extent_mapping(em_tree, start, len);
 568                 if (!em) {
 569                         write_unlock(&em_tree->lock);
 570                         break;
 571                 }
 572                 flags = em->flags;
 573                 gen = em->generation;
 574                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 575                         if (testend && em->start + em->len >= start + len) {
 576                                 free_extent_map(em);
 577                                 write_unlock(&em_tree->lock);
 578                                 break;
 579                         }
 580                         start = em->start + em->len;
 581                         if (testend)
 582                                 len = start + len - (em->start + em->len);
 583                         free_extent_map(em);
 584                         write_unlock(&em_tree->lock);
 585                         continue;
 586                 }
 587                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 588                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 589                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 590                 modified = !list_empty(&em->list);
 591                 if (no_splits)
 592                         goto next;
 593
 594                 if (em->start < start) {
 595                         split->start = em->start;
 596                         split->len = start - em->start;
 597
 598                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 599                                 split->orig_start = em->orig_start;
 600                                 split->block_start = em->block_start;
 601
 602                                 if (compressed)
 603                                         split->block_len = em->block_len;
 604                                 else
 605                                         split->block_len = split->len;
 606                                 split->orig_block_len = max(split->block_len,
 607                                                 em->orig_block_len);
 608                                 split->ram_bytes = em->ram_bytes;
 609                         } else {
 610                                 split->orig_start = split->start;
 611                                 split->block_len = 0;
 612                                 split->block_start = em->block_start;
 613                                 split->orig_block_len = 0;
 614                                 split->ram_bytes = split->len;
 615                         }
 616
 617                         split->generation = gen;
 618                         split->bdev = em->bdev;
 619                         split->flags = flags;
 620                         split->compress_type = em->compress_type;
 621                         replace_extent_mapping(em_tree, em, split, modified);
 622                         free_extent_map(split);
 623                         split = split2;
 624                         split2 = NULL;
 625                 }
 626                 if (testend && em->start + em->len > start + len) {
 627                         u64 diff = start + len - em->start;
 628
 629                         split->start = start + len;
 630                         split->len = em->start + em->len - (start + len);
 631                         split->bdev = em->bdev;
 632                         split->flags = flags;
 633                         split->compress_type = em->compress_type;
 634                         split->generation = gen;
 635
 636                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 637                                 split->orig_block_len = max(em->block_len,
 638                                                     em->orig_block_len);
 639
 640                                 split->ram_bytes = em->ram_bytes;
 641                                 if (compressed) {
 642                                         split->block_len = em->block_len;
 643                                         split->block_start = em->block_start;
 644                                         split->orig_start = em->orig_start;
 645                                 } else {
 646                                         split->block_len = split->len;
 647                                         split->block_start = em->block_start
 648                                                 + diff;
 649                                         split->orig_start = em->orig_start;
 650                                 }
 651                         } else {
 652                                 split->ram_bytes = split->len;
 653                                 split->orig_start = split->start;
 654                                 split->block_len = 0;
 655                                 split->block_start = em->block_start;
 656                                 split->orig_block_len = 0;
 657                         }
 658
 659                         if (extent_map_in_tree(em)) {
 660                                 replace_extent_mapping(em_tree, em, split,
 661                                                        modified);
 662                         } else {
 663                                 ret = add_extent_mapping(em_tree, split,
 664                                                          modified);
 665                                 ASSERT(ret == 0); /* Logic error */
 666                         }
 667                         free_extent_map(split);
 668                         split = NULL;
 669                 }
 670 next:
 671                 if (extent_map_in_tree(em))
 672                         remove_extent_mapping(em_tree, em);
 673                 write_unlock(&em_tree->lock);
 674
 675                 /* once for us */
 676                 free_extent_map(em);
 677                 /* once for the tree*/
 678                 free_extent_map(em);
 679         }
 680         if (split)
 681                 free_extent_map(split);
 682         if (split2)
 683                 free_extent_map(split2);
 684 }
 685
 686 /*
 687  * this is very complex, but the basic idea is to drop all extents
 688  * in the range start - end.  hint_block is filled in with a block number
 689  * that would be a good hint to the block allocator for this file.
 690  *
 691  * If an extent intersects the range but is not entirely inside the range
 692  * it is either truncated or split.  Anything entirely inside the range
 693  * is deleted from the tree.
 694  */
 695 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 696                          struct btrfs_root *root, struct inode *inode,
 697                          struct btrfs_path *path, u64 start, u64 end,
 698                          u64 *drop_end, int drop_cache,
 699                          int replace_extent,
 700                          u32 extent_item_size,
 701                          int *key_inserted)
 702 {
 703         struct extent_buffer *leaf;
 704         struct btrfs_file_extent_item *fi;
 705         struct btrfs_key key;
 706         struct btrfs_key new_key;
 707         u64 ino = btrfs_ino(inode);
 708         u64 search_start = start;
 709         u64 disk_bytenr = 0;
 710         u64 num_bytes = 0;
 711         u64 extent_offset = 0;
 712         u64 extent_end = 0;
 713         int del_nr = 0;
 714         int del_slot = 0;
 715         int extent_type;
 716         int recow;
 717         int ret;
 718         int modify_tree = -1;
 719         int update_refs;
 720         int found = 0;
 721         int leafs_visited = 0;
 722
 723         if (drop_cache)
 724                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 725
 726         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 727                 modify_tree = 0;
 728
 729         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 730                        root == root->fs_info->tree_root);
 731         while (1) {
 732                 recow = 0;
 733                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 734                                                search_start, modify_tree);
 735                 if (ret < 0)
 736                         break;
 737                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 738                         leaf = path->nodes[0];
 739                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 740                         if (key.objectid == ino &&
 741                             key.type == BTRFS_EXTENT_DATA_KEY)
 742                                 path->slots[0]--;
 743                 }
 744                 ret = 0;
 745                 leafs_visited++;
 746 next_slot:
 747                 leaf = path->nodes[0];
 748                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 749                         BUG_ON(del_nr > 0);
 750                         ret = btrfs_next_leaf(root, path);
 751                         if (ret < 0)
 752                                 break;
 753                         if (ret > 0) {
 754                                 ret = 0;
 755                                 break;
 756                         }
 757                         leafs_visited++;
 758                         leaf = path->nodes[0];
 759                         recow = 1;
 760                 }
 761
 762                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 763                 if (key.objectid > ino ||
 764                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 765                         break;
 766
 767                 fi = btrfs_item_ptr(leaf, path->slots[0],
 768                                     struct btrfs_file_extent_item);
 769                 extent_type = btrfs_file_extent_type(leaf, fi);
 770
 771                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 772                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 773                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 774                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 775                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 776                         extent_end = key.offset +
 777                                 btrfs_file_extent_num_bytes(leaf, fi);
 778                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 779                         extent_end = key.offset +
 780                                 btrfs_file_extent_inline_len(leaf,
 781                                                      path->slots[0], fi);
 782                 } else {
 783                         WARN_ON(1);
 784                         extent_end = search_start;
 785                 }
 786
 787                 /*
 788                  * Don't skip extent items representing 0 byte lengths. They
 789                  * used to be created (bug) if while punching holes we hit
 790                  * -ENOSPC condition. So if we find one here, just ensure we
 791                  * delete it, otherwise we would insert a new file extent item
 792                  * with the same key (offset) as that 0 bytes length file
 793                  * extent item in the call to setup_items_for_insert() later
 794                  * in this function.
