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/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/compat.h>
  31 #include <linux/slab.h>
  32 #include <linux/btrfs.h>
  33 #include <linux/uio.h>
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "volumes.h"
  42 #include "qgroup.h"
  43 #include "compression.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 btrfs_inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &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, &inode->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 133 {
 134         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(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 btrfs_inode *inode)
 149 {
 150         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 151         struct btrfs_root *root = inode->root;
 152         struct inode_defrag *defrag;
 153         u64 transid;
 154         int ret;
 155
 156         if (!__need_auto_defrag(fs_info))
 157                 return 0;
 158
 159         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 160                 return 0;
 161
 162         if (trans)
 163                 transid = trans->transid;
 164         else
 165                 transid = inode->root->last_trans;
 166
 167         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 168         if (!defrag)
 169                 return -ENOMEM;
 170
 171         defrag->ino = btrfs_ino(inode);
 172         defrag->transid = transid;
 173         defrag->root = root->root_key.objectid;
 174
 175         spin_lock(&fs_info->defrag_inodes_lock);
 176         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 177                 /*
 178                  * If we set IN_DEFRAG flag and evict the inode from memory,
 179                  * and then re-read this inode, this new inode doesn't have
 180                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 181                  */
 182                 ret = __btrfs_add_inode_defrag(inode, defrag);
 183                 if (ret)
 184                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         } else {
 186                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 187         }
 188         spin_unlock(&fs_info->defrag_inodes_lock);
 189         return 0;
 190 }
 191
 192 /*
 193  * Requeue the defrag object. If there is a defrag object that points to
 194  * the same inode in the tree, we will merge them together (by
 195  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 196  */
 197 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 198                                        struct inode_defrag *defrag)
 199 {
 200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 201         int ret;
 202
 203         if (!__need_auto_defrag(fs_info))
 204                 goto out;
 205
 206         /*
 207          * Here we don't check the IN_DEFRAG flag, because we need merge
 208          * them together.
 209          */
 210         spin_lock(&fs_info->defrag_inodes_lock);
 211         ret = __btrfs_add_inode_defrag(inode, defrag);
 212         spin_unlock(&fs_info->defrag_inodes_lock);
 213         if (ret)
 214                 goto out;
 215         return;
 216 out:
 217         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 218 }
 219
 220 /*
 221  * pick the defragable inode that we want, if it doesn't exist, we will get
 222  * the next one.
 223  */
 224 static struct inode_defrag *
 225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 226 {
 227         struct inode_defrag *entry = NULL;
 228         struct inode_defrag tmp;
 229         struct rb_node *p;
 230         struct rb_node *parent = NULL;
 231         int ret;
 232
 233         tmp.ino = ino;
 234         tmp.root = root;
 235
 236         spin_lock(&fs_info->defrag_inodes_lock);
 237         p = fs_info->defrag_inodes.rb_node;
 238         while (p) {
 239                 parent = p;
 240                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 241
 242                 ret = __compare_inode_defrag(&tmp, entry);
 243                 if (ret < 0)
 244                         p = parent->rb_left;
 245                 else if (ret > 0)
 246                         p = parent->rb_right;
 247                 else
 248                         goto out;
 249         }
 250
 251         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 252                 parent = rb_next(parent);
 253                 if (parent)
 254                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 255                 else
 256                         entry = NULL;
 257         }
 258 out:
 259         if (entry)
 260                 rb_erase(parent, &fs_info->defrag_inodes);
 261         spin_unlock(&fs_info->defrag_inodes_lock);
 262         return entry;
 263 }
 264
 265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 266 {
 267         struct inode_defrag *defrag;
 268         struct rb_node *node;
 269
 270         spin_lock(&fs_info->defrag_inodes_lock);
 271         node = rb_first(&fs_info->defrag_inodes);
 272         while (node) {
 273                 rb_erase(node, &fs_info->defrag_inodes);
 274                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 275                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 276
 277                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 278
 279                 node = rb_first(&fs_info->defrag_inodes);
 280         }
 281         spin_unlock(&fs_info->defrag_inodes_lock);
 282 }
 283
 284 #define BTRFS_DEFRAG_BATCH      1024
 285
 286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 287                                     struct inode_defrag *defrag)
 288 {
 289         struct btrfs_root *inode_root;
 290         struct inode *inode;
 291         struct btrfs_key key;
 292         struct btrfs_ioctl_defrag_range_args range;
 293         int num_defrag;
 294         int index;
 295         int ret;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         key.type = BTRFS_ROOT_ITEM_KEY;
 300         key.offset = (u64)-1;
 301
 302         index = srcu_read_lock(&fs_info->subvol_srcu);
 303
 304         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 305         if (IS_ERR(inode_root)) {
 306                 ret = PTR_ERR(inode_root);
 307                 goto cleanup;
 308         }
 309
 310         key.objectid = defrag->ino;
 311         key.type = BTRFS_INODE_ITEM_KEY;
 312         key.offset = 0;
 313         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 314         if (IS_ERR(inode)) {
 315                 ret = PTR_ERR(inode);
 316                 goto cleanup;
 317         }
 318         srcu_read_unlock(&fs_info->subvol_srcu, index);
 319
 320         /* do a chunk of defrag */
 321         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 322         memset(&range, 0, sizeof(range));
 323         range.len = (u64)-1;
 324         range.start = defrag->last_offset;
 325
 326         sb_start_write(fs_info->sb);
 327         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 328                                        BTRFS_DEFRAG_BATCH);
 329         sb_end_write(fs_info->sb);
 330         /*
 331          * if we filled the whole defrag batch, there
 332          * must be more work to do.  Queue this defrag
 333          * again
 334          */
 335         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 336                 defrag->last_offset = range.start;
 337                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 338         } else if (defrag->last_offset && !defrag->cycled) {
 339                 /*
 340                  * we didn't fill our defrag batch, but
 341                  * we didn't start at zero.  Make sure we loop
 342                  * around to the start of the file.
 343                  */
 344                 defrag->last_offset = 0;
 345                 defrag->cycled = 1;
 346                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 347         } else {
 348                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349         }
 350
 351         iput(inode);
 352         return 0;
 353 cleanup:
 354         srcu_read_unlock(&fs_info->subvol_srcu, index);
 355         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 356         return ret;
 357 }
 358
 359 /*
 360  * run through the list of inodes in the FS that need
 361  * defragging
 362  */
 363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 364 {
 365         struct inode_defrag *defrag;
 366         u64 first_ino = 0;
 367         u64 root_objectid = 0;
 368
 369         atomic_inc(&fs_info->defrag_running);
 370         while (1) {
 371                 /* Pause the auto defragger. */
 372                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 373                              &fs_info->fs_state))
 374                         break;
 375
 376                 if (!__need_auto_defrag(fs_info))
 377                         break;
 378
 379                 /* find an inode to defrag */
 380                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 381                                                  first_ino);
 382                 if (!defrag) {
 383                         if (root_objectid || first_ino) {
 384                                 root_objectid = 0;
 385                                 first_ino = 0;
 386                                 continue;
 387                         } else {
 388                                 break;
 389                         }
 390                 }
 391
 392                 first_ino = defrag->ino + 1;
 393                 root_objectid = defrag->root;
 394
 395                 __btrfs_run_defrag_inode(fs_info, defrag);
 396         }
 397         atomic_dec(&fs_info->defrag_running);
 398
 399         /*
 400          * during unmount, we use the transaction_wait queue to
 401          * wait for the defragger to stop
 402          */
 403         wake_up(&fs_info->transaction_wait);
 404         return 0;
 405 }
 406
 407 /* simple helper to fault in pages and copy.  This should go away
 408  * and be replaced with calls into generic code.
 409  */
 410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 411                                          struct page **prepared_pages,
 412                                          struct iov_iter *i)
 413 {
 414         size_t copied = 0;
 415         size_t total_copied = 0;
 416         int pg = 0;
 417         int offset = pos & (PAGE_SIZE - 1);
 418
 419         while (write_bytes > 0) {
 420                 size_t count = min_t(size_t,
 421                                      PAGE_SIZE - offset, write_bytes);
 422                 struct page *page = prepared_pages[pg];
 423                 /*
 424                  * Copy data from userspace to the current page
 425                  */
 426                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 427
 428                 /* Flush processor's dcache for this page */
 429                 flush_dcache_page(page);
 430
 431                 /*
 432                  * if we get a partial write, we can end up with
 433                  * partially up to date pages.  These add
 434                  * a lot of complexity, so make sure they don't
 435                  * happen by forcing this copy to be retried.
 436                  *
 437                  * The rest of the btrfs_file_write code will fall
 438                  * back to page at a time copies after we return 0.
 439                  */
 440                 if (!PageUptodate(page) && copied < count)
 441                         copied = 0;
 442
 443                 iov_iter_advance(i, copied);
 444                 write_bytes -= copied;
 445                 total_copied += copied;
 446
 447                 /* Return to btrfs_file_write_iter to fault page */
 448                 if (unlikely(copied == 0))
 449                         break;
 450
 451                 if (copied < PAGE_SIZE - offset) {
 452                         offset += copied;
 453                 } else {
 454                         pg++;
 455                         offset = 0;
 456                 }
 457         }
 458         return total_copied;
 459 }
 460
 461 /*
 462  * unlocks pages after btrfs_file_write is done with them
 463  */
 464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 465 {
 466         size_t i;
 467         for (i = 0; i < num_pages; i++) {
 468                 /* page checked is some magic around finding pages that
 469                  * have been modified without going through btrfs_set_page_dirty
 470                  * clear it here. There should be no need to mark the pages
 471                  * accessed as prepare_pages should have marked them accessed
 472                  * in prepare_pages via find_or_create_page()
 473                  */
 474                 ClearPageChecked(pages[i]);
 475                 unlock_page(pages[i]);
 476                 put_page(pages[i]);
 477         }
 478 }
 479
 480 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 481                                          const u64 start,
 482                                          const u64 len,
 483                                          struct extent_state **cached_state)
 484 {
 485         u64 search_start = start;
 486         const u64 end = start + len - 1;
 487
 488         while (search_start < end) {
 489                 const u64 search_len = end - search_start + 1;
 490                 struct extent_map *em;
 491                 u64 em_len;
 492                 int ret = 0;
 493
 494                 em = btrfs_get_extent(inode, NULL, 0, search_start,
 495                                       search_len, 0);
 496                 if (IS_ERR(em))
 497                         return PTR_ERR(em);
 498
 499                 if (em->block_start != EXTENT_MAP_HOLE)
 500                         goto next;
 501
 502                 em_len = em->len;
 503                 if (em->start < search_start)
 504                         em_len -= search_start - em->start;
 505                 if (em_len > search_len)
 506                         em_len = search_len;
 507
 508                 ret = set_extent_bit(&inode->io_tree, search_start,
 509                                      search_start + em_len - 1,
 510                                      EXTENT_DELALLOC_NEW,
 511                                      NULL, cached_state, GFP_NOFS);
 512 next:
 513                 search_start = extent_map_end(em);
 514                 free_extent_map(em);
 515                 if (ret)
 516                         return ret;
 517         }
 518         return 0;
 519 }
 520
 521 /*
 522  * after copy_from_user, pages need to be dirtied and we need to make
 523  * sure holes are created between the current EOF and the start of
 524  * any next extents (if required).
 525  *
 526  * this also makes the decision about creating an inline extent vs
 527  * doing real data extents, marking pages dirty and delalloc as required.
 528  */
 529 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 530                       size_t num_pages, loff_t pos, size_t write_bytes,
 531                       struct extent_state **cached)
 532 {
 533         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 534         int err = 0;
 535         int i;
 536         u64 num_bytes;
 537         u64 start_pos;
 538         u64 end_of_last_block;
 539         u64 end_pos = pos + write_bytes;
 540         loff_t isize = i_size_read(inode);
 541         unsigned int extra_bits = 0;
 542
 543         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 544         num_bytes = round_up(write_bytes + pos - start_pos,
 545                              fs_info->sectorsize);
 546
 547         end_of_last_block = start_pos + num_bytes - 1;
 548
 549         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
 550                 if (start_pos >= isize &&
 551                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
 552                         /*
 553                          * There can't be any extents following eof in this case
 554                          * so just set the delalloc new bit for the range
 555                          * directly.
