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 static inline bool bio_full(struct bio *bio, unsigned len)
  39 {
  40         if (bio->bi_vcnt >= bio->bi_max_vecs)
  41                 return true;
  42         if (bio->bi_iter.bi_size > UINT_MAX - len)
  43                 return true;
  44         return false;
  45 }
  46
  47 static inline struct address_space *faults_disabled_mapping(void)
  48 {
  49         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  50 }
  51
  52 static inline void set_fdm_dropped_locks(void)
  53 {
  54         current->faults_disabled_mapping =
  55                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  56 }
  57
  58 static inline bool fdm_dropped_locks(void)
  59 {
  60         return ((unsigned long) current->faults_disabled_mapping) & 1;
  61 }
  62
  63 struct quota_res {
  64         u64                             sectors;
  65 };
  66
  67 struct bch_writepage_io {
  68         struct closure                  cl;
  69         struct bch_inode_info           *inode;
  70
  71         /* must be last: */
  72         struct bch_write_op             op;
  73 };
  74
  75 struct dio_write {
  76         struct completion               done;
  77         struct kiocb                    *req;
  78         struct mm_struct                *mm;
  79         unsigned                        loop:1,
  80                                         sync:1,
  81                                         free_iov:1;
  82         struct quota_res                quota_res;
  83         u64                             written;
  84
  85         struct iov_iter                 iter;
  86         struct iovec                    inline_vecs[2];
  87
  88         /* must be last: */
  89         struct bch_write_op             op;
  90 };
  91
  92 struct dio_read {
  93         struct closure                  cl;
  94         struct kiocb                    *req;
  95         long                            ret;
  96         bool                            should_dirty;
  97         struct bch_read_bio             rbio;
  98 };
  99
 100 /* pagecache_block must be held */
 101 static int write_invalidate_inode_pages_range(struct address_space *mapping,
 102                                               loff_t start, loff_t end)
 103 {
 104         int ret;
 105
 106         /*
 107          * XXX: the way this is currently implemented, we can spin if a process
 108          * is continually redirtying a specific page
 109          */
 110         do {
 111                 if (!mapping->nrpages)
 112                         return 0;
 113
 114                 ret = filemap_write_and_wait_range(mapping, start, end);
 115                 if (ret)
 116                         break;
 117
 118                 if (!mapping->nrpages)
 119                         return 0;
 120
 121                 ret = invalidate_inode_pages2_range(mapping,
 122                                 start >> PAGE_SHIFT,
 123                                 end >> PAGE_SHIFT);
 124         } while (ret == -EBUSY);
 125
 126         return ret;
 127 }
 128
 129 /* quotas */
 130
 131 #ifdef CONFIG_BCACHEFS_QUOTA
 132
 133 static void bch2_quota_reservation_put(struct bch_fs *c,
 134                                        struct bch_inode_info *inode,
 135                                        struct quota_res *res)
 136 {
 137         if (!res->sectors)
 138                 return;
 139
 140         mutex_lock(&inode->ei_quota_lock);
 141         BUG_ON(res->sectors > inode->ei_quota_reserved);
 142
 143         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 144                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 145         inode->ei_quota_reserved -= res->sectors;
 146         mutex_unlock(&inode->ei_quota_lock);
 147
 148         res->sectors = 0;
 149 }
 150
 151 static int bch2_quota_reservation_add(struct bch_fs *c,
 152                                       struct bch_inode_info *inode,
 153                                       struct quota_res *res,
 154                                       u64 sectors,
 155                                       bool check_enospc)
 156 {
 157         int ret;
 158
 159         mutex_lock(&inode->ei_quota_lock);
 160         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 161                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 162         if (likely(!ret)) {
 163                 inode->ei_quota_reserved += sectors;
 164                 res->sectors += sectors;
 165         }
 166         mutex_unlock(&inode->ei_quota_lock);
 167
 168         return ret;
 169 }
 170
 171 #else
 172
 173 static void bch2_quota_reservation_put(struct bch_fs *c,
 174                                        struct bch_inode_info *inode,
 175                                        struct quota_res *res)
 176 {
 177 }
 178
 179 static int bch2_quota_reservation_add(struct bch_fs *c,
 180                                       struct bch_inode_info *inode,
 181                                       struct quota_res *res,
 182                                       unsigned sectors,
 183                                       bool check_enospc)
 184 {
 185         return 0;
 186 }
 187
 188 #endif
 189
 190 /* i_size updates: */
 191
 192 struct inode_new_size {
 193         loff_t          new_size;
 194         u64             now;
 195         unsigned        fields;
 196 };
 197
 198 static int inode_set_size(struct bch_inode_info *inode,
 199                           struct bch_inode_unpacked *bi,
 200                           void *p)
 201 {
 202         struct inode_new_size *s = p;
 203
 204         bi->bi_size = s->new_size;
 205         if (s->fields & ATTR_ATIME)
 206                 bi->bi_atime = s->now;
 207         if (s->fields & ATTR_MTIME)
 208                 bi->bi_mtime = s->now;
 209         if (s->fields & ATTR_CTIME)
 210                 bi->bi_ctime = s->now;
 211
 212         return 0;
 213 }
 214
 215 int __must_check bch2_write_inode_size(struct bch_fs *c,
 216                                        struct bch_inode_info *inode,
 217                                        loff_t new_size, unsigned fields)
 218 {
 219         struct inode_new_size s = {
 220                 .new_size       = new_size,
 221                 .now            = bch2_current_time(c),
 222                 .fields         = fields,
 223         };
 224
 225         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 226 }
 227
 228 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 229                            struct quota_res *quota_res, s64 sectors)
 230 {
 231         if (!sectors)
 232                 return;
 233
 234         mutex_lock(&inode->ei_quota_lock);
 235         bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
 236                                 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
 237                                 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
 238                                 inode->ei_inode.bi_sectors);
 239         inode->v.i_blocks += sectors;
 240
 241 #ifdef CONFIG_BCACHEFS_QUOTA
 242         if (quota_res && sectors > 0) {
 243                 BUG_ON(sectors > quota_res->sectors);
 244                 BUG_ON(sectors > inode->ei_quota_reserved);
 245
 246                 quota_res->sectors -= sectors;
 247                 inode->ei_quota_reserved -= sectors;
 248         } else {
 249                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 250         }
 251 #endif
 252         mutex_unlock(&inode->ei_quota_lock);
 253 }
 254
 255 /* page state: */
 256
 257 /* stored in page->private: */
 258
 259 struct bch_page_sector {
 260         /* Uncompressed, fully allocated replicas (or on disk reservation): */
 261         unsigned                nr_replicas:4;
 262
 263         /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
 264         unsigned                replicas_reserved:4;
 265
 266         /* i_sectors: */
 267         enum {
 268                 SECTOR_UNALLOCATED,
 269                 SECTOR_RESERVED,
 270                 SECTOR_DIRTY,
 271                 SECTOR_DIRTY_RESERVED,
 272                 SECTOR_ALLOCATED,
 273         }                       state:8;
 274 };
 275
 276 struct bch_page_state {
 277         spinlock_t              lock;
 278         atomic_t                write_count;
 279         bool                    uptodate;
 280         struct bch_page_sector  s[PAGE_SECTORS];
 281 };
 282
 283 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 284 {
 285         return page_has_private(page)
 286                 ? (struct bch_page_state *) page_private(page)
 287                 : NULL;
 288 }
 289
 290 static inline struct bch_page_state *bch2_page_state(struct page *page)
 291 {
 292         EBUG_ON(!PageLocked(page));
 293
 294         return __bch2_page_state(page);
 295 }
 296
 297 /* for newly allocated pages: */
 298 static void __bch2_page_state_release(struct page *page)
 299 {
 300         kfree(detach_page_private(page));
 301 }
 302
 303 static void bch2_page_state_release(struct page *page)
 304 {
 305         EBUG_ON(!PageLocked(page));
 306         __bch2_page_state_release(page);
 307 }
 308
 309 /* for newly allocated pages: */
 310 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 311                                                        gfp_t gfp)
 312 {
 313         struct bch_page_state *s;
 314
 315         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 316         if (!s)
 317                 return NULL;
 318
 319         spin_lock_init(&s->lock);
 320         attach_page_private(page, s);
 321         return s;
 322 }
 323
 324 static struct bch_page_state *bch2_page_state_create(struct page *page,
 325                                                      gfp_t gfp)
 326 {
 327         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 328 }
 329
 330 static unsigned bkey_to_sector_state(const struct bkey *k)
 331 {
 332         if (k->type == KEY_TYPE_reservation)
 333                 return SECTOR_RESERVED;
 334         if (bkey_extent_is_allocation(k))
 335                 return SECTOR_ALLOCATED;
 336         return SECTOR_UNALLOCATED;
 337 }
 338
 339 static void __bch2_page_state_set(struct page *page,
 340                                   unsigned pg_offset, unsigned pg_len,
 341                                   unsigned nr_ptrs, unsigned state)
 342 {
 343         struct bch_page_state *s = bch2_page_state_create(page, __GFP_NOFAIL);
 344         unsigned i;
 345
 346         BUG_ON(pg_offset >= PAGE_SECTORS);
 347         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 348
 349         spin_lock(&s->lock);
 350
 351         for (i = pg_offset; i < pg_offset + pg_len; i++) {
 352                 s->s[i].nr_replicas = nr_ptrs;
 353                 s->s[i].state = state;
 354         }
 355
 356         if (i == PAGE_SECTORS)
 357                 s->uptodate = true;
 358
 359         spin_unlock(&s->lock);
 360 }
 361
 362 static int bch2_page_state_set(struct bch_fs *c, subvol_inum inum,
 363                                struct page **pages, unsigned nr_pages)
 364 {
 365         struct btree_trans trans;
 366         struct btree_iter iter;
 367         struct bkey_s_c k;
 368         u64 offset = pages[0]->index << PAGE_SECTORS_SHIFT;
 369         unsigned pg_idx = 0;
 370         u32 snapshot;
 371         int ret;
 372
 373         bch2_trans_init(&trans, c, 0, 0);
 374 retry:
 375         bch2_trans_begin(&trans);
 376
 377         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 378         if (ret)
 379                 goto err;
 380
 381         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
 382                            SPOS(inum.inum, offset, snapshot),
 383                            BTREE_ITER_SLOTS, k, ret) {
 384                 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
 385                 unsigned state = bkey_to_sector_state(k.k);
 386
 387                 while (pg_idx < nr_pages) {
 388                         struct page *page = pages[pg_idx];
 389                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 390                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 391                         unsigned pg_offset = max(bkey_start_offset(k.k), pg_start) - pg_start;
 392                         unsigned pg_len = min(k.k->p.offset, pg_end) - pg_offset - pg_start;
 393
 394                         BUG_ON(k.k->p.offset < pg_start);
 395                         BUG_ON(bkey_start_offset(k.k) > pg_end);
 396
 397                         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate)
 398                                 __bch2_page_state_set(page, pg_offset, pg_len, nr_ptrs, state);
 399
 400                         if (k.k->p.offset < pg_end)
 401                                 break;
 402                         pg_idx++;
 403                 }
 404
 405                 if (pg_idx == nr_pages)
 406                         break;
 407         }
 408
 409         offset = iter.pos.offset;
 410         bch2_trans_iter_exit(&trans, &iter);
 411 err:
 412         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 413                 goto retry;
 414         bch2_trans_exit(&trans);
 415
 416         return ret;
 417 }
 418
 419 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
 420 {
 421         struct bvec_iter iter;
 422         struct bio_vec bv;
 423         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 424                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 425         unsigned state = bkey_to_sector_state(k.k);
 426
 427         bio_for_each_segment(bv, bio, iter)
 428                 __bch2_page_state_set(bv.bv_page, bv.bv_offset >> 9,
 429                                       bv.bv_len >> 9, nr_ptrs, state);
 430 }
 431
 432 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
 433                                        u64 start, u64 end)
 434 {
 435         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 436         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 437         struct folio_batch fbatch;
 438         unsigned i, j;
 439
 440         if (end <= start)
 441                 return;
 442
 443         folio_batch_init(&fbatch);
 444
 445         while (filemap_get_folios(inode->v.i_mapping,
 446                                   &index, end_index, &fbatch)) {
 447                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 448                         struct folio *folio = fbatch.folios[i];
 449                         u64 pg_start = folio->index << PAGE_SECTORS_SHIFT;
 450                         u64 pg_end = (folio->index + 1) << PAGE_SECTORS_SHIFT;
 451                         unsigned pg_offset = max(start, pg_start) - pg_start;
 452                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 453                         struct bch_page_state *s;
 454
 455                         BUG_ON(end <= pg_start);
 456                         BUG_ON(pg_offset >= PAGE_SECTORS);
 457                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 458
