git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/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 "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22 #include "trace.h"
  23
  24 #include <linux/aio.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/falloc.h>
  27 #include <linux/migrate.h>
  28 #include <linux/mmu_context.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/rmap.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/task_io_accounting_ops.h>
  33 #include <linux/uio.h>
  34 #include <linux/writeback.h>
  35
  36 #include <trace/events/writeback.h>
  37
  38 struct folio_vec {
  39         struct folio    *fv_folio;
  40         size_t          fv_offset;
  41         size_t          fv_len;
  42 };
  43
  44 static inline struct folio_vec biovec_to_foliovec(struct bio_vec bv)
  45 {
  46
  47         struct folio *folio     = page_folio(bv.bv_page);
  48         size_t offset           = (folio_page_idx(folio, bv.bv_page) << PAGE_SHIFT) +
  49                 bv.bv_offset;
  50         size_t len = min_t(size_t, folio_size(folio) - offset, bv.bv_len);
  51
  52         return (struct folio_vec) {
  53                 .fv_folio       = folio,
  54                 .fv_offset      = offset,
  55                 .fv_len         = len,
  56         };
  57 }
  58
  59 static inline struct folio_vec bio_iter_iovec_folio(struct bio *bio,
  60                                                     struct bvec_iter iter)
  61 {
  62         return biovec_to_foliovec(bio_iter_iovec(bio, iter));
  63 }
  64
  65 #define __bio_for_each_folio(bvl, bio, iter, start)                     \
  66         for (iter = (start);                                            \
  67              (iter).bi_size &&                                          \
  68                 ((bvl = bio_iter_iovec_folio((bio), (iter))), 1);       \
  69              bio_advance_iter_single((bio), &(iter), (bvl).fv_len))
  70
  71 /**
  72  * bio_for_each_folio - iterate over folios within a bio
  73  *
  74  * Like other non-_all versions, this iterates over what bio->bi_iter currently
  75  * points to. This version is for drivers, where the bio may have previously
  76  * been split or cloned.
  77  */
  78 #define bio_for_each_folio(bvl, bio, iter)                              \
  79         __bio_for_each_folio(bvl, bio, iter, (bio)->bi_iter)
  80
  81 /*
  82  * Use u64 for the end pos and sector helpers because if the folio covers the
  83  * max supported range of the mapping, the start offset of the next folio
  84  * overflows loff_t. This breaks much of the range based processing in the
  85  * buffered write path.
  86  */
  87 static inline u64 folio_end_pos(struct folio *folio)
  88 {
  89         return folio_pos(folio) + folio_size(folio);
  90 }
  91
  92 static inline size_t folio_sectors(struct folio *folio)
  93 {
  94         return PAGE_SECTORS << folio_order(folio);
  95 }
  96
  97 static inline loff_t folio_sector(struct folio *folio)
  98 {
  99         return folio_pos(folio) >> 9;
 100 }
 101
 102 static inline u64 folio_end_sector(struct folio *folio)
 103 {
 104         return folio_end_pos(folio) >> 9;
 105 }
 106
 107 typedef DARRAY(struct folio *) folios;
 108
 109 static int filemap_get_contig_folios_d(struct address_space *mapping,
 110                                        loff_t start, u64 end,
 111                                        int fgp_flags, gfp_t gfp,
 112                                        folios *folios)
 113 {
 114         struct folio *f;
 115         u64 pos = start;
 116         int ret = 0;
 117
 118         while (pos < end) {
 119                 if ((u64) pos >= (u64) start + (1ULL << 20))
 120                         fgp_flags &= ~FGP_CREAT;
 121
 122                 ret = darray_make_room_gfp(folios, 1, gfp & GFP_KERNEL);
 123                 if (ret)
 124                         break;
 125
 126                 f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
 127                 if (IS_ERR_OR_NULL(f))
 128                         break;
 129
 130                 BUG_ON(folios->nr && folio_pos(f) != pos);
 131
 132                 pos = folio_end_pos(f);
 133                 darray_push(folios, f);
 134         }
 135
 136         if (!folios->nr && !ret && (fgp_flags & FGP_CREAT))
 137                 ret = -ENOMEM;
 138
 139         return folios->nr ? 0 : ret;
 140 }
 141
 142 struct nocow_flush {
 143         struct closure  *cl;
 144         struct bch_dev  *ca;
 145         struct bio      bio;
 146 };
 147
 148 static void nocow_flush_endio(struct bio *_bio)
 149 {
 150
 151         struct nocow_flush *bio = container_of(_bio, struct nocow_flush, bio);
 152
 153         closure_put(bio->cl);
 154         percpu_ref_put(&bio->ca->io_ref);
 155         bio_put(&bio->bio);
 156 }
 157
 158 static void bch2_inode_flush_nocow_writes_async(struct bch_fs *c,
 159                                                 struct bch_inode_info *inode,
 160                                                 struct closure *cl)
 161 {
 162         struct nocow_flush *bio;
 163         struct bch_dev *ca;
 164         struct bch_devs_mask devs;
 165         unsigned dev;
 166
 167         dev = find_first_bit(inode->ei_devs_need_flush.d, BCH_SB_MEMBERS_MAX);
 168         if (dev == BCH_SB_MEMBERS_MAX)
 169                 return;
 170
 171         devs = inode->ei_devs_need_flush;
 172         memset(&inode->ei_devs_need_flush, 0, sizeof(inode->ei_devs_need_flush));
 173
 174         for_each_set_bit(dev, devs.d, BCH_SB_MEMBERS_MAX) {
 175                 rcu_read_lock();
 176                 ca = rcu_dereference(c->devs[dev]);
 177                 if (ca && !percpu_ref_tryget(&ca->io_ref))
 178                         ca = NULL;
 179                 rcu_read_unlock();
 180
 181                 if (!ca)
 182                         continue;
 183
 184                 bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev, 0,
 185                                                     REQ_OP_FLUSH,
 186                                                     GFP_KERNEL,
 187                                                     &c->nocow_flush_bioset),
 188                                    struct nocow_flush, bio);
 189                 bio->cl                 = cl;
 190                 bio->ca                 = ca;
 191                 bio->bio.bi_end_io      = nocow_flush_endio;
 192                 closure_bio_submit(&bio->bio, cl);
 193         }
 194 }
 195
 196 static int bch2_inode_flush_nocow_writes(struct bch_fs *c,
 197                                          struct bch_inode_info *inode)
 198 {
 199         struct closure cl;
 200
 201         closure_init_stack(&cl);
 202         bch2_inode_flush_nocow_writes_async(c, inode, &cl);
 203         closure_sync(&cl);
 204
 205         return 0;
 206 }
 207
 208 static inline bool bio_full(struct bio *bio, unsigned len)
 209 {
 210         if (bio->bi_vcnt >= bio->bi_max_vecs)
 211                 return true;
 212         if (bio->bi_iter.bi_size > UINT_MAX - len)
 213                 return true;
 214         return false;
 215 }
 216
 217 static inline struct address_space *faults_disabled_mapping(void)
 218 {
 219         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
 220 }
 221
 222 static inline void set_fdm_dropped_locks(void)
 223 {
 224         current->faults_disabled_mapping =
 225                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
 226 }
 227
 228 static inline bool fdm_dropped_locks(void)
 229 {
 230         return ((unsigned long) current->faults_disabled_mapping) & 1;
 231 }
 232
 233 struct quota_res {
 234         u64                             sectors;
 235 };
 236
 237 struct bch_writepage_io {
 238         struct bch_inode_info           *inode;
 239
 240         /* must be last: */
 241         struct bch_write_op             op;
 242 };
 243
 244 struct dio_write {
 245         struct kiocb                    *req;
 246         struct address_space            *mapping;
 247         struct bch_inode_info           *inode;
 248         struct mm_struct                *mm;
 249         unsigned                        loop:1,
 250                                         extending:1,
 251                                         sync:1,
 252                                         flush:1,
 253                                         free_iov:1;
 254         struct quota_res                quota_res;
 255         u64                             written;
 256
 257         struct iov_iter                 iter;
 258         struct iovec                    inline_vecs[2];
 259
 260         /* must be last: */
 261         struct bch_write_op             op;
 262 };
 263
 264 struct dio_read {
 265         struct closure                  cl;
 266         struct kiocb                    *req;
 267         long                            ret;
 268         bool                            should_dirty;
 269         struct bch_read_bio             rbio;
 270 };
 271
 272 /* pagecache_block must be held */
 273 static noinline int write_invalidate_inode_pages_range(struct address_space *mapping,
 274                                               loff_t start, loff_t end)
 275 {
 276         int ret;
 277
 278         /*
 279          * XXX: the way this is currently implemented, we can spin if a process
 280          * is continually redirtying a specific page
 281          */
 282         do {
 283                 if (!mapping->nrpages)
 284                         return 0;
 285
 286                 ret = filemap_write_and_wait_range(mapping, start, end);
 287                 if (ret)
 288                         break;
 289
 290                 if (!mapping->nrpages)
 291                         return 0;
 292
 293                 ret = invalidate_inode_pages2_range(mapping,
 294                                 start >> PAGE_SHIFT,
 295                                 end >> PAGE_SHIFT);
 296         } while (ret == -EBUSY);
 297
 298         return ret;
 299 }
 300
 301 /* quotas */
 302
 303 #ifdef CONFIG_BCACHEFS_QUOTA
 304
 305 static void __bch2_quota_reservation_put(struct bch_fs *c,
 306                                          struct bch_inode_info *inode,
 307                                          struct quota_res *res)
 308 {
 309         BUG_ON(res->sectors > inode->ei_quota_reserved);
 310
 311         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 312                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 313         inode->ei_quota_reserved -= res->sectors;
 314         res->sectors = 0;
 315 }
 316
 317 static void bch2_quota_reservation_put(struct bch_fs *c,
 318                                        struct bch_inode_info *inode,
 319                                        struct quota_res *res)
 320 {
 321         if (res->sectors) {
 322                 mutex_lock(&inode->ei_quota_lock);
 323                 __bch2_quota_reservation_put(c, inode, res);
 324                 mutex_unlock(&inode->ei_quota_lock);
 325         }
 326 }
 327
 328 static int bch2_quota_reservation_add(struct bch_fs *c,
 329                                       struct bch_inode_info *inode,
 330                                       struct quota_res *res,
 331                                       u64 sectors,
 332                                       bool check_enospc)
 333 {
 334         int ret;
 335
 336         if (test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags))
 337                 return 0;
 338
 339         mutex_lock(&inode->ei_quota_lock);
 340         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 341                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 342         if (likely(!ret)) {
 343                 inode->ei_quota_reserved += sectors;
 344                 res->sectors += sectors;
 345         }
 346         mutex_unlock(&inode->ei_quota_lock);
 347
 348         return ret;
 349 }
 350
 351 #else
 352
 353 static void __bch2_quota_reservation_put(struct bch_fs *c,
 354                                          struct bch_inode_info *inode,
 355                                          struct quota_res *res) {}
 356
 357 static void bch2_quota_reservation_put(struct bch_fs *c,
 358                                        struct bch_inode_info *inode,
 359                                        struct quota_res *res) {}
 360
 361 static int bch2_quota_reservation_add(struct bch_fs *c,
 362                                       struct bch_inode_info *inode,
 363                                       struct quota_res *res,
 364                                       unsigned sectors,
 365                                       bool check_enospc)
 366 {
 367         return 0;
 368 }
 369
 370 #endif
 371
 372 /* i_size updates: */
 373
 374 struct inode_new_size {
 375         loff_t          new_size;
 376         u64             now;
 377         unsigned        fields;
 378 };
 379
 380 static int inode_set_size(struct bch_inode_info *inode,
 381                           struct bch_inode_unpacked *bi,
 382                           void *p)
 383 {
 384         struct inode_new_size *s = p;
 385
 386         bi->bi_size = s->new_size;
 387         if (s->fields & ATTR_ATIME)
 388                 bi->bi_atime = s->now;
 389         if (s->fields & ATTR_MTIME)
 390                 bi->bi_mtime = s->now;
 391         if (s->fields & ATTR_CTIME)
 392                 bi->bi_ctime = s->now;
 393
 394         return 0;
 395 }
 396
 397 int __must_check bch2_write_inode_size(struct bch_fs *c,
 398                                        struct bch_inode_info *inode,
 399                                        loff_t new_size, unsigned fields)
 400 {
 401         struct inode_new_size s = {
 402                 .new_size       = new_size,
 403                 .now            = bch2_current_time(c),
 404                 .fields         = fields,
 405         };
 406
 407         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 408 }
 409
 410 static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 411                            struct quota_res *quota_res, s64 sectors)
 412 {
 413         bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
 414                                 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
 415                                 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
 416                                 inode->ei_inode.bi_sectors);
 417         inode->v.i_blocks += sectors;
 418
 419 #ifdef CONFIG_BCACHEFS_QUOTA
 420         if (quota_res &&
 421             !test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags) &&
 422             sectors > 0) {
 423                 BUG_ON(sectors > quota_res->sectors);
 424                 BUG_ON(sectors > inode->ei_quota_reserved);
 425
 426                 quota_res->sectors -= sectors;
 427                 inode->ei_quota_reserved -= sectors;
 428         } else {
 429                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 430         }
 431 #endif
 432 }
 433
 434 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 435                            struct quota_res *quota_res, s64 sectors)
 436 {
 437         if (sectors) {
 438                 mutex_lock(&inode->ei_quota_lock);
 439                 __i_sectors_acct(c, inode, quota_res, sectors);
 440                 mutex_unlock(&inode->ei_quota_lock);
 441         }
 442 }
 443
 444 /* page state: */
 445
 446 /* stored in page->private: */
 447
 448 #define BCH_FOLIO_SECTOR_STATE()        \
 449         x(unallocated)                  \
 450         x(reserved)                     \
 451         x(dirty)                        \
 452         x(dirty_reserved)               \
 453         x(allocated)
 454
 455 enum bch_folio_sector_state {
 456 #define x(n)    SECTOR_##n,
 457         BCH_FOLIO_SECTOR_STATE()
 458 #undef x
 459 };
 460
 461 const char * const bch2_folio_sector_states[] = {
 462 #define x(n)    #n,
 463         BCH_FOLIO_SECTOR_STATE()
 464 #undef x
 465         NULL
 466 };
 467
 468 static inline enum bch_folio_sector_state
 469 folio_sector_dirty(enum bch_folio_sector_state state)
 470 {
 471         switch (state) {
 472         case SECTOR_unallocated:
 473                 return SECTOR_dirty;
 474         case SECTOR_reserved:
 475                 return SECTOR_dirty_reserved;
 476         default:
 477                 return state;
 478         }
 479 }
 480
 481 static inline enum bch_folio_sector_state
 482 folio_sector_undirty(enum bch_folio_sector_state state)
 483 {
 484         switch (state) {
 485         case SECTOR_dirty:
 486                 return SECTOR_unallocated;
 487         case SECTOR_dirty_reserved:
 488                 return SECTOR_reserved;
 489         default:
 490                 return state;
 491         }
 492 }
 493
 494 static inline enum bch_folio_sector_state
 495 folio_sector_reserve(enum bch_folio_sector_state state)
 496 {
 497         switch (state) {
 498         case SECTOR_unallocated:
 499                 return SECTOR_reserved;
 500         case SECTOR_dirty:
 501                 return SECTOR_dirty_reserved;
 502         default:
 503                 return state;
 504         }
 505 }
 506
 507 struct bch_folio_sector {
 508         /* Uncompressed, fully allocated replicas (or on disk reservation): */
 509         unsigned                nr_replicas:4;
 510
 511         /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
 512         unsigned                replicas_reserved:4;
 513
 514         /* i_sectors: */
 515         enum bch_folio_sector_state state:8;
 516 };
 517
 518 struct bch_folio {
 519         spinlock_t              lock;
 520         atomic_t                write_count;
 521         /*
 522          * Is the sector state up to date with the btree?
