fs/bcachefs/fs-io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fs-io-buffered.h"
  16 #include "fs-io-pagecache.h"
  17 #include "fsck.h"
  18 #include "inode.h"
  19 #include "journal.h"
  20 #include "io_misc.h"
  21 #include "keylist.h"
  22 #include "quota.h"
  23 #include "reflink.h"
  24 #include "trace.h"
  25
  26 #include <linux/aio.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/falloc.h>
  29 #include <linux/migrate.h>
  30 #include <linux/mmu_context.h>
  31 #include <linux/pagevec.h>
  32 #include <linux/rmap.h>
  33 #include <linux/sched/signal.h>
  34 #include <linux/task_io_accounting_ops.h>
  35 #include <linux/uio.h>
  36
  37 #include <trace/events/writeback.h>
  38
  39 struct nocow_flush {
  40         struct closure  *cl;
  41         struct bch_dev  *ca;
  42         struct bio      bio;
  43 };
  44
  45 static void nocow_flush_endio(struct bio *_bio)
  46 {
  47
  48         struct nocow_flush *bio = container_of(_bio, struct nocow_flush, bio);
  49
  50         closure_put(bio->cl);
  51         percpu_ref_put(&bio->ca->io_ref);
  52         bio_put(&bio->bio);
  53 }
  54
  55 void bch2_inode_flush_nocow_writes_async(struct bch_fs *c,
  56                                          struct bch_inode_info *inode,
  57                                          struct closure *cl)
  58 {
  59         struct nocow_flush *bio;
  60         struct bch_dev *ca;
  61         struct bch_devs_mask devs;
  62         unsigned dev;
  63
  64         dev = find_first_bit(inode->ei_devs_need_flush.d, BCH_SB_MEMBERS_MAX);
  65         if (dev == BCH_SB_MEMBERS_MAX)
  66                 return;
  67
  68         devs = inode->ei_devs_need_flush;
  69         memset(&inode->ei_devs_need_flush, 0, sizeof(inode->ei_devs_need_flush));
  70
  71         for_each_set_bit(dev, devs.d, BCH_SB_MEMBERS_MAX) {
  72                 rcu_read_lock();
  73                 ca = rcu_dereference(c->devs[dev]);
  74                 if (ca && !percpu_ref_tryget(&ca->io_ref))
  75                         ca = NULL;
  76                 rcu_read_unlock();
  77
  78                 if (!ca)
  79                         continue;
  80
  81                 bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev, 0,
  82                                                     REQ_OP_WRITE|REQ_PREFLUSH,
  83                                                     GFP_KERNEL,
  84                                                     &c->nocow_flush_bioset),
  85                                    struct nocow_flush, bio);
  86                 bio->cl                 = cl;
  87                 bio->ca                 = ca;
  88                 bio->bio.bi_end_io      = nocow_flush_endio;
  89                 closure_bio_submit(&bio->bio, cl);
  90         }
  91 }
  92
  93 static int bch2_inode_flush_nocow_writes(struct bch_fs *c,
  94                                          struct bch_inode_info *inode)
  95 {
  96         struct closure cl;
  97
  98         closure_init_stack(&cl);
  99         bch2_inode_flush_nocow_writes_async(c, inode, &cl);
 100         closure_sync(&cl);
 101
 102         return 0;
 103 }
 104
 105 /* i_size updates: */
 106
 107 struct inode_new_size {
 108         loff_t          new_size;
 109         u64             now;
 110         unsigned        fields;
 111 };
 112
 113 static int inode_set_size(struct btree_trans *trans,
 114                           struct bch_inode_info *inode,
 115                           struct bch_inode_unpacked *bi,
 116                           void *p)
 117 {
 118         struct inode_new_size *s = p;
 119
 120         bi->bi_size = s->new_size;
 121         if (s->fields & ATTR_ATIME)
 122                 bi->bi_atime = s->now;
 123         if (s->fields & ATTR_MTIME)
 124                 bi->bi_mtime = s->now;
 125         if (s->fields & ATTR_CTIME)
 126                 bi->bi_ctime = s->now;
 127
 128         return 0;
 129 }
 130
 131 int __must_check bch2_write_inode_size(struct bch_fs *c,
 132                                        struct bch_inode_info *inode,
 133                                        loff_t new_size, unsigned fields)
 134 {
 135         struct inode_new_size s = {
 136                 .new_size       = new_size,
 137                 .now            = bch2_current_time(c),
 138                 .fields         = fields,
 139         };
 140
 141         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 142 }
 143
 144 void __bch2_i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 145                            struct quota_res *quota_res, s64 sectors)
 146 {
 147         bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
 148                                 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
 149                                 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
 150                                 inode->ei_inode.bi_sectors);
 151         inode->v.i_blocks += sectors;
 152
 153 #ifdef CONFIG_BCACHEFS_QUOTA
 154         if (quota_res &&
 155             !test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags) &&
 156             sectors > 0) {
 157                 BUG_ON(sectors > quota_res->sectors);
 158                 BUG_ON(sectors > inode->ei_quota_reserved);
 159
 160                 quota_res->sectors -= sectors;
 161                 inode->ei_quota_reserved -= sectors;
 162         } else {