 459                         folio_lock(folio);
 460                         s = bch2_page_state(&folio->page);
 461
 462                         if (s) {
 463                                 spin_lock(&s->lock);
 464                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 465                                         s->s[j].nr_replicas = 0;
 466                                 spin_unlock(&s->lock);
 467                         }
 468
 469                         folio_unlock(folio);
 470                 }
 471                 folio_batch_release(&fbatch);
 472                 cond_resched();
 473         }
 474 }
 475
 476 static void mark_pagecache_reserved(struct bch_inode_info *inode,
 477                                     u64 start, u64 end)
 478 {
 479         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 480         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 481         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 482         struct folio_batch fbatch;
 483         s64 i_sectors_delta = 0;
 484         unsigned i, j;
 485
 486         if (end <= start)
 487                 return;
 488
 489         folio_batch_init(&fbatch);
 490
 491         while (filemap_get_folios(inode->v.i_mapping,
 492                                   &index, end_index, &fbatch)) {
 493                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 494                         struct folio *folio = fbatch.folios[i];
 495                         u64 pg_start = folio->index << PAGE_SECTORS_SHIFT;
 496                         u64 pg_end = (folio->index + 1) << PAGE_SECTORS_SHIFT;
 497                         unsigned pg_offset = max(start, pg_start) - pg_start;
 498                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 499                         struct bch_page_state *s;
 500
 501                         BUG_ON(end <= pg_start);
 502                         BUG_ON(pg_offset >= PAGE_SECTORS);
 503                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 504
 505                         folio_lock(folio);
 506                         s = bch2_page_state(&folio->page);
 507
 508                         if (s) {
 509                                 spin_lock(&s->lock);
 510                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 511                                         switch (s->s[j].state) {
 512                                         case SECTOR_UNALLOCATED:
 513                                                 s->s[j].state = SECTOR_RESERVED;
 514                                                 break;
 515                                         case SECTOR_DIRTY:
 516                                                 s->s[j].state = SECTOR_DIRTY_RESERVED;
 517                                                 i_sectors_delta--;
 518                                                 break;
 519                                         default:
 520                                                 break;
 521                                         }
 522                                 spin_unlock(&s->lock);
 523                         }
 524
 525                         folio_unlock(folio);
 526                 }
 527                 folio_batch_release(&fbatch);
 528                 cond_resched();
 529         }
 530
 531         i_sectors_acct(c, inode, NULL, i_sectors_delta);
 532 }
 533
 534 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 535 {
 536         /* XXX: this should not be open coded */
 537         return inode->ei_inode.bi_data_replicas
 538                 ? inode->ei_inode.bi_data_replicas - 1
 539                 : c->opts.data_replicas;
 540 }
 541
 542 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 543                                                   unsigned nr_replicas)
 544 {
 545         return max(0, (int) nr_replicas -
 546                    s->nr_replicas -
 547                    s->replicas_reserved);
 548 }
 549
 550 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 551                                 struct bch_inode_info *inode,
 552                                 struct page *page, bool check_enospc)
 553 {
 554         struct bch_page_state *s = bch2_page_state_create(page, 0);
 555         unsigned nr_replicas = inode_nr_replicas(c, inode);
 556         struct disk_reservation disk_res = { 0 };
 557         unsigned i, disk_res_sectors = 0;
 558         int ret;
 559
 560         if (!s)
 561                 return -ENOMEM;
 562
 563         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 564                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 565
 566         if (!disk_res_sectors)
 567                 return 0;
 568
 569         ret = bch2_disk_reservation_get(c, &disk_res,
 570                                         disk_res_sectors, 1,
 571                                         !check_enospc
 572                                         ? BCH_DISK_RESERVATION_NOFAIL
 573                                         : 0);
 574         if (unlikely(ret))
 575                 return ret;
 576
 577         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 578                 s->s[i].replicas_reserved +=
 579                         sectors_to_reserve(&s->s[i], nr_replicas);
 580
 581         return 0;
 582 }
 583
 584 struct bch2_page_reservation {
 585         struct disk_reservation disk;
 586         struct quota_res        quota;
 587 };
 588
 589 static void bch2_page_reservation_init(struct bch_fs *c,
 590                         struct bch_inode_info *inode,
 591                         struct bch2_page_reservation *res)
 592 {
 593         memset(res, 0, sizeof(*res));
 594
 595         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 596 }
 597
 598 static void bch2_page_reservation_put(struct bch_fs *c,
 599                         struct bch_inode_info *inode,
 600                         struct bch2_page_reservation *res)
 601 {
 602         bch2_disk_reservation_put(c, &res->disk);
 603         bch2_quota_reservation_put(c, inode, &res->quota);
 604 }
 605
 606 static int bch2_page_reservation_get(struct bch_fs *c,
 607                         struct bch_inode_info *inode, struct page *page,
 608                         struct bch2_page_reservation *res,
 609                         unsigned offset, unsigned len)
 610 {
 611         struct bch_page_state *s = bch2_page_state_create(page, 0);
 612         unsigned i, disk_sectors = 0, quota_sectors = 0;
 613         int ret;
 614
 615         if (!s)
 616                 return -ENOMEM;
 617
 618         BUG_ON(!s->uptodate);
 619
 620         for (i = round_down(offset, block_bytes(c)) >> 9;
 621              i < round_up(offset + len, block_bytes(c)) >> 9;
 622              i++) {
 623                 disk_sectors += sectors_to_reserve(&s->s[i],
 624                                                 res->disk.nr_replicas);
 625                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 626         }
 627
 628         if (disk_sectors) {
 629                 ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
 630                 if (unlikely(ret))
 631                         return ret;
 632         }
 633
 634         if (quota_sectors) {
 635                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 636                                                  quota_sectors, true);
 637                 if (unlikely(ret)) {
 638                         struct disk_reservation tmp = {
 639                                 .sectors = disk_sectors
 640                         };
 641
 642                         bch2_disk_reservation_put(c, &tmp);
 643                         res->disk.sectors -= disk_sectors;
 644                         return ret;
 645                 }
 646         }
 647
 648         return 0;
 649 }
 650
 651 static void bch2_clear_page_bits(struct page *page)
 652 {
 653         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 654         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 655         struct bch_page_state *s = bch2_page_state(page);
 656         struct disk_reservation disk_res = { 0 };
 657         int i, dirty_sectors = 0;
 658
 659         if (!s)
 660                 return;
 661
 662         EBUG_ON(!PageLocked(page));
 663         EBUG_ON(PageWriteback(page));
 664
 665         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 666                 disk_res.sectors += s->s[i].replicas_reserved;
 667                 s->s[i].replicas_reserved = 0;
 668
 669                 switch (s->s[i].state) {
 670                 case SECTOR_DIRTY:
 671                         s->s[i].state = SECTOR_UNALLOCATED;
 672                         --dirty_sectors;
 673                         break;
 674                 case SECTOR_DIRTY_RESERVED:
 675                         s->s[i].state = SECTOR_RESERVED;
 676                         break;
 677                 default:
 678                         break;
 679                 }
 680         }
 681
 682         bch2_disk_reservation_put(c, &disk_res);
 683
 684         i_sectors_acct(c, inode, NULL, dirty_sectors);
 685
 686         bch2_page_state_release(page);
 687 }
 688
 689 static void bch2_set_page_dirty(struct bch_fs *c,
 690                         struct bch_inode_info *inode, struct page *page,
 691                         struct bch2_page_reservation *res,
 692                         unsigned offset, unsigned len)
 693 {
 694         struct bch_page_state *s = bch2_page_state(page);
 695         unsigned i, dirty_sectors = 0;
 696
 697         WARN_ON((u64) page_offset(page) + offset + len >
 698                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 699
 700         spin_lock(&s->lock);
 701
 702         for (i = round_down(offset, block_bytes(c)) >> 9;
 703              i < round_up(offset + len, block_bytes(c)) >> 9;
 704              i++) {
 705                 unsigned sectors = sectors_to_reserve(&s->s[i],
 706                                                 res->disk.nr_replicas);
 707
 708                 /*
 709                  * This can happen if we race with the error path in
 710                  * bch2_writepage_io_done():
 711                  */
 712                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 713
 714                 s->s[i].replicas_reserved += sectors;
 715                 res->disk.sectors -= sectors;
 716
 717                 switch (s->s[i].state) {
 718                 case SECTOR_UNALLOCATED:
 719                         s->s[i].state = SECTOR_DIRTY;
 720                         dirty_sectors++;
 721                         break;
 722                 case SECTOR_RESERVED:
 723                         s->s[i].state = SECTOR_DIRTY_RESERVED;
 724                         break;
 725                 default:
 726                         break;
 727                 }
 728         }
 729
 730         spin_unlock(&s->lock);
 731
 732         i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 733
 734         if (!PageDirty(page))
 735                 __set_page_dirty_nobuffers(page);
 736 }
 737
 738 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 739 {
 740         struct file *file = vmf->vma->vm_file;
 741         struct address_space *mapping = file->f_mapping;
 742         struct address_space *fdm = faults_disabled_mapping();
 743         struct bch_inode_info *inode = file_bch_inode(file);
 744         int ret;
 745
 746         if (fdm == mapping)
 747                 return VM_FAULT_SIGBUS;
 748
 749         /* Lock ordering: */
 750         if (fdm > mapping) {
 751                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 752
 753                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 754                         goto got_lock;
 755
 756                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 757
 758                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 759                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 760
 761                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 762
 763                 /* Signal that lock has been dropped: */
 764                 set_fdm_dropped_locks();
 765                 return VM_FAULT_SIGBUS;
 766         }
 767
 768         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 769 got_lock:
 770         ret = filemap_fault(vmf);
 771         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 772
 773         return ret;
 774 }
 775
 776 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 777 {
 778         struct page *page = vmf->page;
 779         struct file *file = vmf->vma->vm_file;
 780         struct bch_inode_info *inode = file_bch_inode(file);
 781         struct address_space *mapping = file->f_mapping;
 782         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 783         struct bch2_page_reservation res;
 784         unsigned len;
 785         loff_t isize;
 786         int ret;
 787
 788         bch2_page_reservation_init(c, inode, &res);
 789
 790         sb_start_pagefault(inode->v.i_sb);
 791         file_update_time(file);
 792
 793         /*
 794          * Not strictly necessary, but helps avoid dio writes livelocking in
 795          * write_invalidate_inode_pages_range() - can drop this if/when we get
 796          * a write_invalidate_inode_pages_range() that works without dropping
 797          * page lock before invalidating page
 798          */
 799         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 800
 801         lock_page(page);
 802         isize = i_size_read(&inode->v);
 803
 804         if (page->mapping != mapping || page_offset(page) >= isize) {
 805                 unlock_page(page);
 806                 ret = VM_FAULT_NOPAGE;
 807                 goto out;
 808         }
 809
 810         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 811
 812         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
 813                 if (bch2_page_state_set(c, inode_inum(inode), &page, 1)) {
 814                         unlock_page(page);
 815                         ret = VM_FAULT_SIGBUS;
 816                         goto out;
 817                 }
 818         }
 819
 820         if (bch2_page_reservation_get(c, inode, page, &res, 0, len)) {
 821                 unlock_page(page);
 822                 ret = VM_FAULT_SIGBUS;
 823                 goto out;
 824         }
 825
 826         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 827         bch2_page_reservation_put(c, inode, &res);
 828
 829         wait_for_stable_page(page);
 830         ret = VM_FAULT_LOCKED;
 831 out:
 832         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 833         sb_end_pagefault(inode->v.i_sb);
 834
 835         return ret;
 836 }
 837
 838 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
 839 {
 840         if (offset || length < folio_size(folio))
 841                 return;
 842
 843         bch2_clear_page_bits(&folio->page);
 844 }
 845
 846 bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
 847 {
 848         if (folio_test_dirty(folio) || folio_test_writeback(folio))
 849                 return false;