 523          * (Not the data itself)
 524          */
 525         bool                    uptodate;
 526         struct bch_folio_sector s[];
 527 };
 528
 529 static inline void folio_sector_set(struct folio *folio,
 530                              struct bch_folio *s,
 531                              unsigned i, unsigned n)
 532 {
 533         s->s[i].state = n;
 534 }
 535
 536 /* file offset (to folio offset) to bch_folio_sector index */
 537 static inline int folio_pos_to_s(struct folio *folio, loff_t pos)
 538 {
 539         u64 f_offset = pos - folio_pos(folio);
 540         BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio));
 541         return f_offset >> SECTOR_SHIFT;
 542 }
 543
 544 static inline struct bch_folio *__bch2_folio(struct folio *folio)
 545 {
 546         return folio_has_private(folio)
 547                 ? (struct bch_folio *) folio_get_private(folio)
 548                 : NULL;
 549 }
 550
 551 static inline struct bch_folio *bch2_folio(struct folio *folio)
 552 {
 553         EBUG_ON(!folio_test_locked(folio));
 554
 555         return __bch2_folio(folio);
 556 }
 557
 558 /* for newly allocated folios: */
 559 static void __bch2_folio_release(struct folio *folio)
 560 {
 561         kfree(folio_detach_private(folio));
 562 }
 563
 564 static void bch2_folio_release(struct folio *folio)
 565 {
 566         EBUG_ON(!folio_test_locked(folio));
 567         __bch2_folio_release(folio);
 568 }
 569
 570 /* for newly allocated folios: */
 571 static struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
 572 {
 573         struct bch_folio *s;
 574
 575         s = kzalloc(sizeof(*s) +
 576                     sizeof(struct bch_folio_sector) *
 577                     folio_sectors(folio), gfp);
 578         if (!s)
 579                 return NULL;
 580
 581         spin_lock_init(&s->lock);
 582         folio_attach_private(folio, s);
 583         return s;
 584 }
 585
 586 static struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
 587 {
 588         return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
 589 }
 590
 591 static unsigned bkey_to_sector_state(struct bkey_s_c k)
 592 {
 593         if (bkey_extent_is_reservation(k))
 594                 return SECTOR_reserved;
 595         if (bkey_extent_is_allocation(k.k))
 596                 return SECTOR_allocated;
 597         return SECTOR_unallocated;
 598 }
 599
 600 static void __bch2_folio_set(struct folio *folio,
 601                              unsigned pg_offset, unsigned pg_len,
 602                              unsigned nr_ptrs, unsigned state)
 603 {
 604         struct bch_folio *s = bch2_folio(folio);
 605         unsigned i, sectors = folio_sectors(folio);
 606
 607         BUG_ON(pg_offset >= sectors);
 608         BUG_ON(pg_offset + pg_len > sectors);
 609
 610         spin_lock(&s->lock);
 611
 612         for (i = pg_offset; i < pg_offset + pg_len; i++) {
 613                 s->s[i].nr_replicas     = nr_ptrs;
 614                 folio_sector_set(folio, s, i, state);
 615         }
 616
 617         if (i == sectors)
 618                 s->uptodate = true;
 619
 620         spin_unlock(&s->lock);
 621 }
 622
 623 /*
 624  * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
 625  * extents btree:
 626  */
 627 static int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
 628                           struct folio **folios, unsigned nr_folios)
 629 {
 630         struct btree_trans trans;
 631         struct btree_iter iter;
 632         struct bkey_s_c k;
 633         struct bch_folio *s;
 634         u64 offset = folio_sector(folios[0]);
 635         unsigned folio_idx;
 636         u32 snapshot;
 637         bool need_set = false;
 638         int ret;
 639
 640         for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) {
 641                 s = bch2_folio_create(folios[folio_idx], GFP_KERNEL);
 642                 if (!s)
 643                         return -ENOMEM;
 644
 645                 need_set |= !s->uptodate;
 646         }
 647
 648         if (!need_set)
 649                 return 0;
 650
 651         folio_idx = 0;
 652         bch2_trans_init(&trans, c, 0, 0);
 653 retry:
 654         bch2_trans_begin(&trans);
 655
 656         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 657         if (ret)
 658                 goto err;
 659
 660         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
 661                            SPOS(inum.inum, offset, snapshot),
 662                            BTREE_ITER_SLOTS, k, ret) {
 663                 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
 664                 unsigned state = bkey_to_sector_state(k);
 665
 666                 while (folio_idx < nr_folios) {
 667                         struct folio *folio = folios[folio_idx];
 668                         u64 folio_start = folio_sector(folio);
 669                         u64 folio_end   = folio_end_sector(folio);
 670                         unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start;
 671                         unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start;
 672
 673                         BUG_ON(k.k->p.offset < folio_start);
 674                         BUG_ON(bkey_start_offset(k.k) > folio_end);
 675
 676                         if (!bch2_folio(folio)->uptodate)
 677                                 __bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
 678
 679                         if (k.k->p.offset < folio_end)
 680                                 break;
 681                         folio_idx++;
 682                 }
 683
 684                 if (folio_idx == nr_folios)
 685                         break;
 686         }
 687
 688         offset = iter.pos.offset;
 689         bch2_trans_iter_exit(&trans, &iter);
 690 err:
 691         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 692                 goto retry;
 693         bch2_trans_exit(&trans);
 694
 695         return ret;
 696 }
 697
 698 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
 699 {
 700         struct bvec_iter iter;
 701         struct folio_vec fv;
 702         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 703                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 704         unsigned state = bkey_to_sector_state(k);
 705
 706         bio_for_each_folio(fv, bio, iter)
 707                 __bch2_folio_set(fv.fv_folio,
 708                                  fv.fv_offset >> 9,
 709                                  fv.fv_len >> 9,
 710                                  nr_ptrs, state);
 711 }
 712
 713 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
 714                                        u64 start, u64 end)
 715 {
 716         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 717         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 718         struct folio_batch fbatch;
 719         unsigned i, j;
 720
 721         if (end <= start)
 722                 return;
 723
 724         folio_batch_init(&fbatch);
 725
 726         while (filemap_get_folios(inode->v.i_mapping,
 727                                   &index, end_index, &fbatch)) {
 728                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 729                         struct folio *folio = fbatch.folios[i];
 730                         u64 folio_start = folio_sector(folio);
 731                         u64 folio_end = folio_end_sector(folio);
 732                         unsigned folio_offset = max(start, folio_start) - folio_start;
 733                         unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
 734                         struct bch_folio *s;
 735
 736                         BUG_ON(end <= folio_start);
 737
 738                         folio_lock(folio);
 739                         s = bch2_folio(folio);
 740
 741                         if (s) {
 742                                 spin_lock(&s->lock);
 743                                 for (j = folio_offset; j < folio_offset + folio_len; j++)
 744                                         s->s[j].nr_replicas = 0;
 745                                 spin_unlock(&s->lock);
 746                         }
 747
 748                         folio_unlock(folio);
 749                 }
 750                 folio_batch_release(&fbatch);
 751                 cond_resched();
 752         }
 753 }
 754
 755 static void mark_pagecache_reserved(struct bch_inode_info *inode,
 756                                     u64 start, u64 end)
 757 {
 758         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 759         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 760         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 761         struct folio_batch fbatch;
 762         s64 i_sectors_delta = 0;
 763         unsigned i, j;
 764
 765         if (end <= start)
 766                 return;
 767
 768         folio_batch_init(&fbatch);
 769
 770         while (filemap_get_folios(inode->v.i_mapping,
 771                                   &index, end_index, &fbatch)) {
 772                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 773                         struct folio *folio = fbatch.folios[i];
 774                         u64 folio_start = folio_sector(folio);
 775                         u64 folio_end = folio_end_sector(folio);
 776                         unsigned folio_offset = max(start, folio_start) - folio_start;
 777                         unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
 778                         struct bch_folio *s;
 779
 780                         BUG_ON(end <= folio_start);
 781
 782                         folio_lock(folio);
 783                         s = bch2_folio(folio);
 784
 785                         if (s) {
 786                                 spin_lock(&s->lock);
 787                                 for (j = folio_offset; j < folio_offset + folio_len; j++) {
 788                                         i_sectors_delta -= s->s[j].state == SECTOR_dirty;
 789                                         folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state));
 790                                 }
 791                                 spin_unlock(&s->lock);
 792                         }
 793
 794                         folio_unlock(folio);
 795                 }
 796                 folio_batch_release(&fbatch);
 797                 cond_resched();
 798         }
 799
 800         i_sectors_acct(c, inode, NULL, i_sectors_delta);
 801 }
 802
 803 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 804 {
 805         /* XXX: this should not be open coded */
 806         return inode->ei_inode.bi_data_replicas
 807                 ? inode->ei_inode.bi_data_replicas - 1
 808                 : c->opts.data_replicas;
 809 }
 810
 811 static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
 812                                           unsigned nr_replicas)
 813 {
 814         return max(0, (int) nr_replicas -
 815                    s->nr_replicas -
 816                    s->replicas_reserved);
 817 }
 818
 819 static int bch2_get_folio_disk_reservation(struct bch_fs *c,
 820                                 struct bch_inode_info *inode,
 821                                 struct folio *folio, bool check_enospc)
 822 {
 823         struct bch_folio *s = bch2_folio_create(folio, 0);
 824         unsigned nr_replicas = inode_nr_replicas(c, inode);
 825         struct disk_reservation disk_res = { 0 };
 826         unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
 827         int ret;
 828
 829         if (!s)
 830                 return -ENOMEM;
 831
 832         for (i = 0; i < sectors; i++)
 833                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 834
 835         if (!disk_res_sectors)
 836                 return 0;
 837
 838         ret = bch2_disk_reservation_get(c, &disk_res,
 839                                         disk_res_sectors, 1,
 840                                         !check_enospc
 841                                         ? BCH_DISK_RESERVATION_NOFAIL
 842                                         : 0);
 843         if (unlikely(ret))
 844                 return ret;
 845
 846         for (i = 0; i < sectors; i++)
 847                 s->s[i].replicas_reserved +=
 848                         sectors_to_reserve(&s->s[i], nr_replicas);
 849
 850         return 0;
 851 }
 852
 853 struct bch2_folio_reservation {
 854         struct disk_reservation disk;
 855         struct quota_res        quota;
 856 };
 857
 858 static void bch2_folio_reservation_init(struct bch_fs *c,
 859                         struct bch_inode_info *inode,
 860                         struct bch2_folio_reservation *res)
 861 {
 862         memset(res, 0, sizeof(*res));
 863
 864         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 865 }
 866
 867 static void bch2_folio_reservation_put(struct bch_fs *c,
 868                         struct bch_inode_info *inode,
 869                         struct bch2_folio_reservation *res)
 870 {
 871         bch2_disk_reservation_put(c, &res->disk);
 872         bch2_quota_reservation_put(c, inode, &res->quota);
 873 }
 874
 875 static int bch2_folio_reservation_get(struct bch_fs *c,
 876                         struct bch_inode_info *inode,
 877                         struct folio *folio,
 878                         struct bch2_folio_reservation *res,
 879                         unsigned offset, unsigned len)
 880 {
 881         struct bch_folio *s = bch2_folio_create(folio, 0);
 882         unsigned i, disk_sectors = 0, quota_sectors = 0;
 883         int ret;
 884
 885         if (!s)
 886                 return -ENOMEM;
 887
 888         BUG_ON(!s->uptodate);
 889
 890         for (i = round_down(offset, block_bytes(c)) >> 9;
 891              i < round_up(offset + len, block_bytes(c)) >> 9;
 892              i++) {
 893                 disk_sectors += sectors_to_reserve(&s->s[i],
 894                                                 res->disk.nr_replicas);
 895                 quota_sectors += s->s[i].state == SECTOR_unallocated;
 896         }
 897
 898         if (disk_sectors) {
 899                 ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
 900                 if (unlikely(ret))
 901                         return ret;
 902         }
 903
 904         if (quota_sectors) {
 905                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 906                                                  quota_sectors, true);
 907                 if (unlikely(ret)) {
 908                         struct disk_reservation tmp = {
 909                                 .sectors = disk_sectors
 910                         };
 911
 912                         bch2_disk_reservation_put(c, &tmp);
 913                         res->disk.sectors -= disk_sectors;
 914                         return ret;
 915                 }
 916         }
 917
 918         return 0;
 919 }
 920
 921 static void bch2_clear_folio_bits(struct folio *folio)
 922 {
 923         struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
 924         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 925         struct bch_folio *s = bch2_folio(folio);
 926         struct disk_reservation disk_res = { 0 };
 927         int i, sectors = folio_sectors(folio), dirty_sectors = 0;
 928
 929         if (!s)
 930                 return;
 931
 932         EBUG_ON(!folio_test_locked(folio));
 933         EBUG_ON(folio_test_writeback(folio));
 934
 935         for (i = 0; i < sectors; i++) {
 936                 disk_res.sectors += s->s[i].replicas_reserved;
 937                 s->s[i].replicas_reserved = 0;
 938
 939                 dirty_sectors -= s->s[i].state == SECTOR_dirty;
 940                 folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
 941         }
 942
 943         bch2_disk_reservation_put(c, &disk_res);
 944
 945         i_sectors_acct(c, inode, NULL, dirty_sectors);
 946
 947         bch2_folio_release(folio);
 948 }
 949
 950 static void bch2_set_folio_dirty(struct bch_fs *c,
 951                         struct bch_inode_info *inode,
 952                         struct folio *folio,
 953                         struct bch2_folio_reservation *res,
 954                         unsigned offset, unsigned len)
 955 {
 956         struct bch_folio *s = bch2_folio(folio);
 957         unsigned i, dirty_sectors = 0;
 958
 959         WARN_ON((u64) folio_pos(folio) + offset + len >
 960                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 961
 962         BUG_ON(!s->uptodate);
 963
 964         spin_lock(&s->lock);
 965
 966         for (i = round_down(offset, block_bytes(c)) >> 9;
 967              i < round_up(offset + len, block_bytes(c)) >> 9;
 968              i++) {
 969                 unsigned sectors = sectors_to_reserve(&s->s[i],
 970                                                 res->disk.nr_replicas);
 971
 972                 /*
 973                  * This can happen if we race with the error path in
 974                  * bch2_writepage_io_done():
 975                  */
 976                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 977
 978                 s->s[i].replicas_reserved += sectors;
 979                 res->disk.sectors -= sectors;
 980
 981                 dirty_sectors += s->s[i].state == SECTOR_unallocated;
 982
 983                 folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
 984         }
 985
 986         spin_unlock(&s->lock);
 987
 988         i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 989
 990         if (!folio_test_dirty(folio))
 991                 filemap_dirty_folio(inode->v.i_mapping, folio);
 992 }
 993
 994 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 995 {
 996         struct file *file = vmf->vma->vm_file;
 997         struct address_space *mapping = file->f_mapping;
 998         struct address_space *fdm = faults_disabled_mapping();
 999         struct bch_inode_info *inode = file_bch_inode(file);
1000         int ret;
1001
1002         if (fdm == mapping)
1003                 return VM_FAULT_SIGBUS;
1004
1005         /* Lock ordering: */
1006         if (fdm > mapping) {
1007                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
1008
1009                 if (bch2_pagecache_add_tryget(inode))
1010                         goto got_lock;