 795                  */
 796                 if (extent_end == key.offset && extent_end >= search_start)
 797                         goto delete_extent_item;
 798
 799                 if (extent_end <= search_start) {
 800                         path->slots[0]++;
 801                         goto next_slot;
 802                 }
 803
 804                 found = 1;
 805                 search_start = max(key.offset, start);
 806                 if (recow || !modify_tree) {
 807                         modify_tree = -1;
 808                         btrfs_release_path(path);
 809                         continue;
 810                 }
 811
 812                 /*
 813                  *     | - range to drop - |
 814                  *  | -------- extent -------- |
 815                  */
 816                 if (start > key.offset && end < extent_end) {
 817                         BUG_ON(del_nr > 0);
 818                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 819                                 ret = -EOPNOTSUPP;
 820                                 break;
 821                         }
 822
 823                         memcpy(&new_key, &key, sizeof(new_key));
 824                         new_key.offset = start;
 825                         ret = btrfs_duplicate_item(trans, root, path,
 826                                                    &new_key);
 827                         if (ret == -EAGAIN) {
 828                                 btrfs_release_path(path);
 829                                 continue;
 830                         }
 831                         if (ret < 0)
 832                                 break;
 833
 834                         leaf = path->nodes[0];
 835                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 836                                             struct btrfs_file_extent_item);
 837                         btrfs_set_file_extent_num_bytes(leaf, fi,
 838                                                         start - key.offset);
 839
 840                         fi = btrfs_item_ptr(leaf, path->slots[0],
 841                                             struct btrfs_file_extent_item);
 842
 843                         extent_offset += start - key.offset;
 844                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 845                         btrfs_set_file_extent_num_bytes(leaf, fi,
 846                                                         extent_end - start);
 847                         btrfs_mark_buffer_dirty(leaf);
 848
 849                         if (update_refs && disk_bytenr > 0) {
 850                                 ret = btrfs_inc_extent_ref(trans, root,
 851                                                 disk_bytenr, num_bytes, 0,
 852                                                 root->root_key.objectid,
 853                                                 new_key.objectid,
 854                                                 start - extent_offset, 1);
 855                                 BUG_ON(ret); /* -ENOMEM */
 856                         }
 857                         key.offset = start;
 858                 }
 859                 /*
 860                  *  | ---- range to drop ----- |
 861                  *      | -------- extent -------- |
 862                  */
 863                 if (start <= key.offset && end < extent_end) {
 864                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 865                                 ret = -EOPNOTSUPP;
 866                                 break;
 867                         }
 868
 869                         memcpy(&new_key, &key, sizeof(new_key));
 870                         new_key.offset = end;
 871                         btrfs_set_item_key_safe(root, path, &new_key);
 872
 873                         extent_offset += end - key.offset;
 874                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 875                         btrfs_set_file_extent_num_bytes(leaf, fi,
 876                                                         extent_end - end);
 877                         btrfs_mark_buffer_dirty(leaf);
 878                         if (update_refs && disk_bytenr > 0)
 879                                 inode_sub_bytes(inode, end - key.offset);
 880                         break;
 881                 }
 882
 883                 search_start = extent_end;
 884                 /*
 885                  *       | ---- range to drop ----- |
 886                  *  | -------- extent -------- |
 887                  */
 888                 if (start > key.offset && end >= extent_end) {
 889                         BUG_ON(del_nr > 0);
 890                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 891                                 ret = -EOPNOTSUPP;
 892                                 break;
 893                         }
 894
 895                         btrfs_set_file_extent_num_bytes(leaf, fi,
 896                                                         start - key.offset);
 897                         btrfs_mark_buffer_dirty(leaf);
 898                         if (update_refs && disk_bytenr > 0)
 899                                 inode_sub_bytes(inode, extent_end - start);
 900                         if (end == extent_end)
 901                                 break;
 902
 903                         path->slots[0]++;
 904                         goto next_slot;
 905                 }
 906
 907                 /*
 908                  *  | ---- range to drop ----- |
 909                  *    | ------ extent ------ |
 910                  */
 911                 if (start <= key.offset && end >= extent_end) {
 912 delete_extent_item:
 913                         if (del_nr == 0) {
 914                                 del_slot = path->slots[0];
 915                                 del_nr = 1;
 916                         } else {
 917                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 918                                 del_nr++;
 919                         }
 920
 921                         if (update_refs &&
 922                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 923                                 inode_sub_bytes(inode,
 924                                                 extent_end - key.offset);
 925                                 extent_end = ALIGN(extent_end,
 926                                                    root->sectorsize);
 927                         } else if (update_refs && disk_bytenr > 0) {
 928                                 ret = btrfs_free_extent(trans, root,
 929                                                 disk_bytenr, num_bytes, 0,
 930                                                 root->root_key.objectid,
 931                                                 key.objectid, key.offset -
 932                                                 extent_offset, 0);
 933                                 BUG_ON(ret); /* -ENOMEM */
 934                                 inode_sub_bytes(inode,
 935                                                 extent_end - key.offset);
 936                         }
 937
 938                         if (end == extent_end)
 939                                 break;
 940
 941                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 942                                 path->slots[0]++;
 943                                 goto next_slot;
 944                         }
 945
 946                         ret = btrfs_del_items(trans, root, path, del_slot,
 947                                               del_nr);
 948                         if (ret) {
 949                                 btrfs_abort_transaction(trans, root, ret);
 950                                 break;
 951                         }
 952
 953                         del_nr = 0;
 954                         del_slot = 0;
 955
 956                         btrfs_release_path(path);
 957                         continue;
 958                 }
 959
 960                 BUG_ON(1);
 961         }
 962
 963         if (!ret && del_nr > 0) {
 964                 /*
 965                  * Set path->slots[0] to first slot, so that after the delete
 966                  * if items are move off from our leaf to its immediate left or
 967                  * right neighbor leafs, we end up with a correct and adjusted
 968                  * path->slots[0] for our insertion (if replace_extent != 0).
 969                  */
 970                 path->slots[0] = del_slot;
 971                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 972                 if (ret)
 973                         btrfs_abort_transaction(trans, root, ret);
 974         }
 975
 976         leaf = path->nodes[0];
 977         /*
 978          * If btrfs_del_items() was called, it might have deleted a leaf, in
 979          * which case it unlocked our path, so check path->locks[0] matches a
 980          * write lock.
 981          */
 982         if (!ret && replace_extent && leafs_visited == 1 &&
 983             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 984              path->locks[0] == BTRFS_WRITE_LOCK) &&
 985             btrfs_leaf_free_space(root, leaf) >=
 986             sizeof(struct btrfs_item) + extent_item_size) {
 987
 988                 key.objectid = ino;
 989                 key.type = BTRFS_EXTENT_DATA_KEY;
 990                 key.offset = start;
 991                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 992                         struct btrfs_key slot_key;
 993
 994                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 995                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
 996                                 path->slots[0]++;
 997                 }
 998                 setup_items_for_insert(root, path, &key,
 999                                        &extent_item_size,
1000                                        extent_item_size,
1001                                        sizeof(struct btrfs_item) +
1002                                        extent_item_size, 1);
1003                 *key_inserted = 1;
1004         }
1005
1006         if (!replace_extent || !(*key_inserted))
1007                 btrfs_release_path(path);
1008         if (drop_end)
1009                 *drop_end = found ? min(end, extent_end) : end;
1010         return ret;
1011 }
1012
1013 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1014                        struct btrfs_root *root, struct inode *inode, u64 start,
1015                        u64 end, int drop_cache)
1016 {
1017         struct btrfs_path *path;
1018         int ret;
1019
1020         path = btrfs_alloc_path();
1021         if (!path)
1022                 return -ENOMEM;
1023         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1024                                    drop_cache, 0, 0, NULL);
1025         btrfs_free_path(path);
1026         return ret;
1027 }
1028
1029 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1030                             u64 objectid, u64 bytenr, u64 orig_offset,
1031                             u64 *start, u64 *end)
1032 {
1033         struct btrfs_file_extent_item *fi;
1034         struct btrfs_key key;
1035         u64 extent_end;
1036
1037         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1038                 return 0;
1039
1040         btrfs_item_key_to_cpu(leaf, &key, slot);
1041         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1042                 return 0;
1043
1044         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1045         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1046             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1047             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1048             btrfs_file_extent_compression(leaf, fi) ||
1049             btrfs_file_extent_encryption(leaf, fi) ||
1050             btrfs_file_extent_other_encoding(leaf, fi))
1051                 return 0;
1052
1053         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1054         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1055                 return 0;
1056
1057         *start = key.offset;
1058         *end = extent_end;
1059         return 1;
1060 }
1061
1062 /*
1063  * Mark extent in the range start - end as written.
1064  *
1065  * This changes extent type from 'pre-allocated' to 'regular'. If only
1066  * part of extent is marked as written, the extent will be split into
1067  * two or three.