 556                          */
 557                         extra_bits |= EXTENT_DELALLOC_NEW;
 558                 } else {
 559                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
 560                                                             start_pos,
 561                                                             num_bytes, cached);
 562                         if (err)
 563                                 return err;
 564                 }
 565         }
 566
 567         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 568                                         extra_bits, cached, 0);
 569         if (err)
 570                 return err;
 571
 572         for (i = 0; i < num_pages; i++) {
 573                 struct page *p = pages[i];
 574                 SetPageUptodate(p);
 575                 ClearPageChecked(p);
 576                 set_page_dirty(p);
 577         }
 578
 579         /*
 580          * we've only changed i_size in ram, and we haven't updated
 581          * the disk i_size.  There is no need to log the inode
 582          * at this time.
 583          */
 584         if (end_pos > isize)
 585                 i_size_write(inode, end_pos);
 586         return 0;
 587 }
 588
 589 /*
 590  * this drops all the extents in the cache that intersect the range
 591  * [start, end].  Existing extents are split as required.
 592  */
 593 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 594                              int skip_pinned)
 595 {
 596         struct extent_map *em;
 597         struct extent_map *split = NULL;
 598         struct extent_map *split2 = NULL;
 599         struct extent_map_tree *em_tree = &inode->extent_tree;
 600         u64 len = end - start + 1;
 601         u64 gen;
 602         int ret;
 603         int testend = 1;
 604         unsigned long flags;
 605         int compressed = 0;
 606         bool modified;
 607
 608         WARN_ON(end < start);
 609         if (end == (u64)-1) {
 610                 len = (u64)-1;
 611                 testend = 0;
 612         }
 613         while (1) {
 614                 int no_splits = 0;
 615
 616                 modified = false;
 617                 if (!split)
 618                         split = alloc_extent_map();
 619                 if (!split2)
 620                         split2 = alloc_extent_map();
 621                 if (!split || !split2)
 622                         no_splits = 1;
 623
 624                 write_lock(&em_tree->lock);
 625                 em = lookup_extent_mapping(em_tree, start, len);
 626                 if (!em) {
 627                         write_unlock(&em_tree->lock);
 628                         break;
 629                 }
 630                 flags = em->flags;
 631                 gen = em->generation;
 632                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 633                         if (testend && em->start + em->len >= start + len) {
 634                                 free_extent_map(em);
 635                                 write_unlock(&em_tree->lock);
 636                                 break;
 637                         }
 638                         start = em->start + em->len;
 639                         if (testend)
 640                                 len = start + len - (em->start + em->len);
 641                         free_extent_map(em);
 642                         write_unlock(&em_tree->lock);
 643                         continue;
 644                 }
 645                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 646                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 647                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 648                 modified = !list_empty(&em->list);
 649                 if (no_splits)
 650                         goto next;
 651
 652                 if (em->start < start) {
 653                         split->start = em->start;
 654                         split->len = start - em->start;
 655
 656                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 657                                 split->orig_start = em->orig_start;
 658                                 split->block_start = em->block_start;
 659
 660                                 if (compressed)
 661                                         split->block_len = em->block_len;
 662                                 else
 663                                         split->block_len = split->len;
 664                                 split->orig_block_len = max(split->block_len,
 665                                                 em->orig_block_len);
 666                                 split->ram_bytes = em->ram_bytes;
 667                         } else {
 668                                 split->orig_start = split->start;
 669                                 split->block_len = 0;
 670                                 split->block_start = em->block_start;
 671                                 split->orig_block_len = 0;
 672                                 split->ram_bytes = split->len;
 673                         }
 674
 675                         split->generation = gen;
 676                         split->bdev = em->bdev;
 677                         split->flags = flags;
 678                         split->compress_type = em->compress_type;
 679                         replace_extent_mapping(em_tree, em, split, modified);
 680                         free_extent_map(split);
 681                         split = split2;
 682                         split2 = NULL;
 683                 }
 684                 if (testend && em->start + em->len > start + len) {
 685                         u64 diff = start + len - em->start;
 686
 687                         split->start = start + len;
 688                         split->len = em->start + em->len - (start + len);
 689                         split->bdev = em->bdev;
 690                         split->flags = flags;
 691                         split->compress_type = em->compress_type;
 692                         split->generation = gen;
 693
 694                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 695                                 split->orig_block_len = max(em->block_len,
 696                                                     em->orig_block_len);
 697
 698                                 split->ram_bytes = em->ram_bytes;
 699                                 if (compressed) {
 700                                         split->block_len = em->block_len;
 701                                         split->block_start = em->block_start;
 702                                         split->orig_start = em->orig_start;
 703                                 } else {
 704                                         split->block_len = split->len;
 705                                         split->block_start = em->block_start
 706                                                 + diff;
 707                                         split->orig_start = em->orig_start;
 708                                 }
 709                         } else {
 710                                 split->ram_bytes = split->len;
 711                                 split->orig_start = split->start;
 712                                 split->block_len = 0;
 713                                 split->block_start = em->block_start;
 714                                 split->orig_block_len = 0;
 715                         }
 716
 717                         if (extent_map_in_tree(em)) {
 718                                 replace_extent_mapping(em_tree, em, split,
 719                                                        modified);
 720                         } else {
 721                                 ret = add_extent_mapping(em_tree, split,
 722                                                          modified);
 723                                 ASSERT(ret == 0); /* Logic error */
 724                         }
 725                         free_extent_map(split);
 726                         split = NULL;
 727                 }
 728 next:
 729                 if (extent_map_in_tree(em))
 730                         remove_extent_mapping(em_tree, em);
 731                 write_unlock(&em_tree->lock);
 732
 733                 /* once for us */
 734                 free_extent_map(em);
 735                 /* once for the tree*/
 736                 free_extent_map(em);
 737         }
 738         if (split)
 739                 free_extent_map(split);
 740         if (split2)
 741                 free_extent_map(split2);
 742 }
 743
 744 /*
 745  * this is very complex, but the basic idea is to drop all extents
 746  * in the range start - end.  hint_block is filled in with a block number
 747  * that would be a good hint to the block allocator for this file.
 748  *
 749  * If an extent intersects the range but is not entirely inside the range
 750  * it is either truncated or split.  Anything entirely inside the range
 751  * is deleted from the tree.
 752  */
 753 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 754                          struct btrfs_root *root, struct inode *inode,
 755                          struct btrfs_path *path, u64 start, u64 end,
 756                          u64 *drop_end, int drop_cache,
 757                          int replace_extent,
 758                          u32 extent_item_size,
 759                          int *key_inserted)
 760 {
 761         struct btrfs_fs_info *fs_info = root->fs_info;
 762         struct extent_buffer *leaf;
 763         struct btrfs_file_extent_item *fi;
 764         struct btrfs_key key;
 765         struct btrfs_key new_key;
 766         u64 ino = btrfs_ino(BTRFS_I(inode));
 767         u64 search_start = start;
 768         u64 disk_bytenr = 0;
 769         u64 num_bytes = 0;
 770         u64 extent_offset = 0;
 771         u64 extent_end = 0;
 772         u64 last_end = start;
 773         int del_nr = 0;
 774         int del_slot = 0;
 775         int extent_type;
 776         int recow;
 777         int ret;
 778         int modify_tree = -1;
 779         int update_refs;
 780         int found = 0;
 781         int leafs_visited = 0;
 782
 783         if (drop_cache)
 784                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 785
 786         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 787                 modify_tree = 0;
 788
 789         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 790                        root == fs_info->tree_root);
 791         while (1) {
 792                 recow = 0;
 793                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 794                                                search_start, modify_tree);
 795                 if (ret < 0)
 796                         break;
 797                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 798                         leaf = path->nodes[0];
 799                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 800                         if (key.objectid == ino &&
 801                             key.type == BTRFS_EXTENT_DATA_KEY)
 802                                 path->slots[0]--;
 803                 }
 804                 ret = 0;
 805                 leafs_visited++;
 806 next_slot:
 807                 leaf = path->nodes[0];
 808                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 809                         BUG_ON(del_nr > 0);
 810                         ret = btrfs_next_leaf(root, path);
 811                         if (ret < 0)
 812                                 break;
 813                         if (ret > 0) {
 814                                 ret = 0;
 815                                 break;
 816                         }
 817                         leafs_visited++;
 818                         leaf = path->nodes[0];
 819                         recow = 1;
 820                 }
 821
 822                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 823
 824                 if (key.objectid > ino)
 825                         break;
 826                 if (WARN_ON_ONCE(key.objectid < ino) ||
 827                     key.type < BTRFS_EXTENT_DATA_KEY) {
 828                         ASSERT(del_nr == 0);
 829                         path->slots[0]++;
 830                         goto next_slot;
 831                 }
 832                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 833                         break;
 834
 835                 fi = btrfs_item_ptr(leaf, path->slots[0],
 836                                     struct btrfs_file_extent_item);
 837                 extent_type = btrfs_file_extent_type(leaf, fi);
 838
 839                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 840                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 841                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 842                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 843                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 844                         extent_end = key.offset +
 845                                 btrfs_file_extent_num_bytes(leaf, fi);
 846                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 847                         extent_end = key.offset +
 848                                 btrfs_file_extent_inline_len(leaf,
 849                                                      path->slots[0], fi);
 850                 } else {
 851                         /* can't happen */
 852                         BUG();
 853                 }
 854
 855                 /*
 856                  * Don't skip extent items representing 0 byte lengths. They
 857                  * used to be created (bug) if while punching holes we hit
 858                  * -ENOSPC condition. So if we find one here, just ensure we
 859                  * delete it, otherwise we would insert a new file extent item
 860                  * with the same key (offset) as that 0 bytes length file
 861                  * extent item in the call to setup_items_for_insert() later
 862                  * in this function.
 863                  */
 864                 if (extent_end == key.offset && extent_end >= search_start) {
 865                         last_end = extent_end;
 866                         goto delete_extent_item;
 867                 }
 868
 869                 if (extent_end <= search_start) {
 870                         path->slots[0]++;
 871                         goto next_slot;
 872                 }
 873
 874                 found = 1;
 875                 search_start = max(key.offset, start);
 876                 if (recow || !modify_tree) {
 877                         modify_tree = -1;
 878                         btrfs_release_path(path);
 879                         continue;
 880                 }
 881
 882                 /*
 883                  *     | - range to drop - |
 884                  *  | -------- extent -------- |
 885                  */
 886                 if (start > key.offset && end < extent_end) {
 887                         BUG_ON(del_nr > 0);
 888                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 889                                 ret = -EOPNOTSUPP;
 890                                 break;
 891                         }
 892
 893                         memcpy(&new_key, &key, sizeof(new_key));
 894                         new_key.offset = start;
 895                         ret = btrfs_duplicate_item(trans, root, path,
 896                                                    &new_key);
 897                         if (ret == -EAGAIN) {
 898                                 btrfs_release_path(path);
 899                                 continue;
 900                         }
 901                         if (ret < 0)
 902                                 break;
 903
 904                         leaf = path->nodes[0];
 905                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 906                                             struct btrfs_file_extent_item);
 907                         btrfs_set_file_extent_num_bytes(leaf, fi,
 908                                                         start - key.offset);
 909
 910                         fi = btrfs_item_ptr(leaf, path->slots[0],
 911                                             struct btrfs_file_extent_item);
 912
 913                         extent_offset += start - key.offset;
 914                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 915                         btrfs_set_file_extent_num_bytes(leaf, fi,
 916                                                         extent_end - start);
 917                         btrfs_mark_buffer_dirty(leaf);
 918
 919                         if (update_refs && disk_bytenr > 0) {
 920                                 ret = btrfs_inc_extent_ref(trans, root,
 921                                                 disk_bytenr, num_bytes, 0,
 922                                                 root->root_key.objectid,
 923                                                 new_key.objectid,
 924                                                 start - extent_offset);
 925                                 BUG_ON(ret); /* -ENOMEM */
 926                         }
 927                         key.offset = start;
 928                 }
 929                 /*
 930                  * From here on out we will have actually dropped something, so
 931                  * last_end can be updated.
