fs/btrfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/time.h>
   9 #include <linux/init.h>
  10 #include <linux/string.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/falloc.h>
  13 #include <linux/writeback.h>
  14 #include <linux/compat.h>
  15 #include <linux/slab.h>
  16 #include <linux/btrfs.h>
  17 #include <linux/uio.h>
  18 #include <linux/iversion.h>
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "btrfs_inode.h"
  23 #include "print-tree.h"
  24 #include "tree-log.h"
  25 #include "locking.h"
  26 #include "volumes.h"
  27 #include "qgroup.h"
  28 #include "compression.h"
  29 #include "delalloc-space.h"
  30 #include "reflink.h"
  31
  32 static struct kmem_cache *btrfs_inode_defrag_cachep;
  33 /*
  34  * when auto defrag is enabled we
  35  * queue up these defrag structs to remember which
  36  * inodes need defragging passes
  37  */
  38 struct inode_defrag {
  39         struct rb_node rb_node;
  40         /* objectid */
  41         u64 ino;
  42         /*
  43          * transid where the defrag was added, we search for
  44          * extents newer than this
  45          */
  46         u64 transid;
  47
  48         /* root objectid */
  49         u64 root;
  50
  51         /* last offset we were able to defrag */
  52         u64 last_offset;
  53
  54         /* if we've wrapped around back to zero once already */
  55         int cycled;
  56 };
  57
  58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  59                                   struct inode_defrag *defrag2)
  60 {
  61         if (defrag1->root > defrag2->root)
  62                 return 1;
  63         else if (defrag1->root < defrag2->root)
  64                 return -1;
  65         else if (defrag1->ino > defrag2->ino)
  66                 return 1;
  67         else if (defrag1->ino < defrag2->ino)
  68                 return -1;
  69         else
  70                 return 0;
  71 }
  72
  73 /* pop a record for an inode into the defrag tree.  The lock
  74  * must be held already
  75  *
  76  * If you're inserting a record for an older transid than an
  77  * existing record, the transid already in the tree is lowered
  78  *
  79  * If an existing record is found the defrag item you
  80  * pass in is freed
  81  */
  82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  83                                     struct inode_defrag *defrag)
  84 {
  85         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  86         struct inode_defrag *entry;
  87         struct rb_node **p;
  88         struct rb_node *parent = NULL;
  89         int ret;
  90
  91         p = &fs_info->defrag_inodes.rb_node;
  92         while (*p) {
  93                 parent = *p;
  94                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  95
  96                 ret = __compare_inode_defrag(defrag, entry);
  97                 if (ret < 0)
  98                         p = &parent->rb_left;
  99                 else if (ret > 0)
 100                         p = &parent->rb_right;
 101                 else {
 102                         /* if we're reinserting an entry for
 103                          * an old defrag run, make sure to
 104                          * lower the transid of our existing record
 105                          */
 106                         if (defrag->transid < entry->transid)
 107                                 entry->transid = defrag->transid;
 108                         if (defrag->last_offset > entry->last_offset)
 109                                 entry->last_offset = defrag->last_offset;
 110                         return -EEXIST;
 111                 }
 112         }
 113         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 114         rb_link_node(&defrag->rb_node, parent, p);
 115         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 116         return 0;
 117 }
 118
 119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 120 {
 121         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 122                 return 0;
 123
 124         if (btrfs_fs_closing(fs_info))
 125                 return 0;
 126
 127         return 1;
 128 }
 129
 130 /*
 131  * insert a defrag record for this inode if auto defrag is
 132  * enabled
 133  */
 134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 135                            struct btrfs_inode *inode)
 136 {
 137         struct btrfs_root *root = inode->root;
 138         struct btrfs_fs_info *fs_info = root->fs_info;
 139         struct inode_defrag *defrag;
 140         u64 transid;
 141         int ret;
 142
 143         if (!__need_auto_defrag(fs_info))
 144                 return 0;
 145
 146         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 147                 return 0;
 148
 149         if (trans)
 150                 transid = trans->transid;
 151         else
 152                 transid = inode->root->last_trans;
 153
 154         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 155         if (!defrag)
 156                 return -ENOMEM;
 157
 158         defrag->ino = btrfs_ino(inode);
 159         defrag->transid = transid;
 160         defrag->root = root->root_key.objectid;
 161
 162         spin_lock(&fs_info->defrag_inodes_lock);
 163         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 164                 /*
 165                  * If we set IN_DEFRAG flag and evict the inode from memory,
 166                  * and then re-read this inode, this new inode doesn't have
 167                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 168                  */
 169                 ret = __btrfs_add_inode_defrag(inode, defrag);
 170                 if (ret)
 171                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         } else {
 173                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         }
 175         spin_unlock(&fs_info->defrag_inodes_lock);
 176         return 0;
 177 }
 178
 179 /*
 180  * Requeue the defrag object. If there is a defrag object that points to
 181  * the same inode in the tree, we will merge them together (by
 182  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 183  */
 184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 185                                        struct inode_defrag *defrag)
 186 {
 187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 188         int ret;
 189
 190         if (!__need_auto_defrag(fs_info))
 191                 goto out;
 192
 193         /*
 194          * Here we don't check the IN_DEFRAG flag, because we need merge
 195          * them together.
 196          */
 197         spin_lock(&fs_info->defrag_inodes_lock);
 198         ret = __btrfs_add_inode_defrag(inode, defrag);
 199         spin_unlock(&fs_info->defrag_inodes_lock);
 200         if (ret)
 201                 goto out;
 202         return;
 203 out:
 204         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 205 }
 206
 207 /*
 208  * pick the defragable inode that we want, if it doesn't exist, we will get
 209  * the next one.
 210  */
 211 static struct inode_defrag *
 212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 213 {
 214         struct inode_defrag *entry = NULL;
 215         struct inode_defrag tmp;
 216         struct rb_node *p;
 217         struct rb_node *parent = NULL;
 218         int ret;
 219
 220         tmp.ino = ino;
 221         tmp.root = root;
 222
 223         spin_lock(&fs_info->defrag_inodes_lock);
 224         p = fs_info->defrag_inodes.rb_node;
 225         while (p) {
 226                 parent = p;
 227                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 228
 229                 ret = __compare_inode_defrag(&tmp, entry);
 230                 if (ret < 0)
 231                         p = parent->rb_left;
 232                 else if (ret > 0)
 233                         p = parent->rb_right;
 234                 else
 235                         goto out;
 236         }
 237
 238         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 239                 parent = rb_next(parent);
 240                 if (parent)
 241                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 242                 else
 243                         entry = NULL;
 244         }
 245 out:
 246         if (entry)
 247                 rb_erase(parent, &fs_info->defrag_inodes);
 248         spin_unlock(&fs_info->defrag_inodes_lock);
 249         return entry;
 250 }
 251
 252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 253 {
 254         struct inode_defrag *defrag;
 255         struct rb_node *node;
 256
 257         spin_lock(&fs_info->defrag_inodes_lock);
 258         node = rb_first(&fs_info->defrag_inodes);
 259         while (node) {
 260                 rb_erase(node, &fs_info->defrag_inodes);
 261                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 262                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 263
 264                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 265
 266                 node = rb_first(&fs_info->defrag_inodes);
 267         }
 268         spin_unlock(&fs_info->defrag_inodes_lock);
 269 }
 270
 271 #define BTRFS_DEFRAG_BATCH      1024
 272
 273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 274                                     struct inode_defrag *defrag)
 275 {
 276         struct btrfs_root *inode_root;
 277         struct inode *inode;
 278         struct btrfs_ioctl_defrag_range_args range;
 279         int num_defrag;
 280         int ret;
 281
 282         /* get the inode */
 283         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 284         if (IS_ERR(inode_root)) {
 285                 ret = PTR_ERR(inode_root);
 286                 goto cleanup;
 287         }
 288
 289         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 290         btrfs_put_root(inode_root);
 291         if (IS_ERR(inode)) {
 292                 ret = PTR_ERR(inode);
 293                 goto cleanup;
 294         }
 295
 296         /* do a chunk of defrag */
 297         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 298         memset(&range, 0, sizeof(range));
 299         range.len = (u64)-1;
 300         range.start = defrag->last_offset;
 301
 302         sb_start_write(fs_info->sb);
 303         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 304                                        BTRFS_DEFRAG_BATCH);
 305         sb_end_write(fs_info->sb);
 306         /*
 307          * if we filled the whole defrag batch, there
 308          * must be more work to do.  Queue this defrag
 309          * again
 310          */
 311         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 312                 defrag->last_offset = range.start;
 313                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 314         } else if (defrag->last_offset && !defrag->cycled) {
 315                 /*
 316                  * we didn't fill our defrag batch, but
 317                  * we didn't start at zero.  Make sure we loop
 318                  * around to the start of the file.
 319                  */
 320                 defrag->last_offset = 0;
 321                 defrag->cycled = 1;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else {
 324                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 325         }
 326
 327         iput(inode);
 328         return 0;
 329 cleanup:
 330         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 331         return ret;
 332 }
 333
 334 /*
 335  * run through the list of inodes in the FS that need
 336  * defragging
 337  */
 338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 339 {
 340         struct inode_defrag *defrag;
 341         u64 first_ino = 0;
 342         u64 root_objectid = 0;
 343
 344         atomic_inc(&fs_info->defrag_running);
 345         while (1) {
 346                 /* Pause the auto defragger. */
 347                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 348                              &fs_info->fs_state))
 349                         break;
 350
 351                 if (!__need_auto_defrag(fs_info))
 352                         break;
 353
 354                 /* find an inode to defrag */
 355                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 356                                                  first_ino);
 357                 if (!defrag) {
 358                         if (root_objectid || first_ino) {
 359                                 root_objectid = 0;
 360                                 first_ino = 0;
 361                                 continue;
 362                         } else {
 363                                 break;
 364                         }
 365                 }
 366
 367                 first_ino = defrag->ino + 1;
 368                 root_objectid = defrag->root;
 369
 370                 __btrfs_run_defrag_inode(fs_info, defrag);
 371         }
 372         atomic_dec(&fs_info->defrag_running);
 373
 374         /*
 375          * during unmount, we use the transaction_wait queue to
 376          * wait for the defragger to stop
 377          */
 378         wake_up(&fs_info->transaction_wait);
 379         return 0;
 380 }
 381
 382 /* simple helper to fault in pages and copy.  This should go away
 383  * and be replaced with calls into generic code.
 384  */
 385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 386                                          struct page **prepared_pages,
 387                                          struct iov_iter *i)
 388 {
 389         size_t copied = 0;
 390         size_t total_copied = 0;
 391         int pg = 0;
 392         int offset = offset_in_page(pos);
 393
 394         while (write_bytes > 0) {
 395                 size_t count = min_t(size_t,
 396                                      PAGE_SIZE - offset, write_bytes);
 397                 struct page *page = prepared_pages[pg];
 398                 /*
 399                  * Copy data from userspace to the current page
 400                  */
 401                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 402
 403                 /* Flush processor's dcache for this page */
 404                 flush_dcache_page(page);
 405
 406                 /*
 407                  * if we get a partial write, we can end up with
 408                  * partially up to date pages.  These add
 409                  * a lot of complexity, so make sure they don't
 410                  * happen by forcing this copy to be retried.
 411                  *
 412                  * The rest of the btrfs_file_write code will fall
 413                  * back to page at a time copies after we return 0.
 414                  */
 415                 if (!PageUptodate(page) && copied < count)
 416                         copied = 0;
 417
 418                 iov_iter_advance(i, copied);
 419                 write_bytes -= copied;
 420                 total_copied += copied;
 421
 422                 /* Return to btrfs_file_write_iter to fault page */
 423                 if (unlikely(copied == 0))
 424                         break;
 425
 426                 if (copied < PAGE_SIZE - offset) {
 427                         offset += copied;
 428                 } else {
 429                         pg++;
 430                         offset = 0;
 431                 }
 432         }
 433         return total_copied;
 434 }
 435
 436 /*
 437  * unlocks pages after btrfs_file_write is done with them
 438  */
 439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 440 {
 441         size_t i;
 442         for (i = 0; i < num_pages; i++) {
 443                 /* page checked is some magic around finding pages that
 444                  * have been modified without going through btrfs_set_page_dirty
 445                  * clear it here. There should be no need to mark the pages
 446                  * accessed as prepare_pages should have marked them accessed
 447                  * in prepare_pages via find_or_create_page()
 448                  */
 449                 ClearPageChecked(pages[i]);
 450                 unlock_page(pages[i]);
 451                 put_page(pages[i]);
 452         }
 453 }
 454
 455 /*
 456  * after copy_from_user, pages need to be dirtied and we need to make
 457  * sure holes are created between the current EOF and the start of
 458  * any next extents (if required).
 459  *
 460  * this also makes the decision about creating an inline extent vs
 461  * doing real data extents, marking pages dirty and delalloc as required.
 462  */
 463 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 464                       size_t num_pages, loff_t pos, size_t write_bytes,
 465                       struct extent_state **cached)
 466 {
 467         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 468         int err = 0;
 469         int i;
 470         u64 num_bytes;
 471         u64 start_pos;
 472         u64 end_of_last_block;
 473         u64 end_pos = pos + write_bytes;
 474         loff_t isize = i_size_read(&inode->vfs_inode);
 475         unsigned int extra_bits = 0;
 476
 477         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 478         num_bytes = round_up(write_bytes + pos - start_pos,
 479                              fs_info->sectorsize);
 480
 481         end_of_last_block = start_pos + num_bytes - 1;
 482
 483         /*
 484          * The pages may have already been dirty, clear out old accounting so
 485          * we can set things up properly
 486          */
 487         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 488                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 489                          0, 0, cached);
 490
 491         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 492                                         extra_bits, cached);
 493         if (err)
 494                 return err;
 495
 496         for (i = 0; i < num_pages; i++) {
 497                 struct page *p = pages[i];
 498                 SetPageUptodate(p);
 499                 ClearPageChecked(p);
 500                 set_page_dirty(p);
 501         }
 502
 503         /*
 504          * we've only changed i_size in ram, and we haven't updated
 505          * the disk i_size.  There is no need to log the inode
 506          * at this time.
 507          */
 508         if (end_pos > isize)
 509                 i_size_write(&inode->vfs_inode, end_pos);
 510         return 0;
 511 }
 512
 513 /*
 514  * this drops all the extents in the cache that intersect the range
 515  * [start, end].  Existing extents are split as required.