 163                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 164         }
 165 #endif
 166 }
 167
 168 /* fsync: */
 169
 170 /*
 171  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
 172  * insert trigger: look up the btree inode instead
 173  */
 174 static int bch2_flush_inode(struct bch_fs *c,
 175                             struct bch_inode_info *inode)
 176 {
 177         if (c->opts.journal_flush_disabled)
 178                 return 0;
 179
 180         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fsync))
 181                 return -EROFS;
 182
 183         struct bch_inode_unpacked u;
 184         int ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u) ?:
 185                   bch2_journal_flush_seq(&c->journal, u.bi_journal_seq, TASK_INTERRUPTIBLE) ?:
 186                   bch2_inode_flush_nocow_writes(c, inode);
 187         bch2_write_ref_put(c, BCH_WRITE_REF_fsync);
 188         return ret;
 189 }
 190
 191 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
 192 {
 193         struct bch_inode_info *inode = file_bch_inode(file);
 194         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 195         int ret, err;
 196
 197         trace_bch2_fsync(file, datasync);
 198
 199         ret = file_write_and_wait_range(file, start, end);
 200         if (ret)
 201                 goto out;
 202         ret = sync_inode_metadata(&inode->v, 1);
 203         if (ret)
 204                 goto out;
 205         ret = bch2_flush_inode(c, inode);
 206 out:
 207         ret = bch2_err_class(ret);
 208         if (ret == -EROFS)
 209                 ret = -EIO;
 210
 211         err = file_check_and_advance_wb_err(file);
 212         if (!ret)
 213                 ret = err;
 214
 215         return ret;
 216 }
 217
 218 /* truncate: */
 219
 220 static inline int range_has_data(struct bch_fs *c, u32 subvol,
 221                                  struct bpos start,
 222                                  struct bpos end)
 223 {
 224         return bch2_trans_run(c,
 225                 for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents, start, end,
 226                                                     subvol, 0, k, ({
 227                         bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k);
 228                 })));
 229 }
 230
 231 static int __bch2_truncate_folio(struct bch_inode_info *inode,
 232                                  pgoff_t index, loff_t start, loff_t end)
 233 {
 234         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 235         struct address_space *mapping = inode->v.i_mapping;
 236         struct bch_folio *s;
 237         unsigned start_offset;
 238         unsigned end_offset;
 239         unsigned i;
 240         struct folio *folio;
 241         s64 i_sectors_delta = 0;
 242         int ret = 0;
 243         u64 end_pos;
 244
 245         folio = filemap_lock_folio(mapping, index);
 246         if (IS_ERR_OR_NULL(folio)) {
 247                 /*
 248                  * XXX: we're doing two index lookups when we end up reading the
 249                  * folio
 250                  */
 251                 ret = range_has_data(c, inode->ei_inum.subvol,
 252                                 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT)),
 253                                 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT) + PAGE_SECTORS));
 254                 if (ret <= 0)
 255                         return ret;
 256
 257                 folio = __filemap_get_folio(mapping, index,
 258                                             FGP_LOCK|FGP_CREAT, GFP_KERNEL);
 259                 if (IS_ERR_OR_NULL(folio)) {
 260                         ret = -ENOMEM;
 261                         goto out;
 262                 }
 263         }
 264
 265         BUG_ON(start    >= folio_end_pos(folio));
 266         BUG_ON(end      <= folio_pos(folio));
 267
 268         start_offset    = max(start, folio_pos(folio)) - folio_pos(folio);
 269         end_offset      = min_t(u64, end, folio_end_pos(folio)) - folio_pos(folio);
 270
 271         /* Folio boundary? Nothing to do */
 272         if (start_offset == 0 &&
 273             end_offset == folio_size(folio)) {
 274                 ret = 0;
 275                 goto unlock;
 276         }
 277
 278         s = bch2_folio_create(folio, 0);
 279         if (!s) {
 280                 ret = -ENOMEM;
 281                 goto unlock;
 282         }
 283
 284         if (!folio_test_uptodate(folio)) {
 285                 ret = bch2_read_single_folio(folio, mapping);
 286                 if (ret)
 287                         goto unlock;
 288         }
 289
 290         ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
 291         if (ret)
 292                 goto unlock;
 293
 294         for (i = round_up(start_offset, block_bytes(c)) >> 9;
 295              i < round_down(end_offset, block_bytes(c)) >> 9;
 296              i++) {
 297                 s->s[i].nr_replicas     = 0;
 298
 299                 i_sectors_delta -= s->s[i].state == SECTOR_dirty;
 300                 bch2_folio_sector_set(folio, s, i, SECTOR_unallocated);
 301         }
 302
 303         bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 304
 305         /*
 306          * Caller needs to know whether this folio will be written out by
 307          * writeback - doing an i_size update if necessary - or whether it will
 308          * be responsible for the i_size update.