 932                  */
 933                 last_end = extent_end;
 934
 935                 /*
 936                  *  | ---- range to drop ----- |
 937                  *      | -------- extent -------- |
 938                  */
 939                 if (start <= key.offset && end < extent_end) {
 940                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 941                                 ret = -EOPNOTSUPP;
 942                                 break;
 943                         }
 944
 945                         memcpy(&new_key, &key, sizeof(new_key));
 946                         new_key.offset = end;
 947                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 948
 949                         extent_offset += end - key.offset;
 950                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 951                         btrfs_set_file_extent_num_bytes(leaf, fi,
 952                                                         extent_end - end);
 953                         btrfs_mark_buffer_dirty(leaf);
 954                         if (update_refs && disk_bytenr > 0)
 955                                 inode_sub_bytes(inode, end - key.offset);
 956                         break;
 957                 }
 958
 959                 search_start = extent_end;
 960                 /*
 961                  *       | ---- range to drop ----- |
 962                  *  | -------- extent -------- |
 963                  */
 964                 if (start > key.offset && end >= extent_end) {
 965                         BUG_ON(del_nr > 0);
 966                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 967                                 ret = -EOPNOTSUPP;
 968                                 break;
 969                         }
 970
 971                         btrfs_set_file_extent_num_bytes(leaf, fi,
 972                                                         start - key.offset);
 973                         btrfs_mark_buffer_dirty(leaf);
 974                         if (update_refs && disk_bytenr > 0)
 975                                 inode_sub_bytes(inode, extent_end - start);
 976                         if (end == extent_end)
 977                                 break;
 978
 979                         path->slots[0]++;
 980                         goto next_slot;
 981                 }
 982
 983                 /*
 984                  *  | ---- range to drop ----- |
 985                  *    | ------ extent ------ |
 986                  */
 987                 if (start <= key.offset && end >= extent_end) {
 988 delete_extent_item:
 989                         if (del_nr == 0) {
 990                                 del_slot = path->slots[0];
 991                                 del_nr = 1;
 992                         } else {
 993                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 994                                 del_nr++;
 995                         }
 996
 997                         if (update_refs &&
 998                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 999                                 inode_sub_bytes(inode,
1000                                                 extent_end - key.offset);
1001                                 extent_end = ALIGN(extent_end,
1002                                                    fs_info->sectorsize);
1003                         } else if (update_refs && disk_bytenr > 0) {
1004                                 ret = btrfs_free_extent(trans, root,
1005                                                 disk_bytenr, num_bytes, 0,
1006                                                 root->root_key.objectid,
1007                                                 key.objectid, key.offset -
1008                                                 extent_offset);
1009                                 BUG_ON(ret); /* -ENOMEM */
1010                                 inode_sub_bytes(inode,
1011                                                 extent_end - key.offset);
1012                         }
1013
1014                         if (end == extent_end)
1015                                 break;
1016
1017                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1018                                 path->slots[0]++;
1019                                 goto next_slot;
1020                         }
1021
1022                         ret = btrfs_del_items(trans, root, path, del_slot,
1023                                               del_nr);
1024                         if (ret) {
1025                                 btrfs_abort_transaction(trans, ret);
1026                                 break;
1027                         }
1028
1029                         del_nr = 0;
1030                         del_slot = 0;
1031
1032                         btrfs_release_path(path);
1033                         continue;
1034                 }
1035
1036                 BUG_ON(1);
1037         }
1038
1039         if (!ret && del_nr > 0) {
1040                 /*
1041                  * Set path->slots[0] to first slot, so that after the delete
1042                  * if items are move off from our leaf to its immediate left or
1043                  * right neighbor leafs, we end up with a correct and adjusted
1044                  * path->slots[0] for our insertion (if replace_extent != 0).
1045                  */
1046                 path->slots[0] = del_slot;
1047                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1048                 if (ret)
1049                         btrfs_abort_transaction(trans, ret);
1050         }
1051
1052         leaf = path->nodes[0];
1053         /*
1054          * If btrfs_del_items() was called, it might have deleted a leaf, in
1055          * which case it unlocked our path, so check path->locks[0] matches a
1056          * write lock.
1057          */
1058         if (!ret && replace_extent && leafs_visited == 1 &&
1059             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1060              path->locks[0] == BTRFS_WRITE_LOCK) &&
1061             btrfs_leaf_free_space(fs_info, leaf) >=
1062             sizeof(struct btrfs_item) + extent_item_size) {
1063
1064                 key.objectid = ino;
1065                 key.type = BTRFS_EXTENT_DATA_KEY;
1066                 key.offset = start;
1067                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1068                         struct btrfs_key slot_key;
1069
1070                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1071                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1072                                 path->slots[0]++;
1073                 }
1074                 setup_items_for_insert(root, path, &key,
1075                                        &extent_item_size,
1076                                        extent_item_size,
1077                                        sizeof(struct btrfs_item) +
1078                                        extent_item_size, 1);
1079                 *key_inserted = 1;
1080         }
1081
1082         if (!replace_extent || !(*key_inserted))
1083                 btrfs_release_path(path);
1084         if (drop_end)
1085                 *drop_end = found ? min(end, last_end) : end;
1086         return ret;
1087 }
1088
1089 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1090                        struct btrfs_root *root, struct inode *inode, u64 start,
1091                        u64 end, int drop_cache)
1092 {
1093         struct btrfs_path *path;
1094         int ret;
1095
1096         path = btrfs_alloc_path();
1097         if (!path)
1098                 return -ENOMEM;
1099         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1100                                    drop_cache, 0, 0, NULL);
1101         btrfs_free_path(path);
1102         return ret;
1103 }
1104
1105 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1106                             u64 objectid, u64 bytenr, u64 orig_offset,
1107                             u64 *start, u64 *end)
1108 {
1109         struct btrfs_file_extent_item *fi;
1110         struct btrfs_key key;
1111         u64 extent_end;
1112
1113         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1114                 return 0;
1115
1116         btrfs_item_key_to_cpu(leaf, &key, slot);
1117         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1118                 return 0;
1119
1120         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1121         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1122             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1123             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1124             btrfs_file_extent_compression(leaf, fi) ||
1125             btrfs_file_extent_encryption(leaf, fi) ||
1126             btrfs_file_extent_other_encoding(leaf, fi))
1127                 return 0;
1128
1129         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1130         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1131                 return 0;
1132
1133         *start = key.offset;
1134         *end = extent_end;
1135         return 1;
1136 }
1137
1138 /*
1139  * Mark extent in the range start - end as written.
1140  *
1141  * This changes extent type from 'pre-allocated' to 'regular'. If only
1142  * part of extent is marked as written, the extent will be split into
1143  * two or three.
1144  */
1145 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1146                               struct btrfs_inode *inode, u64 start, u64 end)
1147 {
1148         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1149         struct btrfs_root *root = inode->root;
1150         struct extent_buffer *leaf;
1151         struct btrfs_path *path;
1152         struct btrfs_file_extent_item *fi;
1153         struct btrfs_key key;
1154         struct btrfs_key new_key;
1155         u64 bytenr;
1156         u64 num_bytes;
1157         u64 extent_end;
1158         u64 orig_offset;
1159         u64 other_start;
1160         u64 other_end;
1161         u64 split;
1162         int del_nr = 0;
1163         int del_slot = 0;
1164         int recow;
1165         int ret;
1166         u64 ino = btrfs_ino(inode);
1167
1168         path = btrfs_alloc_path();
1169         if (!path)
1170                 return -ENOMEM;
1171 again:
1172         recow = 0;
1173         split = start;
1174         key.objectid = ino;
1175         key.type = BTRFS_EXTENT_DATA_KEY;
1176         key.offset = split;
1177
1178         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1179         if (ret < 0)
1180                 goto out;
1181         if (ret > 0 && path->slots[0] > 0)
1182                 path->slots[0]--;
1183
1184         leaf = path->nodes[0];
1185         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1186         if (key.objectid != ino ||
1187             key.type != BTRFS_EXTENT_DATA_KEY) {
1188                 ret = -EINVAL;
1189                 btrfs_abort_transaction(trans, ret);
1190                 goto out;
1191         }
1192         fi = btrfs_item_ptr(leaf, path->slots[0],
1193                             struct btrfs_file_extent_item);
1194         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1195                 ret = -EINVAL;
1196                 btrfs_abort_transaction(trans, ret);
1197                 goto out;
1198         }
1199         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1200         if (key.offset > start || extent_end < end) {
1201                 ret = -EINVAL;
1202                 btrfs_abort_transaction(trans, ret);
1203                 goto out;
1204         }
1205
1206         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1207         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1208         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1209         memcpy(&new_key, &key, sizeof(new_key));
1210
1211         if (start == key.offset && end < extent_end) {
1212                 other_start = 0;
1213                 other_end = start;
1214                 if (extent_mergeable(leaf, path->slots[0] - 1,
1215                                      ino, bytenr, orig_offset,
1216                                      &other_start, &other_end)) {
1217                         new_key.offset = end;
1218                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1219                         fi = btrfs_item_ptr(leaf, path->slots[0],
1220                                             struct btrfs_file_extent_item);
1221                         btrfs_set_file_extent_generation(leaf, fi,
1222                                                          trans->transid);
1223                         btrfs_set_file_extent_num_bytes(leaf, fi,
1224                                                         extent_end - end);
1225                         btrfs_set_file_extent_offset(leaf, fi,
1226                                                      end - orig_offset);
1227                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1228                                             struct btrfs_file_extent_item);
1229                         btrfs_set_file_extent_generation(leaf, fi,
1230                                                          trans->transid);
1231                         btrfs_set_file_extent_num_bytes(leaf, fi,
1232                                                         end - other_start);
1233                         btrfs_mark_buffer_dirty(leaf);
1234                         goto out;
1235                 }
1236         }
1237
1238         if (start > key.offset && end == extent_end) {
1239                 other_start = end;
1240                 other_end = 0;
1241                 if (extent_mergeable(leaf, path->slots[0] + 1,
1242                                      ino, bytenr, orig_offset,
1243                                      &other_start, &other_end)) {
1244                         fi = btrfs_item_ptr(leaf, path->slots[0],
1245                                             struct btrfs_file_extent_item);
1246                         btrfs_set_file_extent_num_bytes(leaf, fi,
1247                                                         start - key.offset);
1248                         btrfs_set_file_extent_generation(leaf, fi,
1249                                                          trans->transid);
1250                         path->slots[0]++;
1251                         new_key.offset = start;
1252                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1253
1254                         fi = btrfs_item_ptr(leaf, path->slots[0],
1255                                             struct btrfs_file_extent_item);
1256                         btrfs_set_file_extent_generation(leaf, fi,
1257                                                          trans->transid);
1258                         btrfs_set_file_extent_num_bytes(leaf, fi,
1259                                                         other_end - start);
1260                         btrfs_set_file_extent_offset(leaf, fi,
1261                                                      start - orig_offset);
1262                         btrfs_mark_buffer_dirty(leaf);
1263                         goto out;
1264                 }
1265         }
1266
1267         while (start > key.offset || end < extent_end) {
1268                 if (key.offset == start)
1269                         split = end;
1270
1271                 new_key.offset = split;
1272                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1273                 if (ret == -EAGAIN) {
1274                         btrfs_release_path(path);
1275                         goto again;
1276                 }
1277                 if (ret < 0) {
1278                         btrfs_abort_transaction(trans, ret);
1279                         goto out;
1280                 }
1281
1282                 leaf = path->nodes[0];
1283                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1284                                     struct btrfs_file_extent_item);
1285                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1286                 btrfs_set_file_extent_num_bytes(leaf, fi,
1287                                                 split - key.offset);
1288
1289                 fi = btrfs_item_ptr(leaf, path->slots[0],
1290                                     struct btrfs_file_extent_item);
1291
1292                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1293                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1294                 btrfs_set_file_extent_num_bytes(leaf, fi,
1295                                                 extent_end - split);
1296                 btrfs_mark_buffer_dirty(leaf);
1297
1298                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1299                                            0, root->root_key.objectid,
1300                                            ino, orig_offset);
1301                 if (ret) {
1302                         btrfs_abort_transaction(trans, ret);
1303                         goto out;
1304                 }
1305
1306                 if (split == start) {
1307                         key.offset = start;
1308                 } else {
1309                         if (start != key.offset) {
1310                                 ret = -EINVAL;
1311                                 btrfs_abort_transaction(trans, ret);
1312                                 goto out;
1313                         }
1314                         path->slots[0]--;
1315                         extent_end = end;
1316                 }
1317                 recow = 1;
1318         }
1319
1320         other_start = end;
1321         other_end = 0;
1322         if (extent_mergeable(leaf, path->slots[0] + 1,
1323                              ino, bytenr, orig_offset,
1324                              &other_start, &other_end)) {
1325                 if (recow) {
1326                         btrfs_release_path(path);
1327                         goto again;
1328                 }
1329                 extent_end = other_end;
1330                 del_slot = path->slots[0] + 1;
1331                 del_nr++;
1332                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1333                                         0, root->root_key.objectid,
1334                                         ino, orig_offset);
1335                 if (ret) {
1336                         btrfs_abort_transaction(trans, ret);
1337                         goto out;
1338                 }
1339         }
1340         other_start = 0;
1341         other_end = start;
1342         if (extent_mergeable(leaf, path->slots[0] - 1,
1343                              ino, bytenr, orig_offset,
1344                              &other_start, &other_end)) {
1345                 if (recow) {
1346                         btrfs_release_path(path);
1347                         goto again;
1348                 }
1349                 key.offset = other_start;
1350                 del_slot = path->slots[0];
1351                 del_nr++;
1352                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1353                                         0, root->root_key.objectid,
1354                                         ino, orig_offset);
1355                 if (ret) {
1356                         btrfs_abort_transaction(trans, ret);
1357                         goto out;
1358                 }
1359         }
1360         if (del_nr == 0) {
1361                 fi = btrfs_item_ptr(leaf, path->slots[0],
1362                            struct btrfs_file_extent_item);
1363                 btrfs_set_file_extent_type(leaf, fi,
1364                                            BTRFS_FILE_EXTENT_REG);
1365                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1366                 btrfs_mark_buffer_dirty(leaf);
1367         } else {
1368                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1369                            struct btrfs_file_extent_item);
1370                 btrfs_set_file_extent_type(leaf, fi,
1371                                            BTRFS_FILE_EXTENT_REG);
1372                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1373                 btrfs_set_file_extent_num_bytes(leaf, fi,
1374                                                 extent_end - key.offset);
1375                 btrfs_mark_buffer_dirty(leaf);
1376
1377                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1378                 if (ret < 0) {
1379                         btrfs_abort_transaction(trans, ret);
1380                         goto out;
1381                 }
1382         }
1383 out:
1384         btrfs_free_path(path);
1385         return 0;
1386 }
1387
1388 /*
1389  * on error we return an unlocked page and the error value
1390  * on success we return a locked page and 0
1391  */
1392 static int prepare_uptodate_page(struct inode *inode,
1393                                  struct page *page, u64 pos,
1394                                  bool force_uptodate)
1395 {
1396         int ret = 0;
1397
1398         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1399             !PageUptodate(page)) {
1400                 ret = btrfs_readpage(NULL, page);
1401                 if (ret)
1402                         return ret;
1403                 lock_page(page);
1404                 if (!PageUptodate(page)) {
1405                         unlock_page(page);
1406                         return -EIO;
1407                 }
1408                 if (page->mapping != inode->i_mapping) {
1409                         unlock_page(page);
1410                         return -EAGAIN;
1411                 }
1412         }
1413         return 0;
1414 }
1415
1416 /*
1417  * this just gets pages into the page cache and locks them down.