 516  */
 517 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 518                              int skip_pinned)
 519 {
 520         struct extent_map *em;
 521         struct extent_map *split = NULL;
 522         struct extent_map *split2 = NULL;
 523         struct extent_map_tree *em_tree = &inode->extent_tree;
 524         u64 len = end - start + 1;
 525         u64 gen;
 526         int ret;
 527         int testend = 1;
 528         unsigned long flags;
 529         int compressed = 0;
 530         bool modified;
 531
 532         WARN_ON(end < start);
 533         if (end == (u64)-1) {
 534                 len = (u64)-1;
 535                 testend = 0;
 536         }
 537         while (1) {
 538                 int no_splits = 0;
 539
 540                 modified = false;
 541                 if (!split)
 542                         split = alloc_extent_map();
 543                 if (!split2)
 544                         split2 = alloc_extent_map();
 545                 if (!split || !split2)
 546                         no_splits = 1;
 547
 548                 write_lock(&em_tree->lock);
 549                 em = lookup_extent_mapping(em_tree, start, len);
 550                 if (!em) {
 551                         write_unlock(&em_tree->lock);
 552                         break;
 553                 }
 554                 flags = em->flags;
 555                 gen = em->generation;
 556                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 557                         if (testend && em->start + em->len >= start + len) {
 558                                 free_extent_map(em);
 559                                 write_unlock(&em_tree->lock);
 560                                 break;
 561                         }
 562                         start = em->start + em->len;
 563                         if (testend)
 564                                 len = start + len - (em->start + em->len);
 565                         free_extent_map(em);
 566                         write_unlock(&em_tree->lock);
 567                         continue;
 568                 }
 569                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 570                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 571                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 572                 modified = !list_empty(&em->list);
 573                 if (no_splits)
 574                         goto next;
 575
 576                 if (em->start < start) {
 577                         split->start = em->start;
 578                         split->len = start - em->start;
 579
 580                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 581                                 split->orig_start = em->orig_start;
 582                                 split->block_start = em->block_start;
 583
 584                                 if (compressed)
 585                                         split->block_len = em->block_len;
 586                                 else
 587                                         split->block_len = split->len;
 588                                 split->orig_block_len = max(split->block_len,
 589                                                 em->orig_block_len);
 590                                 split->ram_bytes = em->ram_bytes;
 591                         } else {
 592                                 split->orig_start = split->start;
 593                                 split->block_len = 0;
 594                                 split->block_start = em->block_start;
 595                                 split->orig_block_len = 0;
 596                                 split->ram_bytes = split->len;
 597                         }
 598
 599                         split->generation = gen;
 600                         split->flags = flags;
 601                         split->compress_type = em->compress_type;
 602                         replace_extent_mapping(em_tree, em, split, modified);
 603                         free_extent_map(split);
 604                         split = split2;
 605                         split2 = NULL;
 606                 }
 607                 if (testend && em->start + em->len > start + len) {
 608                         u64 diff = start + len - em->start;
 609
 610                         split->start = start + len;
 611                         split->len = em->start + em->len - (start + len);
 612                         split->flags = flags;
 613                         split->compress_type = em->compress_type;
 614                         split->generation = gen;
 615
 616                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 617                                 split->orig_block_len = max(em->block_len,
 618                                                     em->orig_block_len);
 619
 620                                 split->ram_bytes = em->ram_bytes;
 621                                 if (compressed) {
 622                                         split->block_len = em->block_len;
 623                                         split->block_start = em->block_start;
 624                                         split->orig_start = em->orig_start;
 625                                 } else {
 626                                         split->block_len = split->len;
 627                                         split->block_start = em->block_start
 628                                                 + diff;
 629                                         split->orig_start = em->orig_start;
 630                                 }
 631                         } else {
 632                                 split->ram_bytes = split->len;
 633                                 split->orig_start = split->start;
 634                                 split->block_len = 0;
 635                                 split->block_start = em->block_start;
 636                                 split->orig_block_len = 0;
 637                         }
 638
 639                         if (extent_map_in_tree(em)) {
 640                                 replace_extent_mapping(em_tree, em, split,
 641                                                        modified);
 642                         } else {
 643                                 ret = add_extent_mapping(em_tree, split,
 644                                                          modified);
 645                                 ASSERT(ret == 0); /* Logic error */
 646                         }
 647                         free_extent_map(split);
 648                         split = NULL;
 649                 }
 650 next:
 651                 if (extent_map_in_tree(em))
 652                         remove_extent_mapping(em_tree, em);
 653                 write_unlock(&em_tree->lock);
 654
 655                 /* once for us */
 656                 free_extent_map(em);
 657                 /* once for the tree*/
 658                 free_extent_map(em);
 659         }
 660         if (split)
 661                 free_extent_map(split);
 662         if (split2)
 663                 free_extent_map(split2);
 664 }
 665
 666 /*
 667  * this is very complex, but the basic idea is to drop all extents
 668  * in the range start - end.  hint_block is filled in with a block number
 669  * that would be a good hint to the block allocator for this file.
 670  *
 671  * If an extent intersects the range but is not entirely inside the range
 672  * it is either truncated or split.  Anything entirely inside the range
 673  * is deleted from the tree.
 674  */
 675 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 676                          struct btrfs_root *root, struct btrfs_inode *inode,
 677                          struct btrfs_path *path, u64 start, u64 end,
 678                          u64 *drop_end, int drop_cache,
 679                          int replace_extent,
 680                          u32 extent_item_size,
 681                          int *key_inserted)
 682 {
 683         struct btrfs_fs_info *fs_info = root->fs_info;
 684         struct extent_buffer *leaf;
 685         struct btrfs_file_extent_item *fi;
 686         struct btrfs_ref ref = { 0 };
 687         struct btrfs_key key;
 688         struct btrfs_key new_key;
 689         struct inode *vfs_inode = &inode->vfs_inode;
 690         u64 ino = btrfs_ino(inode);
 691         u64 search_start = start;
 692         u64 disk_bytenr = 0;
 693         u64 num_bytes = 0;
 694         u64 extent_offset = 0;
 695         u64 extent_end = 0;
 696         u64 last_end = start;
 697         int del_nr = 0;
 698         int del_slot = 0;
 699         int extent_type;
 700         int recow;
 701         int ret;
 702         int modify_tree = -1;
 703         int update_refs;
 704         int found = 0;
 705         int leafs_visited = 0;
 706
 707         if (drop_cache)
 708                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 709
 710         if (start >= inode->disk_i_size && !replace_extent)
 711                 modify_tree = 0;
 712
 713         update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
 714         while (1) {
 715                 recow = 0;
 716                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 717                                                search_start, modify_tree);
 718                 if (ret < 0)
 719                         break;
 720                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 721                         leaf = path->nodes[0];
 722                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 723                         if (key.objectid == ino &&
 724                             key.type == BTRFS_EXTENT_DATA_KEY)
 725                                 path->slots[0]--;
 726                 }
 727                 ret = 0;
 728                 leafs_visited++;
 729 next_slot:
 730                 leaf = path->nodes[0];
 731                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 732                         BUG_ON(del_nr > 0);
 733                         ret = btrfs_next_leaf(root, path);
 734                         if (ret < 0)
 735                                 break;
 736                         if (ret > 0) {
 737                                 ret = 0;
 738                                 break;
 739                         }
 740                         leafs_visited++;
 741                         leaf = path->nodes[0];
 742                         recow = 1;
 743                 }
 744
 745                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 746
 747                 if (key.objectid > ino)
 748                         break;
 749                 if (WARN_ON_ONCE(key.objectid < ino) ||
 750                     key.type < BTRFS_EXTENT_DATA_KEY) {
 751                         ASSERT(del_nr == 0);
 752                         path->slots[0]++;
 753                         goto next_slot;
 754                 }
 755                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 756                         break;
 757
 758                 fi = btrfs_item_ptr(leaf, path->slots[0],
 759                                     struct btrfs_file_extent_item);
 760                 extent_type = btrfs_file_extent_type(leaf, fi);
 761
 762                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 763                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 764                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 765                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 766                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 767                         extent_end = key.offset +
 768                                 btrfs_file_extent_num_bytes(leaf, fi);
 769                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 770                         extent_end = key.offset +
 771                                 btrfs_file_extent_ram_bytes(leaf, fi);
 772                 } else {
 773                         /* can't happen */
 774                         BUG();
 775                 }
 776
 777                 /*
 778                  * Don't skip extent items representing 0 byte lengths. They
 779                  * used to be created (bug) if while punching holes we hit
 780                  * -ENOSPC condition. So if we find one here, just ensure we
 781                  * delete it, otherwise we would insert a new file extent item
 782                  * with the same key (offset) as that 0 bytes length file
 783                  * extent item in the call to setup_items_for_insert() later
 784                  * in this function.
 785                  */
 786                 if (extent_end == key.offset && extent_end >= search_start) {
 787                         last_end = extent_end;
 788                         goto delete_extent_item;
 789                 }
 790
 791                 if (extent_end <= search_start) {
 792                         path->slots[0]++;
 793                         goto next_slot;
 794                 }
 795
 796                 found = 1;
 797                 search_start = max(key.offset, start);
 798                 if (recow || !modify_tree) {
 799                         modify_tree = -1;
 800                         btrfs_release_path(path);
 801                         continue;
 802                 }
 803
 804                 /*
 805                  *     | - range to drop - |
 806                  *  | -------- extent -------- |
 807                  */
 808                 if (start > key.offset && end < extent_end) {
 809                         BUG_ON(del_nr > 0);
 810                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 811                                 ret = -EOPNOTSUPP;
 812                                 break;
 813                         }
 814
 815                         memcpy(&new_key, &key, sizeof(new_key));
 816                         new_key.offset = start;
 817                         ret = btrfs_duplicate_item(trans, root, path,
 818                                                    &new_key);
 819                         if (ret == -EAGAIN) {
 820                                 btrfs_release_path(path);
 821                                 continue;
 822                         }
 823                         if (ret < 0)
 824                                 break;
 825
 826                         leaf = path->nodes[0];
 827                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 828                                             struct btrfs_file_extent_item);
 829                         btrfs_set_file_extent_num_bytes(leaf, fi,
 830                                                         start - key.offset);
 831
 832                         fi = btrfs_item_ptr(leaf, path->slots[0],
 833                                             struct btrfs_file_extent_item);
 834
 835                         extent_offset += start - key.offset;
 836                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 837                         btrfs_set_file_extent_num_bytes(leaf, fi,
 838                                                         extent_end - start);
 839                         btrfs_mark_buffer_dirty(leaf);
 840
 841                         if (update_refs && disk_bytenr > 0) {
 842                                 btrfs_init_generic_ref(&ref,
 843                                                 BTRFS_ADD_DELAYED_REF,
 844                                                 disk_bytenr, num_bytes, 0);
 845                                 btrfs_init_data_ref(&ref,
 846                                                 root->root_key.objectid,
 847                                                 new_key.objectid,
 848                                                 start - extent_offset);
 849                                 ret = btrfs_inc_extent_ref(trans, &ref);
 850                                 BUG_ON(ret); /* -ENOMEM */
 851                         }
 852                         key.offset = start;
 853                 }
 854                 /*
 855                  * From here on out we will have actually dropped something, so
 856                  * last_end can be updated.
 857                  */
 858                 last_end = extent_end;
 859
 860                 /*
 861                  *  | ---- range to drop ----- |
 862                  *      | -------- extent -------- |
 863                  */
 864                 if (start <= key.offset && end < extent_end) {
 865                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 866                                 ret = -EOPNOTSUPP;
 867                                 break;
 868                         }
 869
 870                         memcpy(&new_key, &key, sizeof(new_key));
 871                         new_key.offset = end;
 872                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 873
 874                         extent_offset += end - key.offset;
 875                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 876                         btrfs_set_file_extent_num_bytes(leaf, fi,
 877                                                         extent_end - end);
 878                         btrfs_mark_buffer_dirty(leaf);
 879                         if (update_refs && disk_bytenr > 0)
 880                                 inode_sub_bytes(vfs_inode, end - key.offset);
 881                         break;
 882                 }
 883
 884                 search_start = extent_end;
 885                 /*
 886                  *       | ---- range to drop ----- |
 887                  *  | -------- extent -------- |
 888                  */
 889                 if (start > key.offset && end >= extent_end) {
 890                         BUG_ON(del_nr > 0);
 891                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 892                                 ret = -EOPNOTSUPP;
 893                                 break;
 894                         }
 895
 896                         btrfs_set_file_extent_num_bytes(leaf, fi,
 897                                                         start - key.offset);
 898                         btrfs_mark_buffer_dirty(leaf);
 899                         if (update_refs && disk_bytenr > 0)
 900                                 inode_sub_bytes(vfs_inode, extent_end - start);
 901                         if (end == extent_end)
 902                                 break;
 903
 904                         path->slots[0]++;
 905                         goto next_slot;
 906                 }
 907
 908                 /*
 909                  *  | ---- range to drop ----- |
 910                  *    | ------ extent ------ |
 911                  */
 912                 if (start <= key.offset && end >= extent_end) {
 913 delete_extent_item:
 914                         if (del_nr == 0) {
 915                                 del_slot = path->slots[0];
 916                                 del_nr = 1;
 917                         } else {
 918                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 919                                 del_nr++;
 920                         }
 921
 922                         if (update_refs &&
 923                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 924                                 inode_sub_bytes(vfs_inode,
 925                                                 extent_end - key.offset);
 926                                 extent_end = ALIGN(extent_end,
 927                                                    fs_info->sectorsize);
 928                         } else if (update_refs && disk_bytenr > 0) {
 929                                 btrfs_init_generic_ref(&ref,
 930                                                 BTRFS_DROP_DELAYED_REF,
 931                                                 disk_bytenr, num_bytes, 0);
 932                                 btrfs_init_data_ref(&ref,
 933                                                 root->root_key.objectid,
 934                                                 key.objectid,
 935                                                 key.offset - extent_offset);
 936                                 ret = btrfs_free_extent(trans, &ref);
 937                                 BUG_ON(ret); /* -ENOMEM */
 938                                 inode_sub_bytes(vfs_inode,
 939                                                 extent_end - key.offset);
 940                         }
 941
 942                         if (end == extent_end)
 943                                 break;
 944
 945                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 946                                 path->slots[0]++;
 947                                 goto next_slot;
 948                         }
 949
 950                         ret = btrfs_del_items(trans, root, path, del_slot,
 951                                               del_nr);
 952                         if (ret) {
 953                                 btrfs_abort_transaction(trans, ret);
 954                                 break;
 955                         }
 956
 957                         del_nr = 0;
 958                         del_slot = 0;
 959
 960                         btrfs_release_path(path);
 961                         continue;
 962                 }
 963
 964                 BUG();
 965         }
 966
 967         if (!ret && del_nr > 0) {
 968                 /*
 969                  * Set path->slots[0] to first slot, so that after the delete
 970                  * if items are move off from our leaf to its immediate left or
 971                  * right neighbor leafs, we end up with a correct and adjusted
 972                  * path->slots[0] for our insertion (if replace_extent != 0).