 309          *
 310          * Note that we shouldn't ever see a folio beyond EOF, but check and
 311          * warn if so. This has been observed by failure to clean up folios
 312          * after a short write and there's still a chance reclaim will fix
 313          * things up.
 314          */
 315         WARN_ON_ONCE(folio_pos(folio) >= inode->v.i_size);
 316         end_pos = folio_end_pos(folio);
 317         if (inode->v.i_size > folio_pos(folio))
 318                 end_pos = min_t(u64, inode->v.i_size, end_pos);
 319         ret = s->s[folio_pos_to_s(folio, end_pos - 1)].state >= SECTOR_dirty;
 320
 321         folio_zero_segment(folio, start_offset, end_offset);
 322
 323         /*
 324          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
 325          *
 326          * XXX: because we aren't currently tracking whether the folio has actual
 327          * data in it (vs. just 0s, or only partially written) this wrong. ick.
 328          */
 329         BUG_ON(bch2_get_folio_disk_reservation(c, inode, folio, false));
 330
 331         /*
 332          * This removes any writeable userspace mappings; we need to force
 333          * .page_mkwrite to be called again before any mmapped writes, to
 334          * redirty the full page:
 335          */
 336         folio_mkclean(folio);
 337         filemap_dirty_folio(mapping, folio);
 338 unlock:
 339         folio_unlock(folio);
 340         folio_put(folio);
 341 out:
 342         return ret;
 343 }
 344
 345 static int bch2_truncate_folio(struct bch_inode_info *inode, loff_t from)
 346 {
 347         return __bch2_truncate_folio(inode, from >> PAGE_SHIFT,
 348                                      from, ANYSINT_MAX(loff_t));
 349 }
 350
 351 static int bch2_truncate_folios(struct bch_inode_info *inode,
 352                                 loff_t start, loff_t end)
 353 {
 354         int ret = __bch2_truncate_folio(inode, start >> PAGE_SHIFT,
 355                                         start, end);
 356
 357         if (ret >= 0 &&
 358             start >> PAGE_SHIFT != end >> PAGE_SHIFT)
 359                 ret = __bch2_truncate_folio(inode,
 360                                         (end - 1) >> PAGE_SHIFT,
 361                                         start, end);
 362         return ret;
 363 }
 364
 365 static int bch2_extend(struct mnt_idmap *idmap,
 366                        struct bch_inode_info *inode,
 367                        struct bch_inode_unpacked *inode_u,
 368                        struct iattr *iattr)
 369 {
 370         struct address_space *mapping = inode->v.i_mapping;
 371         int ret;
 372
 373         /*
 374          * sync appends:
 375          *
 376          * this has to be done _before_ extending i_size:
 377          */
 378         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
 379         if (ret)
 380                 return ret;
 381
 382         truncate_setsize(&inode->v, iattr->ia_size);
 383
 384         return bch2_setattr_nonsize(idmap, inode, iattr);
 385 }
 386
 387 int bchfs_truncate(struct mnt_idmap *idmap,
 388                   struct bch_inode_info *inode, struct iattr *iattr)
 389 {
 390         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 391         struct address_space *mapping = inode->v.i_mapping;
 392         struct bch_inode_unpacked inode_u;
 393         s64 i_sectors_delta = 0;
 394         int ret = 0;
 395
 396         /*
 397          * If the truncate call with change the size of the file, the
 398          * cmtimes should be updated. If the size will not change, we
 399          * do not need to update the cmtimes.