1011
1012                 bch2_pagecache_block_put(fdm_host);
1013
1014                 bch2_pagecache_add_get(inode);
1015                 bch2_pagecache_add_put(inode);
1016
1017                 bch2_pagecache_block_get(fdm_host);
1018
1019                 /* Signal that lock has been dropped: */
1020                 set_fdm_dropped_locks();
1021                 return VM_FAULT_SIGBUS;
1022         }
1023
1024         bch2_pagecache_add_get(inode);
1025 got_lock:
1026         ret = filemap_fault(vmf);
1027         bch2_pagecache_add_put(inode);
1028
1029         return ret;
1030 }
1031
1032 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
1033 {
1034         struct folio *folio = page_folio(vmf->page);
1035         struct file *file = vmf->vma->vm_file;
1036         struct bch_inode_info *inode = file_bch_inode(file);
1037         struct address_space *mapping = file->f_mapping;
1038         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1039         struct bch2_folio_reservation res;
1040         unsigned len;
1041         loff_t isize;
1042         int ret;
1043
1044         bch2_folio_reservation_init(c, inode, &res);
1045
1046         sb_start_pagefault(inode->v.i_sb);
1047         file_update_time(file);
1048
1049         /*
1050          * Not strictly necessary, but helps avoid dio writes livelocking in
1051          * write_invalidate_inode_pages_range() - can drop this if/when we get
1052          * a write_invalidate_inode_pages_range() that works without dropping
1053          * page lock before invalidating page
1054          */
1055         bch2_pagecache_add_get(inode);
1056
1057         folio_lock(folio);
1058         isize = i_size_read(&inode->v);
1059
1060         if (folio->mapping != mapping || folio_pos(folio) >= isize) {
1061                 folio_unlock(folio);
1062                 ret = VM_FAULT_NOPAGE;
1063                 goto out;
1064         }
1065
1066         len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
1067
1068         if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
1069             bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
1070                 folio_unlock(folio);
1071                 ret = VM_FAULT_SIGBUS;
1072                 goto out;
1073         }
1074
1075         bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
1076         bch2_folio_reservation_put(c, inode, &res);
1077
1078         folio_wait_stable(folio);
1079         ret = VM_FAULT_LOCKED;
1080 out:
1081         bch2_pagecache_add_put(inode);
1082         sb_end_pagefault(inode->v.i_sb);
1083
1084         return ret;
1085 }
1086
1087 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1088 {
1089         if (offset || length < folio_size(folio))
1090                 return;
1091
1092         bch2_clear_folio_bits(folio);
1093 }
1094
1095 bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
1096 {
1097         if (folio_test_dirty(folio) || folio_test_writeback(folio))
1098                 return false;
1099
1100         bch2_clear_folio_bits(folio);
1101         return true;
1102 }
1103
1104 /* readpage(s): */
1105
1106 static void bch2_readpages_end_io(struct bio *bio)
1107 {
1108         struct folio_iter fi;
1109
1110         bio_for_each_folio_all(fi, bio) {
1111                 if (!bio->bi_status) {
1112                         folio_mark_uptodate(fi.folio);
1113                 } else {
1114                         folio_clear_uptodate(fi.folio);
1115                         folio_set_error(fi.folio);
1116                 }
1117                 folio_unlock(fi.folio);
1118         }
1119
1120         bio_put(bio);
1121 }
1122
1123 struct readpages_iter {
1124         struct address_space    *mapping;
1125         unsigned                idx;
1126         folios                  folios;
1127 };
1128
1129 static int readpages_iter_init(struct readpages_iter *iter,
1130                                struct readahead_control *ractl)
1131 {
1132         struct folio **fi;
1133         int ret;
1134
1135         memset(iter, 0, sizeof(*iter));
1136
1137         iter->mapping = ractl->mapping;
1138
1139         ret = filemap_get_contig_folios_d(iter->mapping,
1140                                 ractl->_index << PAGE_SHIFT,
1141                                 (ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
1142                                 0, mapping_gfp_mask(iter->mapping),
1143                                 &iter->folios);
1144         if (ret)
1145                 return ret;
1146
1147         darray_for_each(iter->folios, fi) {
1148                 ractl->_nr_pages -= 1U << folio_order(*fi);
1149                 __bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL);
1150                 folio_put(*fi);
1151                 folio_put(*fi);
1152         }
1153
1154         return 0;
1155 }
1156
1157 static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
1158 {
1159         if (iter->idx >= iter->folios.nr)
1160                 return NULL;
1161         return iter->folios.data[iter->idx];
1162 }
1163
1164 static inline void readpage_iter_advance(struct readpages_iter *iter)
1165 {
1166         iter->idx++;
1167 }
1168
1169 static bool extent_partial_reads_expensive(struct bkey_s_c k)
1170 {
1171         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1172         struct bch_extent_crc_unpacked crc;
1173         const union bch_extent_entry *i;
1174
1175         bkey_for_each_crc(k.k, ptrs, crc, i)
1176                 if (crc.csum_type || crc.compression_type)
1177                         return true;
1178         return false;
1179 }
1180
1181 static int readpage_bio_extend(struct btree_trans *trans,
1182                                struct readpages_iter *iter,
1183                                struct bio *bio,
1184                                unsigned sectors_this_extent,
1185                                bool get_more)
1186 {
1187         /* Don't hold btree locks while allocating memory: */
1188         bch2_trans_unlock(trans);
1189
1190         while (bio_sectors(bio) < sectors_this_extent &&
1191                bio->bi_vcnt < bio->bi_max_vecs) {
1192                 struct folio *folio = readpage_iter_peek(iter);
1193                 int ret;
1194
1195                 if (folio) {
1196                         readpage_iter_advance(iter);
1197                 } else {
1198                         pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
1199
1200                         if (!get_more)
1201                                 break;
1202
1203                         folio = xa_load(&iter->mapping->i_pages, folio_offset);
1204                         if (folio && !xa_is_value(folio))
1205                                 break;
1206
1207                         folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
1208                         if (!folio)
1209                                 break;
1210
1211                         if (!__bch2_folio_create(folio, GFP_KERNEL)) {
1212                                 folio_put(folio);
1213                                 break;
1214                         }
1215
1216                         ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
1217                         if (ret) {
1218                                 __bch2_folio_release(folio);
1219                                 folio_put(folio);
1220                                 break;
1221                         }
1222
1223                         folio_put(folio);
1224                 }
1225
1226                 BUG_ON(folio_sector(folio) != bio_end_sector(bio));
1227
1228                 BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
1229         }
1230
1231         return bch2_trans_relock(trans);
1232 }
1233
1234 static void bchfs_read(struct btree_trans *trans,
1235                        struct bch_read_bio *rbio,
1236                        subvol_inum inum,
1237                        struct readpages_iter *readpages_iter)
1238 {
1239         struct bch_fs *c = trans->c;
1240         struct btree_iter iter;
1241         struct bkey_buf sk;
1242         int flags = BCH_READ_RETRY_IF_STALE|
1243                 BCH_READ_MAY_PROMOTE;
1244         u32 snapshot;
1245         int ret = 0;
1246
1247         rbio->c = c;
1248         rbio->start_time = local_clock();
1249         rbio->subvol = inum.subvol;
1250
1251         bch2_bkey_buf_init(&sk);
1252 retry:
1253         bch2_trans_begin(trans);
1254         iter = (struct btree_iter) { NULL };
1255
1256         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1257         if (ret)
1258                 goto err;
1259
1260         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1261                              SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1262                              BTREE_ITER_SLOTS);
1263         while (1) {
1264                 struct bkey_s_c k;
1265                 unsigned bytes, sectors, offset_into_extent;
1266                 enum btree_id data_btree = BTREE_ID_extents;
1267
1268                 /*
1269                  * read_extent -> io_time_reset may cause a transaction restart
1270                  * without returning an error, we need to check for that here:
1271                  */
1272                 ret = bch2_trans_relock(trans);
1273                 if (ret)
1274                         break;
1275
1276                 bch2_btree_iter_set_pos(&iter,
1277                                 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1278
1279                 k = bch2_btree_iter_peek_slot(&iter);
1280                 ret = bkey_err(k);
1281                 if (ret)
1282                         break;
1283
1284                 offset_into_extent = iter.pos.offset -
1285                         bkey_start_offset(k.k);
1286                 sectors = k.k->size - offset_into_extent;
1287
1288                 bch2_bkey_buf_reassemble(&sk, c, k);
1289
1290                 ret = bch2_read_indirect_extent(trans, &data_btree,
1291                                         &offset_into_extent, &sk);
1292                 if (ret)
1293                         break;
1294
1295                 k = bkey_i_to_s_c(sk.k);
1296
1297                 sectors = min(sectors, k.k->size - offset_into_extent);
1298
1299                 if (readpages_iter) {
1300                         ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
1301                                                   extent_partial_reads_expensive(k));
1302                         if (ret)
1303                                 break;
1304                 }
1305
1306                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1307                 swap(rbio->bio.bi_iter.bi_size, bytes);
1308
1309                 if (rbio->bio.bi_iter.bi_size == bytes)
1310                         flags |= BCH_READ_LAST_FRAGMENT;
1311
1312                 bch2_bio_page_state_set(&rbio->bio, k);
1313
1314                 bch2_read_extent(trans, rbio, iter.pos,
1315                                  data_btree, k, offset_into_extent, flags);
1316
1317                 if (flags & BCH_READ_LAST_FRAGMENT)
1318                         break;
1319
1320                 swap(rbio->bio.bi_iter.bi_size, bytes);
1321                 bio_advance(&rbio->bio, bytes);
1322
1323                 ret = btree_trans_too_many_iters(trans);
1324                 if (ret)
1325                         break;
1326         }
1327 err:
1328         bch2_trans_iter_exit(trans, &iter);
1329
1330         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1331                 goto retry;
1332
1333         if (ret) {
1334                 bch_err_inum_offset_ratelimited(c,
1335                                 iter.pos.inode,
1336                                 iter.pos.offset << 9,
1337                                 "read error %i from btree lookup", ret);
1338                 rbio->bio.bi_status = BLK_STS_IOERR;
1339                 bio_endio(&rbio->bio);
1340         }
1341
1342         bch2_bkey_buf_exit(&sk, c);
1343 }
1344
1345 void bch2_readahead(struct readahead_control *ractl)
1346 {
1347         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1348         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1349         struct bch_io_opts opts;
1350         struct btree_trans trans;
1351         struct folio *folio;
1352         struct readpages_iter readpages_iter;
1353         int ret;
1354
1355         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1356
1357         ret = readpages_iter_init(&readpages_iter, ractl);
1358         BUG_ON(ret);
1359
1360         bch2_trans_init(&trans, c, 0, 0);
1361
1362         bch2_pagecache_add_get(inode);
1363
1364         while ((folio = readpage_iter_peek(&readpages_iter))) {
1365                 unsigned n = min_t(unsigned,
1366                                    readpages_iter.folios.nr -
1367                                    readpages_iter.idx,
1368                                    BIO_MAX_VECS);
1369                 struct bch_read_bio *rbio =
1370                         rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
1371                                                    GFP_KERNEL, &c->bio_read),
1372                                   opts);
1373
1374                 readpage_iter_advance(&readpages_iter);
1375
1376                 rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1377                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1378                 BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1379
1380                 bchfs_read(&trans, rbio, inode_inum(inode),
1381                            &readpages_iter);
1382                 bch2_trans_unlock(&trans);
1383         }
1384
1385         bch2_pagecache_add_put(inode);
1386
1387         bch2_trans_exit(&trans);
1388         darray_exit(&readpages_iter.folios);
1389 }
1390
1391 static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
1392                              subvol_inum inum, struct folio *folio)
1393 {
1394         struct btree_trans trans;
1395
1396         bch2_folio_create(folio, __GFP_NOFAIL);
1397
1398         rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
1399         rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1400         BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1401
1402         bch2_trans_init(&trans, c, 0, 0);
1403         bchfs_read(&trans, rbio, inum, NULL);
1404         bch2_trans_exit(&trans);
1405 }
1406
1407 static void bch2_read_single_folio_end_io(struct bio *bio)
1408 {
1409         complete(bio->bi_private);
1410 }
1411
1412 static int bch2_read_single_folio(struct folio *folio,
1413                                   struct address_space *mapping)
1414 {
1415         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1416         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1417         struct bch_read_bio *rbio;
1418         struct bch_io_opts opts;
1419         int ret;
1420         DECLARE_COMPLETION_ONSTACK(done);
1421
1422         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1423
1424         rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
1425                          opts);
1426         rbio->bio.bi_private = &done;
1427         rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
1428
1429         __bchfs_readfolio(c, rbio, inode_inum(inode), folio);
1430         wait_for_completion(&done);
1431
1432         ret = blk_status_to_errno(rbio->bio.bi_status);
1433         bio_put(&rbio->bio);
1434
1435         if (ret < 0)
1436                 return ret;
1437
1438         folio_mark_uptodate(folio);
1439         return 0;
1440 }
1441
1442 int bch2_read_folio(struct file *file, struct folio *folio)
1443 {
1444         int ret;
1445
1446         ret = bch2_read_single_folio(folio, folio->mapping);
1447         folio_unlock(folio);
1448         return bch2_err_class(ret);
1449 }
1450
1451 /* writepages: */
1452
1453 struct bch_writepage_state {
1454         struct bch_writepage_io *io;
1455         struct bch_io_opts      opts;
1456         struct bch_folio_sector *tmp;
1457         unsigned                tmp_sectors;
1458 };
1459
1460 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1461                                                                   struct bch_inode_info *inode)
1462 {
1463         struct bch_writepage_state ret = { 0 };
1464
1465         bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
1466         return ret;
1467 }
1468
1469 static void bch2_writepage_io_done(struct bch_write_op *op)
1470 {
1471         struct bch_writepage_io *io =
1472                 container_of(op, struct bch_writepage_io, op);
1473         struct bch_fs *c = io->op.c;
1474         struct bio *bio = &io->op.wbio.bio;
1475         struct folio_iter fi;
1476         unsigned i;
1477
1478         if (io->op.error) {
1479                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1480
1481                 bio_for_each_folio_all(fi, bio) {
1482                         struct bch_folio *s;
1483
1484                         folio_set_error(fi.folio);
1485                         mapping_set_error(fi.folio->mapping, -EIO);
1486
1487                         s = __bch2_folio(fi.folio);
1488                         spin_lock(&s->lock);
1489                         for (i = 0; i < folio_sectors(fi.folio); i++)
1490                                 s->s[i].nr_replicas = 0;
1491                         spin_unlock(&s->lock);
1492                 }
1493         }
1494
1495         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1496                 bio_for_each_folio_all(fi, bio) {
1497                         struct bch_folio *s;
1498
1499                         s = __bch2_folio(fi.folio);
1500                         spin_lock(&s->lock);
1501                         for (i = 0; i < folio_sectors(fi.folio); i++)
1502                                 s->s[i].nr_replicas = 0;
1503                         spin_unlock(&s->lock);
1504                 }
1505         }
1506
1507         /*
1508          * racing with fallocate can cause us to add fewer sectors than
1509          * expected - but we shouldn't add more sectors than expected:
1510          */
1511         WARN_ON_ONCE(io->op.i_sectors_delta > 0);
1512
1513         /*
1514          * (error (due to going RO) halfway through a page can screw that up
1515          * slightly)
1516          * XXX wtf?