 850
 851         bch2_clear_page_bits(&folio->page);
 852         return true;
 853 }
 854
 855 /* readpage(s): */
 856
 857 static void bch2_readpages_end_io(struct bio *bio)
 858 {
 859         struct bvec_iter_all iter;
 860         struct bio_vec *bv;
 861
 862         bio_for_each_segment_all(bv, bio, iter) {
 863                 struct page *page = bv->bv_page;
 864
 865                 if (!bio->bi_status) {
 866                         SetPageUptodate(page);
 867                 } else {
 868                         ClearPageUptodate(page);
 869                         SetPageError(page);
 870                 }
 871                 unlock_page(page);
 872         }
 873
 874         bio_put(bio);
 875 }
 876
 877 struct readpages_iter {
 878         struct address_space    *mapping;
 879         struct page             **pages;
 880         unsigned                nr_pages;
 881         unsigned                idx;
 882         pgoff_t                 offset;
 883 };
 884
 885 static int readpages_iter_init(struct readpages_iter *iter,
 886                                struct readahead_control *ractl)
 887 {
 888         unsigned i, nr_pages = readahead_count(ractl);
 889
 890         memset(iter, 0, sizeof(*iter));
 891
 892         iter->mapping   = ractl->mapping;
 893         iter->offset    = readahead_index(ractl);
 894         iter->nr_pages  = nr_pages;
 895
 896         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 897         if (!iter->pages)
 898                 return -ENOMEM;
 899
 900         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 901         for (i = 0; i < nr_pages; i++) {
 902                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 903                 put_page(iter->pages[i]);
 904         }
 905
 906         return 0;
 907 }
 908
 909 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 910 {
 911         if (iter->idx >= iter->nr_pages)
 912                 return NULL;
 913
 914         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 915
 916         return iter->pages[iter->idx];
 917 }
 918
 919 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 920 {
 921         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 922         struct bch_extent_crc_unpacked crc;
 923         const union bch_extent_entry *i;
 924
 925         bkey_for_each_crc(k.k, ptrs, crc, i)
 926                 if (crc.csum_type || crc.compression_type)
 927                         return true;
 928         return false;
 929 }
 930
 931 static void readpage_bio_extend(struct readpages_iter *iter,
 932                                 struct bio *bio,
 933                                 unsigned sectors_this_extent,
 934                                 bool get_more)
 935 {
 936         while (bio_sectors(bio) < sectors_this_extent &&
 937                bio->bi_vcnt < bio->bi_max_vecs) {
 938                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
 939                 struct page *page = readpage_iter_next(iter);
 940                 int ret;
 941
 942                 if (page) {
 943                         if (iter->offset + iter->idx != page_offset)
 944                                 break;
 945
 946                         iter->idx++;
 947                 } else {
 948                         if (!get_more)
 949                                 break;
 950
 951                         page = xa_load(&iter->mapping->i_pages, page_offset);
 952                         if (page && !xa_is_value(page))
 953                                 break;
 954
 955                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 956                         if (!page)
 957                                 break;
 958
 959                         if (!__bch2_page_state_create(page, 0)) {
 960                                 put_page(page);
 961                                 break;
 962                         }
 963
 964                         ret = add_to_page_cache_lru(page, iter->mapping,
 965                                                     page_offset, GFP_NOFS);
 966                         if (ret) {
 967                                 __bch2_page_state_release(page);
 968                                 put_page(page);
 969                                 break;
 970                         }
 971
 972                         put_page(page);
 973                 }
 974
 975                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 976         }
 977 }
 978
 979 static void bchfs_read(struct btree_trans *trans,
 980                        struct bch_read_bio *rbio,
 981                        subvol_inum inum,
 982                        struct readpages_iter *readpages_iter)
 983 {
 984         struct bch_fs *c = trans->c;
 985         struct btree_iter iter;
 986         struct bkey_buf sk;
 987         int flags = BCH_READ_RETRY_IF_STALE|
 988                 BCH_READ_MAY_PROMOTE;
 989         u32 snapshot;
 990         int ret = 0;
 991
 992         rbio->c = c;
 993         rbio->start_time = local_clock();
 994         rbio->subvol = inum.subvol;
 995
 996         bch2_bkey_buf_init(&sk);
 997 retry:
 998         bch2_trans_begin(trans);
 999         iter = (struct btree_iter) { NULL };
1000
1001         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1002         if (ret)
1003                 goto err;
1004
1005         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1006                              SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1007                              BTREE_ITER_SLOTS);
1008         while (1) {
1009                 struct bkey_s_c k;
1010                 unsigned bytes, sectors, offset_into_extent;
1011                 enum btree_id data_btree = BTREE_ID_extents;
1012
1013                 /*
1014                  * read_extent -> io_time_reset may cause a transaction restart
1015                  * without returning an error, we need to check for that here:
1016                  */
1017                 ret = bch2_trans_relock(trans);
1018                 if (ret)
1019                         break;
1020
1021                 bch2_btree_iter_set_pos(&iter,
1022                                 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1023
1024                 k = bch2_btree_iter_peek_slot(&iter);
1025                 ret = bkey_err(k);
1026                 if (ret)
1027                         break;
1028
1029                 offset_into_extent = iter.pos.offset -
1030                         bkey_start_offset(k.k);
1031                 sectors = k.k->size - offset_into_extent;
1032
1033                 bch2_bkey_buf_reassemble(&sk, c, k);
1034
1035                 ret = bch2_read_indirect_extent(trans, &data_btree,
1036                                         &offset_into_extent, &sk);
1037                 if (ret)
1038                         break;
1039
1040                 k = bkey_i_to_s_c(sk.k);
1041
1042                 sectors = min(sectors, k.k->size - offset_into_extent);
1043
1044                 if (readpages_iter)
1045                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
1046                                             extent_partial_reads_expensive(k));
1047
1048                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1049                 swap(rbio->bio.bi_iter.bi_size, bytes);
1050
1051                 if (rbio->bio.bi_iter.bi_size == bytes)
1052                         flags |= BCH_READ_LAST_FRAGMENT;
1053
1054                 bch2_bio_page_state_set(&rbio->bio, k);
1055
1056                 bch2_read_extent(trans, rbio, iter.pos,
1057                                  data_btree, k, offset_into_extent, flags);
1058
1059                 if (flags & BCH_READ_LAST_FRAGMENT)
1060                         break;
1061
1062                 swap(rbio->bio.bi_iter.bi_size, bytes);
1063                 bio_advance(&rbio->bio, bytes);
1064
1065                 ret = btree_trans_too_many_iters(trans);
1066                 if (ret)
1067                         break;
1068         }
1069 err:
1070         bch2_trans_iter_exit(trans, &iter);
1071
1072         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1073                 goto retry;
1074
1075         if (ret) {
1076                 bch_err_inum_ratelimited(c, inum.inum,
1077                                 "read error %i from btree lookup", ret);
1078                 rbio->bio.bi_status = BLK_STS_IOERR;
1079                 bio_endio(&rbio->bio);
1080         }
1081
1082         bch2_bkey_buf_exit(&sk, c);
1083 }
1084
1085 void bch2_readahead(struct readahead_control *ractl)
1086 {
1087         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1088         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1089         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1090         struct btree_trans trans;
1091         struct page *page;
1092         struct readpages_iter readpages_iter;
1093         int ret;
1094
1095         ret = readpages_iter_init(&readpages_iter, ractl);
1096         BUG_ON(ret);
1097
1098         bch2_trans_init(&trans, c, 0, 0);
1099
1100         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1101
1102         while ((page = readpage_iter_next(&readpages_iter))) {
1103                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1104                 unsigned n = min_t(unsigned,
1105                                    readpages_iter.nr_pages -
1106                                    readpages_iter.idx,
1107                                    BIO_MAX_VECS);
1108                 struct bch_read_bio *rbio =
1109                         rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
1110                                                    GFP_NOFS, &c->bio_read),
1111                                   opts);
1112
1113                 readpages_iter.idx++;
1114
1115                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTORS_SHIFT;
1116                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1117                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1118
1119                 bchfs_read(&trans, rbio, inode_inum(inode),
1120                            &readpages_iter);
1121         }
1122
1123         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1124
1125         bch2_trans_exit(&trans);
1126         kfree(readpages_iter.pages);
1127 }
1128
1129 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1130                              subvol_inum inum, struct page *page)
1131 {
1132         struct btree_trans trans;
1133
1134         bch2_page_state_create(page, __GFP_NOFAIL);
1135
1136         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1137         rbio->bio.bi_iter.bi_sector =
1138                 (sector_t) page->index << PAGE_SECTORS_SHIFT;
1139         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1140
1141         bch2_trans_init(&trans, c, 0, 0);
1142         bchfs_read(&trans, rbio, inum, NULL);
1143         bch2_trans_exit(&trans);
1144 }
1145
1146 static void bch2_read_single_page_end_io(struct bio *bio)
1147 {
1148         complete(bio->bi_private);
1149 }
1150
1151 static int bch2_read_single_page(struct page *page,
1152                                  struct address_space *mapping)
1153 {
1154         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1155         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1156         struct bch_read_bio *rbio;
1157         int ret;
1158         DECLARE_COMPLETION_ONSTACK(done);
1159
1160         rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS, &c->bio_read),
1161                          io_opts(c, &inode->ei_inode));
1162         rbio->bio.bi_private = &done;
1163         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1164
1165         __bchfs_readpage(c, rbio, inode_inum(inode), page);
1166         wait_for_completion(&done);
1167
1168         ret = blk_status_to_errno(rbio->bio.bi_status);
1169         bio_put(&rbio->bio);
1170
1171         if (ret < 0)
1172                 return ret;
1173
1174         SetPageUptodate(page);
1175         return 0;
1176 }
1177
1178 int bch2_read_folio(struct file *file, struct folio *folio)
1179 {
1180         struct page *page = &folio->page;
1181         int ret;
1182
1183         ret = bch2_read_single_page(page, page->mapping);
1184         folio_unlock(folio);
1185         return bch2_err_class(ret);
1186 }
1187
1188 /* writepages: */
1189
1190 struct bch_writepage_state {
1191         struct bch_writepage_io *io;
1192         struct bch_io_opts      opts;
1193 };
1194
1195 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1196                                                                   struct bch_inode_info *inode)
1197 {
1198         return (struct bch_writepage_state) {
1199                 .opts = io_opts(c, &inode->ei_inode)
1200         };
1201 }
1202
1203 static void bch2_writepage_io_free(struct closure *cl)
1204 {
1205         struct bch_writepage_io *io = container_of(cl,
1206                                         struct bch_writepage_io, cl);
1207
1208         bio_put(&io->op.wbio.bio);
1209 }
1210
1211 static void bch2_writepage_io_done(struct closure *cl)
1212 {
1213         struct bch_writepage_io *io = container_of(cl,
1214                                         struct bch_writepage_io, cl);
1215         struct bch_fs *c = io->op.c;
1216         struct bio *bio = &io->op.wbio.bio;
1217         struct bvec_iter_all iter;
1218         struct bio_vec *bvec;
1219         unsigned i;
1220
1221         if (io->op.error) {
1222                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1223
1224                 bio_for_each_segment_all(bvec, bio, iter) {
1225                         struct bch_page_state *s;
1226
1227                         SetPageError(bvec->bv_page);
1228                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1229
1230                         s = __bch2_page_state(bvec->bv_page);
1231                         spin_lock(&s->lock);
1232                         for (i = 0; i < PAGE_SECTORS; i++)
1233                                 s->s[i].nr_replicas = 0;
1234                         spin_unlock(&s->lock);
1235                 }
1236         }
1237
1238         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1239                 bio_for_each_segment_all(bvec, bio, iter) {
1240                         struct bch_page_state *s;
1241
1242                         s = __bch2_page_state(bvec->bv_page);
1243                         spin_lock(&s->lock);
1244                         for (i = 0; i < PAGE_SECTORS; i++)
1245                                 s->s[i].nr_replicas = 0;
1246                         spin_unlock(&s->lock);
1247                 }
1248         }
1249
1250         /*
1251          * racing with fallocate can cause us to add fewer sectors than
1252          * expected - but we shouldn't add more sectors than expected:
1253          */
1254         WARN_ON_ONCE(io->op.i_sectors_delta > 0);
1255
1256         /*
1257          * (error (due to going RO) halfway through a page can screw that up
1258          * slightly)
1259          * XXX wtf?