 400          */
 401         if (iattr->ia_size != inode->v.i_size) {
 402                 if (!(iattr->ia_valid & ATTR_MTIME))
 403                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
 404                 if (!(iattr->ia_valid & ATTR_CTIME))
 405                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
 406                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
 407         }
 408
 409         inode_dio_wait(&inode->v);
 410         bch2_pagecache_block_get(inode);
 411
 412         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
 413         if (ret)
 414                 goto err;
 415
 416         /*
 417          * check this before next assertion; on filesystem error our normal
 418          * invariants are a bit broken (truncate has to truncate the page cache
 419          * before the inode).
 420          */
 421         ret = bch2_journal_error(&c->journal);
 422         if (ret)
 423                 goto err;
 424
 425         WARN_ONCE(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
 426                   inode->v.i_size < inode_u.bi_size,
 427                   "truncate spotted in mem i_size < btree i_size: %llu < %llu\n",
 428                   (u64) inode->v.i_size, inode_u.bi_size);
 429
 430         if (iattr->ia_size > inode->v.i_size) {
 431                 ret = bch2_extend(idmap, inode, &inode_u, iattr);
 432                 goto err;
 433         }
 434
 435         iattr->ia_valid &= ~ATTR_SIZE;
 436
 437         ret = bch2_truncate_folio(inode, iattr->ia_size);
 438         if (unlikely(ret < 0))
 439                 goto err;
 440
 441         truncate_setsize(&inode->v, iattr->ia_size);
 442
 443         /*
 444          * When extending, we're going to write the new i_size to disk
 445          * immediately so we need to flush anything above the current on disk
 446          * i_size first:
 447          *
 448          * Also, when extending we need to flush the page that i_size currently
 449          * straddles - if it's mapped to userspace, we need to ensure that
 450          * userspace has to redirty it and call .mkwrite -> set_page_dirty
 451          * again to allocate the part of the page that was extended.
 452          */
 453         if (iattr->ia_size > inode_u.bi_size)
 454                 ret = filemap_write_and_wait_range(mapping,
 455                                 inode_u.bi_size,
 456                                 iattr->ia_size - 1);
 457         else if (iattr->ia_size & (PAGE_SIZE - 1))
 458                 ret = filemap_write_and_wait_range(mapping,
 459                                 round_down(iattr->ia_size, PAGE_SIZE),
 460                                 iattr->ia_size - 1);
 461         if (ret)
 462                 goto err;
 463
 464         ret = bch2_truncate(c, inode_inum(inode), iattr->ia_size, &i_sectors_delta);
 465         bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 466
 467         if (unlikely(ret)) {
 468                 /*
 469                  * If we error here, VFS caches are now inconsistent with btree
 470                  */
 471                 set_bit(EI_INODE_ERROR, &inode->ei_flags);
 472                 goto err;
 473         }
 474
 475         bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
 476                                 !bch2_journal_error(&c->journal), c,
 477                                 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
 478                                 inode->v.i_ino, (u64) inode->v.i_blocks,
 479                                 inode->ei_inode.bi_sectors);
 480
 481         ret = bch2_setattr_nonsize(idmap, inode, iattr);
 482 err:
 483         bch2_pagecache_block_put(inode);
 484         return bch2_err_class(ret);
 485 }
 486
 487 /* fallocate: */
 488
 489 static int inode_update_times_fn(struct btree_trans *trans,
 490                                  struct bch_inode_info *inode,
 491                                  struct bch_inode_unpacked *bi, void *p)
 492 {
 493         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 494
 495         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
 496         return 0;
 497 }
 498
 499 static noinline long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
 500 {
 501         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 502         u64 end         = offset + len;
 503         u64 block_start = round_up(offset, block_bytes(c));
 504         u64 block_end   = round_down(end, block_bytes(c));
 505         bool truncated_last_page;
 506         int ret = 0;
 507
 508         ret = bch2_truncate_folios(inode, offset, end);
 509         if (unlikely(ret < 0))
 510                 goto err;
 511
 512         truncated_last_page = ret;
 513
 514         truncate_pagecache_range(&inode->v, offset, end - 1);
 515
 516         if (block_start < block_end) {
 517                 s64 i_sectors_delta = 0;
 518
 519                 ret = bch2_fpunch(c, inode_inum(inode),
 520                                   block_start >> 9, block_end >> 9,
 521                                   &i_sectors_delta);
 522                 bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 523         }