 973                  */
 974                 path->slots[0] = del_slot;
 975                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 976                 if (ret)
 977                         btrfs_abort_transaction(trans, ret);
 978         }
 979
 980         leaf = path->nodes[0];
 981         /*
 982          * If btrfs_del_items() was called, it might have deleted a leaf, in
 983          * which case it unlocked our path, so check path->locks[0] matches a
 984          * write lock.
 985          */
 986         if (!ret && replace_extent && leafs_visited == 1 &&
 987             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 988              path->locks[0] == BTRFS_WRITE_LOCK) &&
 989             btrfs_leaf_free_space(leaf) >=
 990             sizeof(struct btrfs_item) + extent_item_size) {
 991
 992                 key.objectid = ino;
 993                 key.type = BTRFS_EXTENT_DATA_KEY;
 994                 key.offset = start;
 995                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 996                         struct btrfs_key slot_key;
 997
 998                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 999                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1000                                 path->slots[0]++;
1001                 }
1002                 setup_items_for_insert(root, path, &key, &extent_item_size, 1);
1003                 *key_inserted = 1;
1004         }
1005
1006         if (!replace_extent || !(*key_inserted))
1007                 btrfs_release_path(path);
1008         if (drop_end)
1009                 *drop_end = found ? min(end, last_end) : end;
1010         return ret;
1011 }
1012
1013 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1014                        struct btrfs_root *root, struct inode *inode, u64 start,
1015                        u64 end, int drop_cache)
1016 {
1017         struct btrfs_path *path;
1018         int ret;
1019
1020         path = btrfs_alloc_path();
1021         if (!path)
1022                 return -ENOMEM;
1023         ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1024                                    end, NULL, drop_cache, 0, 0, NULL);
1025         btrfs_free_path(path);
1026         return ret;
1027 }
1028
1029 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1030                             u64 objectid, u64 bytenr, u64 orig_offset,
1031                             u64 *start, u64 *end)
1032 {
1033         struct btrfs_file_extent_item *fi;
1034         struct btrfs_key key;
1035         u64 extent_end;
1036
1037         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1038                 return 0;
1039
1040         btrfs_item_key_to_cpu(leaf, &key, slot);
1041         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1042                 return 0;
1043
1044         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1045         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1046             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1047             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1048             btrfs_file_extent_compression(leaf, fi) ||
1049             btrfs_file_extent_encryption(leaf, fi) ||
1050             btrfs_file_extent_other_encoding(leaf, fi))
1051                 return 0;
1052
1053         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1054         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1055                 return 0;
1056
1057         *start = key.offset;
1058         *end = extent_end;
1059         return 1;
1060 }
1061
1062 /*
1063  * Mark extent in the range start - end as written.
1064  *
1065  * This changes extent type from 'pre-allocated' to 'regular'. If only
1066  * part of extent is marked as written, the extent will be split into
1067  * two or three.
1068  */
1069 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1070                               struct btrfs_inode *inode, u64 start, u64 end)
1071 {
1072         struct btrfs_fs_info *fs_info = trans->fs_info;
1073         struct btrfs_root *root = inode->root;
1074         struct extent_buffer *leaf;
1075         struct btrfs_path *path;
1076         struct btrfs_file_extent_item *fi;
1077         struct btrfs_ref ref = { 0 };
1078         struct btrfs_key key;
1079         struct btrfs_key new_key;
1080         u64 bytenr;
1081         u64 num_bytes;
1082         u64 extent_end;
1083         u64 orig_offset;
1084         u64 other_start;
1085         u64 other_end;
1086         u64 split;
1087         int del_nr = 0;
1088         int del_slot = 0;
1089         int recow;
1090         int ret = 0;
1091         u64 ino = btrfs_ino(inode);
1092
1093         path = btrfs_alloc_path();
1094         if (!path)
1095                 return -ENOMEM;
1096 again:
1097         recow = 0;
1098         split = start;
1099         key.objectid = ino;
1100         key.type = BTRFS_EXTENT_DATA_KEY;
1101         key.offset = split;
1102
1103         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1104         if (ret < 0)
1105                 goto out;
1106         if (ret > 0 && path->slots[0] > 0)
1107                 path->slots[0]--;
1108
1109         leaf = path->nodes[0];
1110         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1111         if (key.objectid != ino ||
1112             key.type != BTRFS_EXTENT_DATA_KEY) {
1113                 ret = -EINVAL;
1114                 btrfs_abort_transaction(trans, ret);
1115                 goto out;
1116         }
1117         fi = btrfs_item_ptr(leaf, path->slots[0],
1118                             struct btrfs_file_extent_item);
1119         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1120                 ret = -EINVAL;
1121                 btrfs_abort_transaction(trans, ret);
1122                 goto out;
1123         }
1124         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1125         if (key.offset > start || extent_end < end) {
1126                 ret = -EINVAL;
1127                 btrfs_abort_transaction(trans, ret);
1128                 goto out;
1129         }
1130
1131         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1132         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1133         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1134         memcpy(&new_key, &key, sizeof(new_key));
1135
1136         if (start == key.offset && end < extent_end) {
1137                 other_start = 0;
1138                 other_end = start;
1139                 if (extent_mergeable(leaf, path->slots[0] - 1,
1140                                      ino, bytenr, orig_offset,
1141                                      &other_start, &other_end)) {
1142                         new_key.offset = end;
1143                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1144                         fi = btrfs_item_ptr(leaf, path->slots[0],
1145                                             struct btrfs_file_extent_item);
1146                         btrfs_set_file_extent_generation(leaf, fi,
1147                                                          trans->transid);
1148                         btrfs_set_file_extent_num_bytes(leaf, fi,
1149                                                         extent_end - end);
1150                         btrfs_set_file_extent_offset(leaf, fi,
1151                                                      end - orig_offset);
1152                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1153                                             struct btrfs_file_extent_item);
1154                         btrfs_set_file_extent_generation(leaf, fi,
1155                                                          trans->transid);
1156                         btrfs_set_file_extent_num_bytes(leaf, fi,
1157                                                         end - other_start);
1158                         btrfs_mark_buffer_dirty(leaf);
1159                         goto out;
1160                 }
1161         }
1162
1163         if (start > key.offset && end == extent_end) {
1164                 other_start = end;
1165                 other_end = 0;
1166                 if (extent_mergeable(leaf, path->slots[0] + 1,
1167                                      ino, bytenr, orig_offset,
1168                                      &other_start, &other_end)) {
1169                         fi = btrfs_item_ptr(leaf, path->slots[0],
1170                                             struct btrfs_file_extent_item);
1171                         btrfs_set_file_extent_num_bytes(leaf, fi,
1172                                                         start - key.offset);
1173                         btrfs_set_file_extent_generation(leaf, fi,
1174                                                          trans->transid);
1175                         path->slots[0]++;
1176                         new_key.offset = start;
1177                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1178
1179                         fi = btrfs_item_ptr(leaf, path->slots[0],
1180                                             struct btrfs_file_extent_item);
1181                         btrfs_set_file_extent_generation(leaf, fi,
1182                                                          trans->transid);
1183                         btrfs_set_file_extent_num_bytes(leaf, fi,
1184                                                         other_end - start);
1185                         btrfs_set_file_extent_offset(leaf, fi,
1186                                                      start - orig_offset);
1187                         btrfs_mark_buffer_dirty(leaf);
1188                         goto out;
1189                 }
1190         }
1191
1192         while (start > key.offset || end < extent_end) {
1193                 if (key.offset == start)
1194                         split = end;
1195
1196                 new_key.offset = split;
1197                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1198                 if (ret == -EAGAIN) {
1199                         btrfs_release_path(path);
1200                         goto again;
1201                 }
1202                 if (ret < 0) {
1203                         btrfs_abort_transaction(trans, ret);
1204                         goto out;
1205                 }
1206
1207                 leaf = path->nodes[0];
1208                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1209                                     struct btrfs_file_extent_item);
1210                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1211                 btrfs_set_file_extent_num_bytes(leaf, fi,
1212                                                 split - key.offset);
1213
1214                 fi = btrfs_item_ptr(leaf, path->slots[0],
1215                                     struct btrfs_file_extent_item);
1216
1217                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1218                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1219                 btrfs_set_file_extent_num_bytes(leaf, fi,
1220                                                 extent_end - split);
1221                 btrfs_mark_buffer_dirty(leaf);
1222
1223                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1224                                        num_bytes, 0);
1225                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1226                                     orig_offset);
1227                 ret = btrfs_inc_extent_ref(trans, &ref);
1228                 if (ret) {
1229                         btrfs_abort_transaction(trans, ret);
1230                         goto out;
1231                 }
1232
1233                 if (split == start) {
1234                         key.offset = start;
1235                 } else {
1236                         if (start != key.offset) {
1237                                 ret = -EINVAL;
1238                                 btrfs_abort_transaction(trans, ret);
1239                                 goto out;
1240                         }
1241                         path->slots[0]--;
1242                         extent_end = end;
1243                 }
1244                 recow = 1;
1245         }
1246
1247         other_start = end;
1248         other_end = 0;
1249         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1250                                num_bytes, 0);
1251         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1252         if (extent_mergeable(leaf, path->slots[0] + 1,
1253                              ino, bytenr, orig_offset,
1254                              &other_start, &other_end)) {
1255                 if (recow) {
1256                         btrfs_release_path(path);
1257                         goto again;
1258                 }
1259                 extent_end = other_end;
1260                 del_slot = path->slots[0] + 1;
1261                 del_nr++;
1262                 ret = btrfs_free_extent(trans, &ref);
1263                 if (ret) {
1264                         btrfs_abort_transaction(trans, ret);
1265                         goto out;
1266                 }
1267         }
1268         other_start = 0;
1269         other_end = start;
1270         if (extent_mergeable(leaf, path->slots[0] - 1,
1271                              ino, bytenr, orig_offset,
1272                              &other_start, &other_end)) {
1273                 if (recow) {
1274                         btrfs_release_path(path);
1275                         goto again;
1276                 }
1277                 key.offset = other_start;
1278                 del_slot = path->slots[0];
1279                 del_nr++;
1280                 ret = btrfs_free_extent(trans, &ref);
1281                 if (ret) {
1282                         btrfs_abort_transaction(trans, ret);
1283                         goto out;
1284                 }
1285         }
1286         if (del_nr == 0) {
1287                 fi = btrfs_item_ptr(leaf, path->slots[0],
1288                            struct btrfs_file_extent_item);
1289                 btrfs_set_file_extent_type(leaf, fi,
1290                                            BTRFS_FILE_EXTENT_REG);
1291                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1292                 btrfs_mark_buffer_dirty(leaf);
1293         } else {
1294                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1295                            struct btrfs_file_extent_item);
1296                 btrfs_set_file_extent_type(leaf, fi,
1297                                            BTRFS_FILE_EXTENT_REG);
1298                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1299                 btrfs_set_file_extent_num_bytes(leaf, fi,
1300                                                 extent_end - key.offset);
1301                 btrfs_mark_buffer_dirty(leaf);
1302
1303                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1304                 if (ret < 0) {
1305                         btrfs_abort_transaction(trans, ret);
1306                         goto out;
1307                 }
1308         }
1309 out:
1310         btrfs_free_path(path);
1311         return ret;
1312 }
1313
1314 /*
1315  * on error we return an unlocked page and the error value
1316  * on success we return a locked page and 0
1317  */
1318 static int prepare_uptodate_page(struct inode *inode,
1319                                  struct page *page, u64 pos,
1320                                  bool force_uptodate)
1321 {
1322         int ret = 0;
1323
1324         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1325             !PageUptodate(page)) {
1326                 ret = btrfs_readpage(NULL, page);
1327                 if (ret)
1328                         return ret;
1329                 lock_page(page);
1330                 if (!PageUptodate(page)) {
1331                         unlock_page(page);
1332                         return -EIO;
1333                 }
1334                 if (page->mapping != inode->i_mapping) {
1335                         unlock_page(page);
1336                         return -EAGAIN;
1337                 }
1338         }
1339         return 0;
1340 }
1341
1342 /*
1343  * this just gets pages into the page cache and locks them down.
1344  */
1345 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1346                                   size_t num_pages, loff_t pos,
1347                                   size_t write_bytes, bool force_uptodate)
1348 {
1349         int i;
1350         unsigned long index = pos >> PAGE_SHIFT;
1351         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1352         int err = 0;
1353         int faili;
1354
1355         for (i = 0; i < num_pages; i++) {
1356 again:
1357                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1358                                                mask | __GFP_WRITE);
1359                 if (!pages[i]) {
1360                         faili = i - 1;
1361                         err = -ENOMEM;
1362                         goto fail;
1363                 }
1364
1365                 if (i == 0)
1366                         err = prepare_uptodate_page(inode, pages[i], pos,
1367                                                     force_uptodate);
1368                 if (!err && i == num_pages - 1)
1369                         err = prepare_uptodate_page(inode, pages[i],
1370                                                     pos + write_bytes, false);
1371                 if (err) {
1372                         put_page(pages[i]);
1373                         if (err == -EAGAIN) {
1374                                 err = 0;
1375                                 goto again;
1376                         }
1377                         faili = i - 1;
1378                         goto fail;
1379                 }
1380                 wait_on_page_writeback(pages[i]);
1381         }
1382
1383         return 0;
1384 fail:
1385         while (faili >= 0) {
1386                 unlock_page(pages[faili]);
1387                 put_page(pages[faili]);
1388                 faili--;
1389         }
1390         return err;
1391
1392 }
1393
1394 /*
1395  * This function locks the extent and properly waits for data=ordered extents
1396  * to finish before allowing the pages to be modified if need.
