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