 524
 525         mutex_lock(&inode->ei_update_lock);
 526         if (end >= inode->v.i_size && !truncated_last_page) {
 527                 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
 528                                             ATTR_MTIME|ATTR_CTIME);
 529         } else {
 530                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
 531                                        ATTR_MTIME|ATTR_CTIME);
 532         }
 533         mutex_unlock(&inode->ei_update_lock);
 534 err:
 535         return ret;
 536 }
 537
 538 static noinline long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 539                                    loff_t offset, loff_t len,
 540                                    bool insert)
 541 {
 542         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 543         struct address_space *mapping = inode->v.i_mapping;
 544         s64 i_sectors_delta = 0;
 545         int ret = 0;
 546
 547         if ((offset | len) & (block_bytes(c) - 1))
 548                 return -EINVAL;
 549
 550         if (insert) {
 551                 if (offset >= inode->v.i_size)
 552                         return -EINVAL;
 553         } else {
 554                 if (offset + len >= inode->v.i_size)
 555                         return -EINVAL;
 556         }
 557
 558         ret = bch2_write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
 559         if (ret)
 560                 return ret;
 561
 562         if (insert)
 563                 i_size_write(&inode->v, inode->v.i_size + len);
 564
 565         ret = bch2_fcollapse_finsert(c, inode_inum(inode), offset >> 9, len >> 9,
 566                                      insert, &i_sectors_delta);
 567         if (!ret && !insert)
 568                 i_size_write(&inode->v, inode->v.i_size - len);
 569         bch2_i_sectors_acct(c, inode, NULL, i_sectors_delta);
 570
 571         return ret;
 572 }
 573
 574 static noinline int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
 575                              u64 start_sector, u64 end_sector)
 576 {
 577         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 578         struct btree_trans *trans = bch2_trans_get(c);
 579         struct btree_iter iter;
 580         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
 581         struct bch_io_opts opts;
 582         int ret = 0;
 583
 584         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
 585
 586         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 587                         POS(inode->v.i_ino, start_sector),
 588                         BTREE_ITER_slots|BTREE_ITER_intent);
 589
 590         while (!ret) {
 591                 s64 i_sectors_delta = 0;
 592                 struct quota_res quota_res = { 0 };
 593                 struct bkey_s_c k;
 594                 unsigned sectors;
 595                 bool is_allocation;
 596                 u64 hole_start, hole_end;
 597                 u32 snapshot;
 598
 599                 bch2_trans_begin(trans);
 600
 601                 if (bkey_ge(iter.pos, end_pos))
 602                         break;
 603
 604                 ret = bch2_subvolume_get_snapshot(trans,
 605                                         inode->ei_inum.subvol, &snapshot);
 606                 if (ret)
 607                         goto bkey_err;
 608
 609                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 610
 611                 k = bch2_btree_iter_peek_slot(&iter);
 612                 if ((ret = bkey_err(k)))
 613                         goto bkey_err;
 614
 615                 hole_start      = iter.pos.offset;
 616                 hole_end        = bpos_min(k.k->p, end_pos).offset;
 617                 is_allocation   = bkey_extent_is_allocation(k.k);
 618
 619                 /* already reserved */
 620                 if (bkey_extent_is_reservation(k) &&
 621                     bch2_bkey_nr_ptrs_fully_allocated(k) >= opts.data_replicas) {
 622                         bch2_btree_iter_advance(&iter);
 623                         continue;
 624                 }
 625
 626                 if (bkey_extent_is_data(k.k) &&
 627                     !(mode & FALLOC_FL_ZERO_RANGE)) {
 628                         bch2_btree_iter_advance(&iter);
 629                         continue;
 630                 }
 631
 632                 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
 633                         /*
 634                          * Lock ordering - can't be holding btree locks while
 635                          * blocking on a folio lock:
 636                          */
 637                         if (bch2_clamp_data_hole(&inode->v,
 638                                                  &hole_start,
 639                                                  &hole_end,
 640                                                  opts.data_replicas, true)) {
 641                                 ret = drop_locks_do(trans,
 642                                         (bch2_clamp_data_hole(&inode->v,
 643                                                               &hole_start,
 644                                                               &hole_end,