1397  *
1398  * The return value:
1399  * 1 - the extent is locked
1400  * 0 - the extent is not locked, and everything is OK
1401  * -EAGAIN - need re-prepare the pages
1402  * the other < 0 number - Something wrong happens
1403  */
1404 static noinline int
1405 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1406                                 size_t num_pages, loff_t pos,
1407                                 size_t write_bytes,
1408                                 u64 *lockstart, u64 *lockend,
1409                                 struct extent_state **cached_state)
1410 {
1411         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1412         u64 start_pos;
1413         u64 last_pos;
1414         int i;
1415         int ret = 0;
1416
1417         start_pos = round_down(pos, fs_info->sectorsize);
1418         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1419
1420         if (start_pos < inode->vfs_inode.i_size) {
1421                 struct btrfs_ordered_extent *ordered;
1422
1423                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1424                                 cached_state);
1425                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1426                                                      last_pos - start_pos + 1);
1427                 if (ordered &&
1428                     ordered->file_offset + ordered->num_bytes > start_pos &&
1429                     ordered->file_offset <= last_pos) {
1430                         unlock_extent_cached(&inode->io_tree, start_pos,
1431                                         last_pos, cached_state);
1432                         for (i = 0; i < num_pages; i++) {
1433                                 unlock_page(pages[i]);
1434                                 put_page(pages[i]);
1435                         }
1436                         btrfs_start_ordered_extent(ordered, 1);
1437                         btrfs_put_ordered_extent(ordered);
1438                         return -EAGAIN;
1439                 }
1440                 if (ordered)
1441                         btrfs_put_ordered_extent(ordered);
1442
1443                 *lockstart = start_pos;
1444                 *lockend = last_pos;
1445                 ret = 1;
1446         }
1447
1448         /*
1449          * It's possible the pages are dirty right now, but we don't want
1450          * to clean them yet because copy_from_user may catch a page fault
1451          * and we might have to fall back to one page at a time.  If that
1452          * happens, we'll unlock these pages and we'd have a window where
1453          * reclaim could sneak in and drop the once-dirty page on the floor
1454          * without writing it.
1455          *
1456          * We have the pages locked and the extent range locked, so there's
1457          * no way someone can start IO on any dirty pages in this range.
1458          *
1459          * We'll call btrfs_dirty_pages() later on, and that will flip around
1460          * delalloc bits and dirty the pages as required.
1461          */
1462         for (i = 0; i < num_pages; i++) {
1463                 set_page_extent_mapped(pages[i]);
1464                 WARN_ON(!PageLocked(pages[i]));
1465         }
1466
1467         return ret;
1468 }
1469
1470 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1471                            size_t *write_bytes, bool nowait)
1472 {
1473         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1474         struct btrfs_root *root = inode->root;
1475         u64 lockstart, lockend;
1476         u64 num_bytes;
1477         int ret;
1478
1479         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1480                 return 0;
1481
1482         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1483                 return -EAGAIN;
1484
1485         lockstart = round_down(pos, fs_info->sectorsize);
1486         lockend = round_up(pos + *write_bytes,
1487                            fs_info->sectorsize) - 1;
1488         num_bytes = lockend - lockstart + 1;
1489
1490         if (nowait) {
1491                 struct btrfs_ordered_extent *ordered;
1492
1493                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1494                         return -EAGAIN;
1495
1496                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1497                                                      num_bytes);
1498                 if (ordered) {
1499                         btrfs_put_ordered_extent(ordered);
1500                         ret = -EAGAIN;
1501                         goto out_unlock;
1502                 }
1503         } else {
1504                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1505                                                    lockend, NULL);
1506         }
1507
1508         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1509                         NULL, NULL, NULL, false);
1510         if (ret <= 0) {
1511                 ret = 0;
1512                 if (!nowait)
1513                         btrfs_drew_write_unlock(&root->snapshot_lock);
1514         } else {
1515                 *write_bytes = min_t(size_t, *write_bytes ,
1516                                      num_bytes - pos + lockstart);
1517         }
1518 out_unlock:
1519         unlock_extent(&inode->io_tree, lockstart, lockend);
1520
1521         return ret;
1522 }
1523
1524 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1525                               size_t *write_bytes)
1526 {
1527         return check_can_nocow(inode, pos, write_bytes, true);
1528 }
1529
1530 /*
1531  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1532  *
1533  * @pos:         File offset
1534  * @write_bytes: The length to write, will be updated to the nocow writeable
1535  *               range
1536  *
1537  * This function will flush ordered extents in the range to ensure proper
1538  * nocow checks.
1539  *
1540  * Return:
1541  * >0           and update @write_bytes if we can do nocow write
1542  *  0           if we can't do nocow write
1543  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1544  *              for * (nowait == true) case
1545  * <0           if other error happened
1546  *
1547  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1548  */
1549 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1550                            size_t *write_bytes)
1551 {
1552         return check_can_nocow(inode, pos, write_bytes, false);
1553 }
1554
1555 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1556 {
1557         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1558 }
1559
1560 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1561                                                struct iov_iter *i)
1562 {
1563         struct file *file = iocb->ki_filp;
1564         loff_t pos = iocb->ki_pos;
1565         struct inode *inode = file_inode(file);
1566         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1567         struct page **pages = NULL;
1568         struct extent_changeset *data_reserved = NULL;
1569         u64 release_bytes = 0;
1570         u64 lockstart;
1571         u64 lockend;
1572         size_t num_written = 0;
1573         int nrptrs;
1574         int ret = 0;
1575         bool only_release_metadata = false;
1576         bool force_page_uptodate = false;
1577
1578         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1579                         PAGE_SIZE / (sizeof(struct page *)));
1580         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1581         nrptrs = max(nrptrs, 8);
1582         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1583         if (!pages)
1584                 return -ENOMEM;
1585
1586         while (iov_iter_count(i) > 0) {
1587                 struct extent_state *cached_state = NULL;
1588                 size_t offset = offset_in_page(pos);
1589                 size_t sector_offset;
1590                 size_t write_bytes = min(iov_iter_count(i),
1591                                          nrptrs * (size_t)PAGE_SIZE -
1592                                          offset);
1593                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1594                                                 PAGE_SIZE);
1595                 size_t reserve_bytes;
1596                 size_t dirty_pages;
1597                 size_t copied;
1598                 size_t dirty_sectors;
1599                 size_t num_sectors;
1600                 int extents_locked;
1601
1602                 WARN_ON(num_pages > nrptrs);
1603
1604                 /*
1605                  * Fault pages before locking them in prepare_pages
1606                  * to avoid recursive lock
1607                  */
1608                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1609                         ret = -EFAULT;
1610                         break;
1611                 }
1612
1613                 only_release_metadata = false;
1614                 sector_offset = pos & (fs_info->sectorsize - 1);
1615                 reserve_bytes = round_up(write_bytes + sector_offset,
1616                                 fs_info->sectorsize);
1617
1618                 extent_changeset_release(data_reserved);
1619                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1620                                                   &data_reserved, pos,
1621                                                   write_bytes);
1622                 if (ret < 0) {
1623                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1624                                                    &write_bytes) > 0) {
1625                                 /*
1626                                  * For nodata cow case, no need to reserve
1627                                  * data space.
1628                                  */
1629                                 only_release_metadata = true;
1630                                 /*
1631                                  * our prealloc extent may be smaller than
1632                                  * write_bytes, so scale down.
1633                                  */
1634                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1635                                                          PAGE_SIZE);
1636                                 reserve_bytes = round_up(write_bytes +
1637                                                          sector_offset,
1638                                                          fs_info->sectorsize);
1639                         } else {
1640                                 break;
1641                         }
1642                 }
1643
1644                 WARN_ON(reserve_bytes == 0);
1645                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1646                                 reserve_bytes);
1647                 if (ret) {
1648                         if (!only_release_metadata)
1649                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1650                                                 data_reserved, pos,
1651                                                 write_bytes);
1652                         else
1653                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1654                         break;
1655                 }
1656
1657                 release_bytes = reserve_bytes;
1658 again:
1659                 /*
1660                  * This is going to setup the pages array with the number of
1661                  * pages we want, so we don't really need to worry about the
1662                  * contents of pages from loop to loop
1663                  */
1664                 ret = prepare_pages(inode, pages, num_pages,
1665                                     pos, write_bytes,
1666                                     force_page_uptodate);
1667                 if (ret) {
1668                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1669                                                        reserve_bytes);
1670                         break;
1671                 }
1672
1673                 extents_locked = lock_and_cleanup_extent_if_need(
1674                                 BTRFS_I(inode), pages,
1675                                 num_pages, pos, write_bytes, &lockstart,
1676                                 &lockend, &cached_state);
1677                 if (extents_locked < 0) {
1678                         if (extents_locked == -EAGAIN)
1679                                 goto again;
1680                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1681                                                        reserve_bytes);
1682                         ret = extents_locked;
1683                         break;
1684                 }
1685
1686                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1687
1688                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1689                 dirty_sectors = round_up(copied + sector_offset,
1690                                         fs_info->sectorsize);
1691                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1692
1693                 /*
1694                  * if we have trouble faulting in the pages, fall
1695                  * back to one page at a time
1696                  */
1697                 if (copied < write_bytes)
1698                         nrptrs = 1;
1699
1700                 if (copied == 0) {
1701                         force_page_uptodate = true;
1702                         dirty_sectors = 0;
1703                         dirty_pages = 0;
1704                 } else {
1705                         force_page_uptodate = false;
1706                         dirty_pages = DIV_ROUND_UP(copied + offset,
1707                                                    PAGE_SIZE);
1708                 }
1709
1710                 if (num_sectors > dirty_sectors) {
1711                         /* release everything except the sectors we dirtied */
1712                         release_bytes -= dirty_sectors <<
1713                                                 fs_info->sb->s_blocksize_bits;
1714                         if (only_release_metadata) {
1715                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1716                                                         release_bytes, true);
1717                         } else {
1718                                 u64 __pos;
1719
1720                                 __pos = round_down(pos,
1721                                                    fs_info->sectorsize) +
1722                                         (dirty_pages << PAGE_SHIFT);
1723                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1724                                                 data_reserved, __pos,
1725                                                 release_bytes, true);
1726                         }
1727                 }
1728
1729                 release_bytes = round_up(copied + sector_offset,
1730                                         fs_info->sectorsize);
1731
1732                 if (copied > 0)
1733                         ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1734                                                 dirty_pages, pos, copied,
1735                                                 &cached_state);
1736
1737                 /*
1738                  * If we have not locked the extent range, because the range's
1739                  * start offset is >= i_size, we might still have a non-NULL
1740                  * cached extent state, acquired while marking the extent range
1741                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1742                  * possible cached extent state to avoid a memory leak.
1743                  */
1744                 if (extents_locked)
1745                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1746                                              lockstart, lockend, &cached_state);
1747                 else
1748                         free_extent_state(cached_state);
1749
1750                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1751                 if (ret) {
1752                         btrfs_drop_pages(pages, num_pages);
1753                         break;
1754                 }
1755
1756                 release_bytes = 0;
1757                 if (only_release_metadata)
1758                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1759
1760                 if (only_release_metadata && copied > 0) {
1761                         lockstart = round_down(pos,
1762                                                fs_info->sectorsize);
1763                         lockend = round_up(pos + copied,
1764                                            fs_info->sectorsize) - 1;
1765
1766                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1767                                        lockend, EXTENT_NORESERVE, NULL,
1768                                        NULL, GFP_NOFS);
1769                 }
1770
1771                 btrfs_drop_pages(pages, num_pages);
1772
1773                 cond_resched();
1774
1775                 balance_dirty_pages_ratelimited(inode->i_mapping);
1776
1777                 pos += copied;
1778                 num_written += copied;
1779         }
1780
1781         kfree(pages);
1782
1783         if (release_bytes) {
1784                 if (only_release_metadata) {
1785                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1786                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1787                                         release_bytes, true);
1788                 } else {
1789                         btrfs_delalloc_release_space(BTRFS_I(inode),
1790                                         data_reserved,
1791                                         round_down(pos, fs_info->sectorsize),
1792                                         release_bytes, true);
1793                 }
1794         }
1795
1796         extent_changeset_free(data_reserved);
1797         return num_written ? num_written : ret;
1798 }
1799
1800 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1801 {
1802         struct file *file = iocb->ki_filp;
1803         struct inode *inode = file_inode(file);
1804         loff_t pos;
1805         ssize_t written;
1806         ssize_t written_buffered;
1807         loff_t endbyte;
1808         int err;
1809
1810         written = btrfs_direct_IO(iocb, from);
1811
1812         if (written < 0 || !iov_iter_count(from))
1813                 return written;
1814
1815         pos = iocb->ki_pos;
1816         written_buffered = btrfs_buffered_write(iocb, from);
1817         if (written_buffered < 0) {
1818                 err = written_buffered;
1819                 goto out;
1820         }
1821         /*
1822          * Ensure all data is persisted. We want the next direct IO read to be
1823          * able to read what was just written.
1824          */
1825         endbyte = pos + written_buffered - 1;
1826         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1827         if (err)
1828                 goto out;
1829         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1830         if (err)
1831                 goto out;
1832         written += written_buffered;
1833         iocb->ki_pos = pos + written_buffered;
1834         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1835                                  endbyte >> PAGE_SHIFT);
1836 out:
1837         return written ? written : err;
1838 }
1839
1840 static void update_time_for_write(struct inode *inode)
1841 {
1842         struct timespec64 now;
1843
1844         if (IS_NOCMTIME(inode))
1845                 return;
1846
1847         now = current_time(inode);
1848         if (!timespec64_equal(&inode->i_mtime, &now))
1849                 inode->i_mtime = now;
1850
1851         if (!timespec64_equal(&inode->i_ctime, &now))
1852                 inode->i_ctime = now;
1853
1854         if (IS_I_VERSION(inode))
1855                 inode_inc_iversion(inode);
1856 }
1857
1858 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1859                                     struct iov_iter *from)
1860 {
1861         struct file *file = iocb->ki_filp;
1862         struct inode *inode = file_inode(file);
1863         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1864         u64 start_pos;
1865         u64 end_pos;
1866         ssize_t num_written = 0;
1867         const bool sync = iocb->ki_flags & IOCB_DSYNC;
1868         ssize_t err;
1869         loff_t pos;
1870         size_t count;
1871         loff_t oldsize;
1872         int clean_page = 0;
1873
1874         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1875             (iocb->ki_flags & IOCB_NOWAIT))
1876                 return -EOPNOTSUPP;
1877
1878         if (iocb->ki_flags & IOCB_NOWAIT) {
1879                 if (!inode_trylock(inode))
1880                         return -EAGAIN;
1881         } else {
1882                 inode_lock(inode);
1883         }
1884
1885         err = generic_write_checks(iocb, from);
1886         if (err <= 0) {
1887                 inode_unlock(inode);
1888                 return err;
1889         }
1890
1891         pos = iocb->ki_pos;
1892         count = iov_iter_count(from);
1893         if (iocb->ki_flags & IOCB_NOWAIT) {
1894                 size_t nocow_bytes = count;
1895
1896                 /*
1897                  * We will allocate space in case nodatacow is not set,
1898                  * so bail
1899                  */
1900                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes)
1901                     <= 0) {
1902                         inode_unlock(inode);
1903                         return -EAGAIN;
1904                 }
1905                 /*
1906                  * There are holes in the range or parts of the range that must
1907                  * be COWed (shared extents, RO block groups, etc), so just bail
1908                  * out.