1418  */
1419 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1420                                   size_t num_pages, loff_t pos,
1421                                   size_t write_bytes, bool force_uptodate)
1422 {
1423         int i;
1424         unsigned long index = pos >> PAGE_SHIFT;
1425         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1426         int err = 0;
1427         int faili;
1428
1429         for (i = 0; i < num_pages; i++) {
1430 again:
1431                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1432                                                mask | __GFP_WRITE);
1433                 if (!pages[i]) {
1434                         faili = i - 1;
1435                         err = -ENOMEM;
1436                         goto fail;
1437                 }
1438
1439                 if (i == 0)
1440                         err = prepare_uptodate_page(inode, pages[i], pos,
1441                                                     force_uptodate);
1442                 if (!err && i == num_pages - 1)
1443                         err = prepare_uptodate_page(inode, pages[i],
1444                                                     pos + write_bytes, false);
1445                 if (err) {
1446                         put_page(pages[i]);
1447                         if (err == -EAGAIN) {
1448                                 err = 0;
1449                                 goto again;
1450                         }
1451                         faili = i - 1;
1452                         goto fail;
1453                 }
1454                 wait_on_page_writeback(pages[i]);
1455         }
1456
1457         return 0;
1458 fail:
1459         while (faili >= 0) {
1460                 unlock_page(pages[faili]);
1461                 put_page(pages[faili]);
1462                 faili--;
1463         }
1464         return err;
1465
1466 }
1467
1468 /*
1469  * This function locks the extent and properly waits for data=ordered extents
1470  * to finish before allowing the pages to be modified if need.
1471  *
1472  * The return value:
1473  * 1 - the extent is locked
1474  * 0 - the extent is not locked, and everything is OK
1475  * -EAGAIN - need re-prepare the pages
1476  * the other < 0 number - Something wrong happens
1477  */
1478 static noinline int
1479 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1480                                 size_t num_pages, loff_t pos,
1481                                 size_t write_bytes,
1482                                 u64 *lockstart, u64 *lockend,
1483                                 struct extent_state **cached_state)
1484 {
1485         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1486         u64 start_pos;
1487         u64 last_pos;
1488         int i;
1489         int ret = 0;
1490
1491         start_pos = round_down(pos, fs_info->sectorsize);
1492         last_pos = start_pos
1493                 + round_up(pos + write_bytes - start_pos,
1494                            fs_info->sectorsize) - 1;
1495
1496         if (start_pos < inode->vfs_inode.i_size) {
1497                 struct btrfs_ordered_extent *ordered;
1498
1499                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1500                                 cached_state);
1501                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1502                                                      last_pos - start_pos + 1);
1503                 if (ordered &&
1504                     ordered->file_offset + ordered->len > start_pos &&
1505                     ordered->file_offset <= last_pos) {
1506                         unlock_extent_cached(&inode->io_tree, start_pos,
1507                                         last_pos, cached_state, GFP_NOFS);
1508                         for (i = 0; i < num_pages; i++) {
1509                                 unlock_page(pages[i]);
1510                                 put_page(pages[i]);
1511                         }
1512                         btrfs_start_ordered_extent(&inode->vfs_inode,
1513                                         ordered, 1);
1514                         btrfs_put_ordered_extent(ordered);
1515                         return -EAGAIN;
1516                 }
1517                 if (ordered)
1518                         btrfs_put_ordered_extent(ordered);
1519                 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1520                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1521                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1522                                  0, 0, cached_state, GFP_NOFS);
1523                 *lockstart = start_pos;
1524                 *lockend = last_pos;
1525                 ret = 1;
1526         }
1527
1528         for (i = 0; i < num_pages; i++) {
1529                 if (clear_page_dirty_for_io(pages[i]))
1530                         account_page_redirty(pages[i]);
1531                 set_page_extent_mapped(pages[i]);
1532                 WARN_ON(!PageLocked(pages[i]));
1533         }
1534
1535         return ret;
1536 }
1537
1538 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1539                                     size_t *write_bytes)
1540 {
1541         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1542         struct btrfs_root *root = inode->root;
1543         struct btrfs_ordered_extent *ordered;
1544         u64 lockstart, lockend;
1545         u64 num_bytes;
1546         int ret;
1547
1548         ret = btrfs_start_write_no_snapshotting(root);
1549         if (!ret)
1550                 return -ENOSPC;
1551
1552         lockstart = round_down(pos, fs_info->sectorsize);
1553         lockend = round_up(pos + *write_bytes,
1554                            fs_info->sectorsize) - 1;
1555
1556         while (1) {
1557                 lock_extent(&inode->io_tree, lockstart, lockend);
1558                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1559                                                      lockend - lockstart + 1);
1560                 if (!ordered) {
1561                         break;
1562                 }
1563                 unlock_extent(&inode->io_tree, lockstart, lockend);
1564                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1565                 btrfs_put_ordered_extent(ordered);
1566         }
1567
1568         num_bytes = lockend - lockstart + 1;
1569         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1570                         NULL, NULL, NULL);
1571         if (ret <= 0) {
1572                 ret = 0;
1573                 btrfs_end_write_no_snapshotting(root);
1574         } else {
1575                 *write_bytes = min_t(size_t, *write_bytes ,
1576                                      num_bytes - pos + lockstart);
1577         }
1578
1579         unlock_extent(&inode->io_tree, lockstart, lockend);
1580
1581         return ret;
1582 }
1583
1584 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1585                                                struct iov_iter *i,
1586                                                loff_t pos)
1587 {
1588         struct inode *inode = file_inode(file);
1589         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1590         struct btrfs_root *root = BTRFS_I(inode)->root;
1591         struct page **pages = NULL;
1592         struct extent_state *cached_state = NULL;
1593         struct extent_changeset *data_reserved = NULL;
1594         u64 release_bytes = 0;
1595         u64 lockstart;
1596         u64 lockend;
1597         size_t num_written = 0;
1598         int nrptrs;
1599         int ret = 0;
1600         bool only_release_metadata = false;
1601         bool force_page_uptodate = false;
1602
1603         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1604                         PAGE_SIZE / (sizeof(struct page *)));
1605         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1606         nrptrs = max(nrptrs, 8);
1607         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1608         if (!pages)
1609                 return -ENOMEM;
1610
1611         while (iov_iter_count(i) > 0) {
1612                 size_t offset = pos & (PAGE_SIZE - 1);
1613                 size_t sector_offset;
1614                 size_t write_bytes = min(iov_iter_count(i),
1615                                          nrptrs * (size_t)PAGE_SIZE -
1616                                          offset);
1617                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1618                                                 PAGE_SIZE);
1619                 size_t reserve_bytes;
1620                 size_t dirty_pages;
1621                 size_t copied;
1622                 size_t dirty_sectors;
1623                 size_t num_sectors;
1624                 int extents_locked;
1625
1626                 WARN_ON(num_pages > nrptrs);
1627
1628                 /*
1629                  * Fault pages before locking them in prepare_pages
1630                  * to avoid recursive lock
1631                  */
1632                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1633                         ret = -EFAULT;
1634                         break;
1635                 }
1636
1637                 sector_offset = pos & (fs_info->sectorsize - 1);
1638                 reserve_bytes = round_up(write_bytes + sector_offset,
1639                                 fs_info->sectorsize);
1640
1641                 extent_changeset_release(data_reserved);
1642                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1643                                                   write_bytes);
1644                 if (ret < 0) {
1645                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1646                                                       BTRFS_INODE_PREALLOC)) &&
1647                             check_can_nocow(BTRFS_I(inode), pos,
1648                                         &write_bytes) > 0) {
1649                                 /*
1650                                  * For nodata cow case, no need to reserve
1651                                  * data space.
1652                                  */
1653                                 only_release_metadata = true;
1654                                 /*
1655                                  * our prealloc extent may be smaller than
1656                                  * write_bytes, so scale down.
