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