1068  */
1069 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1070                               struct inode *inode, u64 start, u64 end)
1071 {
1072         struct btrfs_root *root = BTRFS_I(inode)->root;
1073         struct extent_buffer *leaf;
1074         struct btrfs_path *path;
1075         struct btrfs_file_extent_item *fi;
1076         struct btrfs_key key;
1077         struct btrfs_key new_key;
1078         u64 bytenr;
1079         u64 num_bytes;
1080         u64 extent_end;
1081         u64 orig_offset;
1082         u64 other_start;
1083         u64 other_end;
1084         u64 split;
1085         int del_nr = 0;
1086         int del_slot = 0;
1087         int recow;
1088         int ret;
1089         u64 ino = btrfs_ino(inode);
1090
1091         path = btrfs_alloc_path();
1092         if (!path)
1093                 return -ENOMEM;
1094 again:
1095         recow = 0;
1096         split = start;
1097         key.objectid = ino;
1098         key.type = BTRFS_EXTENT_DATA_KEY;
1099         key.offset = split;
1100
1101         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1102         if (ret < 0)
1103                 goto out;
1104         if (ret > 0 && path->slots[0] > 0)
1105                 path->slots[0]--;
1106
1107         leaf = path->nodes[0];
1108         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1109         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1110         fi = btrfs_item_ptr(leaf, path->slots[0],
1111                             struct btrfs_file_extent_item);
1112         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1113                BTRFS_FILE_EXTENT_PREALLOC);
1114         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1115         BUG_ON(key.offset > start || extent_end < end);
1116
1117         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1118         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1119         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1120         memcpy(&new_key, &key, sizeof(new_key));
1121
1122         if (start == key.offset && end < extent_end) {
1123                 other_start = 0;
1124                 other_end = start;
1125                 if (extent_mergeable(leaf, path->slots[0] - 1,
1126                                      ino, bytenr, orig_offset,
1127                                      &other_start, &other_end)) {
1128                         new_key.offset = end;
1129                         btrfs_set_item_key_safe(root, path, &new_key);
1130                         fi = btrfs_item_ptr(leaf, path->slots[0],
1131                                             struct btrfs_file_extent_item);
1132                         btrfs_set_file_extent_generation(leaf, fi,
1133                                                          trans->transid);
1134                         btrfs_set_file_extent_num_bytes(leaf, fi,
1135                                                         extent_end - end);
1136                         btrfs_set_file_extent_offset(leaf, fi,
1137                                                      end - orig_offset);
1138                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1139                                             struct btrfs_file_extent_item);
1140                         btrfs_set_file_extent_generation(leaf, fi,
1141                                                          trans->transid);
1142                         btrfs_set_file_extent_num_bytes(leaf, fi,
1143                                                         end - other_start);
1144                         btrfs_mark_buffer_dirty(leaf);
1145                         goto out;
1146                 }
1147         }
1148
1149         if (start > key.offset && end == extent_end) {
1150                 other_start = end;
1151                 other_end = 0;
1152                 if (extent_mergeable(leaf, path->slots[0] + 1,
1153                                      ino, bytenr, orig_offset,
1154                                      &other_start, &other_end)) {
1155                         fi = btrfs_item_ptr(leaf, path->slots[0],
1156                                             struct btrfs_file_extent_item);
1157                         btrfs_set_file_extent_num_bytes(leaf, fi,
1158                                                         start - key.offset);
1159                         btrfs_set_file_extent_generation(leaf, fi,
1160                                                          trans->transid);
1161                         path->slots[0]++;
1162                         new_key.offset = start;
1163                         btrfs_set_item_key_safe(root, path, &new_key);
1164
1165                         fi = btrfs_item_ptr(leaf, path->slots[0],
1166                                             struct btrfs_file_extent_item);
1167                         btrfs_set_file_extent_generation(leaf, fi,
1168                                                          trans->transid);
1169                         btrfs_set_file_extent_num_bytes(leaf, fi,
1170                                                         other_end - start);
1171                         btrfs_set_file_extent_offset(leaf, fi,
1172                                                      start - orig_offset);
1173                         btrfs_mark_buffer_dirty(leaf);
1174                         goto out;
1175                 }
1176         }
1177
1178         while (start > key.offset || end < extent_end) {
1179                 if (key.offset == start)
1180                         split = end;
1181
1182                 new_key.offset = split;
1183                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1184                 if (ret == -EAGAIN) {
1185                         btrfs_release_path(path);
1186                         goto again;
1187                 }
1188                 if (ret < 0) {
1189                         btrfs_abort_transaction(trans, root, ret);
1190                         goto out;
1191                 }
1192
1193                 leaf = path->nodes[0];
1194                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1195                                     struct btrfs_file_extent_item);
1196                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1197                 btrfs_set_file_extent_num_bytes(leaf, fi,
1198                                                 split - key.offset);
1199
1200                 fi = btrfs_item_ptr(leaf, path->slots[0],
1201                                     struct btrfs_file_extent_item);
1202
1203                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1204                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1205                 btrfs_set_file_extent_num_bytes(leaf, fi,
1206                                                 extent_end - split);
1207                 btrfs_mark_buffer_dirty(leaf);
1208
1209                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1210                                            root->root_key.objectid,
1211                                            ino, orig_offset, 1);
1212                 BUG_ON(ret); /* -ENOMEM */
1213
1214                 if (split == start) {
1215                         key.offset = start;
1216                 } else {
1217                         BUG_ON(start != key.offset);
1218                         path->slots[0]--;
1219                         extent_end = end;
1220                 }
1221                 recow = 1;
1222         }
1223
1224         other_start = end;
1225         other_end = 0;
1226         if (extent_mergeable(leaf, path->slots[0] + 1,
1227                              ino, bytenr, orig_offset,
1228                              &other_start, &other_end)) {
1229                 if (recow) {
1230                         btrfs_release_path(path);
1231                         goto again;
1232                 }
1233                 extent_end = other_end;
1234                 del_slot = path->slots[0] + 1;
1235                 del_nr++;
1236                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1237                                         0, root->root_key.objectid,
1238                                         ino, orig_offset, 0);
1239                 BUG_ON(ret); /* -ENOMEM */
1240         }
1241         other_start = 0;
1242         other_end = start;
1243         if (extent_mergeable(leaf, path->slots[0] - 1,
1244                              ino, bytenr, orig_offset,
1245                              &other_start, &other_end)) {
1246                 if (recow) {
1247                         btrfs_release_path(path);
1248                         goto again;
1249                 }
1250                 key.offset = other_start;
1251                 del_slot = path->slots[0];
1252                 del_nr++;
1253                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1254                                         0, root->root_key.objectid,
1255                                         ino, orig_offset, 0);
1256                 BUG_ON(ret); /* -ENOMEM */
1257         }
1258         if (del_nr == 0) {
1259                 fi = btrfs_item_ptr(leaf, path->slots[0],
1260                            struct btrfs_file_extent_item);
1261                 btrfs_set_file_extent_type(leaf, fi,
1262                                            BTRFS_FILE_EXTENT_REG);
1263                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1264                 btrfs_mark_buffer_dirty(leaf);
1265         } else {
1266                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1267                            struct btrfs_file_extent_item);
1268                 btrfs_set_file_extent_type(leaf, fi,
1269                                            BTRFS_FILE_EXTENT_REG);
1270                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1271                 btrfs_set_file_extent_num_bytes(leaf, fi,
1272                                                 extent_end - key.offset);
1273                 btrfs_mark_buffer_dirty(leaf);
1274
1275                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1276                 if (ret < 0) {
1277                         btrfs_abort_transaction(trans, root, ret);
1278                         goto out;
1279                 }
1280         }
1281 out:
1282         btrfs_free_path(path);
1283         return 0;
1284 }
1285
1286 /*
1287  * on error we return an unlocked page and the error value
1288  * on success we return a locked page and 0
1289  */
1290 static int prepare_uptodate_page(struct page *page, u64 pos,
1291                                  bool force_uptodate)
1292 {
1293         int ret = 0;
1294
1295         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1296             !PageUptodate(page)) {
1297                 ret = btrfs_readpage(NULL, page);
1298                 if (ret)
1299                         return ret;
1300                 lock_page(page);
1301                 if (!PageUptodate(page)) {
1302                         unlock_page(page);
1303                         return -EIO;
1304                 }
1305         }
1306         return 0;
1307 }
1308
1309 /*
1310  * this just gets pages into the page cache and locks them down.
1311  */
1312 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1313                                   size_t num_pages, loff_t pos,
1314                                   size_t write_bytes, bool force_uptodate)
1315 {
1316         int i;
1317         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1318         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1319         int err = 0;
1320         int faili;
1321
1322         for (i = 0; i < num_pages; i++) {
1323                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1324                                                mask | __GFP_WRITE);
1325                 if (!pages[i]) {
1326                         faili = i - 1;
1327                         err = -ENOMEM;
1328                         goto fail;
1329                 }
1330
1331                 if (i == 0)
1332                         err = prepare_uptodate_page(pages[i], pos,
1333                                                     force_uptodate);
1334                 if (i == num_pages - 1)
1335                         err = prepare_uptodate_page(pages[i],
1336                                                     pos + write_bytes, false);
1337                 if (err) {
1338                         page_cache_release(pages[i]);
1339                         faili = i - 1;
1340                         goto fail;
1341                 }
1342                 wait_on_page_writeback(pages[i]);
1343         }
1344
1345         return 0;
1346 fail:
1347         while (faili >= 0) {
1348                 unlock_page(pages[faili]);
1349                 page_cache_release(pages[faili]);
1350                 faili--;
1351         }
1352         return err;
1353
1354 }
1355
1356 /*
1357  * This function locks the extent and properly waits for data=ordered extents
1358  * to finish before allowing the pages to be modified if need.