1260            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1261          */
1262
1263         /*
1264          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1265          * before calling end_page_writeback:
1266          */
1267         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1268
1269         bio_for_each_segment_all(bvec, bio, iter) {
1270                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1271
1272                 if (atomic_dec_and_test(&s->write_count))
1273                         end_page_writeback(bvec->bv_page);
1274         }
1275
1276         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1277 }
1278
1279 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1280 {
1281         struct bch_writepage_io *io = w->io;
1282
1283         w->io = NULL;
1284         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1285         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1286 }
1287
1288 /*
1289  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1290  * possible, else allocating a new one:
1291  */
1292 static void bch2_writepage_io_alloc(struct bch_fs *c,
1293                                     struct writeback_control *wbc,
1294                                     struct bch_writepage_state *w,
1295                                     struct bch_inode_info *inode,
1296                                     u64 sector,
1297                                     unsigned nr_replicas)
1298 {
1299         struct bch_write_op *op;
1300
1301         w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1302                                               REQ_OP_WRITE,
1303                                               GFP_NOFS,
1304                                               &c->writepage_bioset),
1305                              struct bch_writepage_io, op.wbio.bio);
1306
1307         closure_init(&w->io->cl, NULL);
1308         w->io->inode            = inode;
1309
1310         op                      = &w->io->op;
1311         bch2_write_op_init(op, c, w->opts);
1312         op->target              = w->opts.foreground_target;
1313         op->nr_replicas         = nr_replicas;
1314         op->res.nr_replicas     = nr_replicas;
1315         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1316         op->subvol              = inode->ei_subvol;
1317         op->pos                 = POS(inode->v.i_ino, sector);
1318         op->wbio.bio.bi_iter.bi_sector = sector;
1319         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1320 }
1321
1322 static int __bch2_writepage(struct page *page,
1323                             struct writeback_control *wbc,
1324                             void *data)
1325 {
1326         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1327         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1328         struct bch_writepage_state *w = data;
1329         struct bch_page_state *s, orig;
1330         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1331         loff_t i_size = i_size_read(&inode->v);
1332         pgoff_t end_index = i_size >> PAGE_SHIFT;
1333         int ret;
1334
1335         EBUG_ON(!PageUptodate(page));
1336
1337         /* Is the page fully inside i_size? */
1338         if (page->index < end_index)
1339                 goto do_io;
1340
1341         /* Is the page fully outside i_size? (truncate in progress) */
1342         offset = i_size & (PAGE_SIZE - 1);
1343         if (page->index > end_index || !offset) {
1344                 unlock_page(page);
1345                 return 0;
1346         }
1347
1348         /*
1349          * The page straddles i_size.  It must be zeroed out on each and every
1350          * writepage invocation because it may be mmapped.  "A file is mapped
1351          * in multiples of the page size.  For a file that is not a multiple of
1352          * the  page size, the remaining memory is zeroed when mapped, and
1353          * writes to that region are not written out to the file."
1354          */
1355         zero_user_segment(page, offset, PAGE_SIZE);
1356 do_io:
1357         s = bch2_page_state_create(page, __GFP_NOFAIL);
1358
1359         /*
1360          * Things get really hairy with errors during writeback:
1361          */
1362         ret = bch2_get_page_disk_reservation(c, inode, page, false);
1363         BUG_ON(ret);
1364
1365         /* Before unlocking the page, get copy of reservations: */
1366         spin_lock(&s->lock);
1367         orig = *s;
1368         spin_unlock(&s->lock);
1369
1370         for (i = 0; i < PAGE_SECTORS; i++) {
1371                 if (s->s[i].state < SECTOR_DIRTY)
1372                         continue;
1373
1374                 nr_replicas_this_write =
1375                         min_t(unsigned, nr_replicas_this_write,
1376                               s->s[i].nr_replicas +
1377                               s->s[i].replicas_reserved);
1378         }
1379
1380         for (i = 0; i < PAGE_SECTORS; i++) {
1381                 if (s->s[i].state < SECTOR_DIRTY)
1382                         continue;
1383
1384                 s->s[i].nr_replicas = w->opts.compression
1385                         ? 0 : nr_replicas_this_write;
1386
1387                 s->s[i].replicas_reserved = 0;
1388                 s->s[i].state = SECTOR_ALLOCATED;
1389         }
1390
1391         BUG_ON(atomic_read(&s->write_count));
1392         atomic_set(&s->write_count, 1);
1393
1394         BUG_ON(PageWriteback(page));
1395         set_page_writeback(page);
1396
1397         unlock_page(page);
1398
1399         offset = 0;
1400         while (1) {
1401                 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1402                 u64 sector;
1403
1404                 while (offset < PAGE_SECTORS &&
1405                        orig.s[offset].state < SECTOR_DIRTY)
1406                         offset++;
1407
1408                 if (offset == PAGE_SECTORS)
1409                         break;
1410
1411                 while (offset + sectors < PAGE_SECTORS &&
1412                        orig.s[offset + sectors].state >= SECTOR_DIRTY) {
1413                         reserved_sectors += orig.s[offset + sectors].replicas_reserved;
1414                         dirty_sectors += orig.s[offset + sectors].state == SECTOR_DIRTY;
1415                         sectors++;
1416                 }
1417                 BUG_ON(!sectors);
1418
1419                 sector = ((u64) page->index << PAGE_SECTORS_SHIFT) + offset;
1420
1421                 if (w->io &&
1422                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1423                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1424                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1425                      (BIO_MAX_VECS * PAGE_SIZE) ||
1426                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1427                         bch2_writepage_do_io(w);
1428
1429                 if (!w->io)
1430                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1431                                                 nr_replicas_this_write);
1432
1433                 atomic_inc(&s->write_count);
1434
1435                 BUG_ON(inode != w->io->inode);
1436                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1437                                      sectors << 9, offset << 9));
1438
1439                 /* Check for writing past i_size: */
1440                 WARN_ON_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1441                              round_up(i_size, block_bytes(c)));
1442
1443                 w->io->op.res.sectors += reserved_sectors;
1444                 w->io->op.i_sectors_delta -= dirty_sectors;
1445                 w->io->op.new_i_size = i_size;
1446
1447                 offset += sectors;
1448         }
1449
1450         if (atomic_dec_and_test(&s->write_count))
1451                 end_page_writeback(page);
1452
1453         return 0;
1454 }
1455
1456 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1457 {
1458         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1459         struct bch_writepage_state w =
1460                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1461         struct blk_plug plug;
1462         int ret;
1463
1464         blk_start_plug(&plug);
1465         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1466         if (w.io)
1467                 bch2_writepage_do_io(&w);
1468         blk_finish_plug(&plug);
1469         return bch2_err_class(ret);
1470 }
1471
1472 /* buffered writes: */
1473
1474 int bch2_write_begin(struct file *file, struct address_space *mapping,
1475                      loff_t pos, unsigned len,
1476                      struct page **pagep, void **fsdata)
1477 {
1478         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1479         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1480         struct bch2_page_reservation *res;
1481         pgoff_t index = pos >> PAGE_SHIFT;
1482         unsigned offset = pos & (PAGE_SIZE - 1);
1483         struct page *page;
1484         int ret = -ENOMEM;
1485
1486         res = kmalloc(sizeof(*res), GFP_KERNEL);
1487         if (!res)
1488                 return -ENOMEM;
1489
1490         bch2_page_reservation_init(c, inode, res);
1491         *fsdata = res;
1492
1493         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1494
1495         page = grab_cache_page_write_begin(mapping, index);
1496         if (!page)
1497                 goto err_unlock;
1498
1499         if (PageUptodate(page))
1500                 goto out;
1501
1502         /* If we're writing entire page, don't need to read it in first: */
1503         if (len == PAGE_SIZE)
1504                 goto out;
1505
1506         if (!offset && pos + len >= inode->v.i_size) {
1507                 zero_user_segment(page, len, PAGE_SIZE);
1508                 flush_dcache_page(page);
1509                 goto out;
1510         }
1511
1512         if (index > inode->v.i_size >> PAGE_SHIFT) {
1513                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1514                 flush_dcache_page(page);
1515                 goto out;
1516         }
1517 readpage:
1518         ret = bch2_read_single_page(page, mapping);
1519         if (ret)
1520                 goto err;
1521 out:
1522         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1523                 ret = bch2_page_state_set(c, inode_inum(inode), &page, 1);
1524                 if (ret)
1525                         goto err;
1526         }
1527
1528         ret = bch2_page_reservation_get(c, inode, page, res, offset, len);
1529         if (ret) {
1530                 if (!PageUptodate(page)) {
1531                         /*
1532                          * If the page hasn't been read in, we won't know if we
1533                          * actually need a reservation - we don't actually need
1534                          * to read here, we just need to check if the page is
1535                          * fully backed by uncompressed data:
1536                          */
1537                         goto readpage;
1538                 }
1539
1540                 goto err;
1541         }
1542
1543         *pagep = page;
1544         return 0;
1545 err:
1546         unlock_page(page);
1547         put_page(page);
1548         *pagep = NULL;
1549 err_unlock:
1550         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1551         kfree(res);
1552         *fsdata = NULL;
1553         return bch2_err_class(ret);
1554 }
1555
1556 int bch2_write_end(struct file *file, struct address_space *mapping,
1557                    loff_t pos, unsigned len, unsigned copied,
1558                    struct page *page, void *fsdata)
1559 {
1560         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1561         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1562         struct bch2_page_reservation *res = fsdata;
1563         unsigned offset = pos & (PAGE_SIZE - 1);
1564
1565         lockdep_assert_held(&inode->v.i_rwsem);
1566
1567         if (unlikely(copied < len && !PageUptodate(page))) {
1568                 /*
1569                  * The page needs to be read in, but that would destroy
1570                  * our partial write - simplest thing is to just force
1571                  * userspace to redo the write:
1572                  */
1573                 zero_user(page, 0, PAGE_SIZE);
1574                 flush_dcache_page(page);
1575                 copied = 0;
1576         }
1577
1578         spin_lock(&inode->v.i_lock);
1579         if (pos + copied > inode->v.i_size)
1580                 i_size_write(&inode->v, pos + copied);
1581         spin_unlock(&inode->v.i_lock);
1582
1583         if (copied) {
1584                 if (!PageUptodate(page))