 645                                                               opts.data_replicas, false), 0));
 646                                 if (ret)
 647                                         goto bkey_err;
 648                         }
 649                         bch2_btree_iter_set_pos(&iter, POS(iter.pos.inode, hole_start));
 650
 651                         if (ret)
 652                                 goto bkey_err;
 653
 654                         if (hole_start == hole_end)
 655                                 continue;
 656                 }
 657
 658                 sectors = hole_end - hole_start;
 659
 660                 if (!is_allocation) {
 661                         ret = bch2_quota_reservation_add(c, inode,
 662                                         &quota_res, sectors, true);
 663                         if (unlikely(ret))
 664                                 goto bkey_err;
 665                 }
 666
 667                 ret = bch2_extent_fallocate(trans, inode_inum(inode), &iter,
 668                                             sectors, opts, &i_sectors_delta,
 669                                             writepoint_hashed((unsigned long) current));
 670                 if (ret)
 671                         goto bkey_err;
 672
 673                 bch2_i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
 674
 675                 if (bch2_mark_pagecache_reserved(inode, &hole_start,
 676                                                  iter.pos.offset, true)) {
 677                         ret = drop_locks_do(trans,
 678                                 bch2_mark_pagecache_reserved(inode, &hole_start,
 679                                                              iter.pos.offset, false));
 680                         if (ret)
 681                                 goto bkey_err;
 682                 }
 683 bkey_err:
 684                 bch2_quota_reservation_put(c, inode, &quota_res);
 685                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 686                         ret = 0;
 687         }
 688
 689         if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) {
 690                 struct quota_res quota_res = { 0 };
 691                 s64 i_sectors_delta = 0;
 692
 693                 bch2_fpunch_at(trans, &iter, inode_inum(inode),
 694                                end_sector, &i_sectors_delta);
 695                 bch2_i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
 696                 bch2_quota_reservation_put(c, inode, &quota_res);
 697         }
 698
 699         bch2_trans_iter_exit(trans, &iter);
 700         bch2_trans_put(trans);
 701         return ret;
 702 }
 703
 704 static noinline long bchfs_fallocate(struct bch_inode_info *inode, int mode,
 705                             loff_t offset, loff_t len)
 706 {
 707         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 708         u64 end         = offset + len;
 709         u64 block_start = round_down(offset,    block_bytes(c));
 710         u64 block_end   = round_up(end,         block_bytes(c));
 711         bool truncated_last_page = false;
 712         int ret, ret2 = 0;
 713
 714         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
 715                 ret = inode_newsize_ok(&inode->v, end);
 716                 if (ret)
 717                         return ret;
 718         }
 719
 720         if (mode & FALLOC_FL_ZERO_RANGE) {
 721                 ret = bch2_truncate_folios(inode, offset, end);
 722                 if (unlikely(ret < 0))
 723                         return ret;
 724
 725                 truncated_last_page = ret;
 726
 727                 truncate_pagecache_range(&inode->v, offset, end - 1);
 728
 729                 block_start     = round_up(offset,      block_bytes(c));
 730                 block_end       = round_down(end,       block_bytes(c));
 731         }
 732
 733         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
 734
 735         /*
 736          * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
 737          * so that the VFS cache i_size is consistent with the btree i_size:
 738          */
 739         if (ret &&
 740             !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)))
 741                 return ret;
 742
 743         if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
 744                 end = inode->v.i_size;
 745
 746         if (end >= inode->v.i_size &&
 747             (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
 748              !(mode & FALLOC_FL_KEEP_SIZE))) {
 749                 spin_lock(&inode->v.i_lock);
 750                 i_size_write(&inode->v, end);
 751                 spin_unlock(&inode->v.i_lock);
 752
 753                 mutex_lock(&inode->ei_update_lock);
 754                 ret2 = bch2_write_inode_size(c, inode, end, 0);
 755                 mutex_unlock(&inode->ei_update_lock);
 756         }
 757
 758         return ret ?: ret2;
 759 }
 760
 761 long bch2_fallocate_dispatch(struct file *file, int mode,
 762                              loff_t offset, loff_t len)
 763 {
 764         struct bch_inode_info *inode = file_bch_inode(file);
 765         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 766         long ret;