1359  *
1360  * The return value:
1361  * 1 - the extent is locked
1362  * 0 - the extent is not locked, and everything is OK
1363  * -EAGAIN - need re-prepare the pages
1364  * the other < 0 number - Something wrong happens
1365  */
1366 static noinline int
1367 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1368                                 size_t num_pages, loff_t pos,
1369                                 u64 *lockstart, u64 *lockend,
1370                                 struct extent_state **cached_state)
1371 {
1372         u64 start_pos;
1373         u64 last_pos;
1374         int i;
1375         int ret = 0;
1376
1377         start_pos = pos & ~((u64)PAGE_CACHE_SIZE - 1);
1378         last_pos = start_pos + ((u64)num_pages << PAGE_CACHE_SHIFT) - 1;
1379
1380         if (start_pos < inode->i_size) {
1381                 struct btrfs_ordered_extent *ordered;
1382                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1383                                  start_pos, last_pos, 0, cached_state);
1384                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1385                                                      last_pos - start_pos + 1);
1386                 if (ordered &&
1387                     ordered->file_offset + ordered->len > start_pos &&
1388                     ordered->file_offset <= last_pos) {
1389                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1390                                              start_pos, last_pos,
1391                                              cached_state, GFP_NOFS);
1392                         for (i = 0; i < num_pages; i++) {
1393                                 unlock_page(pages[i]);
1394                                 page_cache_release(pages[i]);
1395                         }
1396                         btrfs_start_ordered_extent(inode, ordered, 1);
1397                         btrfs_put_ordered_extent(ordered);
1398                         return -EAGAIN;
1399                 }
1400                 if (ordered)
1401                         btrfs_put_ordered_extent(ordered);
1402
1403                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1404                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1405                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1406                                   0, 0, cached_state, GFP_NOFS);
1407                 *lockstart = start_pos;
1408                 *lockend = last_pos;
1409                 ret = 1;
1410         }
1411
1412         for (i = 0; i < num_pages; i++) {
1413                 if (clear_page_dirty_for_io(pages[i]))
1414                         account_page_redirty(pages[i]);
1415                 set_page_extent_mapped(pages[i]);
1416                 WARN_ON(!PageLocked(pages[i]));
1417         }
1418
1419         return ret;
1420 }
1421
1422 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1423                                     size_t *write_bytes)
1424 {
1425         struct btrfs_root *root = BTRFS_I(inode)->root;
1426         struct btrfs_ordered_extent *ordered;
1427         u64 lockstart, lockend;
1428         u64 num_bytes;
1429         int ret;
1430
1431         ret = btrfs_start_nocow_write(root);
1432         if (!ret)
1433                 return -ENOSPC;
1434
1435         lockstart = round_down(pos, root->sectorsize);
1436         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1437
1438         while (1) {
1439                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1440                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1441                                                      lockend - lockstart + 1);
1442                 if (!ordered) {
1443                         break;
1444                 }
1445                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1446                 btrfs_start_ordered_extent(inode, ordered, 1);
1447                 btrfs_put_ordered_extent(ordered);
1448         }
1449
1450         num_bytes = lockend - lockstart + 1;
1451         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1452         if (ret <= 0) {
1453                 ret = 0;
1454                 btrfs_end_nocow_write(root);
1455         } else {
1456                 *write_bytes = min_t(size_t, *write_bytes ,
1457                                      num_bytes - pos + lockstart);
1458         }
1459
1460         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1461
1462         return ret;
1463 }
1464
1465 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1466                                                struct iov_iter *i,
1467                                                loff_t pos)
1468 {
1469         struct inode *inode = file_inode(file);
1470         struct btrfs_root *root = BTRFS_I(inode)->root;
1471         struct page **pages = NULL;
1472         struct extent_state *cached_state = NULL;
1473         u64 release_bytes = 0;
1474         u64 lockstart;
1475         u64 lockend;
1476         unsigned long first_index;
1477         size_t num_written = 0;
1478         int nrptrs;
1479         int ret = 0;
1480         bool only_release_metadata = false;
1481         bool force_page_uptodate = false;
1482         bool need_unlock;
1483
1484         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1485                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1486                      (sizeof(struct page *)));
1487         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1488         nrptrs = max(nrptrs, 8);
1489         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1490         if (!pages)
1491                 return -ENOMEM;
1492
1493         first_index = pos >> PAGE_CACHE_SHIFT;
1494
1495         while (iov_iter_count(i) > 0) {
1496                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1497                 size_t write_bytes = min(iov_iter_count(i),
1498                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1499                                          offset);
1500                 size_t num_pages = (write_bytes + offset +
1501                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1502                 size_t reserve_bytes;
1503                 size_t dirty_pages;
1504                 size_t copied;
1505
1506                 WARN_ON(num_pages > nrptrs);
1507
1508                 /*
1509                  * Fault pages before locking them in prepare_pages
1510                  * to avoid recursive lock
1511                  */
1512                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1513                         ret = -EFAULT;
1514                         break;
1515                 }
1516
1517                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1518                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1519                 if (ret == -ENOSPC &&
1520                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1521                                               BTRFS_INODE_PREALLOC))) {
1522                         ret = check_can_nocow(inode, pos, &write_bytes);
1523                         if (ret > 0) {
1524                                 only_release_metadata = true;
1525                                 /*
1526                                  * our prealloc extent may be smaller than
1527                                  * write_bytes, so scale down.
1528                                  */
1529                                 num_pages = (write_bytes + offset +
1530                                              PAGE_CACHE_SIZE - 1) >>
1531                                         PAGE_CACHE_SHIFT;
1532                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1533                                 ret = 0;
1534                         } else {
1535                                 ret = -ENOSPC;
1536                         }
1537                 }
1538
1539                 if (ret)
1540                         break;
1541
1542                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1543                 if (ret) {
1544                         if (!only_release_metadata)
1545                                 btrfs_free_reserved_data_space(inode,
1546                                                                reserve_bytes);
1547                         else
1548                                 btrfs_end_nocow_write(root);
1549                         break;
1550                 }
1551
1552                 release_bytes = reserve_bytes;
1553                 need_unlock = false;
1554 again:
1555                 /*
1556                  * This is going to setup the pages array with the number of
1557                  * pages we want, so we don't really need to worry about the
1558                  * contents of pages from loop to loop
1559                  */
1560                 ret = prepare_pages(inode, pages, num_pages,
1561                                     pos, write_bytes,
1562                                     force_page_uptodate);
1563                 if (ret)
1564                         break;
1565
1566                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1567                                                       pos, &lockstart, &lockend,
1568                                                       &cached_state);
1569                 if (ret < 0) {
1570                         if (ret == -EAGAIN)
1571                                 goto again;
1572                         break;
1573                 } else if (ret > 0) {
1574                         need_unlock = true;
1575                         ret = 0;
1576                 }
1577
1578                 copied = btrfs_copy_from_user(pos, num_pages,
1579                                            write_bytes, pages, i);
1580
1581                 /*
1582                  * if we have trouble faulting in the pages, fall
1583                  * back to one page at a time
1584                  */
1585                 if (copied < write_bytes)
1586                         nrptrs = 1;
1587
1588                 if (copied == 0) {
1589                         force_page_uptodate = true;
1590                         dirty_pages = 0;
1591                 } else {
1592                         force_page_uptodate = false;
1593                         dirty_pages = (copied + offset +
1594                                        PAGE_CACHE_SIZE - 1) >>
1595                                        PAGE_CACHE_SHIFT;
1596                 }
1597
1598                 /*
1599                  * If we had a short copy we need to release the excess delaloc
1600                  * bytes we reserved.  We need to increment outstanding_extents
1601                  * because btrfs_delalloc_release_space will decrement it, but
1602                  * we still have an outstanding extent for the chunk we actually
1603                  * managed to copy.