1909                  */
1910                 if (nocow_bytes < count) {
1911                         inode_unlock(inode);
1912                         return -EAGAIN;
1913                 }
1914         }
1915
1916         current->backing_dev_info = inode_to_bdi(inode);
1917         err = file_remove_privs(file);
1918         if (err) {
1919                 inode_unlock(inode);
1920                 goto out;
1921         }
1922
1923         /*
1924          * If BTRFS flips readonly due to some impossible error
1925          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1926          * although we have opened a file as writable, we have
1927          * to stop this write operation to ensure FS consistency.
1928          */
1929         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1930                 inode_unlock(inode);
1931                 err = -EROFS;
1932                 goto out;
1933         }
1934
1935         /*
1936          * We reserve space for updating the inode when we reserve space for the
1937          * extent we are going to write, so we will enospc out there.  We don't
1938          * need to start yet another transaction to update the inode as we will
1939          * update the inode when we finish writing whatever data we write.
1940          */
1941         update_time_for_write(inode);
1942
1943         start_pos = round_down(pos, fs_info->sectorsize);
1944         oldsize = i_size_read(inode);
1945         if (start_pos > oldsize) {
1946                 /* Expand hole size to cover write data, preventing empty gap */
1947                 end_pos = round_up(pos + count,
1948                                    fs_info->sectorsize);
1949                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1950                 if (err) {
1951                         inode_unlock(inode);
1952                         goto out;
1953                 }
1954                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1955                         clean_page = 1;
1956         }
1957
1958         if (sync)
1959                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1960
1961         if (iocb->ki_flags & IOCB_DIRECT) {
1962                 /*
1963                  * 1. We must always clear IOCB_DSYNC in order to not deadlock
1964                  *    in iomap, as it calls generic_write_sync() in this case.
1965                  * 2. If we are async, we can call iomap_dio_complete() either
1966                  *    in
1967                  *
1968                  *    2.1. A worker thread from the last bio completed.  In this
1969                  *         case we need to mark the btrfs_dio_data that it is
1970                  *         async in order to call generic_write_sync() properly.
1971                  *         This is handled by setting BTRFS_DIO_SYNC_STUB in the
1972                  *         current->journal_info.
1973                  *    2.2  The submitter context, because all IO completed
1974                  *         before we exited iomap_dio_rw().  In this case we can
1975                  *         just re-set the IOCB_DSYNC on the iocb and we'll do
1976                  *         the sync below.  If our ->end_io() gets called and
1977                  *         current->journal_info is set, then we know we're in
1978                  *         our current context and we will clear
1979                  *         current->journal_info to indicate that we need to
1980                  *         sync below.
1981                  */
1982                 if (sync) {
1983                         ASSERT(current->journal_info == NULL);
1984                         iocb->ki_flags &= ~IOCB_DSYNC;
1985                         current->journal_info = BTRFS_DIO_SYNC_STUB;
1986                 }
1987                 num_written = __btrfs_direct_write(iocb, from);
1988
1989                 /*
1990                  * As stated above, we cleared journal_info, so we need to do
1991                  * the sync ourselves.
1992                  */
1993                 if (sync && current->journal_info == NULL)
1994                         iocb->ki_flags |= IOCB_DSYNC;
1995                 current->journal_info = NULL;
1996         } else {
1997                 num_written = btrfs_buffered_write(iocb, from);
1998                 if (num_written > 0)
1999                         iocb->ki_pos = pos + num_written;
2000                 if (clean_page)
2001                         pagecache_isize_extended(inode, oldsize,
2002                                                 i_size_read(inode));
2003         }
2004
2005         inode_unlock(inode);
2006
2007         btrfs_set_inode_last_sub_trans(BTRFS_I(inode));
2008
2009         if (num_written > 0)
2010                 num_written = generic_write_sync(iocb, num_written);
2011
2012         if (sync)
2013                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2014 out:
2015         current->backing_dev_info = NULL;
2016         return num_written ? num_written : err;
2017 }
2018
2019 int btrfs_release_file(struct inode *inode, struct file *filp)
2020 {
2021         struct btrfs_file_private *private = filp->private_data;
2022
2023         if (private && private->filldir_buf)
2024                 kfree(private->filldir_buf);
2025         kfree(private);
2026         filp->private_data = NULL;
2027
2028         /*
2029          * Set by setattr when we are about to truncate a file from a non-zero
2030          * size to a zero size.  This tries to flush down new bytes that may
2031          * have been written if the application were using truncate to replace
2032          * a file in place.
2033          */
2034         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2035                                &BTRFS_I(inode)->runtime_flags))
2036                         filemap_flush(inode->i_mapping);
2037         return 0;
2038 }
2039
2040 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2041 {
2042         int ret;
2043         struct blk_plug plug;
2044
2045         /*
2046          * This is only called in fsync, which would do synchronous writes, so
2047          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2048          * multiple disks using raid profile, a large IO can be split to
2049          * several segments of stripe length (currently 64K).
2050          */
2051         blk_start_plug(&plug);
2052         atomic_inc(&BTRFS_I(inode)->sync_writers);
2053         ret = btrfs_fdatawrite_range(inode, start, end);
2054         atomic_dec(&BTRFS_I(inode)->sync_writers);
2055         blk_finish_plug(&plug);
2056
2057         return ret;
2058 }
2059
2060 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2061 {
2062         struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2063         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2064
2065         if (btrfs_inode_in_log(inode, fs_info->generation) &&
2066             list_empty(&ctx->ordered_extents))
2067                 return true;
2068
2069         /*
2070          * If we are doing a fast fsync we can not bail out if the inode's
2071          * last_trans is <= then the last committed transaction, because we only
2072          * update the last_trans of the inode during ordered extent completion,
2073          * and for a fast fsync we don't wait for that, we only wait for the
2074          * writeback to complete.
2075          */
2076         if (inode->last_trans <= fs_info->last_trans_committed &&
2077             (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2078              list_empty(&ctx->ordered_extents)))
2079                 return true;
2080
2081         return false;
2082 }
2083
2084 /*
2085  * fsync call for both files and directories.  This logs the inode into
2086  * the tree log instead of forcing full commits whenever possible.
2087  *
2088  * It needs to call filemap_fdatawait so that all ordered extent updates are
2089  * in the metadata btree are up to date for copying to the log.
2090  *
2091  * It drops the inode mutex before doing the tree log commit.  This is an
2092  * important optimization for directories because holding the mutex prevents
2093  * new operations on the dir while we write to disk.
2094  */
2095 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2096 {
2097         struct dentry *dentry = file_dentry(file);
2098         struct inode *inode = d_inode(dentry);
2099         struct btrfs_root *root = BTRFS_I(inode)->root;
2100         struct btrfs_trans_handle *trans;
2101         struct btrfs_log_ctx ctx;
2102         int ret = 0, err;
2103         u64 len;
2104         bool full_sync;
2105
2106         trace_btrfs_sync_file(file, datasync);
2107
2108         btrfs_init_log_ctx(&ctx, inode);
2109
2110         /*
2111          * Always set the range to a full range, otherwise we can get into
2112          * several problems, from missing file extent items to represent holes
2113          * when not using the NO_HOLES feature, to log tree corruption due to
2114          * races between hole detection during logging and completion of ordered
2115          * extents outside the range, to missing checksums due to ordered extents
2116          * for which we flushed only a subset of their pages.
2117          */
2118         start = 0;
2119         end = LLONG_MAX;
2120         len = (u64)LLONG_MAX + 1;
2121
2122         /*
2123          * We write the dirty pages in the range and wait until they complete
2124          * out of the ->i_mutex. If so, we can flush the dirty pages by
2125          * multi-task, and make the performance up.  See
2126          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2127          */
2128         ret = start_ordered_ops(inode, start, end);
2129         if (ret)
2130                 goto out;
2131
2132         inode_lock(inode);
2133
2134         /*
2135          * We take the dio_sem here because the tree log stuff can race with
2136          * lockless dio writes and get an extent map logged for an extent we
2137          * never waited on.  We need it this high up for lockdep reasons.
2138          */
2139         down_write(&BTRFS_I(inode)->dio_sem);
2140
2141         atomic_inc(&root->log_batch);
2142
2143         /*
2144          * Always check for the full sync flag while holding the inode's lock,
2145          * to avoid races with other tasks. The flag must be either set all the
2146          * time during logging or always off all the time while logging.
2147          */
2148         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2149                              &BTRFS_I(inode)->runtime_flags);
2150
2151         /*
2152          * Before we acquired the inode's lock, someone may have dirtied more
2153          * pages in the target range. We need to make sure that writeback for
2154          * any such pages does not start while we are logging the inode, because
2155          * if it does, any of the following might happen when we are not doing a
2156          * full inode sync:
2157          *
2158          * 1) We log an extent after its writeback finishes but before its
2159          *    checksums are added to the csum tree, leading to -EIO errors
2160          *    when attempting to read the extent after a log replay.
2161          *
2162          * 2) We can end up logging an extent before its writeback finishes.
2163          *    Therefore after the log replay we will have a file extent item
2164          *    pointing to an unwritten extent (and no data checksums as well).
2165          *
2166          * So trigger writeback for any eventual new dirty pages and then we
2167          * wait for all ordered extents to complete below.
2168          */
2169         ret = start_ordered_ops(inode, start, end);
2170         if (ret) {
2171                 up_write(&BTRFS_I(inode)->dio_sem);
2172                 inode_unlock(inode);
2173                 goto out;
2174         }
2175
2176         /*
2177          * We have to do this here to avoid the priority inversion of waiting on
2178          * IO of a lower priority task while holding a transaction open.
2179          *
2180          * For a full fsync we wait for the ordered extents to complete while
2181          * for a fast fsync we wait just for writeback to complete, and then
2182          * attach the ordered extents to the transaction so that a transaction
2183          * commit waits for their completion, to avoid data loss if we fsync,
2184          * the current transaction commits before the ordered extents complete
2185          * and a power failure happens right after that.
2186          */
2187         if (full_sync) {
2188                 ret = btrfs_wait_ordered_range(inode, start, len);
2189         } else {
2190                 /*
2191                  * Get our ordered extents as soon as possible to avoid doing
2192                  * checksum lookups in the csum tree, and use instead the
2193                  * checksums attached to the ordered extents.
2194                  */
2195                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2196                                                       &ctx.ordered_extents);
2197                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2198         }
2199
2200         if (ret)
2201                 goto out_release_extents;
2202
2203         atomic_inc(&root->log_batch);
2204
2205         smp_mb();
2206         if (skip_inode_logging(&ctx)) {
2207                 /*
2208                  * We've had everything committed since the last time we were
2209                  * modified so clear this flag in case it was set for whatever
2210                  * reason, it's no longer relevant.
2211                  */
2212                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2213                           &BTRFS_I(inode)->runtime_flags);
2214                 /*
2215                  * An ordered extent might have started before and completed
2216                  * already with io errors, in which case the inode was not
2217                  * updated and we end up here. So check the inode's mapping
2218                  * for any errors that might have happened since we last
2219                  * checked called fsync.
2220                  */
2221                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2222                 goto out_release_extents;
2223         }
2224
2225         /*
2226          * We use start here because we will need to wait on the IO to complete
2227          * in btrfs_sync_log, which could require joining a transaction (for
2228          * example checking cross references in the nocow path).  If we use join
2229          * here we could get into a situation where we're waiting on IO to
2230          * happen that is blocked on a transaction trying to commit.  With start
2231          * we inc the extwriter counter, so we wait for all extwriters to exit
2232          * before we start blocking joiners.  This comment is to keep somebody
2233          * from thinking they are super smart and changing this to
2234          * btrfs_join_transaction *cough*Josef*cough*.
2235          */
2236         trans = btrfs_start_transaction(root, 0);
2237         if (IS_ERR(trans)) {
2238                 ret = PTR_ERR(trans);
2239                 goto out_release_extents;
2240         }
2241
2242         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2243         btrfs_release_log_ctx_extents(&ctx);
2244         if (ret < 0) {
2245                 /* Fallthrough and commit/free transaction. */
2246                 ret = 1;
2247         }
2248
2249         /* we've logged all the items and now have a consistent
2250          * version of the file in the log.  It is possible that
2251          * someone will come in and modify the file, but that's
2252          * fine because the log is consistent on disk, and we
2253          * have references to all of the file's extents
2254          *
2255          * It is possible that someone will come in and log the
2256          * file again, but that will end up using the synchronization
2257          * inside btrfs_sync_log to keep things safe.