1517            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1518          */
1519
1520         /*
1521          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1522          * before calling end_page_writeback:
1523          */
1524         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1525
1526         bio_for_each_folio_all(fi, bio) {
1527                 struct bch_folio *s = __bch2_folio(fi.folio);
1528
1529                 if (atomic_dec_and_test(&s->write_count))
1530                         folio_end_writeback(fi.folio);
1531         }
1532
1533         bio_put(&io->op.wbio.bio);
1534 }
1535
1536 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1537 {
1538         struct bch_writepage_io *io = w->io;
1539
1540         w->io = NULL;
1541         closure_call(&io->op.cl, bch2_write, NULL, NULL);
1542 }
1543
1544 /*
1545  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1546  * possible, else allocating a new one:
1547  */
1548 static void bch2_writepage_io_alloc(struct bch_fs *c,
1549                                     struct writeback_control *wbc,
1550                                     struct bch_writepage_state *w,
1551                                     struct bch_inode_info *inode,
1552                                     u64 sector,
1553                                     unsigned nr_replicas)
1554 {
1555         struct bch_write_op *op;
1556
1557         w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1558                                               REQ_OP_WRITE,
1559                                               GFP_KERNEL,
1560                                               &c->writepage_bioset),
1561                              struct bch_writepage_io, op.wbio.bio);
1562
1563         w->io->inode            = inode;
1564         op                      = &w->io->op;
1565         bch2_write_op_init(op, c, w->opts);
1566         op->target              = w->opts.foreground_target;
1567         op->nr_replicas         = nr_replicas;
1568         op->res.nr_replicas     = nr_replicas;
1569         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1570         op->subvol              = inode->ei_subvol;
1571         op->pos                 = POS(inode->v.i_ino, sector);
1572         op->end_io              = bch2_writepage_io_done;
1573         op->devs_need_flush     = &inode->ei_devs_need_flush;
1574         op->wbio.bio.bi_iter.bi_sector = sector;
1575         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1576 }
1577
1578 static int __bch2_writepage(struct folio *folio,
1579                             struct writeback_control *wbc,
1580                             void *data)
1581 {
1582         struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
1583         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1584         struct bch_writepage_state *w = data;
1585         struct bch_folio *s;
1586         unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
1587         loff_t i_size = i_size_read(&inode->v);
1588         int ret;
1589
1590         EBUG_ON(!folio_test_uptodate(folio));
1591
1592         /* Is the folio fully inside i_size? */
1593         if (folio_end_pos(folio) <= i_size)
1594                 goto do_io;
1595
1596         /* Is the folio fully outside i_size? (truncate in progress) */
1597         if (folio_pos(folio) >= i_size) {
1598                 folio_unlock(folio);
1599                 return 0;
1600         }
1601
1602         /*
1603          * The folio straddles i_size.  It must be zeroed out on each and every
1604          * writepage invocation because it may be mmapped.  "A file is mapped
1605          * in multiples of the folio size.  For a file that is not a multiple of
1606          * the  folio size, the remaining memory is zeroed when mapped, and
1607          * writes to that region are not written out to the file."
1608          */
1609         folio_zero_segment(folio,
1610                            i_size - folio_pos(folio),
1611                            folio_size(folio));
1612 do_io:
1613         f_sectors = folio_sectors(folio);
1614         s = bch2_folio(folio);
1615
1616         if (f_sectors > w->tmp_sectors) {
1617                 kfree(w->tmp);
1618                 w->tmp = kzalloc(sizeof(struct bch_folio_sector) *
1619                                  f_sectors, __GFP_NOFAIL);
1620                 w->tmp_sectors = f_sectors;
1621         }
1622
1623         /*
1624          * Things get really hairy with errors during writeback:
1625          */
1626         ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
1627         BUG_ON(ret);
1628
1629         /* Before unlocking the page, get copy of reservations: */
1630         spin_lock(&s->lock);
1631         memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
1632
1633         for (i = 0; i < f_sectors; i++) {
1634                 if (s->s[i].state < SECTOR_dirty)
1635                         continue;
1636
1637                 nr_replicas_this_write =
1638                         min_t(unsigned, nr_replicas_this_write,
1639                               s->s[i].nr_replicas +
1640                               s->s[i].replicas_reserved);
1641         }
1642
1643         for (i = 0; i < f_sectors; i++) {
1644                 if (s->s[i].state < SECTOR_dirty)
1645                         continue;
1646
1647                 s->s[i].nr_replicas = w->opts.compression
1648                         ? 0 : nr_replicas_this_write;
1649
1650                 s->s[i].replicas_reserved = 0;
1651                 folio_sector_set(folio, s, i, SECTOR_allocated);
1652         }
1653         spin_unlock(&s->lock);
1654
1655         BUG_ON(atomic_read(&s->write_count));
1656         atomic_set(&s->write_count, 1);
1657
1658         BUG_ON(folio_test_writeback(folio));
1659         folio_start_writeback(folio);
1660
1661         folio_unlock(folio);
1662
1663         offset = 0;
1664         while (1) {
1665                 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1666                 u64 sector;
1667
1668                 while (offset < f_sectors &&
1669                        w->tmp[offset].state < SECTOR_dirty)
1670                         offset++;
1671
1672                 if (offset == f_sectors)
1673                         break;
1674
1675                 while (offset + sectors < f_sectors &&
1676                        w->tmp[offset + sectors].state >= SECTOR_dirty) {
1677                         reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
1678                         dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
1679                         sectors++;
1680                 }
1681                 BUG_ON(!sectors);
1682
1683                 sector = folio_sector(folio) + offset;
1684
1685                 if (w->io &&
1686                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1687                      bio_full(&w->io->op.wbio.bio, sectors << 9) ||
1688                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1689                      (BIO_MAX_VECS * PAGE_SIZE) ||
1690                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1691                         bch2_writepage_do_io(w);
1692
1693                 if (!w->io)
1694                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1695                                                 nr_replicas_this_write);
1696
1697                 atomic_inc(&s->write_count);
1698
1699                 BUG_ON(inode != w->io->inode);
1700                 BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
1701                                      sectors << 9, offset << 9));
1702
1703                 /* Check for writing past i_size: */
1704                 WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1705                           round_up(i_size, block_bytes(c)) &&
1706                           !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
1707                           "writing past i_size: %llu > %llu (unrounded %llu)\n",
1708                           bio_end_sector(&w->io->op.wbio.bio) << 9,
1709                           round_up(i_size, block_bytes(c)),
1710                           i_size);
1711
1712                 w->io->op.res.sectors += reserved_sectors;
1713                 w->io->op.i_sectors_delta -= dirty_sectors;
1714                 w->io->op.new_i_size = i_size;
1715
1716                 offset += sectors;
1717         }
1718
1719         if (atomic_dec_and_test(&s->write_count))
1720                 folio_end_writeback(folio);
1721
1722         return 0;
1723 }
1724
1725 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1726 {
1727         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1728         struct bch_writepage_state w =
1729                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1730         struct blk_plug plug;
1731         int ret;
1732
1733         blk_start_plug(&plug);
1734         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1735         if (w.io)
1736                 bch2_writepage_do_io(&w);
1737         blk_finish_plug(&plug);
1738         kfree(w.tmp);
1739         return bch2_err_class(ret);
1740 }
1741
1742 /* buffered writes: */
1743
1744 int bch2_write_begin(struct file *file, struct address_space *mapping,
1745                      loff_t pos, unsigned len,
1746                      struct page **pagep, void **fsdata)
1747 {
1748         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1749         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1750         struct bch2_folio_reservation *res;
1751         struct folio *folio;
1752         unsigned offset;
1753         int ret = -ENOMEM;
1754
1755         res = kmalloc(sizeof(*res), GFP_KERNEL);
1756         if (!res)
1757                 return -ENOMEM;
1758
1759         bch2_folio_reservation_init(c, inode, res);
1760         *fsdata = res;
1761
1762         bch2_pagecache_add_get(inode);
1763
1764         folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
1765                                 FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
1766                                 mapping_gfp_mask(mapping));
1767         if (IS_ERR_OR_NULL(folio))
1768                 goto err_unlock;
1769
1770         if (folio_test_uptodate(folio))
1771                 goto out;
1772
1773         offset = pos - folio_pos(folio);
1774         len = min_t(size_t, len, folio_end_pos(folio) - pos);
1775
1776         /* If we're writing entire folio, don't need to read it in first: */
1777         if (!offset && len == folio_size(folio))
1778                 goto out;
1779
1780         if (!offset && pos + len >= inode->v.i_size) {
1781                 folio_zero_segment(folio, len, folio_size(folio));
1782                 flush_dcache_folio(folio);
1783                 goto out;
1784         }
1785
1786         if (folio_pos(folio) >= inode->v.i_size) {
1787                 folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
1788                 flush_dcache_folio(folio);
1789                 goto out;
1790         }
1791 readpage:
1792         ret = bch2_read_single_folio(folio, mapping);
1793         if (ret)
1794                 goto err;
1795 out:
1796         ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
1797         if (ret)
1798                 goto err;
1799
1800         ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
1801         if (ret) {
1802                 if (!folio_test_uptodate(folio)) {
1803                         /*
1804                          * If the folio hasn't been read in, we won't know if we
1805                          * actually need a reservation - we don't actually need
1806                          * to read here, we just need to check if the folio is
1807                          * fully backed by uncompressed data:
1808                          */
1809                         goto readpage;
1810                 }
1811
1812                 goto err;
1813         }
1814
1815         *pagep = &folio->page;
1816         return 0;
1817 err:
1818         folio_unlock(folio);
1819         folio_put(folio);
1820         *pagep = NULL;
1821 err_unlock:
1822         bch2_pagecache_add_put(inode);
1823         kfree(res);
1824         *fsdata = NULL;
1825         return bch2_err_class(ret);
1826 }
1827
1828 int bch2_write_end(struct file *file, struct address_space *mapping,
1829                    loff_t pos, unsigned len, unsigned copied,
1830                    struct page *page, void *fsdata)
1831 {
1832         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1833         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1834         struct bch2_folio_reservation *res = fsdata;
1835         struct folio *folio = page_folio(page);
1836         unsigned offset = pos - folio_pos(folio);
1837
1838         lockdep_assert_held(&inode->v.i_rwsem);
1839         BUG_ON(offset + copied > folio_size(folio));
1840
1841         if (unlikely(copied < len && !folio_test_uptodate(folio))) {
1842                 /*
1843                  * The folio needs to be read in, but that would destroy
1844                  * our partial write - simplest thing is to just force
1845                  * userspace to redo the write:
1846                  */
1847                 folio_zero_range(folio, 0, folio_size(folio));
1848                 flush_dcache_folio(folio);
1849                 copied = 0;
1850         }
1851
1852         spin_lock(&inode->v.i_lock);
1853         if (pos + copied > inode->v.i_size)
1854                 i_size_write(&inode->v, pos + copied);
1855         spin_unlock(&inode->v.i_lock);
1856
1857         if (copied) {
1858                 if (!folio_test_uptodate(folio))
1859                         folio_mark_uptodate(folio);
1860
1861                 bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
1862
1863                 inode->ei_last_dirtied = (unsigned long) current;
1864         }
1865
1866         folio_unlock(folio);
1867         folio_put(folio);
1868         bch2_pagecache_add_put(inode);
1869
1870         bch2_folio_reservation_put(c, inode, res);
1871         kfree(res);
1872
1873         return copied;
1874 }
1875
1876 static noinline void folios_trunc(folios *folios, struct folio **fi)
1877 {
1878         while (folios->data + folios->nr > fi) {
1879                 struct folio *f = darray_pop(folios);
1880
1881                 folio_unlock(f);
1882                 folio_put(f);
1883         }
1884 }
1885
1886 static int __bch2_buffered_write(struct bch_inode_info *inode,
1887                                  struct address_space *mapping,
1888                                  struct iov_iter *iter,
1889                                  loff_t pos, unsigned len)
1890 {
1891         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1892         struct bch2_folio_reservation res;
1893         folios folios;
1894         struct folio **fi, *f;
1895         unsigned copied = 0, f_offset;
1896         u64 end = pos + len, f_pos;
1897         loff_t last_folio_pos = inode->v.i_size;
1898         int ret = 0;
1899
1900         BUG_ON(!len);
1901
1902         bch2_folio_reservation_init(c, inode, &res);
1903         darray_init(&folios);
1904
1905         ret = filemap_get_contig_folios_d(mapping, pos, end,
1906                                    FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
1907                                    mapping_gfp_mask(mapping),
1908                                    &folios);
1909         if (ret)
1910                 goto out;
1911
1912         BUG_ON(!folios.nr);
1913
1914         f = darray_first(folios);
1915         if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
1916                 ret = bch2_read_single_folio(f, mapping);
1917                 if (ret)
1918                         goto out;
1919         }
1920
1921         f = darray_last(folios);
1922         end = min(end, folio_end_pos(f));
1923         last_folio_pos = folio_pos(f);
1924         if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
1925                 if (end >= inode->v.i_size) {
1926                         folio_zero_range(f, 0, folio_size(f));
1927                 } else {
1928                         ret = bch2_read_single_folio(f, mapping);
1929                         if (ret)
1930                                 goto out;
1931                 }
1932         }
1933
1934         ret = bch2_folio_set(c, inode_inum(inode), folios.data, folios.nr);
1935         if (ret)
1936                 goto out;
1937
1938         f_pos = pos;
1939         f_offset = pos - folio_pos(darray_first(folios));
1940         darray_for_each(folios, fi) {
1941                 struct folio *f = *fi;
1942                 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1943
1944                 /*
1945                  * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
1946                  * supposed to write as much as we have disk space for.
1947                  *
1948                  * On failure here we should still write out a partial page if
1949                  * we aren't completely out of disk space - we don't do that
1950                  * yet:
1951                  */
1952                 ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
1953                 if (unlikely(ret)) {
1954                         folios_trunc(&folios, fi);
1955                         if (!folios.nr)
1956                                 goto out;
1957
1958                         end = min(end, folio_end_pos(darray_last(folios)));
1959                         break;
1960                 }
1961
1962                 f_pos = folio_end_pos(f);
1963                 f_offset = 0;
1964         }
1965
1966         if (mapping_writably_mapped(mapping))
1967                 darray_for_each(folios, fi)
1968                         flush_dcache_folio(*fi);
1969
1970         f_pos = pos;
1971         f_offset = pos - folio_pos(darray_first(folios));
1972         darray_for_each(folios, fi) {
1973                 struct folio *f = *fi;
1974                 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1975                 unsigned f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter);
1976
1977                 if (!f_copied) {
1978                         folios_trunc(&folios, fi);
1979                         break;
1980                 }
1981
1982                 if (!folio_test_uptodate(f) &&
1983                     f_copied != folio_size(f) &&
1984                     pos + copied + f_copied < inode->v.i_size) {
1985                         folio_zero_range(f, 0, folio_size(f));
1986                         folios_trunc(&folios, fi);
1987                         break;
1988                 }
1989
1990                 flush_dcache_folio(f);
1991                 copied += f_copied;
1992
1993                 if (f_copied != f_len) {
1994                         folios_trunc(&folios, fi + 1);
1995                         break;
1996                 }
1997
1998                 f_pos = folio_end_pos(f);
1999                 f_offset = 0;
2000         }
2001
2002         if (!copied)
2003                 goto out;
2004
2005         end = pos + copied;
2006
2007         spin_lock(&inode->v.i_lock);
2008         if (end > inode->v.i_size)
2009                 i_size_write(&inode->v, end);
2010         spin_unlock(&inode->v.i_lock);
2011
2012         f_pos = pos;
2013         f_offset = pos - folio_pos(darray_first(folios));
2014         darray_for_each(folios, fi) {
2015                 struct folio *f = *fi;
2016                 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
2017
2018                 if (!folio_test_uptodate(f))
2019                         folio_mark_uptodate(f);
2020
2021                 bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
2022
2023                 f_pos = folio_end_pos(f);
2024                 f_offset = 0;
2025         }
2026
2027         inode->ei_last_dirtied = (unsigned long) current;
2028 out:
2029         darray_for_each(folios, fi) {
2030                 folio_unlock(*fi);
2031                 folio_put(*fi);
2032         }
2033
2034         /*
2035          * If the last folio added to the mapping starts beyond current EOF, we
2036          * performed a short write but left around at least one post-EOF folio.