1657                                  */
1658                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1659                                                          PAGE_SIZE);
1660                                 reserve_bytes = round_up(write_bytes +
1661                                                          sector_offset,
1662                                                          fs_info->sectorsize);
1663                         } else {
1664                                 break;
1665                         }
1666                 }
1667
1668                 WARN_ON(reserve_bytes == 0);
1669                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1670                                 reserve_bytes);
1671                 if (ret) {
1672                         if (!only_release_metadata)
1673                                 btrfs_free_reserved_data_space(inode,
1674                                                 data_reserved, pos,
1675                                                 write_bytes);
1676                         else
1677                                 btrfs_end_write_no_snapshotting(root);
1678                         break;
1679                 }
1680
1681                 release_bytes = reserve_bytes;
1682 again:
1683                 /*
1684                  * This is going to setup the pages array with the number of
1685                  * pages we want, so we don't really need to worry about the
1686                  * contents of pages from loop to loop
1687                  */
1688                 ret = prepare_pages(inode, pages, num_pages,
1689                                     pos, write_bytes,
1690                                     force_page_uptodate);
1691                 if (ret) {
1692                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1693                                                        reserve_bytes);
1694                         break;
1695                 }
1696
1697                 extents_locked = lock_and_cleanup_extent_if_need(
1698                                 BTRFS_I(inode), pages,
1699                                 num_pages, pos, write_bytes, &lockstart,
1700                                 &lockend, &cached_state);
1701                 if (extents_locked < 0) {
1702                         if (extents_locked == -EAGAIN)
1703                                 goto again;
1704                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1705                                                        reserve_bytes);
1706                         ret = extents_locked;
1707                         break;
1708                 }
1709
1710                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1711
1712                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1713                 dirty_sectors = round_up(copied + sector_offset,
1714                                         fs_info->sectorsize);
1715                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1716
1717                 /*
1718                  * if we have trouble faulting in the pages, fall
1719                  * back to one page at a time
1720                  */
1721                 if (copied < write_bytes)
1722                         nrptrs = 1;
1723
1724                 if (copied == 0) {
1725                         force_page_uptodate = true;
1726                         dirty_sectors = 0;
1727                         dirty_pages = 0;
1728                 } else {
1729                         force_page_uptodate = false;
1730                         dirty_pages = DIV_ROUND_UP(copied + offset,
1731                                                    PAGE_SIZE);
1732                 }
1733
1734                 if (num_sectors > dirty_sectors) {
1735                         /* release everything except the sectors we dirtied */
1736                         release_bytes -= dirty_sectors <<
1737                                                 fs_info->sb->s_blocksize_bits;
1738                         if (only_release_metadata) {
1739                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1740                                                                 release_bytes);
1741                         } else {
1742                                 u64 __pos;
1743
1744                                 __pos = round_down(pos,
1745                                                    fs_info->sectorsize) +
1746                                         (dirty_pages << PAGE_SHIFT);
1747                                 btrfs_delalloc_release_space(inode,
1748                                                 data_reserved, __pos,
1749                                                 release_bytes);
1750                         }
1751                 }
1752
1753                 release_bytes = round_up(copied + sector_offset,
1754                                         fs_info->sectorsize);
1755
1756                 if (copied > 0)
1757                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1758                                                 pos, copied, NULL);
1759                 if (extents_locked)
1760                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1761                                              lockstart, lockend, &cached_state,
1762                                              GFP_NOFS);
1763                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1764                 if (ret) {
1765                         btrfs_drop_pages(pages, num_pages);
1766                         break;
1767                 }
1768
1769                 release_bytes = 0;
1770                 if (only_release_metadata)
1771                         btrfs_end_write_no_snapshotting(root);
1772
1773                 if (only_release_metadata && copied > 0) {
1774                         lockstart = round_down(pos,
1775                                                fs_info->sectorsize);
1776                         lockend = round_up(pos + copied,
1777                                            fs_info->sectorsize) - 1;
1778
1779                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1780                                        lockend, EXTENT_NORESERVE, NULL,
1781                                        NULL, GFP_NOFS);
1782                         only_release_metadata = false;
1783                 }
1784
1785                 btrfs_drop_pages(pages, num_pages);
1786
1787                 cond_resched();
1788
1789                 balance_dirty_pages_ratelimited(inode->i_mapping);
1790                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1791                         btrfs_btree_balance_dirty(fs_info);
1792
1793                 pos += copied;
1794                 num_written += copied;
1795         }
1796
1797         kfree(pages);
1798
1799         if (release_bytes) {
1800                 if (only_release_metadata) {
1801                         btrfs_end_write_no_snapshotting(root);
1802                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1803                                         release_bytes);
1804                 } else {
1805                         btrfs_delalloc_release_space(inode, data_reserved,
1806                                         round_down(pos, fs_info->sectorsize),
1807                                         release_bytes);
1808                 }
1809         }
1810
1811         extent_changeset_free(data_reserved);
1812         return num_written ? num_written : ret;
1813 }
1814
1815 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1816 {
1817         struct file *file = iocb->ki_filp;
1818         struct inode *inode = file_inode(file);
1819         loff_t pos = iocb->ki_pos;
1820         ssize_t written;
1821         ssize_t written_buffered;
1822         loff_t endbyte;
1823         int err;
1824
1825         written = generic_file_direct_write(iocb, from);
1826
1827         if (written < 0 || !iov_iter_count(from))
1828                 return written;
1829
1830         pos += written;
1831         written_buffered = __btrfs_buffered_write(file, from, pos);
1832         if (written_buffered < 0) {
1833                 err = written_buffered;
1834                 goto out;
1835         }
1836         /*
1837          * Ensure all data is persisted. We want the next direct IO read to be
1838          * able to read what was just written.
1839          */
1840         endbyte = pos + written_buffered - 1;
1841         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1842         if (err)
1843                 goto out;
1844         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1845         if (err)
1846                 goto out;
1847         written += written_buffered;
1848         iocb->ki_pos = pos + written_buffered;
1849         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1850                                  endbyte >> PAGE_SHIFT);
1851 out:
1852         return written ? written : err;
1853 }
1854
1855 static void update_time_for_write(struct inode *inode)
1856 {
1857         struct timespec now;
1858
1859         if (IS_NOCMTIME(inode))
1860                 return;
1861
1862         now = current_time(inode);
1863         if (!timespec_equal(&inode->i_mtime, &now))
1864                 inode->i_mtime = now;
1865
1866         if (!timespec_equal(&inode->i_ctime, &now))
1867                 inode->i_ctime = now;
1868
1869         if (IS_I_VERSION(inode))
1870                 inode_inc_iversion(inode);
1871 }
1872
1873 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1874                                     struct iov_iter *from)
1875 {
1876         struct file *file = iocb->ki_filp;
1877         struct inode *inode = file_inode(file);
1878         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1879         struct btrfs_root *root = BTRFS_I(inode)->root;
1880         u64 start_pos;
1881         u64 end_pos;
1882         ssize_t num_written = 0;
1883         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1884         ssize_t err;
1885         loff_t pos;
1886         size_t count = iov_iter_count(from);
1887         loff_t oldsize;
1888         int clean_page = 0;
1889
1890         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1891             (iocb->ki_flags & IOCB_NOWAIT))
1892                 return -EOPNOTSUPP;
1893
1894         if (!inode_trylock(inode)) {
1895                 if (iocb->ki_flags & IOCB_NOWAIT)
1896                         return -EAGAIN;
1897                 inode_lock(inode);
1898         }
1899
1900         err = generic_write_checks(iocb, from);
1901         if (err <= 0) {
1902                 inode_unlock(inode);
1903                 return err;
1904         }
1905
1906         pos = iocb->ki_pos;
1907         if (iocb->ki_flags & IOCB_NOWAIT) {
1908                 /*
1909                  * We will allocate space in case nodatacow is not set,
1910                  * so bail
1911                  */
1912                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1913                                               BTRFS_INODE_PREALLOC)) ||
1914                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1915                         inode_unlock(inode);
1916                         return -EAGAIN;
1917                 }
1918         }
1919
1920         current->backing_dev_info = inode_to_bdi(inode);
1921         err = file_remove_privs(file);
1922         if (err) {
1923                 inode_unlock(inode);
1924                 goto out;
1925         }
1926
1927         /*
1928          * If BTRFS flips readonly due to some impossible error
1929          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1930          * although we have opened a file as writable, we have
1931          * to stop this write operation to ensure FS consistency.
1932          */
1933         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1934                 inode_unlock(inode);
1935                 err = -EROFS;
1936                 goto out;
1937         }
1938
1939         /*
1940          * We reserve space for updating the inode when we reserve space for the
1941          * extent we are going to write, so we will enospc out there.  We don't
1942          * need to start yet another transaction to update the inode as we will
1943          * update the inode when we finish writing whatever data we write.
1944          */
1945         update_time_for_write(inode);
1946
1947         start_pos = round_down(pos, fs_info->sectorsize);
1948         oldsize = i_size_read(inode);
1949         if (start_pos > oldsize) {
1950                 /* Expand hole size to cover write data, preventing empty gap */
1951                 end_pos = round_up(pos + count,
1952                                    fs_info->sectorsize);
1953                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1954                 if (err) {
1955                         inode_unlock(inode);
1956                         goto out;
1957                 }
1958                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1959                         clean_page = 1;
1960         }
1961
1962         if (sync)
1963                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1964
1965         if (iocb->ki_flags & IOCB_DIRECT) {
1966                 num_written = __btrfs_direct_write(iocb, from);
1967         } else {
1968                 num_written = __btrfs_buffered_write(file, from, pos);
1969                 if (num_written > 0)
1970                         iocb->ki_pos = pos + num_written;
1971                 if (clean_page)
1972                         pagecache_isize_extended(inode, oldsize,
1973                                                 i_size_read(inode));
1974         }
1975
1976         inode_unlock(inode);
1977
1978         /*
1979          * We also have to set last_sub_trans to the current log transid,
1980          * otherwise subsequent syncs to a file that's been synced in this
1981          * transaction will appear to have already occurred.
1982          */
1983         spin_lock(&BTRFS_I(inode)->lock);
1984         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1985         spin_unlock(&BTRFS_I(inode)->lock);
1986         if (num_written > 0)
1987                 num_written = generic_write_sync(iocb, num_written);
1988
1989         if (sync)
1990                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1991 out:
1992         current->backing_dev_info = NULL;
1993         return num_written ? num_written : err;
1994 }
1995
1996 int btrfs_release_file(struct inode *inode, struct file *filp)
1997 {
1998         struct btrfs_file_private *private = filp->private_data;
1999
2000         if (private && private->trans)
2001                 btrfs_ioctl_trans_end(filp);
2002         if (private && private->filldir_buf)
2003                 kfree(private->filldir_buf);
2004         kfree(private);
2005         filp->private_data = NULL;
2006
2007         /*
2008          * ordered_data_close is set by settattr when we are about to truncate
2009          * a file from a non-zero size to a zero size.  This tries to
2010          * flush down new bytes that may have been written if the
2011          * application were using truncate to replace a file in place.
2012          */
2013         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014                                &BTRFS_I(inode)->runtime_flags))
2015                         filemap_flush(inode->i_mapping);
2016         return 0;
2017 }
2018
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2020 {
2021         int ret;
2022
2023         atomic_inc(&BTRFS_I(inode)->sync_writers);
2024         ret = btrfs_fdatawrite_range(inode, start, end);
2025         atomic_dec(&BTRFS_I(inode)->sync_writers);
2026
2027         return ret;
2028 }
2029
2030 /*
2031  * fsync call for both files and directories.  This logs the inode into
2032  * the tree log instead of forcing full commits whenever possible.
2033  *
2034  * It needs to call filemap_fdatawait so that all ordered extent updates are
2035  * in the metadata btree are up to date for copying to the log.
2036  *
2037  * It drops the inode mutex before doing the tree log commit.  This is an
2038  * important optimization for directories because holding the mutex prevents
2039  * new operations on the dir while we write to disk.
2040  */
2041 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2042 {
2043         struct dentry *dentry = file_dentry(file);
2044         struct inode *inode = d_inode(dentry);
2045         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2046         struct btrfs_root *root = BTRFS_I(inode)->root;
2047         struct btrfs_trans_handle *trans;
2048         struct btrfs_log_ctx ctx;
2049         int ret = 0, err;
2050         bool full_sync = false;
2051         u64 len;
2052
2053         /*
2054          * The range length can be represented by u64, we have to do the typecasts
2055          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2056          */
2057         len = (u64)end - (u64)start + 1;
2058         trace_btrfs_sync_file(file, datasync);
2059
2060         btrfs_init_log_ctx(&ctx, inode);
2061
2062         /*
2063          * We write the dirty pages in the range and wait until they complete
2064          * out of the ->i_mutex. If so, we can flush the dirty pages by
2065          * multi-task, and make the performance up.  See
2066          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2067          */
2068         ret = start_ordered_ops(inode, start, end);
2069         if (ret)
2070                 goto out;
2071
2072         inode_lock(inode);
2073         atomic_inc(&root->log_batch);
2074         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2075                              &BTRFS_I(inode)->runtime_flags);
2076         /*
2077          * We might have have had more pages made dirty after calling
2078          * start_ordered_ops and before acquiring the inode's i_mutex.