2258          */
2259         up_write(&BTRFS_I(inode)->dio_sem);
2260         inode_unlock(inode);
2261
2262         if (ret != BTRFS_NO_LOG_SYNC) {
2263                 if (!ret) {
2264                         ret = btrfs_sync_log(trans, root, &ctx);
2265                         if (!ret) {
2266                                 ret = btrfs_end_transaction(trans);
2267                                 goto out;
2268                         }
2269                 }
2270                 if (!full_sync) {
2271                         ret = btrfs_wait_ordered_range(inode, start, len);
2272                         if (ret) {
2273                                 btrfs_end_transaction(trans);
2274                                 goto out;
2275                         }
2276                 }
2277                 ret = btrfs_commit_transaction(trans);
2278         } else {
2279                 ret = btrfs_end_transaction(trans);
2280         }
2281 out:
2282         ASSERT(list_empty(&ctx.list));
2283         err = file_check_and_advance_wb_err(file);
2284         if (!ret)
2285                 ret = err;
2286         return ret > 0 ? -EIO : ret;
2287
2288 out_release_extents:
2289         btrfs_release_log_ctx_extents(&ctx);
2290         up_write(&BTRFS_I(inode)->dio_sem);
2291         inode_unlock(inode);
2292         goto out;
2293 }
2294
2295 static const struct vm_operations_struct btrfs_file_vm_ops = {
2296         .fault          = filemap_fault,
2297         .map_pages      = filemap_map_pages,
2298         .page_mkwrite   = btrfs_page_mkwrite,
2299 };
2300
2301 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2302 {
2303         struct address_space *mapping = filp->f_mapping;
2304
2305         if (!mapping->a_ops->readpage)
2306                 return -ENOEXEC;
2307
2308         file_accessed(filp);
2309         vma->vm_ops = &btrfs_file_vm_ops;
2310
2311         return 0;
2312 }
2313
2314 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2315                           int slot, u64 start, u64 end)
2316 {
2317         struct btrfs_file_extent_item *fi;
2318         struct btrfs_key key;
2319
2320         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2321                 return 0;
2322
2323         btrfs_item_key_to_cpu(leaf, &key, slot);
2324         if (key.objectid != btrfs_ino(inode) ||
2325             key.type != BTRFS_EXTENT_DATA_KEY)
2326                 return 0;
2327
2328         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2329
2330         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2331                 return 0;
2332
2333         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2334                 return 0;
2335
2336         if (key.offset == end)
2337                 return 1;
2338         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2339                 return 1;
2340         return 0;
2341 }
2342
2343 static int fill_holes(struct btrfs_trans_handle *trans,
2344                 struct btrfs_inode *inode,
2345                 struct btrfs_path *path, u64 offset, u64 end)
2346 {
2347         struct btrfs_fs_info *fs_info = trans->fs_info;
2348         struct btrfs_root *root = inode->root;
2349         struct extent_buffer *leaf;
2350         struct btrfs_file_extent_item *fi;
2351         struct extent_map *hole_em;
2352         struct extent_map_tree *em_tree = &inode->extent_tree;
2353         struct btrfs_key key;
2354         int ret;
2355
2356         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2357                 goto out;
2358
2359         key.objectid = btrfs_ino(inode);
2360         key.type = BTRFS_EXTENT_DATA_KEY;
2361         key.offset = offset;
2362
2363         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2364         if (ret <= 0) {
2365                 /*
2366                  * We should have dropped this offset, so if we find it then
2367                  * something has gone horribly wrong.
2368                  */
2369                 if (ret == 0)
2370                         ret = -EINVAL;
2371                 return ret;
2372         }
2373
2374         leaf = path->nodes[0];
2375         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2376                 u64 num_bytes;
2377
2378                 path->slots[0]--;
2379                 fi = btrfs_item_ptr(leaf, path->slots[0],
2380                                     struct btrfs_file_extent_item);
2381                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2382                         end - offset;
2383                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2384                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2385                 btrfs_set_file_extent_offset(leaf, fi, 0);
2386                 btrfs_mark_buffer_dirty(leaf);
2387                 goto out;
2388         }
2389
2390         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2391                 u64 num_bytes;
2392
2393                 key.offset = offset;
2394                 btrfs_set_item_key_safe(fs_info, path, &key);
2395                 fi = btrfs_item_ptr(leaf, path->slots[0],
2396                                     struct btrfs_file_extent_item);
2397                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2398                         offset;
2399                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2400                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2401                 btrfs_set_file_extent_offset(leaf, fi, 0);
2402                 btrfs_mark_buffer_dirty(leaf);
2403                 goto out;
2404         }
2405         btrfs_release_path(path);
2406
2407         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2408                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2409         if (ret)
2410                 return ret;
2411
2412 out:
2413         btrfs_release_path(path);
2414
2415         hole_em = alloc_extent_map();
2416         if (!hole_em) {
2417                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2418                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2419         } else {
2420                 hole_em->start = offset;
2421                 hole_em->len = end - offset;
2422                 hole_em->ram_bytes = hole_em->len;
2423                 hole_em->orig_start = offset;
2424
2425                 hole_em->block_start = EXTENT_MAP_HOLE;
2426                 hole_em->block_len = 0;
2427                 hole_em->orig_block_len = 0;
2428                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2429                 hole_em->generation = trans->transid;
2430
2431                 do {
2432                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2433                         write_lock(&em_tree->lock);
2434                         ret = add_extent_mapping(em_tree, hole_em, 1);
2435                         write_unlock(&em_tree->lock);
2436                 } while (ret == -EEXIST);
2437                 free_extent_map(hole_em);
2438                 if (ret)
2439                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2440                                         &inode->runtime_flags);
2441         }
2442
2443         return 0;
2444 }
2445
2446 /*
2447  * Find a hole extent on given inode and change start/len to the end of hole
2448  * extent.(hole/vacuum extent whose em->start <= start &&
2449  *         em->start + em->len > start)
2450  * When a hole extent is found, return 1 and modify start/len.
2451  */
2452 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2453 {
2454         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2455         struct extent_map *em;
2456         int ret = 0;
2457
2458         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2459                               round_down(*start, fs_info->sectorsize),
2460                               round_up(*len, fs_info->sectorsize));
2461         if (IS_ERR(em))
2462                 return PTR_ERR(em);
2463
2464         /* Hole or vacuum extent(only exists in no-hole mode) */
2465         if (em->block_start == EXTENT_MAP_HOLE) {
2466                 ret = 1;
2467                 *len = em->start + em->len > *start + *len ?
2468                        0 : *start + *len - em->start - em->len;
2469                 *start = em->start + em->len;
2470         }
2471         free_extent_map(em);
2472         return ret;
2473 }
2474
2475 static int btrfs_punch_hole_lock_range(struct inode *inode,
2476                                        const u64 lockstart,
2477                                        const u64 lockend,
2478                                        struct extent_state **cached_state)
2479 {
2480         while (1) {
2481                 struct btrfs_ordered_extent *ordered;
2482                 int ret;
2483
2484                 truncate_pagecache_range(inode, lockstart, lockend);
2485
2486                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2487                                  cached_state);
2488                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2489                                                             lockend);
2490
2491                 /*
2492                  * We need to make sure we have no ordered extents in this range
2493                  * and nobody raced in and read a page in this range, if we did
2494                  * we need to try again.
2495                  */
2496                 if ((!ordered ||
2497                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2498                      ordered->file_offset > lockend)) &&
2499                      !filemap_range_has_page(inode->i_mapping,
2500                                              lockstart, lockend)) {
2501                         if (ordered)
2502                                 btrfs_put_ordered_extent(ordered);
2503                         break;
2504                 }
2505                 if (ordered)
2506                         btrfs_put_ordered_extent(ordered);
2507                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2508                                      lockend, cached_state);
2509                 ret = btrfs_wait_ordered_range(inode, lockstart,
2510                                                lockend - lockstart + 1);
2511                 if (ret)
2512                         return ret;
2513         }
2514         return 0;
2515 }
2516
2517 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2518                                      struct inode *inode,
2519                                      struct btrfs_path *path,
2520                                      struct btrfs_replace_extent_info *extent_info,
2521                                      const u64 replace_len)
2522 {
2523         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2524         struct btrfs_root *root = BTRFS_I(inode)->root;
2525         struct btrfs_file_extent_item *extent;
2526         struct extent_buffer *leaf;
2527         struct btrfs_key key;
2528         int slot;
2529         struct btrfs_ref ref = { 0 };
2530         int ret;
2531
2532         if (replace_len == 0)
2533                 return 0;
2534
2535         if (extent_info->disk_offset == 0 &&
2536             btrfs_fs_incompat(fs_info, NO_HOLES))
2537                 return 0;
2538
2539         key.objectid = btrfs_ino(BTRFS_I(inode));
2540         key.type = BTRFS_EXTENT_DATA_KEY;
2541         key.offset = extent_info->file_offset;
2542         ret = btrfs_insert_empty_item(trans, root, path, &key,
2543                                       sizeof(struct btrfs_file_extent_item));
2544         if (ret)
2545                 return ret;
2546         leaf = path->nodes[0];
2547         slot = path->slots[0];
2548         write_extent_buffer(leaf, extent_info->extent_buf,
2549                             btrfs_item_ptr_offset(leaf, slot),
2550                             sizeof(struct btrfs_file_extent_item));
2551         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2552         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2553         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2554         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2555         if (extent_info->is_new_extent)
2556                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2557         btrfs_mark_buffer_dirty(leaf);
2558         btrfs_release_path(path);
2559
2560         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2561                         extent_info->file_offset, replace_len);
2562         if (ret)
2563                 return ret;
2564
2565         /* If it's a hole, nothing more needs to be done. */
2566         if (extent_info->disk_offset == 0)
2567                 return 0;
2568
2569         inode_add_bytes(inode, replace_len);
2570
2571         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2572                 key.objectid = extent_info->disk_offset;
2573                 key.type = BTRFS_EXTENT_ITEM_KEY;
2574                 key.offset = extent_info->disk_len;
2575                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2576                                                        btrfs_ino(BTRFS_I(inode)),
2577                                                        extent_info->file_offset,
2578                                                        extent_info->qgroup_reserved,
2579                                                        &key);
2580         } else {
2581                 u64 ref_offset;
2582
2583                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2584                                        extent_info->disk_offset,
2585                                        extent_info->disk_len, 0);
2586                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2587                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2588                                     btrfs_ino(BTRFS_I(inode)), ref_offset);
2589                 ret = btrfs_inc_extent_ref(trans, &ref);
2590         }
2591
2592         extent_info->insertions++;
2593
2594         return ret;
2595 }
2596
2597 /*
2598  * The respective range must have been previously locked, as well as the inode.
2599  * The end offset is inclusive (last byte of the range).
2600  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2601  * the file range with an extent.
2602  * When not punching a hole, we don't want to end up in a state where we dropped
2603  * extents without inserting a new one, so we must abort the transaction to avoid
2604  * a corruption.
2605  */
2606 int btrfs_replace_file_extents(struct inode *inode, struct btrfs_path *path,
2607                            const u64 start, const u64 end,
2608                            struct btrfs_replace_extent_info *extent_info,
2609                            struct btrfs_trans_handle **trans_out)
2610 {
2611         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2612         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2613         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2614         struct btrfs_root *root = BTRFS_I(inode)->root;
2615         struct btrfs_trans_handle *trans = NULL;
2616         struct btrfs_block_rsv *rsv;
2617         unsigned int rsv_count;
2618         u64 cur_offset;
2619         u64 drop_end;
2620         u64 len = end - start;
2621         int ret = 0;
2622
2623         if (end <= start)
2624                 return -EINVAL;
2625
2626         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2627         if (!rsv) {
2628                 ret = -ENOMEM;
2629                 goto out;
2630         }
2631         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2632         rsv->failfast = 1;
2633
2634         /*
2635          * 1 - update the inode
2636          * 1 - removing the extents in the range
2637          * 1 - adding the hole extent if no_holes isn't set or if we are
2638          *     replacing the range with a new extent
2639          */
2640         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2641                 rsv_count = 3;
2642         else
2643                 rsv_count = 2;
2644
2645         trans = btrfs_start_transaction(root, rsv_count);
2646         if (IS_ERR(trans)) {
2647                 ret = PTR_ERR(trans);
2648                 trans = NULL;
2649                 goto out_free;
2650         }
2651
2652         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2653                                       min_size, false);
2654         BUG_ON(ret);
2655         trans->block_rsv = rsv;
2656
2657         cur_offset = start;
2658         while (cur_offset < end) {
2659                 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2660                                            cur_offset, end + 1, &drop_end,
2661                                            1, 0, 0, NULL);
2662                 if (ret != -ENOSPC) {
2663                         /*
2664                          * The only time we don't want to abort is if we are
2665                          * attempting to clone a partial inline extent, in which
2666                          * case we'll get EOPNOTSUPP.  However if we aren't
2667                          * clone we need to abort no matter what, because if we
2668                          * got EOPNOTSUPP via prealloc then we messed up and
2669                          * need to abort.
2670                          */
2671                         if (ret &&
2672                             (ret != -EOPNOTSUPP ||
2673                              (extent_info && extent_info->is_new_extent)))
2674                                 btrfs_abort_transaction(trans, ret);
2675                         break;
2676                 }
2677
2678                 trans->block_rsv = &fs_info->trans_block_rsv;
2679
2680                 if (!extent_info && cur_offset < drop_end &&
2681                     cur_offset < ino_size) {
2682                         ret = fill_holes(trans, BTRFS_I(inode), path,
2683                                         cur_offset, drop_end);
2684                         if (ret) {
2685                                 /*
2686                                  * If we failed then we didn't insert our hole
2687                                  * entries for the area we dropped, so now the
2688                                  * fs is corrupted, so we must abort the
2689                                  * transaction.
2690                                  */
2691                                 btrfs_abort_transaction(trans, ret);
2692                                 break;
2693                         }
2694                 } else if (!extent_info && cur_offset < drop_end) {
2695                         /*
2696                          * We are past the i_size here, but since we didn't
2697                          * insert holes we need to clear the mapped area so we
2698                          * know to not set disk_i_size in this area until a new
2699                          * file extent is inserted here.
2700                          */
2701                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2702                                         cur_offset, drop_end - cur_offset);
2703                         if (ret) {
2704                                 /*
2705                                  * We couldn't clear our area, so we could
2706                                  * presumably adjust up and corrupt the fs, so
2707                                  * we need to abort.
2708                                  */
2709                                 btrfs_abort_transaction(trans, ret);
2710                                 break;
2711                         }
2712                 }
2713
2714                 if (extent_info && drop_end > extent_info->file_offset) {
2715                         u64 replace_len = drop_end - extent_info->file_offset;
2716
2717                         ret = btrfs_insert_replace_extent(trans, inode, path,
2718                                                         extent_info, replace_len);
2719                         if (ret) {
2720                                 btrfs_abort_transaction(trans, ret);
2721                                 break;
2722                         }
2723                         extent_info->data_len -= replace_len;
2724                         extent_info->data_offset += replace_len;
2725                         extent_info->file_offset += replace_len;
2726                 }
2727
2728                 cur_offset = drop_end;
2729
2730                 ret = btrfs_update_inode(trans, root, inode);
2731                 if (ret)
2732                         break;
2733
2734                 btrfs_end_transaction(trans);
2735                 btrfs_btree_balance_dirty(fs_info);
2736
2737                 trans = btrfs_start_transaction(root, rsv_count);
2738                 if (IS_ERR(trans)) {
2739                         ret = PTR_ERR(trans);
2740                         trans = NULL;
2741                         break;
2742                 }
2743
2744                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2745                                               rsv, min_size, false);
2746                 BUG_ON(ret);    /* shouldn't happen */
2747                 trans->block_rsv = rsv;
2748
2749                 if (!extent_info) {
2750                         ret = find_first_non_hole(inode, &cur_offset, &len);
2751                         if (unlikely(ret < 0))
2752                                 break;
2753                         if (ret && !len) {
2754                                 ret = 0;
2755                                 break;
2756                         }
2757                 }
2758         }
2759
2760         /*
2761          * If we were cloning, force the next fsync to be a full one since we
2762          * we replaced (or just dropped in the case of cloning holes when
2763          * NO_HOLES is enabled) extents and extent maps.