2037          * Clean up the mapping before we return.
2038          */
2039         if (last_folio_pos >= inode->v.i_size)
2040                 truncate_pagecache(&inode->v, inode->v.i_size);
2041
2042         darray_exit(&folios);
2043         bch2_folio_reservation_put(c, inode, &res);
2044
2045         return copied ?: ret;
2046 }
2047
2048 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
2049 {
2050         struct file *file = iocb->ki_filp;
2051         struct address_space *mapping = file->f_mapping;
2052         struct bch_inode_info *inode = file_bch_inode(file);
2053         loff_t pos = iocb->ki_pos;
2054         ssize_t written = 0;
2055         int ret = 0;
2056
2057         bch2_pagecache_add_get(inode);
2058
2059         do {
2060                 unsigned offset = pos & (PAGE_SIZE - 1);
2061                 unsigned bytes = iov_iter_count(iter);
2062 again:
2063                 /*
2064                  * Bring in the user page that we will copy from _first_.
2065                  * Otherwise there's a nasty deadlock on copying from the
2066                  * same page as we're writing to, without it being marked
2067                  * up-to-date.
2068                  *
2069                  * Not only is this an optimisation, but it is also required
2070                  * to check that the address is actually valid, when atomic
2071                  * usercopies are used, below.
2072                  */
2073                 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2074                         bytes = min_t(unsigned long, iov_iter_count(iter),
2075                                       PAGE_SIZE - offset);
2076
2077                         if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2078                                 ret = -EFAULT;
2079                                 break;
2080                         }
2081                 }
2082
2083                 if (unlikely(fatal_signal_pending(current))) {
2084                         ret = -EINTR;
2085                         break;
2086                 }
2087
2088                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
2089                 if (unlikely(ret < 0))
2090                         break;
2091
2092                 cond_resched();
2093
2094                 if (unlikely(ret == 0)) {
2095                         /*
2096                          * If we were unable to copy any data at all, we must
2097                          * fall back to a single segment length write.
2098                          *
2099                          * If we didn't fallback here, we could livelock
2100                          * because not all segments in the iov can be copied at
2101                          * once without a pagefault.
2102                          */
2103                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
2104                                       iov_iter_single_seg_count(iter));
2105                         goto again;
2106                 }
2107                 pos += ret;
2108                 written += ret;
2109                 ret = 0;
2110
2111                 balance_dirty_pages_ratelimited(mapping);
2112         } while (iov_iter_count(iter));
2113
2114         bch2_pagecache_add_put(inode);
2115
2116         return written ? written : ret;
2117 }
2118
2119 /* O_DIRECT reads */
2120
2121 static void bio_check_or_release(struct bio *bio, bool check_dirty)
2122 {
2123         if (check_dirty) {
2124                 bio_check_pages_dirty(bio);
2125         } else {
2126                 bio_release_pages(bio, false);
2127                 bio_put(bio);
2128         }
2129 }
2130
2131 static void bch2_dio_read_complete(struct closure *cl)
2132 {
2133         struct dio_read *dio = container_of(cl, struct dio_read, cl);
2134
2135         dio->req->ki_complete(dio->req, dio->ret);
2136         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2137 }
2138
2139 static void bch2_direct_IO_read_endio(struct bio *bio)
2140 {
2141         struct dio_read *dio = bio->bi_private;
2142
2143         if (bio->bi_status)
2144                 dio->ret = blk_status_to_errno(bio->bi_status);
2145
2146         closure_put(&dio->cl);
2147 }
2148
2149 static void bch2_direct_IO_read_split_endio(struct bio *bio)
2150 {
2151         struct dio_read *dio = bio->bi_private;
2152         bool should_dirty = dio->should_dirty;
2153
2154         bch2_direct_IO_read_endio(bio);
2155         bio_check_or_release(bio, should_dirty);
2156 }
2157
2158 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
2159 {
2160         struct file *file = req->ki_filp;
2161         struct bch_inode_info *inode = file_bch_inode(file);
2162         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2163         struct bch_io_opts opts;
2164         struct dio_read *dio;
2165         struct bio *bio;
2166         loff_t offset = req->ki_pos;
2167         bool sync = is_sync_kiocb(req);
2168         size_t shorten;
2169         ssize_t ret;
2170
2171         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2172
2173         if ((offset|iter->count) & (block_bytes(c) - 1))
2174                 return -EINVAL;
2175
2176         ret = min_t(loff_t, iter->count,
2177                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
2178
2179         if (!ret)
2180                 return ret;
2181
2182         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
2183         iter->count -= shorten;
2184
2185         bio = bio_alloc_bioset(NULL,
2186                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2187                                REQ_OP_READ,
2188                                GFP_KERNEL,
2189                                &c->dio_read_bioset);
2190
2191         bio->bi_end_io = bch2_direct_IO_read_endio;
2192
2193         dio = container_of(bio, struct dio_read, rbio.bio);
2194         closure_init(&dio->cl, NULL);
2195
2196         /*
2197          * this is a _really_ horrible hack just to avoid an atomic sub at the
2198          * end:
2199          */
2200         if (!sync) {
2201                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
2202                 atomic_set(&dio->cl.remaining,
2203                            CLOSURE_REMAINING_INITIALIZER -
2204                            CLOSURE_RUNNING +
2205                            CLOSURE_DESTRUCTOR);
2206         } else {
2207                 atomic_set(&dio->cl.remaining,
2208                            CLOSURE_REMAINING_INITIALIZER + 1);
2209         }
2210
2211         dio->req        = req;
2212         dio->ret        = ret;
2213         /*
2214          * This is one of the sketchier things I've encountered: we have to skip
2215          * the dirtying of requests that are internal from the kernel (i.e. from
2216          * loopback), because we'll deadlock on page_lock.
2217          */
2218         dio->should_dirty = iter_is_iovec(iter);
2219
2220         goto start;
2221         while (iter->count) {
2222                 bio = bio_alloc_bioset(NULL,
2223                                        bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2224                                        REQ_OP_READ,
2225                                        GFP_KERNEL,
2226                                        &c->bio_read);
2227                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
2228 start:
2229                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC;
2230                 bio->bi_iter.bi_sector  = offset >> 9;
2231                 bio->bi_private         = dio;
2232
2233                 ret = bio_iov_iter_get_pages(bio, iter);
2234                 if (ret < 0) {
2235                         /* XXX: fault inject this path */
2236                         bio->bi_status = BLK_STS_RESOURCE;
2237                         bio_endio(bio);
2238                         break;
2239                 }
2240
2241                 offset += bio->bi_iter.bi_size;
2242
2243                 if (dio->should_dirty)
2244                         bio_set_pages_dirty(bio);
2245
2246                 if (iter->count)
2247                         closure_get(&dio->cl);
2248
2249                 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
2250         }
2251
2252         iter->count += shorten;
2253
2254         if (sync) {
2255                 closure_sync(&dio->cl);
2256                 closure_debug_destroy(&dio->cl);
2257                 ret = dio->ret;
2258                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2259                 return ret;
2260         } else {
2261                 return -EIOCBQUEUED;
2262         }
2263 }
2264
2265 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2266 {
2267         struct file *file = iocb->ki_filp;
2268         struct bch_inode_info *inode = file_bch_inode(file);
2269         struct address_space *mapping = file->f_mapping;
2270         size_t count = iov_iter_count(iter);
2271         ssize_t ret;
2272
2273         if (!count)
2274                 return 0; /* skip atime */
2275
2276         if (iocb->ki_flags & IOCB_DIRECT) {
2277                 struct blk_plug plug;
2278
2279                 if (unlikely(mapping->nrpages)) {
2280                         ret = filemap_write_and_wait_range(mapping,
2281                                                 iocb->ki_pos,
2282                                                 iocb->ki_pos + count - 1);
2283                         if (ret < 0)
2284                                 goto out;
2285                 }
2286
2287                 file_accessed(file);
2288
2289                 blk_start_plug(&plug);
2290                 ret = bch2_direct_IO_read(iocb, iter);
2291                 blk_finish_plug(&plug);
2292
2293                 if (ret >= 0)
2294                         iocb->ki_pos += ret;
2295         } else {
2296                 bch2_pagecache_add_get(inode);
2297                 ret = generic_file_read_iter(iocb, iter);
2298                 bch2_pagecache_add_put(inode);
2299         }
2300 out:
2301         return bch2_err_class(ret);
2302 }
2303
2304 /* O_DIRECT writes */
2305
2306 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2307                                        u64 offset, u64 size,
2308                                        unsigned nr_replicas, bool compressed)
2309 {
2310         struct btree_trans trans;
2311         struct btree_iter iter;
2312         struct bkey_s_c k;
2313         u64 end = offset + size;
2314         u32 snapshot;
2315         bool ret = true;
2316         int err;
2317
2318         bch2_trans_init(&trans, c, 0, 0);
2319 retry:
2320         bch2_trans_begin(&trans);
2321
2322         err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2323         if (err)
2324                 goto err;
2325
2326         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2327                            SPOS(inum.inum, offset, snapshot),
2328                            BTREE_ITER_SLOTS, k, err) {
2329                 if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
2330                         break;
2331
2332                 if (k.k->p.snapshot != snapshot ||
2333                     nr_replicas > bch2_bkey_replicas(c, k) ||
2334                     (!compressed && bch2_bkey_sectors_compressed(k))) {
2335                         ret = false;
2336                         break;
2337                 }
2338         }
2339
2340         offset = iter.pos.offset;
2341         bch2_trans_iter_exit(&trans, &iter);
2342 err:
2343         if (bch2_err_matches(err, BCH_ERR_transaction_restart))
2344                 goto retry;
2345         bch2_trans_exit(&trans);
2346
2347         return err ? false : ret;
2348 }
2349
2350 static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
2351 {
2352         struct bch_fs *c = dio->op.c;
2353         struct bch_inode_info *inode = dio->inode;
2354         struct bio *bio = &dio->op.wbio.bio;
2355
2356         return bch2_check_range_allocated(c, inode_inum(inode),
2357                                 dio->op.pos.offset, bio_sectors(bio),
2358                                 dio->op.opts.data_replicas,
2359                                 dio->op.opts.compression != 0);
2360 }
2361
2362 static void bch2_dio_write_loop_async(struct bch_write_op *);
2363 static __always_inline long bch2_dio_write_done(struct dio_write *dio);
2364
2365 /*
2366  * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
2367  * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
2368  * caller's stack, we're not guaranteed that it will live for the duration of
2369  * the IO:
2370  */
2371 static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
2372 {
2373         struct iovec *iov = dio->inline_vecs;
2374
2375         /*
2376          * iov_iter has a single embedded iovec - nothing to do:
2377          */
2378         if (iter_is_ubuf(&dio->iter))
2379                 return 0;
2380
2381         /*
2382          * We don't currently handle non-iovec iov_iters here - return an error,
2383          * and we'll fall back to doing the IO synchronously:
2384          */
2385         if (!iter_is_iovec(&dio->iter))
2386                 return -1;
2387
2388         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2389                 iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
2390                                     GFP_KERNEL);
2391                 if (unlikely(!iov))
2392                         return -ENOMEM;
2393
2394                 dio->free_iov = true;
2395         }
2396
2397         memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov));
2398         dio->iter.__iov = iov;
2399         return 0;
2400 }
2401
2402 static void bch2_dio_write_flush_done(struct closure *cl)
2403 {
2404         struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
2405         struct bch_fs *c = dio->op.c;
2406
2407         closure_debug_destroy(cl);
2408
2409         dio->op.error = bch2_journal_error(&c->journal);
2410
2411         bch2_dio_write_done(dio);
2412 }
2413
2414 static noinline void bch2_dio_write_flush(struct dio_write *dio)
2415 {
2416         struct bch_fs *c = dio->op.c;
2417         struct bch_inode_unpacked inode;
2418         int ret;
2419
2420         dio->flush = 0;
2421
2422         closure_init(&dio->op.cl, NULL);
2423
2424         if (!dio->op.error) {
2425                 ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
2426                 if (ret) {
2427                         dio->op.error = ret;
2428                 } else {
2429                         bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl);
2430                         bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
2431                 }
2432         }
2433
2434         if (dio->sync) {
2435                 closure_sync(&dio->op.cl);
2436                 closure_debug_destroy(&dio->op.cl);
2437         } else {
2438                 continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
2439         }
2440 }
2441
2442 static __always_inline long bch2_dio_write_done(struct dio_write *dio)
2443 {
2444         struct kiocb *req = dio->req;
2445         struct bch_inode_info *inode = dio->inode;
2446         bool sync = dio->sync;
2447         long ret;
2448
2449         if (unlikely(dio->flush)) {
2450                 bch2_dio_write_flush(dio);
2451                 if (!sync)
2452                         return -EIOCBQUEUED;
2453         }
2454
2455         bch2_pagecache_block_put(inode);
2456
2457         if (dio->free_iov)
2458                 kfree(dio->iter.__iov);
2459
2460         ret = dio->op.error ?: ((long) dio->written << 9);
2461         bio_put(&dio->op.wbio.bio);
2462
2463         /* inode->i_dio_count is our ref on inode and thus bch_fs */
2464         inode_dio_end(&inode->v);
2465
2466         if (ret < 0)
2467                 ret = bch2_err_class(ret);
2468
2469         if (!sync) {
2470                 req->ki_complete(req, ret);
2471                 ret = -EIOCBQUEUED;
2472         }
2473         return ret;
2474 }
2475
2476 static __always_inline void bch2_dio_write_end(struct dio_write *dio)
2477 {
2478         struct bch_fs *c = dio->op.c;
2479         struct kiocb *req = dio->req;
2480         struct bch_inode_info *inode = dio->inode;
2481         struct bio *bio = &dio->op.wbio.bio;
2482
2483         req->ki_pos     += (u64) dio->op.written << 9;
2484         dio->written    += dio->op.written;
2485
2486         if (dio->extending) {
2487                 spin_lock(&inode->v.i_lock);
2488                 if (req->ki_pos > inode->v.i_size)
2489                         i_size_write(&inode->v, req->ki_pos);
2490                 spin_unlock(&inode->v.i_lock);
2491         }
2492
2493         if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
2494                 mutex_lock(&inode->ei_quota_lock);
2495                 __i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
2496                 __bch2_quota_reservation_put(c, inode, &dio->quota_res);
2497                 mutex_unlock(&inode->ei_quota_lock);
2498         }
2499
2500         bio_release_pages(bio, false);
2501
2502         if (unlikely(dio->op.error))
2503                 set_bit(EI_INODE_ERROR, &inode->ei_flags);
2504 }
2505
2506 static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
2507 {
2508         struct bch_fs *c = dio->op.c;
2509         struct kiocb *req = dio->req;
2510         struct address_space *mapping = dio->mapping;
2511         struct bch_inode_info *inode = dio->inode;
2512         struct bch_io_opts opts;
2513         struct bio *bio = &dio->op.wbio.bio;
2514         unsigned unaligned, iter_count;
2515         bool sync = dio->sync, dropped_locks;
2516         long ret;
2517
2518         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2519
2520         while (1) {
2521                 iter_count = dio->iter.count;
2522
2523                 EBUG_ON(current->faults_disabled_mapping);
2524                 current->faults_disabled_mapping = mapping;
2525
2526                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2527
2528                 dropped_locks = fdm_dropped_locks();
2529
2530                 current->faults_disabled_mapping = NULL;
2531
2532                 /*
2533                  * If the fault handler returned an error but also signalled
2534                  * that it dropped & retook ei_pagecache_lock, we just need to
2535                  * re-shoot down the page cache and retry:
2536                  */