1585                         SetPageUptodate(page);
1586
1587                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1588
1589                 inode->ei_last_dirtied = (unsigned long) current;
1590         }
1591
1592         unlock_page(page);
1593         put_page(page);
1594         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1595
1596         bch2_page_reservation_put(c, inode, res);
1597         kfree(res);
1598
1599         return copied;
1600 }
1601
1602 #define WRITE_BATCH_PAGES       32
1603
1604 static int __bch2_buffered_write(struct bch_inode_info *inode,
1605                                  struct address_space *mapping,
1606                                  struct iov_iter *iter,
1607                                  loff_t pos, unsigned len)
1608 {
1609         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1610         struct page *pages[WRITE_BATCH_PAGES];
1611         struct bch2_page_reservation res;
1612         unsigned long index = pos >> PAGE_SHIFT;
1613         unsigned offset = pos & (PAGE_SIZE - 1);
1614         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1615         unsigned i, reserved = 0, set_dirty = 0;
1616         unsigned copied = 0, nr_pages_copied = 0;
1617         int ret = 0;
1618
1619         BUG_ON(!len);
1620         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1621
1622         bch2_page_reservation_init(c, inode, &res);
1623
1624         for (i = 0; i < nr_pages; i++) {
1625                 pages[i] = grab_cache_page_write_begin(mapping, index + i);
1626                 if (!pages[i]) {
1627                         nr_pages = i;
1628                         if (!i) {
1629                                 ret = -ENOMEM;
1630                                 goto out;
1631                         }
1632                         len = min_t(unsigned, len,
1633                                     nr_pages * PAGE_SIZE - offset);
1634                         break;
1635                 }
1636         }
1637
1638         if (offset && !PageUptodate(pages[0])) {
1639                 ret = bch2_read_single_page(pages[0], mapping);
1640                 if (ret)
1641                         goto out;
1642         }
1643
1644         if ((pos + len) & (PAGE_SIZE - 1) &&
1645             !PageUptodate(pages[nr_pages - 1])) {
1646                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1647                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1648                 } else {
1649                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1650                         if (ret)
1651                                 goto out;
1652                 }
1653         }
1654
1655         while (reserved < len) {
1656                 unsigned i = (offset + reserved) >> PAGE_SHIFT;
1657                 struct page *page = pages[i];
1658                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1659                 unsigned pg_len = min_t(unsigned, len - reserved,
1660                                         PAGE_SIZE - pg_offset);
1661
1662                 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1663                         ret = bch2_page_state_set(c, inode_inum(inode),
1664                                                   pages + i, nr_pages - i);
1665                         if (ret)
1666                                 goto out;
1667                 }
1668
1669                 ret = bch2_page_reservation_get(c, inode, page, &res,
1670                                                 pg_offset, pg_len);
1671                 if (ret)
1672                         goto out;
1673
1674                 reserved += pg_len;
1675         }
1676
1677         if (mapping_writably_mapped(mapping))
1678                 for (i = 0; i < nr_pages; i++)
1679                         flush_dcache_page(pages[i]);
1680
1681         while (copied < len) {
1682                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1683                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1684                 unsigned pg_len = min_t(unsigned, len - copied,
1685                                         PAGE_SIZE - pg_offset);
1686                 unsigned pg_copied = copy_page_from_iter_atomic(page,
1687                                                 pg_offset, pg_len,iter);
1688
1689                 if (!pg_copied)
1690                         break;
1691
1692                 if (!PageUptodate(page) &&
1693                     pg_copied != PAGE_SIZE &&
1694                     pos + copied + pg_copied < inode->v.i_size) {
1695                         zero_user(page, 0, PAGE_SIZE);
1696                         break;
1697                 }
1698
1699                 flush_dcache_page(page);
1700                 copied += pg_copied;
1701
1702                 if (pg_copied != pg_len)
1703                         break;
1704         }
1705
1706         if (!copied)
1707                 goto out;
1708
1709         spin_lock(&inode->v.i_lock);
1710         if (pos + copied > inode->v.i_size)
1711                 i_size_write(&inode->v, pos + copied);
1712         spin_unlock(&inode->v.i_lock);
1713
1714         while (set_dirty < copied) {
1715                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1716                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1717                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1718                                         PAGE_SIZE - pg_offset);
1719
1720                 if (!PageUptodate(page))
1721                         SetPageUptodate(page);
1722
1723                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1724                 unlock_page(page);
1725                 put_page(page);
1726
1727                 set_dirty += pg_len;
1728         }
1729
1730         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1731         inode->ei_last_dirtied = (unsigned long) current;
1732 out:
1733         for (i = nr_pages_copied; i < nr_pages; i++) {
1734                 unlock_page(pages[i]);
1735                 put_page(pages[i]);
1736         }
1737
1738         bch2_page_reservation_put(c, inode, &res);
1739
1740         return copied ?: ret;
1741 }
1742
1743 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1744 {
1745         struct file *file = iocb->ki_filp;
1746         struct address_space *mapping = file->f_mapping;
1747         struct bch_inode_info *inode = file_bch_inode(file);
1748         loff_t pos = iocb->ki_pos;
1749         ssize_t written = 0;
1750         int ret = 0;
1751
1752         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1753
1754         do {
1755                 unsigned offset = pos & (PAGE_SIZE - 1);
1756                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1757                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1758 again:
1759                 /*
1760                  * Bring in the user page that we will copy from _first_.
1761                  * Otherwise there's a nasty deadlock on copying from the
1762                  * same page as we're writing to, without it being marked
1763                  * up-to-date.
1764                  *
1765                  * Not only is this an optimisation, but it is also required
1766                  * to check that the address is actually valid, when atomic
1767                  * usercopies are used, below.
1768                  */
1769                 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1770                         bytes = min_t(unsigned long, iov_iter_count(iter),
1771                                       PAGE_SIZE - offset);
1772
1773                         if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1774                                 ret = -EFAULT;
1775                                 break;
1776                         }
1777                 }
1778
1779                 if (unlikely(fatal_signal_pending(current))) {
1780                         ret = -EINTR;
1781                         break;
1782                 }
1783
1784                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1785                 if (unlikely(ret < 0))
1786                         break;
1787
1788                 cond_resched();
1789
1790                 if (unlikely(ret == 0)) {
1791                         /*
1792                          * If we were unable to copy any data at all, we must
1793                          * fall back to a single segment length write.
1794                          *
1795                          * If we didn't fallback here, we could livelock
1796                          * because not all segments in the iov can be copied at
1797                          * once without a pagefault.
1798                          */
1799                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1800                                       iov_iter_single_seg_count(iter));
1801                         goto again;
1802                 }
1803                 pos += ret;
1804                 written += ret;
1805                 ret = 0;
1806
1807                 balance_dirty_pages_ratelimited(mapping);
1808         } while (iov_iter_count(iter));
1809
1810         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1811
1812         return written ? written : ret;
1813 }
1814
1815 /* O_DIRECT reads */
1816
1817 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1818 {
1819         if (check_dirty) {
1820                 bio_check_pages_dirty(bio);
1821         } else {
1822                 bio_release_pages(bio, false);
1823                 bio_put(bio);
1824         }
1825 }
1826
1827 static void bch2_dio_read_complete(struct closure *cl)
1828 {
1829         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1830
1831         dio->req->ki_complete(dio->req, dio->ret);
1832         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1833 }
1834
1835 static void bch2_direct_IO_read_endio(struct bio *bio)
1836 {
1837         struct dio_read *dio = bio->bi_private;
1838
1839         if (bio->bi_status)
1840                 dio->ret = blk_status_to_errno(bio->bi_status);
1841
1842         closure_put(&dio->cl);
1843 }
1844
1845 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1846 {
1847         struct dio_read *dio = bio->bi_private;
1848         bool should_dirty = dio->should_dirty;
1849
1850         bch2_direct_IO_read_endio(bio);
1851         bio_check_or_release(bio, should_dirty);
1852 }
1853
1854 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1855 {
1856         struct file *file = req->ki_filp;
1857         struct bch_inode_info *inode = file_bch_inode(file);
1858         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1859         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1860         struct dio_read *dio;
1861         struct bio *bio;
1862         loff_t offset = req->ki_pos;
1863         bool sync = is_sync_kiocb(req);
1864         size_t shorten;
1865         ssize_t ret;
1866
1867         if ((offset|iter->count) & (block_bytes(c) - 1))
1868                 return -EINVAL;
1869
1870         ret = min_t(loff_t, iter->count,
1871                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1872
1873         if (!ret)
1874                 return ret;
1875
1876         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1877         iter->count -= shorten;
1878
1879         bio = bio_alloc_bioset(NULL,
1880                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1881                                REQ_OP_READ,
1882                                GFP_KERNEL,
1883                                &c->dio_read_bioset);
1884
1885         bio->bi_end_io = bch2_direct_IO_read_endio;
1886
1887         dio = container_of(bio, struct dio_read, rbio.bio);
1888         closure_init(&dio->cl, NULL);
1889
1890         /*
1891          * this is a _really_ horrible hack just to avoid an atomic sub at the
1892          * end:
1893          */
1894         if (!sync) {
1895                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1896                 atomic_set(&dio->cl.remaining,
1897                            CLOSURE_REMAINING_INITIALIZER -
1898                            CLOSURE_RUNNING +
1899                            CLOSURE_DESTRUCTOR);
1900         } else {
1901                 atomic_set(&dio->cl.remaining,
1902                            CLOSURE_REMAINING_INITIALIZER + 1);
1903         }
1904
1905         dio->req        = req;
1906         dio->ret        = ret;
1907         /*
1908          * This is one of the sketchier things I've encountered: we have to skip
1909          * the dirtying of requests that are internal from the kernel (i.e. from
1910          * loopback), because we'll deadlock on page_lock.