2079          */
2080         if (full_sync) {
2081                 /*
2082                  * For a full sync, we need to make sure any ordered operations
2083                  * start and finish before we start logging the inode, so that
2084                  * all extents are persisted and the respective file extent
2085                  * items are in the fs/subvol btree.
2086                  */
2087                 ret = btrfs_wait_ordered_range(inode, start, len);
2088         } else {
2089                 /*
2090                  * Start any new ordered operations before starting to log the
2091                  * inode. We will wait for them to finish in btrfs_sync_log().
2092                  *
2093                  * Right before acquiring the inode's mutex, we might have new
2094                  * writes dirtying pages, which won't immediately start the
2095                  * respective ordered operations - that is done through the
2096                  * fill_delalloc callbacks invoked from the writepage and
2097                  * writepages address space operations. So make sure we start
2098                  * all ordered operations before starting to log our inode. Not
2099                  * doing this means that while logging the inode, writeback
2100                  * could start and invoke writepage/writepages, which would call
2101                  * the fill_delalloc callbacks (cow_file_range,
2102                  * submit_compressed_extents). These callbacks add first an
2103                  * extent map to the modified list of extents and then create
2104                  * the respective ordered operation, which means in
2105                  * tree-log.c:btrfs_log_inode() we might capture all existing
2106                  * ordered operations (with btrfs_get_logged_extents()) before
2107                  * the fill_delalloc callback adds its ordered operation, and by
2108                  * the time we visit the modified list of extent maps (with
2109                  * btrfs_log_changed_extents()), we see and process the extent
2110                  * map they created. We then use the extent map to construct a
2111                  * file extent item for logging without waiting for the
2112                  * respective ordered operation to finish - this file extent
2113                  * item points to a disk location that might not have yet been
2114                  * written to, containing random data - so after a crash a log
2115                  * replay will make our inode have file extent items that point
2116                  * to disk locations containing invalid data, as we returned
2117                  * success to userspace without waiting for the respective
2118                  * ordered operation to finish, because it wasn't captured by
2119                  * btrfs_get_logged_extents().
2120                  */
2121                 ret = start_ordered_ops(inode, start, end);
2122         }
2123         if (ret) {
2124                 inode_unlock(inode);
2125                 goto out;
2126         }
2127         atomic_inc(&root->log_batch);
2128
2129         /*
2130          * If the last transaction that changed this file was before the current
2131          * transaction and we have the full sync flag set in our inode, we can
2132          * bail out now without any syncing.
2133          *
2134          * Note that we can't bail out if the full sync flag isn't set. This is
2135          * because when the full sync flag is set we start all ordered extents
2136          * and wait for them to fully complete - when they complete they update
2137          * the inode's last_trans field through:
2138          *
2139          *     btrfs_finish_ordered_io() ->
2140          *         btrfs_update_inode_fallback() ->
2141          *             btrfs_update_inode() ->
2142          *                 btrfs_set_inode_last_trans()
2143          *
2144          * So we are sure that last_trans is up to date and can do this check to
2145          * bail out safely. For the fast path, when the full sync flag is not
2146          * set in our inode, we can not do it because we start only our ordered
2147          * extents and don't wait for them to complete (that is when
2148          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2149          * value might be less than or equals to fs_info->last_trans_committed,
2150          * and setting a speculative last_trans for an inode when a buffered
2151          * write is made (such as fs_info->generation + 1 for example) would not
2152          * be reliable since after setting the value and before fsync is called
2153          * any number of transactions can start and commit (transaction kthread
2154          * commits the current transaction periodically), and a transaction
2155          * commit does not start nor waits for ordered extents to complete.
2156          */
2157         smp_mb();
2158         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2159             (full_sync && BTRFS_I(inode)->last_trans <=
2160              fs_info->last_trans_committed) ||
2161             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2162              BTRFS_I(inode)->last_trans
2163              <= fs_info->last_trans_committed)) {
2164                 /*
2165                  * We've had everything committed since the last time we were
2166                  * modified so clear this flag in case it was set for whatever
2167                  * reason, it's no longer relevant.
2168                  */
2169                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2170                           &BTRFS_I(inode)->runtime_flags);
2171                 /*
2172                  * An ordered extent might have started before and completed
2173                  * already with io errors, in which case the inode was not
2174                  * updated and we end up here. So check the inode's mapping
2175                  * for any errors that might have happened since we last
2176                  * checked called fsync.
2177                  */
2178                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2179                 inode_unlock(inode);
2180                 goto out;
2181         }
2182
2183         /*
2184          * ok we haven't committed the transaction yet, lets do a commit
2185          */
2186         if (file->private_data)
2187                 btrfs_ioctl_trans_end(file);
2188
2189         /*
2190          * We use start here because we will need to wait on the IO to complete
2191          * in btrfs_sync_log, which could require joining a transaction (for
2192          * example checking cross references in the nocow path).  If we use join
2193          * here we could get into a situation where we're waiting on IO to
2194          * happen that is blocked on a transaction trying to commit.  With start
2195          * we inc the extwriter counter, so we wait for all extwriters to exit
2196          * before we start blocking join'ers.  This comment is to keep somebody
2197          * from thinking they are super smart and changing this to
2198          * btrfs_join_transaction *cough*Josef*cough*.
2199          */
2200         trans = btrfs_start_transaction(root, 0);
2201         if (IS_ERR(trans)) {
2202                 ret = PTR_ERR(trans);
2203                 inode_unlock(inode);
2204                 goto out;
2205         }
2206         trans->sync = true;
2207
2208         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2209         if (ret < 0) {
2210                 /* Fallthrough and commit/free transaction. */
2211                 ret = 1;
2212         }
2213
2214         /* we've logged all the items and now have a consistent
2215          * version of the file in the log.  It is possible that
2216          * someone will come in and modify the file, but that's
2217          * fine because the log is consistent on disk, and we
2218          * have references to all of the file's extents
2219          *
2220          * It is possible that someone will come in and log the
2221          * file again, but that will end up using the synchronization
2222          * inside btrfs_sync_log to keep things safe.
2223          */
2224         inode_unlock(inode);
2225
2226         /*
2227          * If any of the ordered extents had an error, just return it to user
2228          * space, so that the application knows some writes didn't succeed and
2229          * can take proper action (retry for e.g.). Blindly committing the
2230          * transaction in this case, would fool userspace that everything was
2231          * successful. And we also want to make sure our log doesn't contain
2232          * file extent items pointing to extents that weren't fully written to -
2233          * just like in the non fast fsync path, where we check for the ordered
2234          * operation's error flag before writing to the log tree and return -EIO
2235          * if any of them had this flag set (btrfs_wait_ordered_range) -
2236          * therefore we need to check for errors in the ordered operations,
2237          * which are indicated by ctx.io_err.
2238          */
2239         if (ctx.io_err) {
2240                 btrfs_end_transaction(trans);
2241                 ret = ctx.io_err;
2242                 goto out;
2243         }
2244
2245         if (ret != BTRFS_NO_LOG_SYNC) {
2246                 if (!ret) {
2247                         ret = btrfs_sync_log(trans, root, &ctx);
2248                         if (!ret) {
2249                                 ret = btrfs_end_transaction(trans);
2250                                 goto out;
2251                         }
2252                 }
2253                 if (!full_sync) {
2254                         ret = btrfs_wait_ordered_range(inode, start, len);
2255                         if (ret) {
2256                                 btrfs_end_transaction(trans);
2257                                 goto out;
2258                         }
2259                 }
2260                 ret = btrfs_commit_transaction(trans);
2261         } else {
2262                 ret = btrfs_end_transaction(trans);
2263         }
2264 out:
2265         ASSERT(list_empty(&ctx.list));
2266         err = file_check_and_advance_wb_err(file);
2267         if (!ret)
2268                 ret = err;
2269         return ret > 0 ? -EIO : ret;
2270 }
2271
2272 static const struct vm_operations_struct btrfs_file_vm_ops = {
2273         .fault          = filemap_fault,
2274         .map_pages      = filemap_map_pages,
2275         .page_mkwrite   = btrfs_page_mkwrite,
2276 };
2277
2278 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2279 {
2280         struct address_space *mapping = filp->f_mapping;
2281
2282         if (!mapping->a_ops->readpage)
2283                 return -ENOEXEC;
2284
2285         file_accessed(filp);
2286         vma->vm_ops = &btrfs_file_vm_ops;
2287
2288         return 0;
2289 }
2290
2291 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2292                           int slot, u64 start, u64 end)
2293 {
2294         struct btrfs_file_extent_item *fi;
2295         struct btrfs_key key;
2296
2297         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2298                 return 0;
2299
2300         btrfs_item_key_to_cpu(leaf, &key, slot);
2301         if (key.objectid != btrfs_ino(inode) ||
2302             key.type != BTRFS_EXTENT_DATA_KEY)
2303                 return 0;
2304
2305         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2306
2307         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2308                 return 0;
2309
2310         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2311                 return 0;
2312
2313         if (key.offset == end)
2314                 return 1;
2315         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2316                 return 1;
2317         return 0;
2318 }
2319
2320 static int fill_holes(struct btrfs_trans_handle *trans,
2321                 struct btrfs_inode *inode,
2322                 struct btrfs_path *path, u64 offset, u64 end)
2323 {
2324         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2325         struct btrfs_root *root = inode->root;
2326         struct extent_buffer *leaf;
2327         struct btrfs_file_extent_item *fi;
2328         struct extent_map *hole_em;
2329         struct extent_map_tree *em_tree = &inode->extent_tree;
2330         struct btrfs_key key;
2331         int ret;
2332
2333         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2334                 goto out;
2335
2336         key.objectid = btrfs_ino(inode);
2337         key.type = BTRFS_EXTENT_DATA_KEY;
2338         key.offset = offset;
2339
2340         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2341         if (ret <= 0) {
2342                 /*
2343                  * We should have dropped this offset, so if we find it then
2344                  * something has gone horribly wrong.
2345                  */
2346                 if (ret == 0)
2347                         ret = -EINVAL;
2348                 return ret;
2349         }
2350
2351         leaf = path->nodes[0];
2352         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2353                 u64 num_bytes;
2354
2355                 path->slots[0]--;
2356                 fi = btrfs_item_ptr(leaf, path->slots[0],
2357                                     struct btrfs_file_extent_item);
2358                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2359                         end - offset;
2360                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2361                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2362                 btrfs_set_file_extent_offset(leaf, fi, 0);
2363                 btrfs_mark_buffer_dirty(leaf);
2364                 goto out;
2365         }
2366
2367         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2368                 u64 num_bytes;
2369
2370                 key.offset = offset;
2371                 btrfs_set_item_key_safe(fs_info, path, &key);
2372                 fi = btrfs_item_ptr(leaf, path->slots[0],
2373                                     struct btrfs_file_extent_item);
2374                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2375                         offset;
2376                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2377                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2378                 btrfs_set_file_extent_offset(leaf, fi, 0);
2379                 btrfs_mark_buffer_dirty(leaf);
2380                 goto out;
2381         }
2382         btrfs_release_path(path);
2383
2384         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2385                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2386         if (ret)
2387                 return ret;
2388
2389 out:
2390         btrfs_release_path(path);
2391
2392         hole_em = alloc_extent_map();
2393         if (!hole_em) {
2394                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2395                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2396         } else {
2397                 hole_em->start = offset;
2398                 hole_em->len = end - offset;
2399                 hole_em->ram_bytes = hole_em->len;
2400                 hole_em->orig_start = offset;
2401
2402                 hole_em->block_start = EXTENT_MAP_HOLE;
2403                 hole_em->block_len = 0;
2404                 hole_em->orig_block_len = 0;
2405                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2406                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2407                 hole_em->generation = trans->transid;
2408
2409                 do {
2410                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2411                         write_lock(&em_tree->lock);
2412                         ret = add_extent_mapping(em_tree, hole_em, 1);
2413                         write_unlock(&em_tree->lock);
2414                 } while (ret == -EEXIST);
2415                 free_extent_map(hole_em);
2416                 if (ret)
2417                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2418                                         &inode->runtime_flags);
2419         }
2420
2421         return 0;
2422 }
2423
2424 /*
2425  * Find a hole extent on given inode and change start/len to the end of hole
2426  * extent.(hole/vacuum extent whose em->start <= start &&
2427  *         em->start + em->len > start)
2428  * When a hole extent is found, return 1 and modify start/len.