2537                 if (dropped_locks && ret)
2538                         ret = 0;
2539
2540                 if (unlikely(ret < 0))
2541                         goto err;
2542
2543                 if (unlikely(dropped_locks)) {
2544                         ret = write_invalidate_inode_pages_range(mapping,
2545                                         req->ki_pos,
2546                                         req->ki_pos + iter_count - 1);
2547                         if (unlikely(ret))
2548                                 goto err;
2549
2550                         if (!bio->bi_iter.bi_size)
2551                                 continue;
2552                 }
2553
2554                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2555                 bio->bi_iter.bi_size -= unaligned;
2556                 iov_iter_revert(&dio->iter, unaligned);
2557
2558                 if (!bio->bi_iter.bi_size) {
2559                         /*
2560                          * bio_iov_iter_get_pages was only able to get <
2561                          * blocksize worth of pages:
2562                          */
2563                         ret = -EFAULT;
2564                         goto err;
2565                 }
2566
2567                 bch2_write_op_init(&dio->op, c, opts);
2568                 dio->op.end_io          = sync
2569                         ? NULL
2570                         : bch2_dio_write_loop_async;
2571                 dio->op.target          = dio->op.opts.foreground_target;
2572                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
2573                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
2574                 dio->op.subvol          = inode->ei_subvol;
2575                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2576                 dio->op.devs_need_flush = &inode->ei_devs_need_flush;
2577
2578                 if (sync)
2579                         dio->op.flags |= BCH_WRITE_SYNC;
2580                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2581
2582                 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2583                                                  bio_sectors(bio), true);
2584                 if (unlikely(ret))
2585                         goto err;
2586
2587                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2588                                                 dio->op.opts.data_replicas, 0);
2589                 if (unlikely(ret) &&
2590                     !bch2_dio_write_check_allocated(dio))
2591                         goto err;
2592
2593                 task_io_account_write(bio->bi_iter.bi_size);
2594
2595                 if (unlikely(dio->iter.count) &&
2596                     !dio->sync &&
2597                     !dio->loop &&
2598                     bch2_dio_write_copy_iov(dio))
2599                         dio->sync = sync = true;
2600
2601                 dio->loop = true;
2602                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2603
2604                 if (!sync)
2605                         return -EIOCBQUEUED;
2606
2607                 bch2_dio_write_end(dio);
2608
2609                 if (likely(!dio->iter.count) || dio->op.error)
2610                         break;
2611
2612                 bio_reset(bio, NULL, REQ_OP_WRITE);
2613         }
2614 out:
2615         return bch2_dio_write_done(dio);
2616 err:
2617         dio->op.error = ret;
2618
2619         bio_release_pages(bio, false);
2620
2621         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2622         goto out;
2623 }
2624
2625 static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
2626 {
2627         struct mm_struct *mm = dio->mm;
2628
2629         bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
2630
2631         if (mm)
2632                 kthread_use_mm(mm);
2633         bch2_dio_write_loop(dio);
2634         if (mm)
2635                 kthread_unuse_mm(mm);
2636 }
2637
2638 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2639 {
2640         struct dio_write *dio = container_of(op, struct dio_write, op);
2641
2642         bch2_dio_write_end(dio);
2643
2644         if (likely(!dio->iter.count) || dio->op.error)
2645                 bch2_dio_write_done(dio);
2646         else
2647                 bch2_dio_write_continue(dio);
2648 }
2649
2650 static noinline
2651 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2652 {
2653         struct file *file = req->ki_filp;
2654         struct address_space *mapping = file->f_mapping;
2655         struct bch_inode_info *inode = file_bch_inode(file);
2656         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2657         struct dio_write *dio;
2658         struct bio *bio;
2659         bool locked = true, extending;
2660         ssize_t ret;
2661
2662         prefetch(&c->opts);
2663         prefetch((void *) &c->opts + 64);
2664         prefetch(&inode->ei_inode);
2665         prefetch((void *) &inode->ei_inode + 64);
2666
2667         inode_lock(&inode->v);
2668
2669         ret = generic_write_checks(req, iter);
2670         if (unlikely(ret <= 0))
2671                 goto err;
2672
2673         ret = file_remove_privs(file);
2674         if (unlikely(ret))
2675                 goto err;
2676
2677         ret = file_update_time(file);
2678         if (unlikely(ret))
2679                 goto err;
2680
2681         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2682                 goto err;
2683
2684         inode_dio_begin(&inode->v);
2685         bch2_pagecache_block_get(inode);
2686
2687         extending = req->ki_pos + iter->count > inode->v.i_size;
2688         if (!extending) {
2689                 inode_unlock(&inode->v);
2690                 locked = false;
2691         }
2692
2693         bio = bio_alloc_bioset(NULL,
2694                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2695                                REQ_OP_WRITE,
2696                                GFP_KERNEL,
2697                                &c->dio_write_bioset);
2698         dio = container_of(bio, struct dio_write, op.wbio.bio);
2699         dio->req                = req;
2700         dio->mapping            = mapping;
2701         dio->inode              = inode;
2702         dio->mm                 = current->mm;
2703         dio->loop               = false;
2704         dio->extending          = extending;
2705         dio->sync               = is_sync_kiocb(req) || extending;
2706         dio->flush              = iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
2707         dio->free_iov           = false;
2708         dio->quota_res.sectors  = 0;
2709         dio->written            = 0;
2710         dio->iter               = *iter;
2711         dio->op.c               = c;
2712
2713         if (unlikely(mapping->nrpages)) {
2714                 ret = write_invalidate_inode_pages_range(mapping,
2715                                                 req->ki_pos,
2716                                                 req->ki_pos + iter->count - 1);
2717                 if (unlikely(ret))
2718                         goto err_put_bio;
2719         }
2720
2721         ret = bch2_dio_write_loop(dio);
2722 err:
2723         if (locked)
2724                 inode_unlock(&inode->v);
2725         return ret;
2726 err_put_bio:
2727         bch2_pagecache_block_put(inode);
2728         bio_put(bio);
2729         inode_dio_end(&inode->v);
2730         goto err;
2731 }
2732
2733 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2734 {
2735         struct file *file = iocb->ki_filp;
2736         struct bch_inode_info *inode = file_bch_inode(file);
2737         ssize_t ret;
2738
2739         if (iocb->ki_flags & IOCB_DIRECT) {
2740                 ret = bch2_direct_write(iocb, from);
2741                 goto out;
2742         }
2743
2744         /* We can write back this queue in page reclaim */
2745         current->backing_dev_info = inode_to_bdi(&inode->v);
2746         inode_lock(&inode->v);
2747
2748         ret = generic_write_checks(iocb, from);
2749         if (ret <= 0)
2750                 goto unlock;
2751
2752         ret = file_remove_privs(file);
2753         if (ret)
2754                 goto unlock;
2755
2756         ret = file_update_time(file);
2757         if (ret)
2758                 goto unlock;
2759
2760         ret = bch2_buffered_write(iocb, from);
2761         if (likely(ret > 0))
2762                 iocb->ki_pos += ret;
2763 unlock:
2764         inode_unlock(&inode->v);
2765         current->backing_dev_info = NULL;
2766
2767         if (ret > 0)
2768                 ret = generic_write_sync(iocb, ret);
2769 out:
2770         return bch2_err_class(ret);
2771 }
2772
2773 /* fsync: */
2774
2775 /*
2776  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2777  * insert trigger: look up the btree inode instead
2778  */
2779 static int bch2_flush_inode(struct bch_fs *c,
2780                             struct bch_inode_info *inode)
2781 {
2782         struct bch_inode_unpacked u;
2783         int ret;
2784
2785         if (c->opts.journal_flush_disabled)
2786                 return 0;
2787
2788         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u);
2789         if (ret)
2790                 return ret;
2791
2792         return bch2_journal_flush_seq(&c->journal, u.bi_journal_seq) ?:
2793                 bch2_inode_flush_nocow_writes(c, inode);
2794 }
2795
2796 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2797 {
2798         struct bch_inode_info *inode = file_bch_inode(file);
2799         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2800         int ret, ret2, ret3;
2801
2802         ret = file_write_and_wait_range(file, start, end);
2803         ret2 = sync_inode_metadata(&inode->v, 1);
2804         ret3 = bch2_flush_inode(c, inode);
2805
2806         return bch2_err_class(ret ?: ret2 ?: ret3);
2807 }
2808
2809 /* truncate: */
2810
2811 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2812                                  struct bpos start,
2813                                  struct bpos end)
2814 {
2815         struct btree_trans trans;
2816         struct btree_iter iter;
2817         struct bkey_s_c k;
2818         int ret = 0;
2819
2820         bch2_trans_init(&trans, c, 0, 0);
2821 retry:
2822         bch2_trans_begin(&trans);
2823
2824         ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2825         if (ret)
2826                 goto err;
2827
2828         for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents, start, end, 0, k, ret)
2829                 if (bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k)) {
2830                         ret = 1;
2831                         break;
2832                 }
2833         start = iter.pos;
2834         bch2_trans_iter_exit(&trans, &iter);
2835 err:
2836         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
2837                 goto retry;
2838
2839         bch2_trans_exit(&trans);
2840         return ret;
2841 }
2842
2843 static int __bch2_truncate_folio(struct bch_inode_info *inode,
2844                                  pgoff_t index, loff_t start, loff_t end)
2845 {
2846         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2847         struct address_space *mapping = inode->v.i_mapping;
2848         struct bch_folio *s;
2849         unsigned start_offset = start & (PAGE_SIZE - 1);
2850         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2851         unsigned i;
2852         struct folio *folio;
2853         s64 i_sectors_delta = 0;
2854         int ret = 0;
2855         u64 end_pos;
2856
2857         folio = filemap_lock_folio(mapping, index);
2858         if (IS_ERR_OR_NULL(folio)) {
2859                 /*
2860                  * XXX: we're doing two index lookups when we end up reading the
2861                  * folio
2862                  */
2863                 ret = range_has_data(c, inode->ei_subvol,
2864                                 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT)),
2865                                 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT) + PAGE_SECTORS));
2866                 if (ret <= 0)
2867                         return ret;
2868
2869                 folio = __filemap_get_folio(mapping, index,
2870                                             FGP_LOCK|FGP_CREAT, GFP_KERNEL);
2871                 if (unlikely(IS_ERR_OR_NULL(folio))) {
2872                         ret = -ENOMEM;
2873                         goto out;
2874                 }
2875         }
2876
2877         BUG_ON(start    >= folio_end_pos(folio));
2878         BUG_ON(end      <= folio_pos(folio));
2879
2880         start_offset    = max(start, folio_pos(folio)) - folio_pos(folio);
2881         end_offset      = min_t(u64, end, folio_end_pos(folio)) - folio_pos(folio);
2882
2883         /* Folio boundary? Nothing to do */
2884         if (start_offset == 0 &&
2885             end_offset == folio_size(folio)) {
2886                 ret = 0;
2887                 goto unlock;
2888         }
2889
2890         s = bch2_folio_create(folio, 0);
2891         if (!s) {
2892                 ret = -ENOMEM;
2893                 goto unlock;
2894         }
2895
2896         if (!folio_test_uptodate(folio)) {
2897                 ret = bch2_read_single_folio(folio, mapping);
2898                 if (ret)
2899                         goto unlock;
2900         }
2901
2902         ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
2903         if (ret)
2904                 goto unlock;
2905
2906         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2907              i < round_down(end_offset, block_bytes(c)) >> 9;
2908              i++) {
2909                 s->s[i].nr_replicas     = 0;
2910
2911                 i_sectors_delta -= s->s[i].state == SECTOR_dirty;
2912                 folio_sector_set(folio, s, i, SECTOR_unallocated);
2913         }
2914
2915         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2916
2917         /*
2918          * Caller needs to know whether this folio will be written out by
2919          * writeback - doing an i_size update if necessary - or whether it will
2920          * be responsible for the i_size update.
2921          *
2922          * Note that we shouldn't ever see a folio beyond EOF, but check and
2923          * warn if so. This has been observed by failure to clean up folios
2924          * after a short write and there's still a chance reclaim will fix
2925          * things up.
2926          */
2927         WARN_ON_ONCE(folio_pos(folio) >= inode->v.i_size);
2928         end_pos = folio_end_pos(folio);
2929         if (inode->v.i_size > folio_pos(folio))
2930                 end_pos = min_t(u64, inode->v.i_size, end_pos);
2931         ret = s->s[folio_pos_to_s(folio, end_pos - 1)].state >= SECTOR_dirty;
2932
2933         folio_zero_segment(folio, start_offset, end_offset);
2934
2935         /*
2936          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2937          *
2938          * XXX: because we aren't currently tracking whether the folio has actual
2939          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2940          */
2941         BUG_ON(bch2_get_folio_disk_reservation(c, inode, folio, false));
2942
2943         /*
2944          * This removes any writeable userspace mappings; we need to force
2945          * .page_mkwrite to be called again before any mmapped writes, to
2946          * redirty the full page:
2947          */
2948         folio_mkclean(folio);
2949         filemap_dirty_folio(mapping, folio);
2950 unlock:
2951         folio_unlock(folio);
2952         folio_put(folio);
2953 out:
2954         return ret;
2955 }
2956
2957 static int bch2_truncate_folio(struct bch_inode_info *inode, loff_t from)
2958 {
2959         return __bch2_truncate_folio(inode, from >> PAGE_SHIFT,
2960                                      from, ANYSINT_MAX(loff_t));
2961 }
2962
2963 static int bch2_truncate_folios(struct bch_inode_info *inode,
2964                                 loff_t start, loff_t end)
2965 {
2966         int ret = __bch2_truncate_folio(inode, start >> PAGE_SHIFT,
2967                                         start, end);
2968
2969         if (ret >= 0 &&
2970             start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2971                 ret = __bch2_truncate_folio(inode,
2972                                         (end - 1) >> PAGE_SHIFT,
2973                                         start, end);
2974         return ret;
2975 }
2976
2977 static int bch2_extend(struct mnt_idmap *idmap,
2978                        struct bch_inode_info *inode,
2979                        struct bch_inode_unpacked *inode_u,
2980                        struct iattr *iattr)
2981 {
2982         struct address_space *mapping = inode->v.i_mapping;
2983         int ret;
2984
2985         /*
2986          * sync appends:
2987          *
2988          * this has to be done _before_ extending i_size:
2989          */
2990         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2991         if (ret)
2992                 return ret;
2993
2994         truncate_setsize(&inode->v, iattr->ia_size);
2995
2996         return bch2_setattr_nonsize(idmap, inode, iattr);
2997 }
2998
2999 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
3000                                    struct bch_inode_unpacked *bi,
3001                                    void *p)
3002 {
3003         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
3004         return 0;
3005 }
3006
3007 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
3008                                   struct bch_inode_unpacked *bi, void *p)
3009 {
3010         u64 *new_i_size = p;
3011
3012         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
3013         bi->bi_size = *new_i_size;
3014         return 0;
3015 }
3016
3017 int bch2_truncate(struct mnt_idmap *idmap,
3018                   struct bch_inode_info *inode, struct iattr *iattr)
3019 {
3020         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3021         struct address_space *mapping = inode->v.i_mapping;
3022         struct bch_inode_unpacked inode_u;
3023         u64 new_i_size = iattr->ia_size;
3024         s64 i_sectors_delta = 0;
3025         int ret = 0;
3026
3027         /*
3028          * If the truncate call with change the size of the file, the
3029          * cmtimes should be updated. If the size will not change, we
3030          * do not need to update the cmtimes.
