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