2764          * This is for the sake of simplicity, and cloning into files larger
2765          * than 16Mb would force the full fsync any way (when
2766          * try_release_extent_mapping() is invoked during page cache truncation.
2767          */
2768         if (extent_info && !extent_info->is_new_extent)
2769                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2770                         &BTRFS_I(inode)->runtime_flags);
2771
2772         if (ret)
2773                 goto out_trans;
2774
2775         trans->block_rsv = &fs_info->trans_block_rsv;
2776         /*
2777          * If we are using the NO_HOLES feature we might have had already an
2778          * hole that overlaps a part of the region [lockstart, lockend] and
2779          * ends at (or beyond) lockend. Since we have no file extent items to
2780          * represent holes, drop_end can be less than lockend and so we must
2781          * make sure we have an extent map representing the existing hole (the
2782          * call to __btrfs_drop_extents() might have dropped the existing extent
2783          * map representing the existing hole), otherwise the fast fsync path
2784          * will not record the existence of the hole region
2785          * [existing_hole_start, lockend].
2786          */
2787         if (drop_end <= end)
2788                 drop_end = end + 1;
2789         /*
2790          * Don't insert file hole extent item if it's for a range beyond eof
2791          * (because it's useless) or if it represents a 0 bytes range (when
2792          * cur_offset == drop_end).
2793          */
2794         if (!extent_info && cur_offset < ino_size && cur_offset < drop_end) {
2795                 ret = fill_holes(trans, BTRFS_I(inode), path,
2796                                 cur_offset, drop_end);
2797                 if (ret) {
2798                         /* Same comment as above. */
2799                         btrfs_abort_transaction(trans, ret);
2800                         goto out_trans;
2801                 }
2802         } else if (!extent_info && cur_offset < drop_end) {
2803                 /* See the comment in the loop above for the reasoning here. */
2804                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2805                                         cur_offset, drop_end - cur_offset);
2806                 if (ret) {
2807                         btrfs_abort_transaction(trans, ret);
2808                         goto out_trans;
2809                 }
2810
2811         }
2812         if (extent_info) {
2813                 ret = btrfs_insert_replace_extent(trans, inode, path, extent_info,
2814                                                 extent_info->data_len);
2815                 if (ret) {
2816                         btrfs_abort_transaction(trans, ret);
2817                         goto out_trans;
2818                 }
2819         }
2820
2821 out_trans:
2822         if (!trans)
2823                 goto out_free;
2824
2825         trans->block_rsv = &fs_info->trans_block_rsv;
2826         if (ret)
2827                 btrfs_end_transaction(trans);
2828         else
2829                 *trans_out = trans;
2830 out_free:
2831         btrfs_free_block_rsv(fs_info, rsv);
2832 out:
2833         return ret;
2834 }
2835
2836 static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2837 {
2838         struct inode *inode = file_inode(file);
2839         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2840         struct btrfs_root *root = BTRFS_I(inode)->root;
2841         struct extent_state *cached_state = NULL;
2842         struct btrfs_path *path;
2843         struct btrfs_trans_handle *trans = NULL;
2844         u64 lockstart;
2845         u64 lockend;
2846         u64 tail_start;
2847         u64 tail_len;
2848         u64 orig_start = offset;
2849         int ret = 0;
2850         bool same_block;
2851         u64 ino_size;
2852         bool truncated_block = false;
2853         bool updated_inode = false;
2854
2855         ret = btrfs_wait_ordered_range(inode, offset, len);
2856         if (ret)
2857                 return ret;
2858
2859         inode_lock(inode);
2860         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2861         ret = find_first_non_hole(inode, &offset, &len);
2862         if (ret < 0)
2863                 goto out_only_mutex;
2864         if (ret && !len) {
2865                 /* Already in a large hole */
2866                 ret = 0;
2867                 goto out_only_mutex;
2868         }
2869
2870         ret = file_modified(file);
2871         if (ret)
2872                 goto out_only_mutex;
2873
2874         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2875         lockend = round_down(offset + len,
2876                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2877         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2878                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2879         /*
2880          * We needn't truncate any block which is beyond the end of the file
2881          * because we are sure there is no data there.
2882          */
2883         /*
2884          * Only do this if we are in the same block and we aren't doing the
2885          * entire block.
2886          */
2887         if (same_block && len < fs_info->sectorsize) {
2888                 if (offset < ino_size) {
2889                         truncated_block = true;
2890                         ret = btrfs_truncate_block(inode, offset, len, 0);
2891                 } else {
2892                         ret = 0;
2893                 }
2894                 goto out_only_mutex;
2895         }
2896
2897         /* zero back part of the first block */
2898         if (offset < ino_size) {
2899                 truncated_block = true;
2900                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2901                 if (ret) {
2902                         inode_unlock(inode);
2903                         return ret;
2904                 }
2905         }
2906
2907         /* Check the aligned pages after the first unaligned page,
2908          * if offset != orig_start, which means the first unaligned page
2909          * including several following pages are already in holes,
2910          * the extra check can be skipped */
2911         if (offset == orig_start) {
2912                 /* after truncate page, check hole again */
2913                 len = offset + len - lockstart;
2914                 offset = lockstart;
2915                 ret = find_first_non_hole(inode, &offset, &len);
2916                 if (ret < 0)
2917                         goto out_only_mutex;
2918                 if (ret && !len) {
2919                         ret = 0;
2920                         goto out_only_mutex;
2921                 }
2922                 lockstart = offset;
2923         }
2924
2925         /* Check the tail unaligned part is in a hole */
2926         tail_start = lockend + 1;
2927         tail_len = offset + len - tail_start;
2928         if (tail_len) {
2929                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2930                 if (unlikely(ret < 0))
2931                         goto out_only_mutex;
2932                 if (!ret) {
2933                         /* zero the front end of the last page */
2934                         if (tail_start + tail_len < ino_size) {
2935                                 truncated_block = true;
2936                                 ret = btrfs_truncate_block(inode,
2937                                                         tail_start + tail_len,
2938                                                         0, 1);
2939                                 if (ret)
2940                                         goto out_only_mutex;
2941                         }
2942                 }
2943         }
2944
2945         if (lockend < lockstart) {
2946                 ret = 0;
2947                 goto out_only_mutex;
2948         }
2949
2950         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2951                                           &cached_state);
2952         if (ret)
2953                 goto out_only_mutex;
2954
2955         path = btrfs_alloc_path();
2956         if (!path) {
2957                 ret = -ENOMEM;
2958                 goto out;
2959         }
2960
2961         ret = btrfs_replace_file_extents(inode, path, lockstart, lockend, NULL,
2962                                      &trans);
2963         btrfs_free_path(path);
2964         if (ret)
2965                 goto out;
2966
2967         ASSERT(trans != NULL);
2968         inode_inc_iversion(inode);
2969         inode->i_mtime = inode->i_ctime = current_time(inode);
2970         ret = btrfs_update_inode(trans, root, inode);
2971         updated_inode = true;
2972         btrfs_end_transaction(trans);
2973         btrfs_btree_balance_dirty(fs_info);
2974 out:
2975         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2976                              &cached_state);
2977 out_only_mutex:
2978         if (!updated_inode && truncated_block && !ret) {
2979                 /*
2980                  * If we only end up zeroing part of a page, we still need to
2981                  * update the inode item, so that all the time fields are
2982                  * updated as well as the necessary btrfs inode in memory fields
2983                  * for detecting, at fsync time, if the inode isn't yet in the
2984                  * log tree or it's there but not up to date.
2985                  */
2986                 struct timespec64 now = current_time(inode);
2987
2988                 inode_inc_iversion(inode);
2989                 inode->i_mtime = now;
2990                 inode->i_ctime = now;
2991                 trans = btrfs_start_transaction(root, 1);
2992                 if (IS_ERR(trans)) {
2993                         ret = PTR_ERR(trans);
2994                 } else {
2995                         int ret2;
2996
2997                         ret = btrfs_update_inode(trans, root, inode);
2998                         ret2 = btrfs_end_transaction(trans);
2999                         if (!ret)
3000                                 ret = ret2;
3001                 }
3002         }
3003         inode_unlock(inode);
3004         return ret;
3005 }
3006
3007 /* Helper structure to record which range is already reserved */
3008 struct falloc_range {
3009         struct list_head list;
3010         u64 start;
3011         u64 len;
3012 };
3013
3014 /*
3015  * Helper function to add falloc range
3016  *
3017  * Caller should have locked the larger range of extent containing
3018  * [start, len)
3019  */
3020 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3021 {
3022         struct falloc_range *prev = NULL;
3023         struct falloc_range *range = NULL;
3024
3025         if (list_empty(head))
3026                 goto insert;
3027
3028         /*
3029          * As fallocate iterate by bytenr order, we only need to check
3030          * the last range.
3031          */
3032         prev = list_entry(head->prev, struct falloc_range, list);
3033         if (prev->start + prev->len == start) {
3034                 prev->len += len;
3035                 return 0;
3036         }
3037 insert:
3038         range = kmalloc(sizeof(*range), GFP_KERNEL);
3039         if (!range)
3040                 return -ENOMEM;
3041         range->start = start;
3042         range->len = len;
3043         list_add_tail(&range->list, head);
3044         return 0;
3045 }
3046
3047 static int btrfs_fallocate_update_isize(struct inode *inode,
3048                                         const u64 end,
3049                                         const int mode)
3050 {
3051         struct btrfs_trans_handle *trans;
3052         struct btrfs_root *root = BTRFS_I(inode)->root;
3053         int ret;
3054         int ret2;
3055
3056         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3057                 return 0;
3058
3059         trans = btrfs_start_transaction(root, 1);
3060         if (IS_ERR(trans))
3061                 return PTR_ERR(trans);
3062
3063         inode->i_ctime = current_time(inode);
3064         i_size_write(inode, end);
3065         btrfs_inode_safe_disk_i_size_write(inode, 0);
3066         ret = btrfs_update_inode(trans, root, inode);
3067         ret2 = btrfs_end_transaction(trans);
3068
3069         return ret ? ret : ret2;
3070 }
3071
3072 enum {
3073         RANGE_BOUNDARY_WRITTEN_EXTENT,
3074         RANGE_BOUNDARY_PREALLOC_EXTENT,
3075         RANGE_BOUNDARY_HOLE,
3076 };
3077
3078 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3079                                                  u64 offset)
3080 {
3081         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3082         struct extent_map *em;
3083         int ret;
3084
3085         offset = round_down(offset, sectorsize);
3086         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3087         if (IS_ERR(em))
3088                 return PTR_ERR(em);
3089
3090         if (em->block_start == EXTENT_MAP_HOLE)
3091                 ret = RANGE_BOUNDARY_HOLE;
3092         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3093                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3094         else
3095                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3096
3097         free_extent_map(em);
3098         return ret;
3099 }
3100
3101 static int btrfs_zero_range(struct inode *inode,
3102                             loff_t offset,
3103                             loff_t len,
3104                             const int mode)
3105 {
3106         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3107         struct extent_map *em;
3108         struct extent_changeset *data_reserved = NULL;
3109         int ret;
3110         u64 alloc_hint = 0;
3111         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3112         u64 alloc_start = round_down(offset, sectorsize);
3113         u64 alloc_end = round_up(offset + len, sectorsize);
3114         u64 bytes_to_reserve = 0;
3115         bool space_reserved = false;
3116
3117         inode_dio_wait(inode);
3118
3119         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3120                               alloc_end - alloc_start);
3121         if (IS_ERR(em)) {
3122                 ret = PTR_ERR(em);
3123                 goto out;
3124         }
3125
3126         /*
3127          * Avoid hole punching and extent allocation for some cases. More cases
3128          * could be considered, but these are unlikely common and we keep things
3129          * as simple as possible for now. Also, intentionally, if the target
3130          * range contains one or more prealloc extents together with regular
3131          * extents and holes, we drop all the existing extents and allocate a
3132          * new prealloc extent, so that we get a larger contiguous disk extent.
3133          */
3134         if (em->start <= alloc_start &&
3135             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3136                 const u64 em_end = em->start + em->len;
3137
3138                 if (em_end >= offset + len) {
3139                         /*
3140                          * The whole range is already a prealloc extent,
3141                          * do nothing except updating the inode's i_size if
3142                          * needed.
3143                          */
3144                         free_extent_map(em);
3145                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3146                                                            mode);
3147                         goto out;
3148                 }
3149                 /*
3150                  * Part of the range is already a prealloc extent, so operate
3151                  * only on the remaining part of the range.
3152                  */
3153                 alloc_start = em_end;
3154                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3155                 len = offset + len - alloc_start;
3156                 offset = alloc_start;
3157                 alloc_hint = em->block_start + em->len;
3158         }
3159         free_extent_map(em);
3160
3161         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3162             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3163                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3164                                       sectorsize);
3165                 if (IS_ERR(em)) {
3166                         ret = PTR_ERR(em);
3167                         goto out;
3168                 }
3169
3170                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3171                         free_extent_map(em);
3172                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3173                                                            mode);
3174                         goto out;
3175                 }
3176                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3177                         free_extent_map(em);
3178                         ret = btrfs_truncate_block(inode, offset, len, 0);
3179                         if (!ret)
3180                                 ret = btrfs_fallocate_update_isize(inode,
3181                                                                    offset + len,
3182                                                                    mode);
3183                         return ret;
3184                 }
3185                 free_extent_map(em);
3186                 alloc_start = round_down(offset, sectorsize);
3187                 alloc_end = alloc_start + sectorsize;
3188                 goto reserve_space;
3189         }
3190
3191         alloc_start = round_up(offset, sectorsize);
3192         alloc_end = round_down(offset + len, sectorsize);
3193
3194         /*
3195          * For unaligned ranges, check the pages at the boundaries, they might
3196          * map to an extent, in which case we need to partially zero them, or
3197          * they might map to a hole, in which case we need our allocation range
3198          * to cover them.