3031          */
3032         if (iattr->ia_size != inode->v.i_size) {
3033                 if (!(iattr->ia_valid & ATTR_MTIME))
3034                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
3035                 if (!(iattr->ia_valid & ATTR_CTIME))
3036                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
3037                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
3038         }
3039
3040         inode_dio_wait(&inode->v);
3041         bch2_pagecache_block_get(inode);
3042
3043         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
3044         if (ret)
3045                 goto err;
3046
3047         /*
3048          * check this before next assertion; on filesystem error our normal
3049          * invariants are a bit broken (truncate has to truncate the page cache
3050          * before the inode).
3051          */
3052         ret = bch2_journal_error(&c->journal);
3053         if (ret)
3054                 goto err;
3055
3056         WARN_ONCE(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
3057                   inode->v.i_size < inode_u.bi_size,
3058                   "truncate spotted in mem i_size < btree i_size: %llu < %llu\n",
3059                   (u64) inode->v.i_size, inode_u.bi_size);
3060
3061         if (iattr->ia_size > inode->v.i_size) {
3062                 ret = bch2_extend(idmap, inode, &inode_u, iattr);
3063                 goto err;
3064         }
3065
3066         iattr->ia_valid &= ~ATTR_SIZE;
3067
3068         ret = bch2_truncate_folio(inode, iattr->ia_size);
3069         if (unlikely(ret < 0))
3070                 goto err;
3071
3072         /*
3073          * When extending, we're going to write the new i_size to disk
3074          * immediately so we need to flush anything above the current on disk
3075          * i_size first:
3076          *
3077          * Also, when extending we need to flush the page that i_size currently
3078          * straddles - if it's mapped to userspace, we need to ensure that
3079          * userspace has to redirty it and call .mkwrite -> set_page_dirty
3080          * again to allocate the part of the page that was extended.
3081          */
3082         if (iattr->ia_size > inode_u.bi_size)
3083                 ret = filemap_write_and_wait_range(mapping,
3084                                 inode_u.bi_size,
3085                                 iattr->ia_size - 1);
3086         else if (iattr->ia_size & (PAGE_SIZE - 1))
3087                 ret = filemap_write_and_wait_range(mapping,
3088                                 round_down(iattr->ia_size, PAGE_SIZE),
3089                                 iattr->ia_size - 1);
3090         if (ret)
3091                 goto err;
3092
3093         mutex_lock(&inode->ei_update_lock);
3094         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
3095                                &new_i_size, 0);
3096         mutex_unlock(&inode->ei_update_lock);
3097
3098         if (unlikely(ret))
3099                 goto err;
3100
3101         truncate_setsize(&inode->v, iattr->ia_size);
3102
3103         ret = bch2_fpunch(c, inode_inum(inode),
3104                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
3105                         U64_MAX, &i_sectors_delta);
3106         i_sectors_acct(c, inode, NULL, i_sectors_delta);
3107
3108         bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
3109                                 !bch2_journal_error(&c->journal), c,
3110                                 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
3111                                 inode->v.i_ino, (u64) inode->v.i_blocks,
3112                                 inode->ei_inode.bi_sectors);
3113         if (unlikely(ret))
3114                 goto err;
3115
3116         mutex_lock(&inode->ei_update_lock);
3117         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
3118         mutex_unlock(&inode->ei_update_lock);
3119
3120         ret = bch2_setattr_nonsize(idmap, inode, iattr);
3121 err:
3122         bch2_pagecache_block_put(inode);
3123         return bch2_err_class(ret);
3124 }
3125
3126 /* fallocate: */
3127
3128 static int inode_update_times_fn(struct bch_inode_info *inode,
3129                                  struct bch_inode_unpacked *bi, void *p)
3130 {
3131         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3132
3133         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
3134         return 0;
3135 }
3136
3137 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
3138 {
3139         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3140         u64 end         = offset + len;
3141         u64 block_start = round_up(offset, block_bytes(c));
3142         u64 block_end   = round_down(end, block_bytes(c));
3143         bool truncated_last_page;
3144         int ret = 0;
3145
3146         ret = bch2_truncate_folios(inode, offset, end);
3147         if (unlikely(ret < 0))
3148                 goto err;
3149
3150         truncated_last_page = ret;
3151
3152         truncate_pagecache_range(&inode->v, offset, end - 1);
3153
3154         if (block_start < block_end) {
3155                 s64 i_sectors_delta = 0;
3156
3157                 ret = bch2_fpunch(c, inode_inum(inode),
3158                                   block_start >> 9, block_end >> 9,
3159                                   &i_sectors_delta);
3160                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3161         }
3162
3163         mutex_lock(&inode->ei_update_lock);
3164         if (end >= inode->v.i_size && !truncated_last_page) {
3165                 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
3166                                             ATTR_MTIME|ATTR_CTIME);
3167         } else {
3168                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3169                                        ATTR_MTIME|ATTR_CTIME);
3170         }
3171         mutex_unlock(&inode->ei_update_lock);
3172 err:
3173         return ret;
3174 }
3175
3176 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
3177                                    loff_t offset, loff_t len,
3178                                    bool insert)
3179 {
3180         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3181         struct address_space *mapping = inode->v.i_mapping;
3182         struct bkey_buf copy;
3183         struct btree_trans trans;
3184         struct btree_iter src, dst, del;
3185         loff_t shift, new_size;
3186         u64 src_start;
3187         int ret = 0;
3188
3189         if ((offset | len) & (block_bytes(c) - 1))
3190                 return -EINVAL;
3191
3192         if (insert) {
3193                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
3194                         return -EFBIG;
3195
3196                 if (offset >= inode->v.i_size)
3197                         return -EINVAL;
3198
3199                 src_start       = U64_MAX;
3200                 shift           = len;
3201         } else {
3202                 if (offset + len >= inode->v.i_size)
3203                         return -EINVAL;
3204
3205                 src_start       = offset + len;
3206                 shift           = -len;
3207         }
3208
3209         new_size = inode->v.i_size + shift;
3210
3211         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
3212         if (ret)
3213                 return ret;
3214
3215         if (insert) {
3216                 i_size_write(&inode->v, new_size);
3217                 mutex_lock(&inode->ei_update_lock);
3218                 ret = bch2_write_inode_size(c, inode, new_size,
3219                                             ATTR_MTIME|ATTR_CTIME);
3220                 mutex_unlock(&inode->ei_update_lock);
3221         } else {
3222                 s64 i_sectors_delta = 0;
3223
3224                 ret = bch2_fpunch(c, inode_inum(inode),
3225                                   offset >> 9, (offset + len) >> 9,
3226                                   &i_sectors_delta);
3227                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3228
3229                 if (ret)
3230                         return ret;
3231         }
3232
3233         bch2_bkey_buf_init(&copy);
3234         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
3235         bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
3236                         POS(inode->v.i_ino, src_start >> 9),
3237                         BTREE_ITER_INTENT);
3238         bch2_trans_copy_iter(&dst, &src);
3239         bch2_trans_copy_iter(&del, &src);
3240
3241         while (ret == 0 ||
3242                bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
3243                 struct disk_reservation disk_res =
3244                         bch2_disk_reservation_init(c, 0);
3245                 struct bkey_i delete;
3246                 struct bkey_s_c k;
3247                 struct bpos next_pos;
3248                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
3249                 struct bpos atomic_end;
3250                 unsigned trigger_flags = 0;
3251                 u32 snapshot;
3252
3253                 bch2_trans_begin(&trans);
3254
3255                 ret = bch2_subvolume_get_snapshot(&trans,
3256                                         inode->ei_subvol, &snapshot);
3257                 if (ret)
3258                         continue;
3259
3260                 bch2_btree_iter_set_snapshot(&src, snapshot);
3261                 bch2_btree_iter_set_snapshot(&dst, snapshot);
3262                 bch2_btree_iter_set_snapshot(&del, snapshot);
3263
3264                 bch2_trans_begin(&trans);
3265
3266                 k = insert
3267                         ? bch2_btree_iter_peek_prev(&src)
3268                         : bch2_btree_iter_peek_upto(&src, POS(inode->v.i_ino, U64_MAX));
3269                 if ((ret = bkey_err(k)))
3270                         continue;
3271
3272                 if (!k.k || k.k->p.inode != inode->v.i_ino)
3273                         break;
3274
3275                 if (insert &&
3276                     bkey_le(k.k->p, POS(inode->v.i_ino, offset >> 9)))
3277                         break;
3278 reassemble:
3279                 bch2_bkey_buf_reassemble(&copy, c, k);
3280
3281                 if (insert &&
3282                     bkey_lt(bkey_start_pos(k.k), move_pos))
3283                         bch2_cut_front(move_pos, copy.k);
3284
3285                 copy.k->k.p.offset += shift >> 9;
3286                 bch2_btree_iter_set_pos(&dst, bkey_start_pos(&copy.k->k));
3287
3288                 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
3289                 if (ret)
3290                         continue;
3291
3292                 if (!bkey_eq(atomic_end, copy.k->k.p)) {
3293                         if (insert) {
3294                                 move_pos = atomic_end;
3295                                 move_pos.offset -= shift >> 9;
3296                                 goto reassemble;
3297                         } else {
3298                                 bch2_cut_back(atomic_end, copy.k);
3299                         }
3300                 }
3301
3302                 bkey_init(&delete.k);
3303                 delete.k.p = copy.k->k.p;
3304                 delete.k.size = copy.k->k.size;
3305                 delete.k.p.offset -= shift >> 9;
3306                 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
3307
3308                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
3309
3310                 if (copy.k->k.size != k.k->size) {
3311                         /* We might end up splitting compressed extents: */
3312                         unsigned nr_ptrs =
3313                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
3314
3315                         ret = bch2_disk_reservation_get(c, &disk_res,
3316                                         copy.k->k.size, nr_ptrs,
3317                                         BCH_DISK_RESERVATION_NOFAIL);
3318                         BUG_ON(ret);
3319                 }
3320
3321                 ret =   bch2_btree_iter_traverse(&del) ?:
3322                         bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
3323                         bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
3324                         bch2_trans_commit(&trans, &disk_res, NULL,
3325                                           BTREE_INSERT_NOFAIL);
3326                 bch2_disk_reservation_put(c, &disk_res);
3327
3328                 if (!ret)
3329                         bch2_btree_iter_set_pos(&src, next_pos);
3330         }
3331         bch2_trans_iter_exit(&trans, &del);
3332         bch2_trans_iter_exit(&trans, &dst);
3333         bch2_trans_iter_exit(&trans, &src);
3334         bch2_trans_exit(&trans);
3335         bch2_bkey_buf_exit(&copy, c);
3336
3337         if (ret)
3338                 return ret;
3339
3340         mutex_lock(&inode->ei_update_lock);
3341         if (!insert) {
3342                 i_size_write(&inode->v, new_size);
3343                 ret = bch2_write_inode_size(c, inode, new_size,
3344                                             ATTR_MTIME|ATTR_CTIME);
3345         } else {
3346                 /* We need an inode update to update bi_journal_seq for fsync: */
3347                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3348                                        ATTR_MTIME|ATTR_CTIME);
3349         }
3350         mutex_unlock(&inode->ei_update_lock);
3351         return ret;
3352 }
3353
3354 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
3355                              u64 start_sector, u64 end_sector)
3356 {
3357         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3358         struct btree_trans trans;
3359         struct btree_iter iter;
3360         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
3361         struct bch_io_opts opts;
3362         int ret = 0;
3363
3364         bch2_inode_opts_get(&opts, c, &inode->ei_inode);
3365         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
3366
3367         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3368                         POS(inode->v.i_ino, start_sector),
3369                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
3370
3371         while (!ret && bkey_lt(iter.pos, end_pos)) {
3372                 s64 i_sectors_delta = 0;
3373                 struct quota_res quota_res = { 0 };
3374                 struct bkey_s_c k;
3375                 unsigned sectors;
3376                 u32 snapshot;
3377
3378                 bch2_trans_begin(&trans);
3379
3380                 ret = bch2_subvolume_get_snapshot(&trans,
3381                                         inode->ei_subvol, &snapshot);
3382                 if (ret)
3383                         goto bkey_err;
3384
3385                 bch2_btree_iter_set_snapshot(&iter, snapshot);
3386
3387                 k = bch2_btree_iter_peek_slot(&iter);
3388                 if ((ret = bkey_err(k)))
3389                         goto bkey_err;
3390
3391                 /* already reserved */
3392                 if (bkey_extent_is_reservation(k) &&
3393                     bch2_bkey_nr_ptrs_fully_allocated(k) >= opts.data_replicas) {
3394                         bch2_btree_iter_advance(&iter);
3395                         continue;
3396                 }
3397
3398                 if (bkey_extent_is_data(k.k) &&
3399                     !(mode & FALLOC_FL_ZERO_RANGE)) {
3400                         bch2_btree_iter_advance(&iter);
3401                         continue;
3402                 }
3403
3404                 /*
3405                  * XXX: for nocow mode, we should promote shared extents to
3406                  * unshared here
3407                  */
3408
3409                 sectors = bpos_min(k.k->p, end_pos).offset - iter.pos.offset;
3410
3411                 if (!bkey_extent_is_allocation(k.k)) {
3412                         ret = bch2_quota_reservation_add(c, inode,
3413                                         &quota_res,
3414                                         sectors, true);
3415                         if (unlikely(ret))
3416                                 goto bkey_err;
3417                 }
3418
3419                 ret = bch2_extent_fallocate(&trans, inode_inum(inode), &iter,
3420                                             sectors, opts, &i_sectors_delta,
3421                                             writepoint_hashed((unsigned long) current));
3422                 if (ret)
3423                         goto bkey_err;
3424
3425                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3426 bkey_err:
3427                 bch2_quota_reservation_put(c, inode, &quota_res);
3428                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3429                         ret = 0;
3430         }
3431
3432         bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3433         mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3434
3435         if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) {
3436                 struct quota_res quota_res = { 0 };
3437                 s64 i_sectors_delta = 0;
3438
3439                 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3440                                end_sector, &i_sectors_delta);
3441                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3442                 bch2_quota_reservation_put(c, inode, &quota_res);
3443         }
3444
3445         bch2_trans_iter_exit(&trans, &iter);
3446         bch2_trans_exit(&trans);
3447         return ret;
3448 }
3449
3450 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3451                             loff_t offset, loff_t len)
3452 {
3453         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3454         u64 end         = offset + len;
3455         u64 block_start = round_down(offset,    block_bytes(c));
3456         u64 block_end   = round_up(end,         block_bytes(c));
3457         bool truncated_last_page = false;
3458         int ret, ret2 = 0;
3459
3460         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3461                 ret = inode_newsize_ok(&inode->v, end);