1911          */
1912         dio->should_dirty = iter_is_iovec(iter);
1913
1914         goto start;
1915         while (iter->count) {
1916                 bio = bio_alloc_bioset(NULL,
1917                                        bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1918                                        REQ_OP_READ,
1919                                        GFP_KERNEL,
1920                                        &c->bio_read);
1921                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1922 start:
1923                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1924                 bio->bi_iter.bi_sector  = offset >> 9;
1925                 bio->bi_private         = dio;
1926
1927                 ret = bio_iov_iter_get_pages(bio, iter);
1928                 if (ret < 0) {
1929                         /* XXX: fault inject this path */
1930                         bio->bi_status = BLK_STS_RESOURCE;
1931                         bio_endio(bio);
1932                         break;
1933                 }
1934
1935                 offset += bio->bi_iter.bi_size;
1936
1937                 if (dio->should_dirty)
1938                         bio_set_pages_dirty(bio);
1939
1940                 if (iter->count)
1941                         closure_get(&dio->cl);
1942
1943                 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
1944         }
1945
1946         iter->count += shorten;
1947
1948         if (sync) {
1949                 closure_sync(&dio->cl);
1950                 closure_debug_destroy(&dio->cl);
1951                 ret = dio->ret;
1952                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1953                 return ret;
1954         } else {
1955                 return -EIOCBQUEUED;
1956         }
1957 }
1958
1959 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1960 {
1961         struct file *file = iocb->ki_filp;
1962         struct bch_inode_info *inode = file_bch_inode(file);
1963         struct address_space *mapping = file->f_mapping;
1964         size_t count = iov_iter_count(iter);
1965         ssize_t ret;
1966
1967         if (!count)
1968                 return 0; /* skip atime */
1969
1970         if (iocb->ki_flags & IOCB_DIRECT) {
1971                 struct blk_plug plug;
1972
1973                 ret = filemap_write_and_wait_range(mapping,
1974                                         iocb->ki_pos,
1975                                         iocb->ki_pos + count - 1);
1976                 if (ret < 0)
1977                         goto out;
1978
1979                 file_accessed(file);
1980
1981                 blk_start_plug(&plug);
1982                 ret = bch2_direct_IO_read(iocb, iter);
1983                 blk_finish_plug(&plug);
1984
1985                 if (ret >= 0)
1986                         iocb->ki_pos += ret;
1987         } else {
1988                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1989                 ret = generic_file_read_iter(iocb, iter);
1990                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1991         }
1992 out:
1993         return bch2_err_class(ret);
1994 }
1995
1996 /* O_DIRECT writes */
1997
1998 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
1999                                        u64 offset, u64 size,
2000                                        unsigned nr_replicas, bool compressed)
2001 {
2002         struct btree_trans trans;
2003         struct btree_iter iter;
2004         struct bkey_s_c k;
2005         u64 end = offset + size;
2006         u32 snapshot;
2007         bool ret = true;
2008         int err;
2009
2010         bch2_trans_init(&trans, c, 0, 0);
2011 retry:
2012         bch2_trans_begin(&trans);
2013
2014         err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2015         if (err)
2016                 goto err;
2017
2018         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2019                            SPOS(inum.inum, offset, snapshot),
2020                            BTREE_ITER_SLOTS, k, err) {
2021                 if (bkey_cmp(bkey_start_pos(k.k), POS(inum.inum, end)) >= 0)
2022                         break;
2023
2024                 if (k.k->p.snapshot != snapshot ||
2025                     nr_replicas > bch2_bkey_replicas(c, k) ||
2026                     (!compressed && bch2_bkey_sectors_compressed(k))) {
2027                         ret = false;
2028                         break;
2029                 }
2030         }
2031
2032         offset = iter.pos.offset;
2033         bch2_trans_iter_exit(&trans, &iter);
2034 err:
2035         if (bch2_err_matches(err, BCH_ERR_transaction_restart))
2036                 goto retry;
2037         bch2_trans_exit(&trans);
2038
2039         return err ? false : ret;
2040 }
2041
2042 static void bch2_dio_write_loop_async(struct bch_write_op *);
2043
2044 static long bch2_dio_write_loop(struct dio_write *dio)
2045 {
2046         bool kthread = (current->flags & PF_KTHREAD) != 0;
2047         struct kiocb *req = dio->req;
2048         struct address_space *mapping = req->ki_filp->f_mapping;
2049         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
2050         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2051         struct bio *bio = &dio->op.wbio.bio;
2052         struct bvec_iter_all iter;
2053         struct bio_vec *bv;
2054         unsigned unaligned, iter_count;
2055         bool sync = dio->sync, dropped_locks;
2056         long ret;
2057
2058         if (dio->loop)
2059                 goto loop;
2060
2061         while (1) {
2062                 iter_count = dio->iter.count;
2063
2064                 if (kthread && dio->mm)
2065                         kthread_use_mm(dio->mm);
2066                 BUG_ON(current->faults_disabled_mapping);
2067                 current->faults_disabled_mapping = mapping;
2068
2069                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2070
2071                 dropped_locks = fdm_dropped_locks();
2072
2073                 current->faults_disabled_mapping = NULL;
2074                 if (kthread && dio->mm)
2075                         kthread_unuse_mm(dio->mm);
2076
2077                 /*
2078                  * If the fault handler returned an error but also signalled
2079                  * that it dropped & retook ei_pagecache_lock, we just need to
2080                  * re-shoot down the page cache and retry:
2081                  */
2082                 if (dropped_locks && ret)
2083                         ret = 0;
2084
2085                 if (unlikely(ret < 0))
2086                         goto err;
2087
2088                 if (unlikely(dropped_locks)) {
2089                         ret = write_invalidate_inode_pages_range(mapping,
2090                                         req->ki_pos,
2091                                         req->ki_pos + iter_count - 1);
2092                         if (unlikely(ret))
2093                                 goto err;
2094
2095                         if (!bio->bi_iter.bi_size)
2096                                 continue;
2097                 }
2098
2099                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2100                 bio->bi_iter.bi_size -= unaligned;
2101                 iov_iter_revert(&dio->iter, unaligned);
2102
2103                 if (!bio->bi_iter.bi_size) {
2104                         /*
2105                          * bio_iov_iter_get_pages was only able to get <
2106                          * blocksize worth of pages:
2107                          */
2108                         ret = -EFAULT;
2109                         goto err;
2110                 }
2111
2112                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
2113                 dio->op.end_io          = bch2_dio_write_loop_async;
2114                 dio->op.target          = dio->op.opts.foreground_target;
2115                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
2116                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
2117                 dio->op.subvol          = inode->ei_subvol;
2118                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2119
2120                 if ((req->ki_flags & IOCB_DSYNC) &&
2121                     !c->opts.journal_flush_disabled)
2122                         dio->op.flags |= BCH_WRITE_FLUSH;
2123                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2124
2125                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2126                                                 dio->op.opts.data_replicas, 0);
2127                 if (unlikely(ret) &&
2128                     !bch2_check_range_allocated(c, inode_inum(inode),
2129                                 dio->op.pos.offset, bio_sectors(bio),
2130                                 dio->op.opts.data_replicas,
2131                                 dio->op.opts.compression != 0))
2132                         goto err;
2133
2134                 task_io_account_write(bio->bi_iter.bi_size);
2135
2136                 if (!dio->sync && !dio->loop && dio->iter.count) {
2137                         struct iovec *iov = dio->inline_vecs;
2138
2139                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2140                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
2141                                               GFP_KERNEL);
2142                                 if (unlikely(!iov)) {
2143                                         dio->sync = sync = true;
2144                                         goto do_io;
2145                                 }
2146
2147                                 dio->free_iov = true;
2148                         }
2149
2150                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2151                         dio->iter.iov = iov;
2152                 }
2153 do_io:
2154                 dio->loop = true;
2155                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2156
2157                 if (sync)
2158                         wait_for_completion(&dio->done);
2159                 else
2160                         return -EIOCBQUEUED;
2161 loop:
2162                 i_sectors_acct(c, inode, &dio->quota_res,
2163                                dio->op.i_sectors_delta);
2164                 req->ki_pos += (u64) dio->op.written << 9;
2165                 dio->written += dio->op.written;
2166
2167                 spin_lock(&inode->v.i_lock);
2168                 if (req->ki_pos > inode->v.i_size)
2169                         i_size_write(&inode->v, req->ki_pos);
2170                 spin_unlock(&inode->v.i_lock);
2171
2172                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2173                         bio_for_each_segment_all(bv, bio, iter)
2174                                 put_page(bv->bv_page);
2175                 bio->bi_vcnt = 0;
2176
2177                 if (dio->op.error) {
2178                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
2179                         break;
2180                 }
2181
2182                 if (!dio->iter.count)
2183                         break;
2184
2185                 bio_reset(bio, NULL, REQ_OP_WRITE);
2186                 reinit_completion(&dio->done);
2187         }
2188
2189         ret = dio->op.error ?: ((long) dio->written << 9);
2190 err:
2191         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2192         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2193
2194         if (dio->free_iov)
2195                 kfree(dio->iter.iov);
2196
2197         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2198                 bio_for_each_segment_all(bv, bio, iter)
2199                         put_page(bv->bv_page);
2200         bio_put(bio);
2201
2202         /* inode->i_dio_count is our ref on inode and thus bch_fs */
2203         inode_dio_end(&inode->v);
2204
2205         if (ret < 0)
2206                 ret = bch2_err_class(ret);
2207
2208         if (!sync) {
2209                 req->ki_complete(req, ret);
2210                 ret = -EIOCBQUEUED;
2211         }
2212         return ret;
2213 }
2214
2215 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2216 {
2217         struct dio_write *dio = container_of(op, struct dio_write, op);
2218
2219         if (dio->sync)
2220                 complete(&dio->done);
2221         else
2222                 bch2_dio_write_loop(dio);
2223 }
2224
2225 static noinline
2226 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2227 {
2228         struct file *file = req->ki_filp;
2229         struct address_space *mapping = file->f_mapping;
2230         struct bch_inode_info *inode = file_bch_inode(file);
2231         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2232         struct dio_write *dio;
2233         struct bio *bio;
2234         bool locked = true, extending;
2235         ssize_t ret;
2236
2237         prefetch(&c->opts);
2238         prefetch((void *) &c->opts + 64);
2239         prefetch(&inode->ei_inode);
2240         prefetch((void *) &inode->ei_inode + 64);
2241
2242         inode_lock(&inode->v);
2243
2244         ret = generic_write_checks(req, iter);
2245         if (unlikely(ret <= 0))
2246                 goto err;
2247
2248         ret = file_remove_privs(file);
2249         if (unlikely(ret))
2250                 goto err;
2251
2252         ret = file_update_time(file);
2253         if (unlikely(ret))
2254                 goto err;
2255
2256         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2257                 goto err;
2258
2259         inode_dio_begin(&inode->v);
2260         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2261
2262         extending = req->ki_pos + iter->count > inode->v.i_size;
2263         if (!extending) {
2264                 inode_unlock(&inode->v);
2265                 locked = false;
2266         }
2267
2268         bio = bio_alloc_bioset(NULL,
2269                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2270                                REQ_OP_WRITE,
2271                                GFP_KERNEL,
2272                                &c->dio_write_bioset);
2273         dio = container_of(bio, struct dio_write, op.wbio.bio);
2274         init_completion(&dio->done);
2275         dio->req                = req;
2276         dio->mm                 = current->mm;
2277         dio->loop               = false;
2278         dio->sync               = is_sync_kiocb(req) || extending;
2279         dio->free_iov           = false;
2280         dio->quota_res.sectors  = 0;
2281         dio->written            = 0;
2282         dio->iter               = *iter;
2283
2284         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2285                                          iter->count >> 9, true);
2286         if (unlikely(ret))
2287                 goto err_put_bio;
2288
2289         ret = write_invalidate_inode_pages_range(mapping,
2290                                         req->ki_pos,
2291                                         req->ki_pos + iter->count - 1);
2292         if (unlikely(ret))
2293                 goto err_put_bio;
2294
2295         ret = bch2_dio_write_loop(dio);
2296 err:
2297         if (locked)
2298                 inode_unlock(&inode->v);
2299         return ret;
2300 err_put_bio:
2301         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2302         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2303         bio_put(bio);
2304         inode_dio_end(&inode->v);
2305         goto err;
2306 }
2307
2308 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2309 {
2310         struct file *file = iocb->ki_filp;
2311         struct bch_inode_info *inode = file_bch_inode(file);
2312         ssize_t ret;
2313
2314         if (iocb->ki_flags & IOCB_DIRECT) {
2315                 ret = bch2_direct_write(iocb, from);
2316                 goto out;
2317         }
2318
2319         /* We can write back this queue in page reclaim */
2320         current->backing_dev_info = inode_to_bdi(&inode->v);
2321         inode_lock(&inode->v);
2322
2323         ret = generic_write_checks(iocb, from);
2324         if (ret <= 0)
2325                 goto unlock;
2326
2327         ret = file_remove_privs(file);
2328         if (ret)