2429  */
2430 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2431 {
2432         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2433         struct extent_map *em;
2434         int ret = 0;
2435
2436         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2437                               round_down(*start, fs_info->sectorsize),
2438                               round_up(*len, fs_info->sectorsize), 0);
2439         if (IS_ERR(em))
2440                 return PTR_ERR(em);
2441
2442         /* Hole or vacuum extent(only exists in no-hole mode) */
2443         if (em->block_start == EXTENT_MAP_HOLE) {
2444                 ret = 1;
2445                 *len = em->start + em->len > *start + *len ?
2446                        0 : *start + *len - em->start - em->len;
2447                 *start = em->start + em->len;
2448         }
2449         free_extent_map(em);
2450         return ret;
2451 }
2452
2453 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2454 {
2455         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2456         struct btrfs_root *root = BTRFS_I(inode)->root;
2457         struct extent_state *cached_state = NULL;
2458         struct btrfs_path *path;
2459         struct btrfs_block_rsv *rsv;
2460         struct btrfs_trans_handle *trans;
2461         u64 lockstart;
2462         u64 lockend;
2463         u64 tail_start;
2464         u64 tail_len;
2465         u64 orig_start = offset;
2466         u64 cur_offset;
2467         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2468         u64 drop_end;
2469         int ret = 0;
2470         int err = 0;
2471         unsigned int rsv_count;
2472         bool same_block;
2473         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2474         u64 ino_size;
2475         bool truncated_block = false;
2476         bool updated_inode = false;
2477
2478         ret = btrfs_wait_ordered_range(inode, offset, len);
2479         if (ret)
2480                 return ret;
2481
2482         inode_lock(inode);
2483         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2484         ret = find_first_non_hole(inode, &offset, &len);
2485         if (ret < 0)
2486                 goto out_only_mutex;
2487         if (ret && !len) {
2488                 /* Already in a large hole */
2489                 ret = 0;
2490                 goto out_only_mutex;
2491         }
2492
2493         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2494         lockend = round_down(offset + len,
2495                              btrfs_inode_sectorsize(inode)) - 1;
2496         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2497                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2498         /*
2499          * We needn't truncate any block which is beyond the end of the file
2500          * because we are sure there is no data there.
2501          */
2502         /*
2503          * Only do this if we are in the same block and we aren't doing the
2504          * entire block.
2505          */
2506         if (same_block && len < fs_info->sectorsize) {
2507                 if (offset < ino_size) {
2508                         truncated_block = true;
2509                         ret = btrfs_truncate_block(inode, offset, len, 0);
2510                 } else {
2511                         ret = 0;
2512                 }
2513                 goto out_only_mutex;
2514         }
2515
2516         /* zero back part of the first block */
2517         if (offset < ino_size) {
2518                 truncated_block = true;
2519                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2520                 if (ret) {
2521                         inode_unlock(inode);
2522                         return ret;
2523                 }
2524         }
2525
2526         /* Check the aligned pages after the first unaligned page,
2527          * if offset != orig_start, which means the first unaligned page
2528          * including several following pages are already in holes,
2529          * the extra check can be skipped */
2530         if (offset == orig_start) {
2531                 /* after truncate page, check hole again */
2532                 len = offset + len - lockstart;
2533                 offset = lockstart;
2534                 ret = find_first_non_hole(inode, &offset, &len);
2535                 if (ret < 0)
2536                         goto out_only_mutex;
2537                 if (ret && !len) {
2538                         ret = 0;
2539                         goto out_only_mutex;
2540                 }
2541                 lockstart = offset;
2542         }
2543
2544         /* Check the tail unaligned part is in a hole */
2545         tail_start = lockend + 1;
2546         tail_len = offset + len - tail_start;
2547         if (tail_len) {
2548                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2549                 if (unlikely(ret < 0))
2550                         goto out_only_mutex;
2551                 if (!ret) {
2552                         /* zero the front end of the last page */
2553                         if (tail_start + tail_len < ino_size) {
2554                                 truncated_block = true;
2555                                 ret = btrfs_truncate_block(inode,
2556                                                         tail_start + tail_len,
2557                                                         0, 1);
2558                                 if (ret)
2559                                         goto out_only_mutex;
2560                         }
2561                 }
2562         }
2563
2564         if (lockend < lockstart) {
2565                 ret = 0;
2566                 goto out_only_mutex;
2567         }
2568
2569         while (1) {
2570                 struct btrfs_ordered_extent *ordered;
2571
2572                 truncate_pagecache_range(inode, lockstart, lockend);
2573
2574                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2575                                  &cached_state);
2576                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2577
2578                 /*
2579                  * We need to make sure we have no ordered extents in this range
2580                  * and nobody raced in and read a page in this range, if we did
2581                  * we need to try again.
2582                  */
2583                 if ((!ordered ||
2584                     (ordered->file_offset + ordered->len <= lockstart ||
2585                      ordered->file_offset > lockend)) &&
2586                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2587                         if (ordered)
2588                                 btrfs_put_ordered_extent(ordered);
2589                         break;
2590                 }
2591                 if (ordered)
2592                         btrfs_put_ordered_extent(ordered);
2593                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2594                                      lockend, &cached_state, GFP_NOFS);
2595                 ret = btrfs_wait_ordered_range(inode, lockstart,
2596                                                lockend - lockstart + 1);
2597                 if (ret) {
2598                         inode_unlock(inode);
2599                         return ret;
2600                 }
2601         }
2602
2603         path = btrfs_alloc_path();
2604         if (!path) {
2605                 ret = -ENOMEM;
2606                 goto out;
2607         }
2608
2609         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2610         if (!rsv) {
2611                 ret = -ENOMEM;
2612                 goto out_free;
2613         }
2614         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2615         rsv->failfast = 1;
2616
2617         /*
2618          * 1 - update the inode
2619          * 1 - removing the extents in the range
2620          * 1 - adding the hole extent if no_holes isn't set
2621          */
2622         rsv_count = no_holes ? 2 : 3;
2623         trans = btrfs_start_transaction(root, rsv_count);
2624         if (IS_ERR(trans)) {
2625                 err = PTR_ERR(trans);
2626                 goto out_free;
2627         }
2628
2629         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2630                                       min_size, 0);
2631         BUG_ON(ret);
2632         trans->block_rsv = rsv;
2633
2634         cur_offset = lockstart;
2635         len = lockend - cur_offset;
2636         while (cur_offset < lockend) {
2637                 ret = __btrfs_drop_extents(trans, root, inode, path,
2638                                            cur_offset, lockend + 1,
2639                                            &drop_end, 1, 0, 0, NULL);
2640                 if (ret != -ENOSPC)
2641                         break;
2642
2643                 trans->block_rsv = &fs_info->trans_block_rsv;
2644
2645                 if (cur_offset < drop_end && cur_offset < ino_size) {
2646                         ret = fill_holes(trans, BTRFS_I(inode), path,
2647                                         cur_offset, drop_end);
2648                         if (ret) {
2649                                 /*
2650                                  * If we failed then we didn't insert our hole
2651                                  * entries for the area we dropped, so now the
2652                                  * fs is corrupted, so we must abort the
2653                                  * transaction.
2654                                  */
2655                                 btrfs_abort_transaction(trans, ret);
2656                                 err = ret;
2657                                 break;
2658                         }
2659                 }
2660
2661                 cur_offset = drop_end;
2662
2663                 ret = btrfs_update_inode(trans, root, inode);
2664                 if (ret) {
2665                         err = ret;
2666                         break;
2667                 }
2668
2669                 btrfs_end_transaction(trans);
2670                 btrfs_btree_balance_dirty(fs_info);
2671
2672                 trans = btrfs_start_transaction(root, rsv_count);
2673                 if (IS_ERR(trans)) {
2674                         ret = PTR_ERR(trans);
2675                         trans = NULL;
2676                         break;
2677                 }
2678
2679                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2680                                               rsv, min_size, 0);
2681                 BUG_ON(ret);    /* shouldn't happen */
2682                 trans->block_rsv = rsv;
2683
2684                 ret = find_first_non_hole(inode, &cur_offset, &len);
2685                 if (unlikely(ret < 0))
2686                         break;
2687                 if (ret && !len) {
2688                         ret = 0;
2689                         break;
2690                 }
2691         }
2692
2693         if (ret) {
2694                 err = ret;
2695                 goto out_trans;
2696         }
2697
2698         trans->block_rsv = &fs_info->trans_block_rsv;
2699         /*
2700          * If we are using the NO_HOLES feature we might have had already an
2701          * hole that overlaps a part of the region [lockstart, lockend] and
2702          * ends at (or beyond) lockend. Since we have no file extent items to
2703          * represent holes, drop_end can be less than lockend and so we must
2704          * make sure we have an extent map representing the existing hole (the
2705          * call to __btrfs_drop_extents() might have dropped the existing extent
2706          * map representing the existing hole), otherwise the fast fsync path
2707          * will not record the existence of the hole region
2708          * [existing_hole_start, lockend].
2709          */
2710         if (drop_end <= lockend)
2711                 drop_end = lockend + 1;
2712         /*
2713          * Don't insert file hole extent item if it's for a range beyond eof
2714          * (because it's useless) or if it represents a 0 bytes range (when
2715          * cur_offset == drop_end).
2716          */
2717         if (cur_offset < ino_size && cur_offset < drop_end) {
2718                 ret = fill_holes(trans, BTRFS_I(inode), path,
2719                                 cur_offset, drop_end);
2720                 if (ret) {
2721                         /* Same comment as above. */
2722                         btrfs_abort_transaction(trans, ret);
2723                         err = ret;
2724                         goto out_trans;
2725                 }
2726         }
2727
2728 out_trans:
2729         if (!trans)
2730                 goto out_free;
2731
2732         inode_inc_iversion(inode);
2733         inode->i_mtime = inode->i_ctime = current_time(inode);
2734
2735         trans->block_rsv = &fs_info->trans_block_rsv;
2736         ret = btrfs_update_inode(trans, root, inode);
2737         updated_inode = true;
2738         btrfs_end_transaction(trans);
2739         btrfs_btree_balance_dirty(fs_info);
2740 out_free:
2741         btrfs_free_path(path);
2742         btrfs_free_block_rsv(fs_info, rsv);
2743 out:
2744         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2745                              &cached_state, GFP_NOFS);
2746 out_only_mutex:
2747         if (!updated_inode && truncated_block && !ret && !err) {
2748                 /*
2749                  * If we only end up zeroing part of a page, we still need to
2750                  * update the inode item, so that all the time fields are
2751                  * updated as well as the necessary btrfs inode in memory fields
2752                  * for detecting, at fsync time, if the inode isn't yet in the
2753                  * log tree or it's there but not up to date.
2754                  */
2755                 trans = btrfs_start_transaction(root, 1);
2756                 if (IS_ERR(trans)) {
2757                         err = PTR_ERR(trans);
2758                 } else {
2759                         err = btrfs_update_inode(trans, root, inode);
2760                         ret = btrfs_end_transaction(trans);
2761                 }
2762         }
2763         inode_unlock(inode);
2764         if (ret && !err)
2765                 err = ret;
2766         return err;
2767 }
2768
2769 /* Helper structure to record which range is already reserved */
2770 struct falloc_range {
2771         struct list_head list;
2772         u64 start;
2773         u64 len;
2774 };
2775
2776 /*
2777  * Helper function to add falloc range
2778  *
2779  * Caller should have locked the larger range of extent containing
2780  * [start, len)
2781  */
2782 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2783 {
2784         struct falloc_range *prev = NULL;
2785         struct falloc_range *range = NULL;
2786
2787         if (list_empty(head))
2788                 goto insert;
2789
2790         /*
2791          * As fallocate iterate by bytenr order, we only need to check
2792          * the last range.