 767
 768         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fallocate))
 769                 return -EROFS;
 770
 771         inode_lock(&inode->v);
 772         inode_dio_wait(&inode->v);
 773         bch2_pagecache_block_get(inode);
 774
 775         ret = file_modified(file);
 776         if (ret)
 777                 goto err;
 778
 779         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
 780                 ret = bchfs_fallocate(inode, mode, offset, len);
 781         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
 782                 ret = bchfs_fpunch(inode, offset, len);
 783         else if (mode == FALLOC_FL_INSERT_RANGE)
 784                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
 785         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
 786                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
 787         else
 788                 ret = -EOPNOTSUPP;
 789 err:
 790         bch2_pagecache_block_put(inode);
 791         inode_unlock(&inode->v);
 792         bch2_write_ref_put(c, BCH_WRITE_REF_fallocate);
 793
 794         return bch2_err_class(ret);
 795 }
 796
 797 /*
 798  * Take a quota reservation for unallocated blocks in a given file range
 799  * Does not check pagecache
 800  */
 801 static int quota_reserve_range(struct bch_inode_info *inode,
 802                                struct quota_res *res,
 803                                u64 start, u64 end)
 804 {
 805         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 806         u64 sectors = end - start;
 807
 808         int ret = bch2_trans_run(c,
 809                 for_each_btree_key_in_subvolume_upto(trans, iter,
 810                                 BTREE_ID_extents,
 811                                 POS(inode->v.i_ino, start),
 812                                 POS(inode->v.i_ino, end - 1),
 813                                 inode->ei_inum.subvol, 0, k, ({
 814                         if (bkey_extent_is_allocation(k.k)) {
 815                                 u64 s = min(end, k.k->p.offset) -
 816                                         max(start, bkey_start_offset(k.k));
 817                                 BUG_ON(s > sectors);
 818                                 sectors -= s;
 819                         }
 820
 821                         0;
 822                 })));
 823
 824         return ret ?: bch2_quota_reservation_add(c, inode, res, sectors, true);
 825 }
 826
 827 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 828                              struct file *file_dst, loff_t pos_dst,
 829                              loff_t len, unsigned remap_flags)
 830 {
 831         struct bch_inode_info *src = file_bch_inode(file_src);
 832         struct bch_inode_info *dst = file_bch_inode(file_dst);
 833         struct bch_fs *c = src->v.i_sb->s_fs_info;
 834         struct quota_res quota_res = { 0 };
 835         s64 i_sectors_delta = 0;
 836         u64 aligned_len;
 837         loff_t ret = 0;
 838
 839         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
 840                 return -EINVAL;
 841
 842         if ((pos_src & (block_bytes(c) - 1)) ||
 843             (pos_dst & (block_bytes(c) - 1)))
 844                 return -EINVAL;
 845
 846         if (src == dst &&
 847             abs(pos_src - pos_dst) < len)
 848                 return -EINVAL;
 849
 850         lock_two_nondirectories(&src->v, &dst->v);
 851         bch2_lock_inodes(INODE_PAGECACHE_BLOCK, src, dst);
 852
 853         inode_dio_wait(&src->v);
 854         inode_dio_wait(&dst->v);
 855
 856         ret = generic_remap_file_range_prep(file_src, pos_src,
 857                                             file_dst, pos_dst,
 858                                             &len, remap_flags);
 859         if (ret < 0 || len == 0)
 860                 goto err;
 861
 862         aligned_len = round_up((u64) len, block_bytes(c));
 863
 864         ret = bch2_write_invalidate_inode_pages_range(dst->v.i_mapping,
 865                                 pos_dst, pos_dst + len - 1);
 866         if (ret)
 867                 goto err;
 868
 869         ret = quota_reserve_range(dst, &quota_res, pos_dst >> 9,
 870                                   (pos_dst + aligned_len) >> 9);
 871         if (ret)
 872                 goto err;
 873
 874         if (!(remap_flags & REMAP_FILE_DEDUP))
 875                 file_update_time(file_dst);
 876
 877         bch2_mark_pagecache_unallocated(src, pos_src >> 9,
 878                                    (pos_src + aligned_len) >> 9);
 879
 880         ret = bch2_remap_range(c,
 881                                inode_inum(dst), pos_dst >> 9,
 882                                inode_inum(src), pos_src >> 9,
 883                                aligned_len >> 9,
 884                                pos_dst + len, &i_sectors_delta);
 885         if (ret < 0)
 886                 goto err;
 887
 888         /*
 889          * due to alignment, we might have remapped slightly more than requsted
 890          */
 891         ret = min((u64) ret << 9, (u64) len);
 892
 893         bch2_i_sectors_acct(c, dst, &quota_res, i_sectors_delta);
 894
 895         spin_lock(&dst->v.i_lock);