2329                 goto unlock;
2330
2331         ret = file_update_time(file);
2332         if (ret)
2333                 goto unlock;
2334
2335         ret = bch2_buffered_write(iocb, from);
2336         if (likely(ret > 0))
2337                 iocb->ki_pos += ret;
2338 unlock:
2339         inode_unlock(&inode->v);
2340         current->backing_dev_info = NULL;
2341
2342         if (ret > 0)
2343                 ret = generic_write_sync(iocb, ret);
2344 out:
2345         return bch2_err_class(ret);
2346 }
2347
2348 /* fsync: */
2349
2350 /*
2351  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2352  * insert trigger: look up the btree inode instead
2353  */
2354 static int bch2_flush_inode(struct bch_fs *c, subvol_inum inum)
2355 {
2356         struct bch_inode_unpacked inode;
2357         int ret;
2358
2359         if (c->opts.journal_flush_disabled)
2360                 return 0;
2361
2362         ret = bch2_inode_find_by_inum(c, inum, &inode);
2363         if (ret)
2364                 return ret;
2365
2366         return bch2_journal_flush_seq(&c->journal, inode.bi_journal_seq);
2367 }
2368
2369 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2370 {
2371         struct bch_inode_info *inode = file_bch_inode(file);
2372         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2373         int ret, ret2, ret3;
2374
2375         ret = file_write_and_wait_range(file, start, end);
2376         ret2 = sync_inode_metadata(&inode->v, 1);
2377         ret3 = bch2_flush_inode(c, inode_inum(inode));
2378
2379         return bch2_err_class(ret ?: ret2 ?: ret3);
2380 }
2381
2382 /* truncate: */
2383
2384 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2385                                  struct bpos start,
2386                                  struct bpos end)
2387 {
2388         struct btree_trans trans;
2389         struct btree_iter iter;
2390         struct bkey_s_c k;
2391         int ret = 0;
2392
2393         bch2_trans_init(&trans, c, 0, 0);
2394 retry:
2395         bch2_trans_begin(&trans);
2396
2397         ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2398         if (ret)
2399                 goto err;
2400
2401         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2402                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2403                         break;
2404
2405                 if (bkey_extent_is_data(k.k)) {
2406                         ret = 1;
2407                         break;
2408                 }
2409         }
2410         start = iter.pos;
2411         bch2_trans_iter_exit(&trans, &iter);
2412 err:
2413         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
2414                 goto retry;
2415
2416         bch2_trans_exit(&trans);
2417         return ret;
2418 }
2419
2420 static int __bch2_truncate_page(struct bch_inode_info *inode,
2421                                 pgoff_t index, loff_t start, loff_t end)
2422 {
2423         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2424         struct address_space *mapping = inode->v.i_mapping;
2425         struct bch_page_state *s;
2426         unsigned start_offset = start & (PAGE_SIZE - 1);
2427         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2428         unsigned i;
2429         struct page *page;
2430         s64 i_sectors_delta = 0;
2431         int ret = 0;
2432
2433         /* Page boundary? Nothing to do */
2434         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2435               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2436                 return 0;
2437
2438         /* Above i_size? */
2439         if (index << PAGE_SHIFT >= inode->v.i_size)
2440                 return 0;
2441
2442         page = find_lock_page(mapping, index);
2443         if (!page) {
2444                 /*
2445                  * XXX: we're doing two index lookups when we end up reading the
2446                  * page
2447                  */
2448                 ret = range_has_data(c, inode->ei_subvol,
2449                                 POS(inode->v.i_ino, index << PAGE_SECTORS_SHIFT),
2450                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTORS_SHIFT));
2451                 if (ret <= 0)
2452                         return ret;
2453
2454                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2455                 if (unlikely(!page)) {
2456                         ret = -ENOMEM;
2457                         goto out;
2458                 }
2459         }
2460
2461         s = bch2_page_state_create(page, 0);
2462         if (!s) {
2463                 ret = -ENOMEM;
2464                 goto unlock;
2465         }
2466
2467         if (!PageUptodate(page)) {
2468                 ret = bch2_read_single_page(page, mapping);
2469                 if (ret)
2470                         goto unlock;
2471         }
2472
2473         if (index != start >> PAGE_SHIFT)
2474                 start_offset = 0;
2475         if (index != end >> PAGE_SHIFT)
2476                 end_offset = PAGE_SIZE;
2477
2478         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2479              i < round_down(end_offset, block_bytes(c)) >> 9;
2480              i++) {
2481                 s->s[i].nr_replicas     = 0;
2482                 if (s->s[i].state == SECTOR_DIRTY)
2483                         i_sectors_delta--;
2484                 s->s[i].state           = SECTOR_UNALLOCATED;
2485         }
2486
2487         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2488
2489         /*
2490          * Caller needs to know whether this page will be written out by
2491          * writeback - doing an i_size update if necessary - or whether it will
2492          * be responsible for the i_size update:
2493          */
2494         ret = s->s[(min_t(u64, inode->v.i_size - (index << PAGE_SHIFT),
2495                           PAGE_SIZE) - 1) >> 9].state >= SECTOR_DIRTY;
2496
2497         zero_user_segment(page, start_offset, end_offset);
2498
2499         /*
2500          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2501          *
2502          * XXX: because we aren't currently tracking whether the page has actual
2503          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2504          */
2505         BUG_ON(bch2_get_page_disk_reservation(c, inode, page, false));
2506
2507         /*
2508          * This removes any writeable userspace mappings; we need to force
2509          * .page_mkwrite to be called again before any mmapped writes, to
2510          * redirty the full page:
2511          */
2512         page_mkclean(page);
2513         __set_page_dirty_nobuffers(page);
2514 unlock:
2515         unlock_page(page);
2516         put_page(page);
2517 out:
2518         return ret;
2519 }
2520
2521 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2522 {
2523         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2524                                     from, round_up(from, PAGE_SIZE));
2525 }
2526
2527 static int bch2_truncate_pages(struct bch_inode_info *inode,
2528                                loff_t start, loff_t end)
2529 {
2530         int ret = __bch2_truncate_page(inode, start >> PAGE_SHIFT,
2531                                        start, end);
2532
2533         if (ret >= 0 &&
2534             start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2535                 ret = __bch2_truncate_page(inode,
2536                                            end >> PAGE_SHIFT,
2537                                            start, end);
2538         return ret;
2539 }
2540
2541 static int bch2_extend(struct user_namespace *mnt_userns,
2542                        struct bch_inode_info *inode,
2543                        struct bch_inode_unpacked *inode_u,
2544                        struct iattr *iattr)
2545 {
2546         struct address_space *mapping = inode->v.i_mapping;
2547         int ret;
2548
2549         /*
2550          * sync appends:
2551          *
2552          * this has to be done _before_ extending i_size:
2553          */
2554         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2555         if (ret)
2556                 return ret;
2557
2558         truncate_setsize(&inode->v, iattr->ia_size);
2559
2560         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2561 }
2562
2563 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2564                                    struct bch_inode_unpacked *bi,
2565                                    void *p)
2566 {
2567         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2568         return 0;
2569 }
2570
2571 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2572                                   struct bch_inode_unpacked *bi, void *p)
2573 {
2574         u64 *new_i_size = p;
2575
2576         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2577         bi->bi_size = *new_i_size;
2578         return 0;
2579 }
2580
2581 int bch2_truncate(struct user_namespace *mnt_userns,
2582                   struct bch_inode_info *inode, struct iattr *iattr)
2583 {
2584         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2585         struct address_space *mapping = inode->v.i_mapping;
2586         struct bch_inode_unpacked inode_u;
2587         u64 new_i_size = iattr->ia_size;
2588         s64 i_sectors_delta = 0;
2589         int ret = 0;
2590
2591         /*
2592          * If the truncate call with change the size of the file, the
2593          * cmtimes should be updated. If the size will not change, we
2594          * do not need to update the cmtimes.
2595          */
2596         if (iattr->ia_size != inode->v.i_size) {
2597                 if (!(iattr->ia_valid & ATTR_MTIME))
2598                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2599                 if (!(iattr->ia_valid & ATTR_CTIME))
2600                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2601                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2602         }
2603
2604         inode_dio_wait(&inode->v);
2605         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2606
2607         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
2608         if (ret)
2609                 goto err;
2610
2611         /*
2612          * check this before next assertion; on filesystem error our normal
2613          * invariants are a bit broken (truncate has to truncate the page cache
2614          * before the inode).
2615          */
2616         ret = bch2_journal_error(&c->journal);
2617         if (ret)
2618                 goto err;
2619
2620         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2621                 inode->v.i_size < inode_u.bi_size);
2622
2623         if (iattr->ia_size > inode->v.i_size) {
2624                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2625                 goto err;
2626         }
2627
2628         iattr->ia_valid &= ~ATTR_SIZE;
2629
2630         ret = bch2_truncate_page(inode, iattr->ia_size);
2631         if (unlikely(ret < 0))
2632                 goto err;
2633
2634         /*
2635          * When extending, we're going to write the new i_size to disk
2636          * immediately so we need to flush anything above the current on disk
2637          * i_size first:
2638          *
2639          * Also, when extending we need to flush the page that i_size currently
2640          * straddles - if it's mapped to userspace, we need to ensure that
2641          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2642          * again to allocate the part of the page that was extended.
2643          */
2644         if (iattr->ia_size > inode_u.bi_size)
2645                 ret = filemap_write_and_wait_range(mapping,
2646                                 inode_u.bi_size,
2647                                 iattr->ia_size - 1);
2648         else if (iattr->ia_size & (PAGE_SIZE - 1))
2649                 ret = filemap_write_and_wait_range(mapping,
2650                                 round_down(iattr->ia_size, PAGE_SIZE),
2651                                 iattr->ia_size - 1);
2652         if (ret)
2653                 goto err;
2654
2655         mutex_lock(&inode->ei_update_lock);
2656         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2657                                &new_i_size, 0);
2658         mutex_unlock(&inode->ei_update_lock);
2659
2660         if (unlikely(ret))
2661                 goto err;
2662
2663         truncate_setsize(&inode->v, iattr->ia_size);
2664
2665         ret = bch2_fpunch(c, inode_inum(inode),
2666                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2667                         U64_MAX, &i_sectors_delta);
2668         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2669
2670         bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
2671                                 !bch2_journal_error(&c->journal), c,
2672                                 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
2673                                 inode->v.i_ino, (u64) inode->v.i_blocks,
2674                                 inode->ei_inode.bi_sectors);
2675         if (unlikely(ret))
2676                 goto err;
2677
2678         mutex_lock(&inode->ei_update_lock);
2679         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2680         mutex_unlock(&inode->ei_update_lock);
2681
2682         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2683 err:
2684         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2685         return bch2_err_class(ret);
2686 }
2687
2688 /* fallocate: */
2689
2690 static int inode_update_times_fn(struct bch_inode_info *inode,
2691                                  struct bch_inode_unpacked *bi, void *p)
2692 {
2693         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2694
2695         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2696         return 0;
2697 }
2698
2699 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2700 {
2701         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2702         u64 end         = offset + len;
2703         u64 block_start = round_up(offset, block_bytes(c));
2704         u64 block_end   = round_down(end, block_bytes(c));
2705         bool truncated_last_page;
2706         int ret = 0;
2707
2708         ret = bch2_truncate_pages(inode, offset, end);
2709         if (unlikely(ret < 0))
2710                 goto err;
2711
2712         truncated_last_page = ret;
2713
2714         truncate_pagecache_range(&inode->v, offset, end - 1);
2715
2716         if (block_start < block_end ) {
2717                 s64 i_sectors_delta = 0;
2718
2719                 ret = bch2_fpunch(c, inode_inum(inode),
2720                                   block_start >> 9, block_end >> 9,
2721                                   &i_sectors_delta);
2722                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2723         }
2724
2725         mutex_lock(&inode->ei_update_lock);
2726         if (end >= inode->v.i_size && !truncated_last_page) {
2727                 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2728                                             ATTR_MTIME|ATTR_CTIME);
2729         } else {
2730                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2731                                        ATTR_MTIME|ATTR_CTIME);
2732         }
2733         mutex_unlock(&inode->ei_update_lock);
2734 err:
2735         return ret;
2736 }
2737
2738 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2739                                    loff_t offset, loff_t len,
2740                                    bool insert)
2741 {
2742         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2743         struct address_space *mapping = inode->v.i_mapping;
2744         struct bkey_buf copy;
2745         struct btree_trans trans;
2746         struct btree_iter src, dst, del;
2747         loff_t shift, new_size;
2748         u64 src_start;
2749         int ret = 0;
2750
2751         if ((offset | len) & (block_bytes(c) - 1))
2752                 return -EINVAL;
2753
2754         if (insert) {
2755                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2756                         return -EFBIG;
2757
2758                 if (offset >= inode->v.i_size)