1604                  */
1605                 if (num_pages > dirty_pages) {
1606                         release_bytes = (num_pages - dirty_pages) <<
1607                                 PAGE_CACHE_SHIFT;
1608                         if (copied > 0) {
1609                                 spin_lock(&BTRFS_I(inode)->lock);
1610                                 BTRFS_I(inode)->outstanding_extents++;
1611                                 spin_unlock(&BTRFS_I(inode)->lock);
1612                         }
1613                         if (only_release_metadata)
1614                                 btrfs_delalloc_release_metadata(inode,
1615                                                                 release_bytes);
1616                         else
1617                                 btrfs_delalloc_release_space(inode,
1618                                                              release_bytes);
1619                 }
1620
1621                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1622
1623                 if (copied > 0)
1624                         ret = btrfs_dirty_pages(root, inode, pages,
1625                                                 dirty_pages, pos, copied,
1626                                                 NULL);
1627                 if (need_unlock)
1628                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1629                                              lockstart, lockend, &cached_state,
1630                                              GFP_NOFS);
1631                 if (ret) {
1632                         btrfs_drop_pages(pages, num_pages);
1633                         break;
1634                 }
1635
1636                 release_bytes = 0;
1637                 if (only_release_metadata)
1638                         btrfs_end_nocow_write(root);
1639
1640                 if (only_release_metadata && copied > 0) {
1641                         u64 lockstart = round_down(pos, root->sectorsize);
1642                         u64 lockend = lockstart +
1643                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1644
1645                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1646                                        lockend, EXTENT_NORESERVE, NULL,
1647                                        NULL, GFP_NOFS);
1648                         only_release_metadata = false;
1649                 }
1650
1651                 btrfs_drop_pages(pages, num_pages);
1652
1653                 cond_resched();
1654
1655                 balance_dirty_pages_ratelimited(inode->i_mapping);
1656                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1657                         btrfs_btree_balance_dirty(root);
1658
1659                 pos += copied;
1660                 num_written += copied;
1661         }
1662
1663         kfree(pages);
1664
1665         if (release_bytes) {
1666                 if (only_release_metadata) {
1667                         btrfs_end_nocow_write(root);
1668                         btrfs_delalloc_release_metadata(inode, release_bytes);
1669                 } else {
1670                         btrfs_delalloc_release_space(inode, release_bytes);
1671                 }
1672         }
1673
1674         return num_written ? num_written : ret;
1675 }
1676
1677 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1678                                     const struct iovec *iov,
1679                                     unsigned long nr_segs, loff_t pos,
1680                                     size_t count, size_t ocount)
1681 {
1682         struct file *file = iocb->ki_filp;
1683         struct iov_iter i;
1684         ssize_t written;
1685         ssize_t written_buffered;
1686         loff_t endbyte;
1687         int err;
1688
1689         written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1690                                             count, ocount);
1691
1692         if (written < 0 || written == count)
1693                 return written;
1694
1695         pos += written;
1696         count -= written;
1697         iov_iter_init(&i, iov, nr_segs, count, written);
1698         written_buffered = __btrfs_buffered_write(file, &i, pos);
1699         if (written_buffered < 0) {
1700                 err = written_buffered;
1701                 goto out;
1702         }
1703         endbyte = pos + written_buffered - 1;
1704         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1705         if (err)
1706                 goto out;
1707         written += written_buffered;
1708         iocb->ki_pos = pos + written_buffered;
1709         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1710                                  endbyte >> PAGE_CACHE_SHIFT);
1711 out:
1712         return written ? written : err;
1713 }
1714
1715 static void update_time_for_write(struct inode *inode)
1716 {
1717         struct timespec now;
1718
1719         if (IS_NOCMTIME(inode))
1720                 return;
1721
1722         now = current_fs_time(inode->i_sb);
1723         if (!timespec_equal(&inode->i_mtime, &now))
1724                 inode->i_mtime = now;
1725
1726         if (!timespec_equal(&inode->i_ctime, &now))
1727                 inode->i_ctime = now;
1728
1729         if (IS_I_VERSION(inode))
1730                 inode_inc_iversion(inode);
1731 }
1732
1733 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1734                                     const struct iovec *iov,
1735                                     unsigned long nr_segs, loff_t pos)
1736 {
1737         struct file *file = iocb->ki_filp;
1738         struct inode *inode = file_inode(file);
1739         struct btrfs_root *root = BTRFS_I(inode)->root;
1740         u64 start_pos;
1741         u64 end_pos;
1742         ssize_t num_written = 0;
1743         ssize_t err = 0;
1744         size_t count, ocount;
1745         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1746
1747         mutex_lock(&inode->i_mutex);
1748
1749         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1750         if (err) {
1751                 mutex_unlock(&inode->i_mutex);
1752                 goto out;
1753         }
1754         count = ocount;
1755
1756         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1757         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1758         if (err) {
1759                 mutex_unlock(&inode->i_mutex);
1760                 goto out;
1761         }
1762
1763         if (count == 0) {
1764                 mutex_unlock(&inode->i_mutex);
1765                 goto out;
1766         }
1767
1768         err = file_remove_suid(file);
1769         if (err) {
1770                 mutex_unlock(&inode->i_mutex);
1771                 goto out;
1772         }
1773
1774         /*
1775          * If BTRFS flips readonly due to some impossible error
1776          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1777          * although we have opened a file as writable, we have
1778          * to stop this write operation to ensure FS consistency.
1779          */
1780         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1781                 mutex_unlock(&inode->i_mutex);
1782                 err = -EROFS;
1783                 goto out;
1784         }
1785
1786         /*
1787          * We reserve space for updating the inode when we reserve space for the
1788          * extent we are going to write, so we will enospc out there.  We don't
1789          * need to start yet another transaction to update the inode as we will
1790          * update the inode when we finish writing whatever data we write.
1791          */
1792         update_time_for_write(inode);
1793
1794         start_pos = round_down(pos, root->sectorsize);
1795         if (start_pos > i_size_read(inode)) {
1796                 /* Expand hole size to cover write data, preventing empty gap */
1797                 end_pos = round_up(pos + count, root->sectorsize);
1798                 err = btrfs_cont_expand(inode, i_size_read(inode), end_pos);
1799                 if (err) {
1800                         mutex_unlock(&inode->i_mutex);
1801                         goto out;
1802                 }
1803         }
1804
1805         if (sync)
1806                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1807
1808         if (unlikely(file->f_flags & O_DIRECT)) {
1809                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1810                                                    pos, count, ocount);
1811         } else {
1812                 struct iov_iter i;
1813
1814                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1815
1816                 num_written = __btrfs_buffered_write(file, &i, pos);
1817                 if (num_written > 0)
1818                         iocb->ki_pos = pos + num_written;
1819         }
1820
1821         mutex_unlock(&inode->i_mutex);
1822
1823         /*
1824          * we want to make sure fsync finds this change
1825          * but we haven't joined a transaction running right now.
1826          *
1827          * Later on, someone is sure to update the inode and get the
1828          * real transid recorded.
1829          *
1830          * We set last_trans now to the fs_info generation + 1,
1831          * this will either be one more than the running transaction
1832          * or the generation used for the next transaction if there isn't
1833          * one running right now.
1834          *
1835          * We also have to set last_sub_trans to the current log transid,
1836          * otherwise subsequent syncs to a file that's been synced in this
1837          * transaction will appear to have already occured.
1838          */
1839         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1840         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1841         if (num_written > 0) {
1842                 err = generic_write_sync(file, pos, num_written);
1843                 if (err < 0)
1844                         num_written = err;
1845         }
1846
1847         if (sync)
1848                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1849 out:
1850         current->backing_dev_info = NULL;
1851         return num_written ? num_written : err;
1852 }
1853
1854 int btrfs_release_file(struct inode *inode, struct file *filp)
1855 {
1856         /*
1857          * ordered_data_close is set by settattr when we are about to truncate
1858          * a file from a non-zero size to a zero size.  This tries to
1859          * flush down new bytes that may have been written if the
1860          * application were using truncate to replace a file in place.
1861          */
1862         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1863                                &BTRFS_I(inode)->runtime_flags)) {
1864                 struct btrfs_trans_handle *trans;
1865                 struct btrfs_root *root = BTRFS_I(inode)->root;
1866
1867                 /*
1868                  * We need to block on a committing transaction to keep us from
1869                  * throwing a ordered operation on to the list and causing
1870                  * something like sync to deadlock trying to flush out this
1871                  * inode.
1872                  */
1873                 trans = btrfs_start_transaction(root, 0);
1874                 if (IS_ERR(trans))
1875                         return PTR_ERR(trans);
1876                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1877                 btrfs_end_transaction(trans, root);
1878                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1879                         filemap_flush(inode->i_mapping);
1880         }
1881         if (filp->private_data)
1882                 btrfs_ioctl_trans_end(filp);
1883         return 0;
1884 }
1885
1886 /*
1887  * fsync call for both files and directories.  This logs the inode into
1888  * the tree log instead of forcing full commits whenever possible.
1889  *
1890  * It needs to call filemap_fdatawait so that all ordered extent updates are
1891  * in the metadata btree are up to date for copying to the log.
1892  *
1893  * It drops the inode mutex before doing the tree log commit.  This is an
1894  * important optimization for directories because holding the mutex prevents
1895  * new operations on the dir while we write to disk.
1896  */
1897 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1898 {
1899         struct dentry *dentry = file->f_path.dentry;
1900         struct inode *inode = dentry->d_inode;
1901         struct btrfs_root *root = BTRFS_I(inode)->root;
1902         struct btrfs_trans_handle *trans;
1903         struct btrfs_log_ctx ctx;
1904         int ret = 0;
1905         bool full_sync = 0;
1906
1907         trace_btrfs_sync_file(file, datasync);
1908
1909         /*
1910          * We write the dirty pages in the range and wait until they complete
1911          * out of the ->i_mutex. If so, we can flush the dirty pages by
1912          * multi-task, and make the performance up.  See
1913          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1914          */
1915         atomic_inc(&BTRFS_I(inode)->sync_writers);
1916         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1917         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1918                              &BTRFS_I(inode)->runtime_flags))
1919                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1920         atomic_dec(&BTRFS_I(inode)->sync_writers);
1921         if (ret)
1922                 return ret;
1923
1924         mutex_lock(&inode->i_mutex);
1925
1926         /*
1927          * We flush the dirty pages again to avoid some dirty pages in the
1928          * range being left.