3199          */
3200         if (!IS_ALIGNED(offset, sectorsize)) {
3201                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3202                                                             offset);
3203                 if (ret < 0)
3204                         goto out;
3205                 if (ret == RANGE_BOUNDARY_HOLE) {
3206                         alloc_start = round_down(offset, sectorsize);
3207                         ret = 0;
3208                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3209                         ret = btrfs_truncate_block(inode, offset, 0, 0);
3210                         if (ret)
3211                                 goto out;
3212                 } else {
3213                         ret = 0;
3214                 }
3215         }
3216
3217         if (!IS_ALIGNED(offset + len, sectorsize)) {
3218                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3219                                                             offset + len);
3220                 if (ret < 0)
3221                         goto out;
3222                 if (ret == RANGE_BOUNDARY_HOLE) {
3223                         alloc_end = round_up(offset + len, sectorsize);
3224                         ret = 0;
3225                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3226                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3227                         if (ret)
3228                                 goto out;
3229                 } else {
3230                         ret = 0;
3231                 }
3232         }
3233
3234 reserve_space:
3235         if (alloc_start < alloc_end) {
3236                 struct extent_state *cached_state = NULL;
3237                 const u64 lockstart = alloc_start;
3238                 const u64 lockend = alloc_end - 1;
3239
3240                 bytes_to_reserve = alloc_end - alloc_start;
3241                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3242                                                       bytes_to_reserve);
3243                 if (ret < 0)
3244                         goto out;
3245                 space_reserved = true;
3246                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3247                                                   &cached_state);
3248                 if (ret)
3249                         goto out;
3250                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3251                                                 alloc_start, bytes_to_reserve);
3252                 if (ret) {
3253                         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3254                                              lockend, &cached_state);
3255                         goto out;
3256                 }
3257                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3258                                                 alloc_end - alloc_start,
3259                                                 i_blocksize(inode),
3260                                                 offset + len, &alloc_hint);
3261                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3262                                      lockend, &cached_state);
3263                 /* btrfs_prealloc_file_range releases reserved space on error */
3264                 if (ret) {
3265                         space_reserved = false;
3266                         goto out;
3267                 }
3268         }
3269         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3270  out:
3271         if (ret && space_reserved)
3272                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3273                                                alloc_start, bytes_to_reserve);
3274         extent_changeset_free(data_reserved);
3275
3276         return ret;
3277 }
3278
3279 static long btrfs_fallocate(struct file *file, int mode,
3280                             loff_t offset, loff_t len)
3281 {
3282         struct inode *inode = file_inode(file);
3283         struct extent_state *cached_state = NULL;
3284         struct extent_changeset *data_reserved = NULL;
3285         struct falloc_range *range;
3286         struct falloc_range *tmp;
3287         struct list_head reserve_list;
3288         u64 cur_offset;
3289         u64 last_byte;
3290         u64 alloc_start;
3291         u64 alloc_end;
3292         u64 alloc_hint = 0;
3293         u64 locked_end;
3294         u64 actual_end = 0;
3295         struct extent_map *em;
3296         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3297         int ret;
3298
3299         alloc_start = round_down(offset, blocksize);
3300         alloc_end = round_up(offset + len, blocksize);
3301         cur_offset = alloc_start;
3302
3303         /* Make sure we aren't being give some crap mode */
3304         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3305                      FALLOC_FL_ZERO_RANGE))
3306                 return -EOPNOTSUPP;
3307
3308         if (mode & FALLOC_FL_PUNCH_HOLE)
3309                 return btrfs_punch_hole(file, offset, len);
3310
3311         /*
3312          * Only trigger disk allocation, don't trigger qgroup reserve
3313          *
3314          * For qgroup space, it will be checked later.
3315          */
3316         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3317                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3318                                                       alloc_end - alloc_start);
3319                 if (ret < 0)
3320                         return ret;
3321         }
3322
3323         inode_lock(inode);
3324
3325         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3326                 ret = inode_newsize_ok(inode, offset + len);
3327                 if (ret)
3328                         goto out;
3329         }
3330
3331         ret = file_modified(file);
3332         if (ret)
3333                 goto out;
3334
3335         /*
3336          * TODO: Move these two operations after we have checked
3337          * accurate reserved space, or fallocate can still fail but
3338          * with page truncated or size expanded.
3339          *
3340          * But that's a minor problem and won't do much harm BTW.
3341          */
3342         if (alloc_start > inode->i_size) {
3343                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3344                                         alloc_start);
3345                 if (ret)
3346                         goto out;
3347         } else if (offset + len > inode->i_size) {
3348                 /*
3349                  * If we are fallocating from the end of the file onward we
3350                  * need to zero out the end of the block if i_size lands in the
3351                  * middle of a block.
3352                  */
3353                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3354                 if (ret)
3355                         goto out;
3356         }
3357
3358         /*
3359          * wait for ordered IO before we have any locks.  We'll loop again
3360          * below with the locks held.
3361          */
3362         ret = btrfs_wait_ordered_range(inode, alloc_start,
3363                                        alloc_end - alloc_start);
3364         if (ret)
3365                 goto out;
3366
3367         if (mode & FALLOC_FL_ZERO_RANGE) {
3368                 ret = btrfs_zero_range(inode, offset, len, mode);
3369                 inode_unlock(inode);
3370                 return ret;
3371         }
3372
3373         locked_end = alloc_end - 1;
3374         while (1) {
3375                 struct btrfs_ordered_extent *ordered;
3376
3377                 /* the extent lock is ordered inside the running
3378                  * transaction
3379                  */
3380                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3381                                  locked_end, &cached_state);
3382                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3383                                                             locked_end);
3384
3385                 if (ordered &&
3386                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3387                     ordered->file_offset < alloc_end) {
3388                         btrfs_put_ordered_extent(ordered);
3389                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3390                                              alloc_start, locked_end,
3391                                              &cached_state);
3392                         /*
3393                          * we can't wait on the range with the transaction
3394                          * running or with the extent lock held
3395                          */
3396                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3397                                                        alloc_end - alloc_start);
3398                         if (ret)
3399                                 goto out;
3400                 } else {
3401                         if (ordered)
3402                                 btrfs_put_ordered_extent(ordered);
3403                         break;
3404                 }
3405         }
3406
3407         /* First, check if we exceed the qgroup limit */
3408         INIT_LIST_HEAD(&reserve_list);
3409         while (cur_offset < alloc_end) {
3410                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3411                                       alloc_end - cur_offset);
3412                 if (IS_ERR(em)) {
3413                         ret = PTR_ERR(em);
3414                         break;
3415                 }
3416                 last_byte = min(extent_map_end(em), alloc_end);
3417                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3418                 last_byte = ALIGN(last_byte, blocksize);
3419                 if (em->block_start == EXTENT_MAP_HOLE ||
3420                     (cur_offset >= inode->i_size &&
3421                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3422                         ret = add_falloc_range(&reserve_list, cur_offset,
3423                                                last_byte - cur_offset);
3424                         if (ret < 0) {
3425                                 free_extent_map(em);
3426                                 break;
3427                         }
3428                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3429                                         &data_reserved, cur_offset,
3430                                         last_byte - cur_offset);
3431                         if (ret < 0) {
3432                                 cur_offset = last_byte;
3433                                 free_extent_map(em);
3434                                 break;
3435                         }
3436                 } else {
3437                         /*
3438                          * Do not need to reserve unwritten extent for this
3439                          * range, free reserved data space first, otherwise
3440                          * it'll result in false ENOSPC error.
3441                          */
3442                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3443                                 data_reserved, cur_offset,
3444                                 last_byte - cur_offset);
3445                 }
3446                 free_extent_map(em);
3447                 cur_offset = last_byte;
3448         }
3449
3450         /*
3451          * If ret is still 0, means we're OK to fallocate.
3452          * Or just cleanup the list and exit.
3453          */
3454         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3455                 if (!ret)
3456                         ret = btrfs_prealloc_file_range(inode, mode,
3457                                         range->start,
3458                                         range->len, i_blocksize(inode),
3459                                         offset + len, &alloc_hint);
3460                 else
3461                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3462                                         data_reserved, range->start,
3463                                         range->len);
3464                 list_del(&range->list);
3465                 kfree(range);
3466         }
3467         if (ret < 0)
3468                 goto out_unlock;
3469
3470         /*
3471          * We didn't need to allocate any more space, but we still extended the
3472          * size of the file so we need to update i_size and the inode item.
3473          */
3474         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3475 out_unlock:
3476         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3477                              &cached_state);
3478 out:
3479         inode_unlock(inode);
3480         /* Let go of our reservation. */
3481         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3482                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3483                                 cur_offset, alloc_end - cur_offset);
3484         extent_changeset_free(data_reserved);
3485         return ret;
3486 }
3487
3488 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3489                                   int whence)
3490 {
3491         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3492         struct extent_map *em = NULL;
3493         struct extent_state *cached_state = NULL;
3494         loff_t i_size = inode->i_size;
3495         u64 lockstart;
3496         u64 lockend;
3497         u64 start;
3498         u64 len;
3499         int ret = 0;
3500
3501         if (i_size == 0 || offset >= i_size)
3502                 return -ENXIO;
3503
3504         /*
3505          * offset can be negative, in this case we start finding DATA/HOLE from
3506          * the very start of the file.
3507          */
3508         start = max_t(loff_t, 0, offset);
3509
3510         lockstart = round_down(start, fs_info->sectorsize);
3511         lockend = round_up(i_size, fs_info->sectorsize);
3512         if (lockend <= lockstart)
3513                 lockend = lockstart + fs_info->sectorsize;
3514         lockend--;
3515         len = lockend - lockstart + 1;
3516
3517         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3518                          &cached_state);
3519
3520         while (start < i_size) {
3521                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3522                 if (IS_ERR(em)) {
3523                         ret = PTR_ERR(em);
3524                         em = NULL;
3525                         break;
3526                 }
3527
3528                 if (whence == SEEK_HOLE &&
3529                     (em->block_start == EXTENT_MAP_HOLE ||
3530                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3531                         break;
3532                 else if (whence == SEEK_DATA &&
3533                            (em->block_start != EXTENT_MAP_HOLE &&
3534                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3535                         break;
3536
3537                 start = em->start + em->len;
3538                 free_extent_map(em);
3539                 em = NULL;
3540                 cond_resched();
3541         }
3542         free_extent_map(em);
3543         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3544                              &cached_state);
3545         if (ret) {
3546                 offset = ret;
3547         } else {
3548                 if (whence == SEEK_DATA && start >= i_size)
3549                         offset = -ENXIO;
3550                 else
3551                         offset = min_t(loff_t, start, i_size);
3552         }
3553
3554         return offset;
3555 }
3556
3557 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3558 {
3559         struct inode *inode = file->f_mapping->host;
3560
3561         switch (whence) {
3562         default:
3563                 return generic_file_llseek(file, offset, whence);
3564         case SEEK_DATA:
3565         case SEEK_HOLE:
3566                 inode_lock_shared(inode);
3567                 offset = find_desired_extent(inode, offset, whence);
3568                 inode_unlock_shared(inode);
3569                 break;
3570         }
3571
3572         if (offset < 0)
3573                 return offset;
3574
3575         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3576 }
3577
3578 static int btrfs_file_open(struct inode *inode, struct file *filp)
3579 {
3580         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3581         return generic_file_open(inode, filp);
3582 }
3583
3584 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3585 {
3586         ssize_t ret = 0;
3587
3588         if (iocb->ki_flags & IOCB_DIRECT) {
3589                 struct inode *inode = file_inode(iocb->ki_filp);
3590
3591                 inode_lock_shared(inode);
3592                 ret = btrfs_direct_IO(iocb, to);
3593                 inode_unlock_shared(inode);
3594                 if (ret < 0 || !iov_iter_count(to) ||
3595                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3596                         return ret;
3597         }
3598
3599         return generic_file_buffered_read(iocb, to, ret);
3600 }
3601
3602 const struct file_operations btrfs_file_operations = {
3603         .llseek         = btrfs_file_llseek,
3604         .read_iter      = btrfs_file_read_iter,
3605         .splice_read    = generic_file_splice_read,
3606         .write_iter     = btrfs_file_write_iter,
3607         .splice_write   = iter_file_splice_write,
3608         .mmap           = btrfs_file_mmap,
3609         .open           = btrfs_file_open,
3610         .release        = btrfs_release_file,
3611         .fsync          = btrfs_sync_file,
3612         .fallocate      = btrfs_fallocate,
3613         .unlocked_ioctl = btrfs_ioctl,
3614 #ifdef CONFIG_COMPAT
3615         .compat_ioctl   = btrfs_compat_ioctl,
3616 #endif
3617         .remap_file_range = btrfs_remap_file_range,
3618 };
3619
3620 void __cold btrfs_auto_defrag_exit(void)
3621 {
3622         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3623 }
3624
3625 int __init btrfs_auto_defrag_init(void)
3626 {
3627         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3628                                         sizeof(struct inode_defrag), 0,
3629                                         SLAB_MEM_SPREAD,
3630                                         NULL);
3631         if (!btrfs_inode_defrag_cachep)
3632                 return -ENOMEM;
3633
3634         return 0;
3635 }
3636
3637 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3638 {
3639         int ret;
3640
3641         /*
3642          * So with compression we will find and lock a dirty page and clear the
3643          * first one as dirty, setup an async extent, and immediately return
3644          * with the entire range locked but with nobody actually marked with
3645          * writeback.  So we can't just filemap_write_and_wait_range() and
3646          * expect it to work since it will just kick off a thread to do the
3647          * actual work.  So we need to call filemap_fdatawrite_range _again_
3648          * since it will wait on the page lock, which won't be unlocked until
3649          * after the pages have been marked as writeback and so we're good to go
3650          * from there.  We have to do this otherwise we'll miss the ordered
3651          * extents and that results in badness.  Please Josef, do not think you
3652          * know better and pull this out at some point in the future, it is
3653          * right and you are wrong.
3654          */
3655         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3656         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3657                              &BTRFS_I(inode)->runtime_flags))
3658                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3659
3660         return ret;
3661 }