3462                 if (ret)
3463                         return ret;
3464         }
3465
3466         if (mode & FALLOC_FL_ZERO_RANGE) {
3467                 ret = bch2_truncate_folios(inode, offset, end);
3468                 if (unlikely(ret < 0))
3469                         return ret;
3470
3471                 truncated_last_page = ret;
3472
3473                 truncate_pagecache_range(&inode->v, offset, end - 1);
3474
3475                 block_start     = round_up(offset,      block_bytes(c));
3476                 block_end       = round_down(end,       block_bytes(c));
3477         }
3478
3479         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3480
3481         /*
3482          * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3483          * so that the VFS cache i_size is consistent with the btree i_size:
3484          */
3485         if (ret &&
3486             !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)))
3487                 return ret;
3488
3489         if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3490                 end = inode->v.i_size;
3491
3492         if (end >= inode->v.i_size &&
3493             (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3494              !(mode & FALLOC_FL_KEEP_SIZE))) {
3495                 spin_lock(&inode->v.i_lock);
3496                 i_size_write(&inode->v, end);
3497                 spin_unlock(&inode->v.i_lock);
3498
3499                 mutex_lock(&inode->ei_update_lock);
3500                 ret2 = bch2_write_inode_size(c, inode, end, 0);
3501                 mutex_unlock(&inode->ei_update_lock);
3502         }
3503
3504         return ret ?: ret2;
3505 }
3506
3507 long bch2_fallocate_dispatch(struct file *file, int mode,
3508                              loff_t offset, loff_t len)
3509 {
3510         struct bch_inode_info *inode = file_bch_inode(file);
3511         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3512         long ret;
3513
3514         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fallocate))
3515                 return -EROFS;
3516
3517         inode_lock(&inode->v);
3518         inode_dio_wait(&inode->v);
3519         bch2_pagecache_block_get(inode);
3520
3521         ret = file_modified(file);
3522         if (ret)
3523                 goto err;
3524
3525         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3526                 ret = bchfs_fallocate(inode, mode, offset, len);
3527         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3528                 ret = bchfs_fpunch(inode, offset, len);
3529         else if (mode == FALLOC_FL_INSERT_RANGE)
3530                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3531         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3532                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3533         else
3534                 ret = -EOPNOTSUPP;
3535 err:
3536         bch2_pagecache_block_put(inode);
3537         inode_unlock(&inode->v);
3538         bch2_write_ref_put(c, BCH_WRITE_REF_fallocate);
3539
3540         return bch2_err_class(ret);
3541 }
3542
3543 /*
3544  * Take a quota reservation for unallocated blocks in a given file range
3545  * Does not check pagecache
3546  */
3547 static int quota_reserve_range(struct bch_inode_info *inode,
3548                                struct quota_res *res,
3549                                u64 start, u64 end)
3550 {
3551         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3552         struct btree_trans trans;
3553         struct btree_iter iter;
3554         struct bkey_s_c k;
3555         u32 snapshot;
3556         u64 sectors = end - start;
3557         u64 pos = start;
3558         int ret;
3559
3560         bch2_trans_init(&trans, c, 0, 0);
3561 retry:
3562         bch2_trans_begin(&trans);
3563
3564         ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot);
3565         if (ret)
3566                 goto err;
3567
3568         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3569                              SPOS(inode->v.i_ino, pos, snapshot), 0);
3570
3571         while (!(ret = btree_trans_too_many_iters(&trans)) &&
3572                (k = bch2_btree_iter_peek_upto(&iter, POS(inode->v.i_ino, end - 1))).k &&
3573                !(ret = bkey_err(k))) {
3574                 if (bkey_extent_is_allocation(k.k)) {
3575                         u64 s = min(end, k.k->p.offset) -
3576                                 max(start, bkey_start_offset(k.k));
3577                         BUG_ON(s > sectors);
3578                         sectors -= s;
3579                 }
3580                 bch2_btree_iter_advance(&iter);
3581         }
3582         pos = iter.pos.offset;
3583         bch2_trans_iter_exit(&trans, &iter);
3584 err:
3585         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3586                 goto retry;
3587
3588         bch2_trans_exit(&trans);
3589
3590         if (ret)
3591                 return ret;
3592
3593         return bch2_quota_reservation_add(c, inode, res, sectors, true);
3594 }
3595
3596 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3597                              struct file *file_dst, loff_t pos_dst,
3598                              loff_t len, unsigned remap_flags)
3599 {
3600         struct bch_inode_info *src = file_bch_inode(file_src);
3601         struct bch_inode_info *dst = file_bch_inode(file_dst);
3602         struct bch_fs *c = src->v.i_sb->s_fs_info;
3603         struct quota_res quota_res = { 0 };
3604         s64 i_sectors_delta = 0;
3605         u64 aligned_len;
3606         loff_t ret = 0;
3607
3608         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3609                 return -EINVAL;
3610
3611         if (remap_flags & REMAP_FILE_DEDUP)
3612                 return -EOPNOTSUPP;
3613
3614         if ((pos_src & (block_bytes(c) - 1)) ||
3615             (pos_dst & (block_bytes(c) - 1)))
3616                 return -EINVAL;
3617
3618         if (src == dst &&
3619             abs(pos_src - pos_dst) < len)
3620                 return -EINVAL;
3621
3622         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3623
3624         inode_dio_wait(&src->v);
3625         inode_dio_wait(&dst->v);
3626
3627         ret = generic_remap_file_range_prep(file_src, pos_src,
3628                                             file_dst, pos_dst,
3629                                             &len, remap_flags);
3630         if (ret < 0 || len == 0)
3631                 goto err;
3632
3633         aligned_len = round_up((u64) len, block_bytes(c));
3634
3635         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3636                                 pos_dst, pos_dst + len - 1);
3637         if (ret)
3638                 goto err;
3639
3640         ret = quota_reserve_range(dst, &quota_res, pos_dst >> 9,
3641                                   (pos_dst + aligned_len) >> 9);
3642         if (ret)
3643                 goto err;
3644
3645         file_update_time(file_dst);
3646
3647         mark_pagecache_unallocated(src, pos_src >> 9,
3648                                    (pos_src + aligned_len) >> 9);
3649
3650         ret = bch2_remap_range(c,
3651                                inode_inum(dst), pos_dst >> 9,
3652                                inode_inum(src), pos_src >> 9,
3653                                aligned_len >> 9,
3654                                pos_dst + len, &i_sectors_delta);
3655         if (ret < 0)
3656                 goto err;
3657
3658         /*
3659          * due to alignment, we might have remapped slightly more than requsted
3660          */
3661         ret = min((u64) ret << 9, (u64) len);
3662
3663         i_sectors_acct(c, dst, &quota_res, i_sectors_delta);
3664
3665         spin_lock(&dst->v.i_lock);
3666         if (pos_dst + ret > dst->v.i_size)
3667                 i_size_write(&dst->v, pos_dst + ret);
3668         spin_unlock(&dst->v.i_lock);
3669
3670         if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3671             IS_SYNC(file_inode(file_dst)))
3672                 ret = bch2_flush_inode(c, dst);
3673 err:
3674         bch2_quota_reservation_put(c, dst, &quota_res);
3675         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3676
3677         return bch2_err_class(ret);
3678 }
3679
3680 /* fseek: */
3681
3682 static int folio_data_offset(struct folio *folio, loff_t pos)
3683 {
3684         struct bch_folio *s = bch2_folio(folio);
3685         unsigned i, sectors = folio_sectors(folio);
3686
3687         if (s)
3688                 for (i = folio_pos_to_s(folio, pos); i < sectors; i++)
3689                         if (s->s[i].state >= SECTOR_dirty)
3690                                 return i << SECTOR_SHIFT;
3691
3692         return -1;
3693 }
3694
3695 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3696                                        loff_t start_offset,
3697                                        loff_t end_offset)
3698 {
3699         struct folio_batch fbatch;
3700         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
3701         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
3702         pgoff_t index           = start_index;
3703         unsigned i;
3704         loff_t ret;
3705         int offset;
3706
3707         folio_batch_init(&fbatch);
3708
3709         while (filemap_get_folios(vinode->i_mapping,
3710                                   &index, end_index, &fbatch)) {
3711                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
3712                         struct folio *folio = fbatch.folios[i];
3713
3714                         folio_lock(folio);
3715                         offset = folio_data_offset(folio,
3716                                         max(folio_pos(folio), start_offset));
3717                         if (offset >= 0) {
3718                                 ret = clamp(folio_pos(folio) + offset,
3719                                             start_offset, end_offset);
3720                                 folio_unlock(folio);
3721                                 folio_batch_release(&fbatch);
3722                                 return ret;
3723                         }
3724                         folio_unlock(folio);
3725                 }
3726                 folio_batch_release(&fbatch);
3727                 cond_resched();
3728         }
3729
3730         return end_offset;
3731 }
3732
3733 static loff_t bch2_seek_data(struct file *file, u64 offset)
3734 {
3735         struct bch_inode_info *inode = file_bch_inode(file);
3736         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3737         struct btree_trans trans;
3738         struct btree_iter iter;
3739         struct bkey_s_c k;
3740         subvol_inum inum = inode_inum(inode);
3741         u64 isize, next_data = MAX_LFS_FILESIZE;
3742         u32 snapshot;
3743         int ret;
3744
3745         isize = i_size_read(&inode->v);
3746         if (offset >= isize)
3747                 return -ENXIO;
3748
3749         bch2_trans_init(&trans, c, 0, 0);
3750 retry:
3751         bch2_trans_begin(&trans);
3752
3753         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3754         if (ret)
3755                 goto err;
3756
3757         for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents,
3758                            SPOS(inode->v.i_ino, offset >> 9, snapshot),
3759                            POS(inode->v.i_ino, U64_MAX),
3760                            0, k, ret) {
3761                 if (bkey_extent_is_data(k.k)) {
3762                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3763                         break;
3764                 } else if (k.k->p.offset >> 9 > isize)
3765                         break;
3766         }
3767         bch2_trans_iter_exit(&trans, &iter);
3768 err:
3769         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3770                 goto retry;
3771
3772         bch2_trans_exit(&trans);
3773         if (ret)
3774                 return ret;
3775
3776         if (next_data > offset)
3777                 next_data = bch2_seek_pagecache_data(&inode->v,
3778                                                      offset, next_data);
3779
3780         if (next_data >= isize)
3781                 return -ENXIO;
3782
3783         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3784 }
3785
3786 static bool folio_hole_offset(struct address_space *mapping, loff_t *offset)
3787 {
3788         struct folio *folio;
3789         struct bch_folio *s;
3790         unsigned i, sectors;
3791         bool ret = true;
3792
3793         folio = filemap_lock_folio(mapping, *offset >> PAGE_SHIFT);
3794         if (IS_ERR_OR_NULL(folio))
3795                 return true;
3796
3797         s = bch2_folio(folio);
3798         if (!s)
3799                 goto unlock;
3800
3801         sectors = folio_sectors(folio);
3802         for (i = folio_pos_to_s(folio, *offset); i < sectors; i++)
3803                 if (s->s[i].state < SECTOR_dirty) {
3804                         *offset = max(*offset,
3805                                       folio_pos(folio) + (i << SECTOR_SHIFT));
3806                         goto unlock;
3807                 }
3808
3809         *offset = folio_end_pos(folio);
3810         ret = false;
3811 unlock:
3812         folio_unlock(folio);
3813         return ret;
3814 }
3815
3816 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3817                                        loff_t start_offset,
3818                                        loff_t end_offset)
3819 {
3820         struct address_space *mapping = vinode->i_mapping;
3821         loff_t offset = start_offset;
3822
3823         while (offset < end_offset &&
3824                !folio_hole_offset(mapping, &offset))
3825                 ;
3826
3827         return min(offset, end_offset);
3828 }
3829
3830 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3831 {
3832         struct bch_inode_info *inode = file_bch_inode(file);
3833         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3834         struct btree_trans trans;
3835         struct btree_iter iter;
3836         struct bkey_s_c k;
3837         subvol_inum inum = inode_inum(inode);
3838         u64 isize, next_hole = MAX_LFS_FILESIZE;
3839         u32 snapshot;
3840         int ret;
3841
3842         isize = i_size_read(&inode->v);
3843         if (offset >= isize)
3844                 return -ENXIO;
3845
3846         bch2_trans_init(&trans, c, 0, 0);
3847 retry:
3848         bch2_trans_begin(&trans);
3849
3850         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3851         if (ret)
3852                 goto err;
3853
3854         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3855                            SPOS(inode->v.i_ino, offset >> 9, snapshot),
3856                            BTREE_ITER_SLOTS, k, ret) {
3857                 if (k.k->p.inode != inode->v.i_ino) {
3858                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3859                                         offset, MAX_LFS_FILESIZE);
3860                         break;
3861                 } else if (!bkey_extent_is_data(k.k)) {
3862                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3863                                         max(offset, bkey_start_offset(k.k) << 9),
3864                                         k.k->p.offset << 9);
3865
3866                         if (next_hole < k.k->p.offset << 9)
3867                                 break;
3868                 } else {
3869                         offset = max(offset, bkey_start_offset(k.k) << 9);
3870                 }
3871         }
3872         bch2_trans_iter_exit(&trans, &iter);
3873 err:
3874         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3875                 goto retry;
3876
3877         bch2_trans_exit(&trans);
3878         if (ret)
3879                 return ret;
3880
3881         if (next_hole > isize)
3882                 next_hole = isize;
3883
3884         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3885 }
3886
3887 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3888 {
3889         loff_t ret;
3890
3891         switch (whence) {
3892         case SEEK_SET:
3893         case SEEK_CUR:
3894         case SEEK_END:
3895                 ret = generic_file_llseek(file, offset, whence);
3896                 break;
3897         case SEEK_DATA:
3898                 ret = bch2_seek_data(file, offset);
3899                 break;
3900         case SEEK_HOLE:
3901                 ret = bch2_seek_hole(file, offset);
3902                 break;
3903         default:
3904                 ret = -EINVAL;
3905                 break;
3906         }
3907
3908         return bch2_err_class(ret);
3909 }
3910
3911 void bch2_fs_fsio_exit(struct bch_fs *c)
3912 {
3913         bioset_exit(&c->nocow_flush_bioset);
3914         bioset_exit(&c->dio_write_bioset);
3915         bioset_exit(&c->dio_read_bioset);
3916         bioset_exit(&c->writepage_bioset);
3917 }
3918
3919 int bch2_fs_fsio_init(struct bch_fs *c)
3920 {
3921         int ret = 0;
3922
3923         pr_verbose_init(c->opts, "");
3924
3925         if (bioset_init(&c->writepage_bioset,
3926                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3927                         BIOSET_NEED_BVECS))
3928                 return -BCH_ERR_ENOMEM_writepage_bioset_init;
3929
3930         if (bioset_init(&c->dio_read_bioset,
3931                         4, offsetof(struct dio_read, rbio.bio),
3932                         BIOSET_NEED_BVECS))
3933                 return -BCH_ERR_ENOMEM_dio_read_bioset_init;
3934
3935         if (bioset_init(&c->dio_write_bioset,
3936                         4, offsetof(struct dio_write, op.wbio.bio),
3937                         BIOSET_NEED_BVECS))
3938                 return -BCH_ERR_ENOMEM_dio_write_bioset_init;
3939
3940         if (bioset_init(&c->nocow_flush_bioset,
3941                         1, offsetof(struct nocow_flush, bio), 0))
3942                 return -BCH_ERR_ENOMEM_nocow_flush_bioset_init;
3943
3944         pr_verbose_init(c->opts, "ret %i", ret);
3945         return ret;
3946 }
3947
3948 #endif /* NO_BCACHEFS_FS */