 896         if (pos_dst + ret > dst->v.i_size)
 897                 i_size_write(&dst->v, pos_dst + ret);
 898         spin_unlock(&dst->v.i_lock);
 899
 900         if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
 901             IS_SYNC(file_inode(file_dst)))
 902                 ret = bch2_flush_inode(c, dst);
 903 err:
 904         bch2_quota_reservation_put(c, dst, &quota_res);
 905         bch2_unlock_inodes(INODE_PAGECACHE_BLOCK, src, dst);
 906         unlock_two_nondirectories(&src->v, &dst->v);
 907
 908         return bch2_err_class(ret);
 909 }
 910
 911 /* fseek: */
 912
 913 static loff_t bch2_seek_data(struct file *file, u64 offset)
 914 {
 915         struct bch_inode_info *inode = file_bch_inode(file);
 916         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 917         subvol_inum inum = inode_inum(inode);
 918         u64 isize, next_data = MAX_LFS_FILESIZE;
 919
 920         isize = i_size_read(&inode->v);
 921         if (offset >= isize)
 922                 return -ENXIO;
 923
 924         int ret = bch2_trans_run(c,
 925                 for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents,
 926                                    POS(inode->v.i_ino, offset >> 9),
 927                                    POS(inode->v.i_ino, U64_MAX),
 928                                    inum.subvol, 0, k, ({
 929                         if (bkey_extent_is_data(k.k)) {
 930                                 next_data = max(offset, bkey_start_offset(k.k) << 9);
 931                                 break;
 932                         } else if (k.k->p.offset >> 9 > isize)
 933                                 break;
 934                         0;
 935                 })));
 936         if (ret)
 937                 return ret;
 938
 939         if (next_data > offset)
 940                 next_data = bch2_seek_pagecache_data(&inode->v,
 941                                         offset, next_data, 0, false);
 942
 943         if (next_data >= isize)
 944                 return -ENXIO;
 945
 946         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
 947 }
 948
 949 static loff_t bch2_seek_hole(struct file *file, u64 offset)
 950 {
 951         struct bch_inode_info *inode = file_bch_inode(file);
 952         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 953         subvol_inum inum = inode_inum(inode);
 954         u64 isize, next_hole = MAX_LFS_FILESIZE;
 955
 956         isize = i_size_read(&inode->v);
 957         if (offset >= isize)
 958                 return -ENXIO;
 959
 960         int ret = bch2_trans_run(c,
 961                 for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents,
 962                                    POS(inode->v.i_ino, offset >> 9),
 963                                    POS(inode->v.i_ino, U64_MAX),
 964                                    inum.subvol, BTREE_ITER_slots, k, ({
 965                         if (k.k->p.inode != inode->v.i_ino) {
 966                                 next_hole = bch2_seek_pagecache_hole(&inode->v,
 967                                                 offset, MAX_LFS_FILESIZE, 0, false);
 968                                 break;
 969                         } else if (!bkey_extent_is_data(k.k)) {
 970                                 next_hole = bch2_seek_pagecache_hole(&inode->v,
 971                                                 max(offset, bkey_start_offset(k.k) << 9),
 972                                                 k.k->p.offset << 9, 0, false);
 973
 974                                 if (next_hole < k.k->p.offset << 9)
 975                                         break;
 976                         } else {
 977                                 offset = max(offset, bkey_start_offset(k.k) << 9);
 978                         }
 979                         0;
 980                 })));
 981         if (ret)
 982                 return ret;
 983
 984         if (next_hole > isize)
 985                 next_hole = isize;
 986
 987         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
 988 }
 989
 990 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
 991 {
 992         loff_t ret;
 993
 994         switch (whence) {
 995         case SEEK_SET:
 996         case SEEK_CUR:
 997         case SEEK_END:
 998                 ret = generic_file_llseek(file, offset, whence);
 999                 break;
1000         case SEEK_DATA:
1001                 ret = bch2_seek_data(file, offset);
1002                 break;
1003         case SEEK_HOLE:
1004                 ret = bch2_seek_hole(file, offset);
1005                 break;
1006         default:
1007                 ret = -EINVAL;
1008                 break;
1009         }
1010
1011         return bch2_err_class(ret);
1012 }
1013
1014 void bch2_fs_fsio_exit(struct bch_fs *c)
1015 {
1016         bioset_exit(&c->nocow_flush_bioset);
1017 }
1018
1019 int bch2_fs_fsio_init(struct bch_fs *c)
1020 {
1021         if (bioset_init(&c->nocow_flush_bioset,
1022                         1, offsetof(struct nocow_flush, bio), 0))
1023                 return -BCH_ERR_ENOMEM_nocow_flush_bioset_init;
1024
1025         return 0;
1026 }
1027
1028 #endif /* NO_BCACHEFS_FS */