1929          */
1930         atomic_inc(&root->log_batch);
1931         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1932                              &BTRFS_I(inode)->runtime_flags);
1933         if (full_sync) {
1934                 ret = btrfs_wait_ordered_range(inode, start, end - start + 1);
1935                 if (ret) {
1936                         mutex_unlock(&inode->i_mutex);
1937                         goto out;
1938                 }
1939         }
1940         atomic_inc(&root->log_batch);
1941
1942         /*
1943          * check the transaction that last modified this inode
1944          * and see if its already been committed
1945          */
1946         if (!BTRFS_I(inode)->last_trans) {
1947                 mutex_unlock(&inode->i_mutex);
1948                 goto out;
1949         }
1950
1951         /*
1952          * if the last transaction that changed this file was before
1953          * the current transaction, we can bail out now without any
1954          * syncing
1955          */
1956         smp_mb();
1957         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1958             BTRFS_I(inode)->last_trans <=
1959             root->fs_info->last_trans_committed) {
1960                 BTRFS_I(inode)->last_trans = 0;
1961
1962                 /*
1963                  * We'v had everything committed since the last time we were
1964                  * modified so clear this flag in case it was set for whatever
1965                  * reason, it's no longer relevant.
1966                  */
1967                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1968                           &BTRFS_I(inode)->runtime_flags);
1969                 mutex_unlock(&inode->i_mutex);
1970                 goto out;
1971         }
1972
1973         /*
1974          * ok we haven't committed the transaction yet, lets do a commit
1975          */
1976         if (file->private_data)
1977                 btrfs_ioctl_trans_end(file);
1978
1979         /*
1980          * We use start here because we will need to wait on the IO to complete
1981          * in btrfs_sync_log, which could require joining a transaction (for
1982          * example checking cross references in the nocow path).  If we use join
1983          * here we could get into a situation where we're waiting on IO to
1984          * happen that is blocked on a transaction trying to commit.  With start
1985          * we inc the extwriter counter, so we wait for all extwriters to exit
1986          * before we start blocking join'ers.  This comment is to keep somebody
1987          * from thinking they are super smart and changing this to
1988          * btrfs_join_transaction *cough*Josef*cough*.
1989          */
1990         trans = btrfs_start_transaction(root, 0);
1991         if (IS_ERR(trans)) {
1992                 ret = PTR_ERR(trans);
1993                 mutex_unlock(&inode->i_mutex);
1994                 goto out;
1995         }
1996         trans->sync = true;
1997
1998         btrfs_init_log_ctx(&ctx);
1999
2000         ret = btrfs_log_dentry_safe(trans, root, dentry, &ctx);
2001         if (ret < 0) {
2002                 /* Fallthrough and commit/free transaction. */
2003                 ret = 1;
2004         }
2005
2006         /* we've logged all the items and now have a consistent
2007          * version of the file in the log.  It is possible that
2008          * someone will come in and modify the file, but that's
2009          * fine because the log is consistent on disk, and we
2010          * have references to all of the file's extents
2011          *
2012          * It is possible that someone will come in and log the
2013          * file again, but that will end up using the synchronization
2014          * inside btrfs_sync_log to keep things safe.
2015          */
2016         mutex_unlock(&inode->i_mutex);
2017
2018         if (ret != BTRFS_NO_LOG_SYNC) {
2019                 if (!ret) {
2020                         ret = btrfs_sync_log(trans, root, &ctx);
2021                         if (!ret) {
2022                                 ret = btrfs_end_transaction(trans, root);
2023                                 goto out;
2024                         }
2025                 }
2026                 if (!full_sync) {
2027                         ret = btrfs_wait_ordered_range(inode, start,
2028                                                        end - start + 1);
2029                         if (ret) {
2030                                 btrfs_end_transaction(trans, root);
2031                                 goto out;
2032                         }
2033                 }
2034                 ret = btrfs_commit_transaction(trans, root);
2035         } else {
2036                 ret = btrfs_end_transaction(trans, root);
2037         }
2038 out:
2039         return ret > 0 ? -EIO : ret;
2040 }
2041
2042 static const struct vm_operations_struct btrfs_file_vm_ops = {
2043         .fault          = filemap_fault,
2044         .map_pages      = filemap_map_pages,
2045         .page_mkwrite   = btrfs_page_mkwrite,
2046         .remap_pages    = generic_file_remap_pages,
2047 };
2048
2049 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2050 {
2051         struct address_space *mapping = filp->f_mapping;
2052
2053         if (!mapping->a_ops->readpage)
2054                 return -ENOEXEC;
2055
2056         file_accessed(filp);
2057         vma->vm_ops = &btrfs_file_vm_ops;
2058
2059         return 0;
2060 }
2061
2062 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2063                           int slot, u64 start, u64 end)
2064 {
2065         struct btrfs_file_extent_item *fi;
2066         struct btrfs_key key;
2067
2068         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2069                 return 0;
2070
2071         btrfs_item_key_to_cpu(leaf, &key, slot);
2072         if (key.objectid != btrfs_ino(inode) ||
2073             key.type != BTRFS_EXTENT_DATA_KEY)
2074                 return 0;
2075
2076         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2077
2078         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2079                 return 0;
2080
2081         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2082                 return 0;
2083
2084         if (key.offset == end)
2085                 return 1;
2086         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2087                 return 1;
2088         return 0;
2089 }
2090
2091 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2092                       struct btrfs_path *path, u64 offset, u64 end)
2093 {
2094         struct btrfs_root *root = BTRFS_I(inode)->root;
2095         struct extent_buffer *leaf;
2096         struct btrfs_file_extent_item *fi;
2097         struct extent_map *hole_em;
2098         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2099         struct btrfs_key key;
2100         int ret;
2101
2102         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2103                 goto out;
2104
2105         key.objectid = btrfs_ino(inode);
2106         key.type = BTRFS_EXTENT_DATA_KEY;
2107         key.offset = offset;
2108
2109         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2110         if (ret < 0)
2111                 return ret;
2112         BUG_ON(!ret);
2113
2114         leaf = path->nodes[0];
2115         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2116                 u64 num_bytes;
2117
2118                 path->slots[0]--;
2119                 fi = btrfs_item_ptr(leaf, path->slots[0],
2120                                     struct btrfs_file_extent_item);
2121                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2122                         end - offset;
2123                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2124                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2125                 btrfs_set_file_extent_offset(leaf, fi, 0);
2126                 btrfs_mark_buffer_dirty(leaf);
2127                 goto out;
2128         }
2129
2130         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
2131                 u64 num_bytes;
2132
2133                 path->slots[0]++;
2134                 key.offset = offset;
2135                 btrfs_set_item_key_safe(root, path, &key);
2136                 fi = btrfs_item_ptr(leaf, path->slots[0],
2137                                     struct btrfs_file_extent_item);
2138                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2139                         offset;
2140                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2141                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2142                 btrfs_set_file_extent_offset(leaf, fi, 0);
2143                 btrfs_mark_buffer_dirty(leaf);
2144                 goto out;
2145         }
2146         btrfs_release_path(path);
2147
2148         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2149                                        0, 0, end - offset, 0, end - offset,
2150                                        0, 0, 0);
2151         if (ret)
2152                 return ret;
2153
2154 out:
2155         btrfs_release_path(path);
2156
2157         hole_em = alloc_extent_map();
2158         if (!hole_em) {
2159                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2160                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2161                         &BTRFS_I(inode)->runtime_flags);
2162         } else {
2163                 hole_em->start = offset;
2164                 hole_em->len = end - offset;
2165                 hole_em->ram_bytes = hole_em->len;
2166                 hole_em->orig_start = offset;
2167
2168                 hole_em->block_start = EXTENT_MAP_HOLE;
2169                 hole_em->block_len = 0;
2170                 hole_em->orig_block_len = 0;
2171                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2172                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2173                 hole_em->generation = trans->transid;
2174
2175                 do {
2176                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2177                         write_lock(&em_tree->lock);
2178                         ret = add_extent_mapping(em_tree, hole_em, 1);
2179                         write_unlock(&em_tree->lock);
2180                 } while (ret == -EEXIST);
2181                 free_extent_map(hole_em);
2182                 if (ret)
2183                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2184                                 &BTRFS_I(inode)->runtime_flags);
2185         }
2186
2187         return 0;
2188 }
2189
2190 /*
2191  * Find a hole extent on given inode and change start/len to the end of hole
2192  * extent.(hole/vacuum extent whose em->start <= start &&
2193  *         em->start + em->len > start)
2194  * When a hole extent is found, return 1 and modify start/len.
2195  */
2196 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2197 {
2198         struct extent_map *em;
2199         int ret = 0;
2200
2201         em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2202         if (IS_ERR_OR_NULL(em)) {
2203                 if (!em)
2204                         ret = -ENOMEM;
2205                 else
2206                         ret = PTR_ERR(em);
2207                 return ret;
2208         }
2209
2210         /* Hole or vacuum extent(only exists in no-hole mode) */
2211         if (em->block_start == EXTENT_MAP_HOLE) {
2212                 ret = 1;
2213                 *len = em->start + em->len > *start + *len ?