2793          */
2794         prev = list_entry(head->prev, struct falloc_range, list);
2795         if (prev->start + prev->len == start) {
2796                 prev->len += len;
2797                 return 0;
2798         }
2799 insert:
2800         range = kmalloc(sizeof(*range), GFP_KERNEL);
2801         if (!range)
2802                 return -ENOMEM;
2803         range->start = start;
2804         range->len = len;
2805         list_add_tail(&range->list, head);
2806         return 0;
2807 }
2808
2809 static long btrfs_fallocate(struct file *file, int mode,
2810                             loff_t offset, loff_t len)
2811 {
2812         struct inode *inode = file_inode(file);
2813         struct extent_state *cached_state = NULL;
2814         struct extent_changeset *data_reserved = NULL;
2815         struct falloc_range *range;
2816         struct falloc_range *tmp;
2817         struct list_head reserve_list;
2818         u64 cur_offset;
2819         u64 last_byte;
2820         u64 alloc_start;
2821         u64 alloc_end;
2822         u64 alloc_hint = 0;
2823         u64 locked_end;
2824         u64 actual_end = 0;
2825         struct extent_map *em;
2826         int blocksize = btrfs_inode_sectorsize(inode);
2827         int ret;
2828
2829         alloc_start = round_down(offset, blocksize);
2830         alloc_end = round_up(offset + len, blocksize);
2831         cur_offset = alloc_start;
2832
2833         /* Make sure we aren't being give some crap mode */
2834         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2835                 return -EOPNOTSUPP;
2836
2837         if (mode & FALLOC_FL_PUNCH_HOLE)
2838                 return btrfs_punch_hole(inode, offset, len);
2839
2840         /*
2841          * Only trigger disk allocation, don't trigger qgroup reserve
2842          *
2843          * For qgroup space, it will be checked later.
2844          */
2845         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2846                         alloc_end - alloc_start);
2847         if (ret < 0)
2848                 return ret;
2849
2850         inode_lock(inode);
2851
2852         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2853                 ret = inode_newsize_ok(inode, offset + len);
2854                 if (ret)
2855                         goto out;
2856         }
2857
2858         /*
2859          * TODO: Move these two operations after we have checked
2860          * accurate reserved space, or fallocate can still fail but
2861          * with page truncated or size expanded.
2862          *
2863          * But that's a minor problem and won't do much harm BTW.
2864          */
2865         if (alloc_start > inode->i_size) {
2866                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2867                                         alloc_start);
2868                 if (ret)
2869                         goto out;
2870         } else if (offset + len > inode->i_size) {
2871                 /*
2872                  * If we are fallocating from the end of the file onward we
2873                  * need to zero out the end of the block if i_size lands in the
2874                  * middle of a block.
2875                  */
2876                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2877                 if (ret)
2878                         goto out;
2879         }
2880
2881         /*
2882          * wait for ordered IO before we have any locks.  We'll loop again
2883          * below with the locks held.
2884          */
2885         ret = btrfs_wait_ordered_range(inode, alloc_start,
2886                                        alloc_end - alloc_start);
2887         if (ret)
2888                 goto out;
2889
2890         locked_end = alloc_end - 1;
2891         while (1) {
2892                 struct btrfs_ordered_extent *ordered;
2893
2894                 /* the extent lock is ordered inside the running
2895                  * transaction
2896                  */
2897                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2898                                  locked_end, &cached_state);
2899                 ordered = btrfs_lookup_first_ordered_extent(inode,
2900                                                             alloc_end - 1);
2901                 if (ordered &&
2902                     ordered->file_offset + ordered->len > alloc_start &&
2903                     ordered->file_offset < alloc_end) {
2904                         btrfs_put_ordered_extent(ordered);
2905                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2906                                              alloc_start, locked_end,
2907                                              &cached_state, GFP_KERNEL);
2908                         /*
2909                          * we can't wait on the range with the transaction
2910                          * running or with the extent lock held
2911                          */
2912                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2913                                                        alloc_end - alloc_start);
2914                         if (ret)
2915                                 goto out;
2916                 } else {
2917                         if (ordered)
2918                                 btrfs_put_ordered_extent(ordered);
2919                         break;
2920                 }
2921         }
2922
2923         /* First, check if we exceed the qgroup limit */
2924         INIT_LIST_HEAD(&reserve_list);
2925         while (1) {
2926                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
2927                                       alloc_end - cur_offset, 0);
2928                 if (IS_ERR(em)) {
2929                         ret = PTR_ERR(em);
2930                         break;
2931                 }
2932                 last_byte = min(extent_map_end(em), alloc_end);
2933                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2934                 last_byte = ALIGN(last_byte, blocksize);
2935                 if (em->block_start == EXTENT_MAP_HOLE ||
2936                     (cur_offset >= inode->i_size &&
2937                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2938                         ret = add_falloc_range(&reserve_list, cur_offset,
2939                                                last_byte - cur_offset);
2940                         if (ret < 0) {
2941                                 free_extent_map(em);
2942                                 break;
2943                         }
2944                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2945                                         cur_offset, last_byte - cur_offset);
2946                         if (ret < 0) {
2947                                 free_extent_map(em);
2948                                 break;
2949                         }
2950                 } else {
2951                         /*
2952                          * Do not need to reserve unwritten extent for this
2953                          * range, free reserved data space first, otherwise
2954                          * it'll result in false ENOSPC error.
2955                          */
2956                         btrfs_free_reserved_data_space(inode, data_reserved,
2957                                         cur_offset, last_byte - cur_offset);
2958                 }
2959                 free_extent_map(em);
2960                 cur_offset = last_byte;
2961                 if (cur_offset >= alloc_end)
2962                         break;
2963         }
2964
2965         /*
2966          * If ret is still 0, means we're OK to fallocate.
2967          * Or just cleanup the list and exit.
2968          */
2969         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2970                 if (!ret)
2971                         ret = btrfs_prealloc_file_range(inode, mode,
2972                                         range->start,
2973                                         range->len, i_blocksize(inode),
2974                                         offset + len, &alloc_hint);
2975                 else
2976                         btrfs_free_reserved_data_space(inode,
2977                                         data_reserved, range->start,
2978                                         range->len);
2979                 list_del(&range->list);
2980                 kfree(range);
2981         }
2982         if (ret < 0)
2983                 goto out_unlock;
2984
2985         if (actual_end > inode->i_size &&
2986             !(mode & FALLOC_FL_KEEP_SIZE)) {
2987                 struct btrfs_trans_handle *trans;
2988                 struct btrfs_root *root = BTRFS_I(inode)->root;
2989
2990                 /*
2991                  * We didn't need to allocate any more space, but we
2992                  * still extended the size of the file so we need to
2993                  * update i_size and the inode item.
2994                  */
2995                 trans = btrfs_start_transaction(root, 1);
2996                 if (IS_ERR(trans)) {
2997                         ret = PTR_ERR(trans);
2998                 } else {
2999                         inode->i_ctime = current_time(inode);
3000                         i_size_write(inode, actual_end);
3001                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
3002                         ret = btrfs_update_inode(trans, root, inode);
3003                         if (ret)
3004                                 btrfs_end_transaction(trans);
3005                         else
3006                                 ret = btrfs_end_transaction(trans);
3007                 }
3008         }
3009 out_unlock:
3010         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3011                              &cached_state, GFP_KERNEL);
3012 out:
3013         inode_unlock(inode);
3014         /* Let go of our reservation. */
3015         if (ret != 0)
3016                 btrfs_free_reserved_data_space(inode, data_reserved,
3017                                 alloc_start, alloc_end - cur_offset);
3018         extent_changeset_free(data_reserved);
3019         return ret;
3020 }
3021
3022 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3023 {
3024         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3025         struct extent_map *em = NULL;
3026         struct extent_state *cached_state = NULL;
3027         u64 lockstart;
3028         u64 lockend;
3029         u64 start;
3030         u64 len;
3031         int ret = 0;
3032
3033         if (inode->i_size == 0)
3034                 return -ENXIO;
3035
3036         /*
3037          * *offset can be negative, in this case we start finding DATA/HOLE from
3038          * the very start of the file.
3039          */
3040         start = max_t(loff_t, 0, *offset);
3041
3042         lockstart = round_down(start, fs_info->sectorsize);
3043         lockend = round_up(i_size_read(inode),
3044                            fs_info->sectorsize);
3045         if (lockend <= lockstart)
3046                 lockend = lockstart + fs_info->sectorsize;
3047         lockend--;
3048         len = lockend - lockstart + 1;
3049
3050         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3051                          &cached_state);
3052
3053         while (start < inode->i_size) {
3054                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3055                                 start, len, 0);
3056                 if (IS_ERR(em)) {
3057                         ret = PTR_ERR(em);
3058                         em = NULL;
3059                         break;
3060                 }
3061
3062                 if (whence == SEEK_HOLE &&
3063                     (em->block_start == EXTENT_MAP_HOLE ||
3064                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3065                         break;
3066                 else if (whence == SEEK_DATA &&
3067                            (em->block_start != EXTENT_MAP_HOLE &&
3068                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3069                         break;
3070
3071                 start = em->start + em->len;
3072                 free_extent_map(em);
3073                 em = NULL;
3074                 cond_resched();
3075         }
3076         free_extent_map(em);
3077         if (!ret) {
3078                 if (whence == SEEK_DATA && start >= inode->i_size)
3079                         ret = -ENXIO;
3080                 else
3081                         *offset = min_t(loff_t, start, inode->i_size);
3082         }
3083         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3084                              &cached_state, GFP_NOFS);
3085         return ret;
3086 }
3087
3088 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3089 {
3090         struct inode *inode = file->f_mapping->host;
3091         int ret;
3092
3093         inode_lock(inode);
3094         switch (whence) {
3095         case SEEK_END:
3096         case SEEK_CUR:
3097                 offset = generic_file_llseek(file, offset, whence);
3098                 goto out;
3099         case SEEK_DATA:
3100         case SEEK_HOLE:
3101                 if (offset >= i_size_read(inode)) {
3102                         inode_unlock(inode);
3103                         return -ENXIO;
3104                 }
3105
3106                 ret = find_desired_extent(inode, &offset, whence);
3107                 if (ret) {
3108                         inode_unlock(inode);
3109                         return ret;
3110                 }
3111         }
3112
3113         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3114 out:
3115         inode_unlock(inode);
3116         return offset;
3117 }
3118
3119 static int btrfs_file_open(struct inode *inode, struct file *filp)
3120 {
3121         filp->f_mode |= FMODE_NOWAIT;
3122         return generic_file_open(inode, filp);
3123 }
3124
3125 const struct file_operations btrfs_file_operations = {
3126         .llseek         = btrfs_file_llseek,
3127         .read_iter      = generic_file_read_iter,
3128         .splice_read    = generic_file_splice_read,
3129         .write_iter     = btrfs_file_write_iter,
3130         .mmap           = btrfs_file_mmap,
3131         .open           = btrfs_file_open,
3132         .release        = btrfs_release_file,
3133         .fsync          = btrfs_sync_file,
3134         .fallocate      = btrfs_fallocate,
3135         .unlocked_ioctl = btrfs_ioctl,
3136 #ifdef CONFIG_COMPAT
3137         .compat_ioctl   = btrfs_compat_ioctl,
3138 #endif
3139         .clone_file_range = btrfs_clone_file_range,
3140         .dedupe_file_range = btrfs_dedupe_file_range,
3141 };
3142
3143 void btrfs_auto_defrag_exit(void)
3144 {
3145         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3146 }
3147
3148 int btrfs_auto_defrag_init(void)
3149 {
3150         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3151                                         sizeof(struct inode_defrag), 0,
3152                                         SLAB_MEM_SPREAD,
3153                                         NULL);
3154         if (!btrfs_inode_defrag_cachep)
3155                 return -ENOMEM;
3156
3157         return 0;
3158 }
3159
3160 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3161 {
3162         int ret;
3163
3164         /*
3165          * So with compression we will find and lock a dirty page and clear the
3166          * first one as dirty, setup an async extent, and immediately return
3167          * with the entire range locked but with nobody actually marked with
3168          * writeback.  So we can't just filemap_write_and_wait_range() and
3169          * expect it to work since it will just kick off a thread to do the
3170          * actual work.  So we need to call filemap_fdatawrite_range _again_
3171          * since it will wait on the page lock, which won't be unlocked until
3172          * after the pages have been marked as writeback and so we're good to go
3173          * from there.  We have to do this otherwise we'll miss the ordered
3174          * extents and that results in badness.  Please Josef, do not think you
3175          * know better and pull this out at some point in the future, it is
3176          * right and you are wrong.
3177          */
3178         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3179         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3180                              &BTRFS_I(inode)->runtime_flags))
3181                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3182
3183         return ret;
3184 }