2214                        0 : *start + *len - em->start - em->len;
2215                 *start = em->start + em->len;
2216         }
2217         free_extent_map(em);
2218         return ret;
2219 }
2220
2221 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2222 {
2223         struct btrfs_root *root = BTRFS_I(inode)->root;
2224         struct extent_state *cached_state = NULL;
2225         struct btrfs_path *path;
2226         struct btrfs_block_rsv *rsv;
2227         struct btrfs_trans_handle *trans;
2228         u64 lockstart;
2229         u64 lockend;
2230         u64 tail_start;
2231         u64 tail_len;
2232         u64 orig_start = offset;
2233         u64 cur_offset;
2234         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2235         u64 drop_end;
2236         int ret = 0;
2237         int err = 0;
2238         int rsv_count;
2239         bool same_page;
2240         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2241         u64 ino_size;
2242
2243         ret = btrfs_wait_ordered_range(inode, offset, len);
2244         if (ret)
2245                 return ret;
2246
2247         mutex_lock(&inode->i_mutex);
2248         ino_size = round_up(inode->i_size, PAGE_CACHE_SIZE);
2249         ret = find_first_non_hole(inode, &offset, &len);
2250         if (ret < 0)
2251                 goto out_only_mutex;
2252         if (ret && !len) {
2253                 /* Already in a large hole */
2254                 ret = 0;
2255                 goto out_only_mutex;
2256         }
2257
2258         lockstart = round_up(offset , BTRFS_I(inode)->root->sectorsize);
2259         lockend = round_down(offset + len,
2260                              BTRFS_I(inode)->root->sectorsize) - 1;
2261         same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2262                     ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2263
2264         /*
2265          * We needn't truncate any page which is beyond the end of the file
2266          * because we are sure there is no data there.
2267          */
2268         /*
2269          * Only do this if we are in the same page and we aren't doing the
2270          * entire page.
2271          */
2272         if (same_page && len < PAGE_CACHE_SIZE) {
2273                 if (offset < ino_size)
2274                         ret = btrfs_truncate_page(inode, offset, len, 0);
2275                 goto out_only_mutex;
2276         }
2277
2278         /* zero back part of the first page */
2279         if (offset < ino_size) {
2280                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2281                 if (ret) {
2282                         mutex_unlock(&inode->i_mutex);
2283                         return ret;
2284                 }
2285         }
2286
2287         /* Check the aligned pages after the first unaligned page,
2288          * if offset != orig_start, which means the first unaligned page
2289          * including serveral following pages are already in holes,
2290          * the extra check can be skipped */
2291         if (offset == orig_start) {
2292                 /* after truncate page, check hole again */
2293                 len = offset + len - lockstart;
2294                 offset = lockstart;
2295                 ret = find_first_non_hole(inode, &offset, &len);
2296                 if (ret < 0)
2297                         goto out_only_mutex;
2298                 if (ret && !len) {
2299                         ret = 0;
2300                         goto out_only_mutex;
2301                 }
2302                 lockstart = offset;
2303         }
2304
2305         /* Check the tail unaligned part is in a hole */
2306         tail_start = lockend + 1;
2307         tail_len = offset + len - tail_start;
2308         if (tail_len) {
2309                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2310                 if (unlikely(ret < 0))
2311                         goto out_only_mutex;
2312                 if (!ret) {
2313                         /* zero the front end of the last page */
2314                         if (tail_start + tail_len < ino_size) {
2315                                 ret = btrfs_truncate_page(inode,
2316                                                 tail_start + tail_len, 0, 1);
2317                                 if (ret)
2318                                         goto out_only_mutex;
2319                                 }
2320                 }
2321         }
2322
2323         if (lockend < lockstart) {
2324                 mutex_unlock(&inode->i_mutex);
2325                 return 0;
2326         }
2327
2328         while (1) {
2329                 struct btrfs_ordered_extent *ordered;
2330
2331                 truncate_pagecache_range(inode, lockstart, lockend);
2332
2333                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2334                                  0, &cached_state);
2335                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2336
2337                 /*
2338                  * We need to make sure we have no ordered extents in this range
2339                  * and nobody raced in and read a page in this range, if we did
2340                  * we need to try again.
2341                  */
2342                 if ((!ordered ||
2343                     (ordered->file_offset + ordered->len <= lockstart ||
2344                      ordered->file_offset > lockend)) &&
2345                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2346                         if (ordered)
2347                                 btrfs_put_ordered_extent(ordered);
2348                         break;
2349                 }
2350                 if (ordered)
2351                         btrfs_put_ordered_extent(ordered);
2352                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2353                                      lockend, &cached_state, GFP_NOFS);
2354                 ret = btrfs_wait_ordered_range(inode, lockstart,
2355                                                lockend - lockstart + 1);
2356                 if (ret) {
2357                         mutex_unlock(&inode->i_mutex);
2358                         return ret;
2359                 }
2360         }
2361
2362         path = btrfs_alloc_path();
2363         if (!path) {
2364                 ret = -ENOMEM;
2365                 goto out;
2366         }
2367
2368         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2369         if (!rsv) {
2370                 ret = -ENOMEM;
2371                 goto out_free;
2372         }
2373         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2374         rsv->failfast = 1;
2375
2376         /*
2377          * 1 - update the inode
2378          * 1 - removing the extents in the range
2379          * 1 - adding the hole extent if no_holes isn't set
2380          */
2381         rsv_count = no_holes ? 2 : 3;
2382         trans = btrfs_start_transaction(root, rsv_count);
2383         if (IS_ERR(trans)) {
2384                 err = PTR_ERR(trans);
2385                 goto out_free;
2386         }
2387
2388         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2389                                       min_size);
2390         BUG_ON(ret);
2391         trans->block_rsv = rsv;
2392
2393         cur_offset = lockstart;
2394         len = lockend - cur_offset;
2395         while (cur_offset < lockend) {
2396                 ret = __btrfs_drop_extents(trans, root, inode, path,
2397                                            cur_offset, lockend + 1,
2398                                            &drop_end, 1, 0, 0, NULL);
2399                 if (ret != -ENOSPC)
2400                         break;
2401
2402                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2403
2404                 if (cur_offset < ino_size) {
2405                         ret = fill_holes(trans, inode, path, cur_offset,
2406                                          drop_end);
2407                         if (ret) {
2408                                 err = ret;
2409                                 break;
2410                         }
2411                 }
2412
2413                 cur_offset = drop_end;
2414
2415                 ret = btrfs_update_inode(trans, root, inode);
2416                 if (ret) {
2417                         err = ret;
2418                         break;
2419                 }
2420
2421                 btrfs_end_transaction(trans, root);
2422                 btrfs_btree_balance_dirty(root);
2423
2424                 trans = btrfs_start_transaction(root, rsv_count);
2425                 if (IS_ERR(trans)) {
2426                         ret = PTR_ERR(trans);
2427                         trans = NULL;
2428                         break;
2429                 }
2430
2431                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2432                                               rsv, min_size);
2433                 BUG_ON(ret);    /* shouldn't happen */
2434                 trans->block_rsv = rsv;
2435
2436                 ret = find_first_non_hole(inode, &cur_offset, &len);
2437                 if (unlikely(ret < 0))
2438                         break;
2439                 if (ret && !len) {
2440                         ret = 0;
2441                         break;
2442                 }
2443         }
2444
2445         if (ret) {
2446                 err = ret;
2447                 goto out_trans;
2448         }
2449
2450         trans->block_rsv = &root->fs_info->trans_block_rsv;
2451         /*
2452          * Don't insert file hole extent item if it's for a range beyond eof
2453          * (because it's useless) or if it represents a 0 bytes range (when
2454          * cur_offset == drop_end).
2455          */
2456         if (cur_offset < ino_size && cur_offset < drop_end) {
2457                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2458                 if (ret) {
2459                         err = ret;
2460                         goto out_trans;
2461                 }
2462         }
2463
2464 out_trans:
2465         if (!trans)
2466                 goto out_free;
2467
2468         inode_inc_iversion(inode);
2469         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2470
2471         trans->block_rsv = &root->fs_info->trans_block_rsv;
2472         ret = btrfs_update_inode(trans, root, inode);
2473         btrfs_end_transaction(trans, root);
2474         btrfs_btree_balance_dirty(root);
2475 out_free:
2476         btrfs_free_path(path);
2477         btrfs_free_block_rsv(root, rsv);
2478 out:
2479         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2480                              &cached_state, GFP_NOFS);
2481 out_only_mutex:
2482         mutex_unlock(&inode->i_mutex);
2483         if (ret && !err)
2484                 err = ret;
2485         return err;
2486 }
2487
2488 static long btrfs_fallocate(struct file *file, int mode,
2489                             loff_t offset, loff_t len)
2490 {
2491         struct inode *inode = file_inode(file);
2492         struct extent_state *cached_state = NULL;
2493         struct btrfs_root *root = BTRFS_I(inode)->root;
2494         u64 cur_offset;
2495         u64 last_byte;
2496         u64 alloc_start;
2497         u64 alloc_end;
2498         u64 alloc_hint = 0;
2499         u64 locked_end;
2500         struct extent_map *em;
2501         int blocksize = BTRFS_I(inode)->root->sectorsize;
2502         int ret;
2503
2504         alloc_start = round_down(offset, blocksize);
2505         alloc_end = round_up(offset + len, blocksize);
2506
2507         /* Make sure we aren't being give some crap mode */
2508         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2509                 return -EOPNOTSUPP;
2510
2511         if (mode & FALLOC_FL_PUNCH_HOLE)
2512                 return btrfs_punch_hole(inode, offset, len);
2513
2514         /*
2515          * Make sure we have enough space before we do the
2516          * allocation.
2517          */
2518         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2519         if (ret)
2520                 return ret;
2521         if (root->fs_info->quota_enabled) {
2522                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2523                 if (ret)
2524                         goto out_reserve_fail;
2525         }
2526
2527         mutex_lock(&inode->i_mutex);
2528         ret = inode_newsize_ok(inode, alloc_end);
2529         if (ret)
2530                 goto out;
2531
2532         if (alloc_start > inode->i_size) {
2533                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2534                                         alloc_start);
2535                 if (ret)
2536                         goto out;
2537         } else {
2538                 /*
2539                  * If we are fallocating from the end of the file onward we
2540                  * need to zero out the end of the page if i_size lands in the
2541                  * middle of a page.
2542                  */
2543                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2544                 if (ret)
2545                         goto out;
2546         }
2547
2548         /*
2549          * wait for ordered IO before we have any locks.  We'll loop again
2550          * below with the locks held.
2551          */
2552         ret = btrfs_wait_ordered_range(inode, alloc_start,
2553                                        alloc_end - alloc_start);
2554         if (ret)
2555                 goto out;
2556
2557         locked_end = alloc_end - 1;
2558         while (1) {
2559                 struct btrfs_ordered_extent *ordered;
2560
2561                 /* the extent lock is ordered inside the running
2562                  * transaction
2563                  */
2564                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2565                                  locked_end, 0, &cached_state);
2566                 ordered = btrfs_lookup_first_ordered_extent(inode,
2567                                                             alloc_end - 1);
2568                 if (ordered &&
2569                     ordered->file_offset + ordered->len > alloc_start &&
2570                     ordered->file_offset < alloc_end) {
2571                         btrfs_put_ordered_extent(ordered);
2572                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2573                                              alloc_start, locked_end,
2574                                              &cached_state, GFP_NOFS);
2575                         /*
2576                          * we can't wait on the range with the transaction
2577                          * running or with the extent lock held
2578                          */
2579                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2580                                                        alloc_end - alloc_start);
2581                         if (ret)
2582                                 goto out;
2583                 } else {
2584                         if (ordered)
2585                                 btrfs_put_ordered_extent(ordered);
2586                         break;
2587                 }
2588         }
2589
2590         cur_offset = alloc_start;
2591         while (1) {
2592                 u64 actual_end;
2593
2594                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2595                                       alloc_end - cur_offset, 0);
2596                 if (IS_ERR_OR_NULL(em)) {
2597                         if (!em)
2598                                 ret = -ENOMEM;
2599                         else
2600                                 ret = PTR_ERR(em);
2601                         break;
2602                 }
2603                 last_byte = min(extent_map_end(em), alloc_end);
2604                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2605                 last_byte = ALIGN(last_byte, blocksize);
2606
2607                 if (em->block_start == EXTENT_MAP_HOLE ||
2608                     (cur_offset >= inode->i_size &&
2609                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2610                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2611                                                         last_byte - cur_offset,
2612                                                         1 << inode->i_blkbits,
2613                                                         offset + len,
2614                                                         &alloc_hint);
2615
2616                         if (ret < 0) {
2617                                 free_extent_map(em);
2618                                 break;
2619                         }
2620                 } else if (actual_end > inode->i_size &&
2621                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2622                         /*
2623                          * We didn't need to allocate any more space, but we
2624                          * still extended the size of the file so we need to
2625                          * update i_size.
2626                          */
2627                         inode->i_ctime = CURRENT_TIME;
2628                         i_size_write(inode, actual_end);
2629                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2630                 }
2631                 free_extent_map(em);
2632
2633                 cur_offset = last_byte;
2634                 if (cur_offset >= alloc_end) {
2635                         ret = 0;
2636                         break;
2637                 }
2638         }
2639         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2640                              &cached_state, GFP_NOFS);
2641 out:
2642         mutex_unlock(&inode->i_mutex);
2643         if (root->fs_info->quota_enabled)
2644                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2645 out_reserve_fail:
2646         /* Let go of our reservation. */
2647         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2648         return ret;
2649 }
2650
2651 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2652 {
2653         struct btrfs_root *root = BTRFS_I(inode)->root;
2654         struct extent_map *em = NULL;
2655         struct extent_state *cached_state = NULL;
2656         u64 lockstart = *offset;
2657         u64 lockend = i_size_read(inode);
2658         u64 start = *offset;
2659         u64 len = i_size_read(inode);
2660         int ret = 0;
2661
2662         lockend = max_t(u64, root->sectorsize, lockend);
2663         if (lockend <= lockstart)
2664                 lockend = lockstart + root->sectorsize;
2665
2666         lockend--;
2667         len = lockend - lockstart + 1;
2668
2669         len = max_t(u64, len, root->sectorsize);
2670         if (inode->i_size == 0)
2671                 return -ENXIO;
2672
2673         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2674                          &cached_state);
2675
2676         while (start < inode->i_size) {
2677                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2678                 if (IS_ERR(em)) {
2679                         ret = PTR_ERR(em);
2680                         em = NULL;
2681                         break;
2682                 }
2683
2684                 if (whence == SEEK_HOLE &&
2685                     (em->block_start == EXTENT_MAP_HOLE ||
2686                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2687                         break;
2688                 else if (whence == SEEK_DATA &&
2689                            (em->block_start != EXTENT_MAP_HOLE &&
2690                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2691                         break;
2692
2693                 start = em->start + em->len;
2694                 free_extent_map(em);
2695                 em = NULL;
2696                 cond_resched();
2697         }
2698         free_extent_map(em);
2699         if (!ret) {
2700                 if (whence == SEEK_DATA && start >= inode->i_size)
2701                         ret = -ENXIO;
2702                 else
2703                         *offset = min_t(loff_t, start, inode->i_size);
2704         }
2705         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2706                              &cached_state, GFP_NOFS);
2707         return ret;
2708 }
2709
2710 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2711 {
2712         struct inode *inode = file->f_mapping->host;
2713         int ret;
2714
2715         mutex_lock(&inode->i_mutex);
2716         switch (whence) {
2717         case SEEK_END:
2718         case SEEK_CUR:
2719                 offset = generic_file_llseek(file, offset, whence);
2720                 goto out;
2721         case SEEK_DATA:
2722         case SEEK_HOLE:
2723                 if (offset >= i_size_read(inode)) {
2724                         mutex_unlock(&inode->i_mutex);
2725                         return -ENXIO;
2726                 }
2727
2728                 ret = find_desired_extent(inode, &offset, whence);
2729                 if (ret) {
2730                         mutex_unlock(&inode->i_mutex);
2731                         return ret;
2732                 }
2733         }
2734
2735         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2736 out:
2737         mutex_unlock(&inode->i_mutex);
2738         return offset;
2739 }
2740
2741 const struct file_operations btrfs_file_operations = {
2742         .llseek         = btrfs_file_llseek,
2743         .read           = do_sync_read,
2744         .write          = do_sync_write,
2745         .aio_read       = generic_file_aio_read,
2746         .splice_read    = generic_file_splice_read,
2747         .aio_write      = btrfs_file_aio_write,
2748         .mmap           = btrfs_file_mmap,
2749         .open           = generic_file_open,
2750         .release        = btrfs_release_file,
2751         .fsync          = btrfs_sync_file,
2752         .fallocate      = btrfs_fallocate,
2753         .unlocked_ioctl = btrfs_ioctl,
2754 #ifdef CONFIG_COMPAT
2755         .compat_ioctl   = btrfs_ioctl,
2756 #endif
2757 };
2758
2759 void btrfs_auto_defrag_exit(void)
2760 {
2761         if (btrfs_inode_defrag_cachep)
2762                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2763 }
2764
2765 int btrfs_auto_defrag_init(void)
2766 {
2767         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2768                                         sizeof(struct inode_defrag), 0,
2769                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2770                                         NULL);
2771         if (!btrfs_inode_defrag_cachep)
2772                 return -ENOMEM;
2773
2774         return 0;
2775 }