2759                         return -EINVAL;
2760
2761                 src_start       = U64_MAX;
2762                 shift           = len;
2763         } else {
2764                 if (offset + len >= inode->v.i_size)
2765                         return -EINVAL;
2766
2767                 src_start       = offset + len;
2768                 shift           = -len;
2769         }
2770
2771         new_size = inode->v.i_size + shift;
2772
2773         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2774         if (ret)
2775                 return ret;
2776
2777         if (insert) {
2778                 i_size_write(&inode->v, new_size);
2779                 mutex_lock(&inode->ei_update_lock);
2780                 ret = bch2_write_inode_size(c, inode, new_size,
2781                                             ATTR_MTIME|ATTR_CTIME);
2782                 mutex_unlock(&inode->ei_update_lock);
2783         } else {
2784                 s64 i_sectors_delta = 0;
2785
2786                 ret = bch2_fpunch(c, inode_inum(inode),
2787                                   offset >> 9, (offset + len) >> 9,
2788                                   &i_sectors_delta);
2789                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2790
2791                 if (ret)
2792                         return ret;
2793         }
2794
2795         bch2_bkey_buf_init(&copy);
2796         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2797         bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
2798                         POS(inode->v.i_ino, src_start >> 9),
2799                         BTREE_ITER_INTENT);
2800         bch2_trans_copy_iter(&dst, &src);
2801         bch2_trans_copy_iter(&del, &src);
2802
2803         while (ret == 0 ||
2804                bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
2805                 struct disk_reservation disk_res =
2806                         bch2_disk_reservation_init(c, 0);
2807                 struct bkey_i delete;
2808                 struct bkey_s_c k;
2809                 struct bpos next_pos;
2810                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2811                 struct bpos atomic_end;
2812                 unsigned trigger_flags = 0;
2813                 u32 snapshot;
2814
2815                 bch2_trans_begin(&trans);
2816
2817                 ret = bch2_subvolume_get_snapshot(&trans,
2818                                         inode->ei_subvol, &snapshot);
2819                 if (ret)
2820                         continue;
2821
2822                 bch2_btree_iter_set_snapshot(&src, snapshot);
2823                 bch2_btree_iter_set_snapshot(&dst, snapshot);
2824                 bch2_btree_iter_set_snapshot(&del, snapshot);
2825
2826                 bch2_trans_begin(&trans);
2827
2828                 k = insert
2829                         ? bch2_btree_iter_peek_prev(&src)
2830                         : bch2_btree_iter_peek(&src);
2831                 if ((ret = bkey_err(k)))
2832                         continue;
2833
2834                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2835                         break;
2836
2837                 if (insert &&
2838                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2839                         break;
2840 reassemble:
2841                 bch2_bkey_buf_reassemble(&copy, c, k);
2842
2843                 if (insert &&
2844                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2845                         bch2_cut_front(move_pos, copy.k);
2846
2847                 copy.k->k.p.offset += shift >> 9;
2848                 bch2_btree_iter_set_pos(&dst, bkey_start_pos(&copy.k->k));
2849
2850                 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
2851                 if (ret)
2852                         continue;
2853
2854                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2855                         if (insert) {
2856                                 move_pos = atomic_end;
2857                                 move_pos.offset -= shift >> 9;
2858                                 goto reassemble;
2859                         } else {
2860                                 bch2_cut_back(atomic_end, copy.k);
2861                         }
2862                 }
2863
2864                 bkey_init(&delete.k);
2865                 delete.k.p = copy.k->k.p;
2866                 delete.k.size = copy.k->k.size;
2867                 delete.k.p.offset -= shift >> 9;
2868                 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
2869
2870                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2871
2872                 if (copy.k->k.size != k.k->size) {
2873                         /* We might end up splitting compressed extents: */
2874                         unsigned nr_ptrs =
2875                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2876
2877                         ret = bch2_disk_reservation_get(c, &disk_res,
2878                                         copy.k->k.size, nr_ptrs,
2879                                         BCH_DISK_RESERVATION_NOFAIL);
2880                         BUG_ON(ret);
2881                 }
2882
2883                 ret =   bch2_btree_iter_traverse(&del) ?:
2884                         bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
2885                         bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
2886                         bch2_trans_commit(&trans, &disk_res, NULL,
2887                                           BTREE_INSERT_NOFAIL);
2888                 bch2_disk_reservation_put(c, &disk_res);
2889
2890                 if (!ret)
2891                         bch2_btree_iter_set_pos(&src, next_pos);
2892         }
2893         bch2_trans_iter_exit(&trans, &del);
2894         bch2_trans_iter_exit(&trans, &dst);
2895         bch2_trans_iter_exit(&trans, &src);
2896         bch2_trans_exit(&trans);
2897         bch2_bkey_buf_exit(&copy, c);
2898
2899         if (ret)
2900                 return ret;
2901
2902         mutex_lock(&inode->ei_update_lock);
2903         if (!insert) {
2904                 i_size_write(&inode->v, new_size);
2905                 ret = bch2_write_inode_size(c, inode, new_size,
2906                                             ATTR_MTIME|ATTR_CTIME);
2907         } else {
2908                 /* We need an inode update to update bi_journal_seq for fsync: */
2909                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2910                                        ATTR_MTIME|ATTR_CTIME);
2911         }
2912         mutex_unlock(&inode->ei_update_lock);
2913         return ret;
2914 }
2915
2916 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2917                              u64 start_sector, u64 end_sector)
2918 {
2919         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2920         struct btree_trans trans;
2921         struct btree_iter iter;
2922         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2923         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2924         int ret = 0;
2925
2926         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2927
2928         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2929                         POS(inode->v.i_ino, start_sector),
2930                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2931
2932         while (!ret && bkey_cmp(iter.pos, end_pos) < 0) {
2933                 s64 i_sectors_delta = 0;
2934                 struct disk_reservation disk_res = { 0 };
2935                 struct quota_res quota_res = { 0 };
2936                 struct bkey_i_reservation reservation;
2937                 struct bkey_s_c k;
2938                 unsigned sectors;
2939                 u32 snapshot;
2940
2941                 bch2_trans_begin(&trans);
2942
2943                 ret = bch2_subvolume_get_snapshot(&trans,
2944                                         inode->ei_subvol, &snapshot);
2945                 if (ret)
2946                         goto bkey_err;
2947
2948                 bch2_btree_iter_set_snapshot(&iter, snapshot);
2949
2950                 k = bch2_btree_iter_peek_slot(&iter);
2951                 if ((ret = bkey_err(k)))
2952                         goto bkey_err;
2953
2954                 /* already reserved */
2955                 if (k.k->type == KEY_TYPE_reservation &&
2956                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2957                         bch2_btree_iter_advance(&iter);
2958                         continue;
2959                 }
2960
2961                 if (bkey_extent_is_data(k.k) &&
2962                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2963                         bch2_btree_iter_advance(&iter);
2964                         continue;
2965                 }
2966
2967                 bkey_reservation_init(&reservation.k_i);
2968                 reservation.k.type      = KEY_TYPE_reservation;
2969                 reservation.k.p         = k.k->p;
2970                 reservation.k.size      = k.k->size;
2971
2972                 bch2_cut_front(iter.pos,        &reservation.k_i);
2973                 bch2_cut_back(end_pos,          &reservation.k_i);
2974
2975                 sectors = reservation.k.size;
2976                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2977
2978                 if (!bkey_extent_is_allocation(k.k)) {
2979                         ret = bch2_quota_reservation_add(c, inode,
2980                                         &quota_res,
2981                                         sectors, true);
2982                         if (unlikely(ret))
2983                                 goto bkey_err;
2984                 }
2985
2986                 if (reservation.v.nr_replicas < replicas ||
2987                     bch2_bkey_sectors_compressed(k)) {
2988                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2989                                                         replicas, 0);
2990                         if (unlikely(ret))
2991                                 goto bkey_err;
2992
2993                         reservation.v.nr_replicas = disk_res.nr_replicas;
2994                 }
2995
2996                 ret = bch2_extent_update(&trans, inode_inum(inode), &iter,
2997                                          &reservation.k_i,
2998                                 &disk_res, NULL,
2999                                 0, &i_sectors_delta, true);
3000                 if (ret)
3001                         goto bkey_err;
3002                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3003 bkey_err:
3004                 bch2_quota_reservation_put(c, inode, &quota_res);
3005                 bch2_disk_reservation_put(c, &disk_res);
3006                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3007                         ret = 0;
3008         }
3009
3010         bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3011         mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3012
3013         if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) {
3014                 struct quota_res quota_res = { 0 };
3015                 s64 i_sectors_delta = 0;
3016
3017                 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3018                                end_sector, &i_sectors_delta);
3019                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3020                 bch2_quota_reservation_put(c, inode, &quota_res);
3021         }
3022
3023         bch2_trans_iter_exit(&trans, &iter);
3024         bch2_trans_exit(&trans);
3025         return ret;
3026 }
3027
3028 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3029                             loff_t offset, loff_t len)
3030 {
3031         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3032         u64 end         = offset + len;
3033         u64 block_start = round_down(offset,    block_bytes(c));
3034         u64 block_end   = round_up(end,         block_bytes(c));
3035         bool truncated_last_page = false;
3036         int ret, ret2 = 0;
3037
3038         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3039                 ret = inode_newsize_ok(&inode->v, end);
3040                 if (ret)
3041                         return ret;
3042         }
3043
3044         if (mode & FALLOC_FL_ZERO_RANGE) {
3045                 ret = bch2_truncate_pages(inode, offset, end);
3046                 if (unlikely(ret < 0))
3047                         return ret;
3048
3049                 truncated_last_page = ret;
3050
3051                 truncate_pagecache_range(&inode->v, offset, end - 1);
3052
3053                 block_start     = round_up(offset,      block_bytes(c));
3054                 block_end       = round_down(end,       block_bytes(c));
3055         }
3056
3057         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3058
3059         /*
3060          * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3061          * so that the VFS cache i_size is consistent with the btree i_size:
3062          */
3063         if (ret &&
3064             !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)))
3065                 return ret;
3066
3067         if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3068                 end = inode->v.i_size;
3069
3070         if (end >= inode->v.i_size &&
3071             (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3072              !(mode & FALLOC_FL_KEEP_SIZE))) {
3073                 spin_lock(&inode->v.i_lock);
3074                 i_size_write(&inode->v, end);
3075                 spin_unlock(&inode->v.i_lock);
3076
3077                 mutex_lock(&inode->ei_update_lock);
3078                 ret2 = bch2_write_inode_size(c, inode, end, 0);
3079                 mutex_unlock(&inode->ei_update_lock);
3080         }
3081
3082         return ret ?: ret2;
3083 }
3084
3085 long bch2_fallocate_dispatch(struct file *file, int mode,
3086                              loff_t offset, loff_t len)
3087 {
3088         struct bch_inode_info *inode = file_bch_inode(file);
3089         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3090         long ret;
3091
3092         if (!percpu_ref_tryget_live(&c->writes))
3093                 return -EROFS;
3094
3095         inode_lock(&inode->v);
3096         inode_dio_wait(&inode->v);
3097         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
3098
3099         ret = file_modified(file);
3100         if (ret)
3101                 goto err;
3102
3103         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3104                 ret = bchfs_fallocate(inode, mode, offset, len);
3105         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3106                 ret = bchfs_fpunch(inode, offset, len);
3107         else if (mode == FALLOC_FL_INSERT_RANGE)
3108                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3109         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3110                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3111         else
3112                 ret = -EOPNOTSUPP;
3113 err:
3114         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
3115         inode_unlock(&inode->v);
3116         percpu_ref_put(&c->writes);
3117
3118         return bch2_err_class(ret);
3119 }
3120
3121 static int quota_reserve_range(struct bch_inode_info *inode,
3122                                struct quota_res *res,
3123                                u64 start, u64 end)
3124 {
3125         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3126         struct btree_trans trans;
3127         struct btree_iter iter;
3128         struct bkey_s_c k;
3129         u32 snapshot;
3130         u64 sectors = end - start;
3131         u64 pos = start;
3132         int ret;
3133
3134         bch2_trans_init(&trans, c, 0, 0);
3135 retry:
3136         bch2_trans_begin(&trans);
3137
3138         ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot);
3139         if (ret)
3140                 goto err;
3141
3142         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3143                              SPOS(inode->v.i_ino, pos, snapshot), 0);
3144
3145         while (!(ret = btree_trans_too_many_iters(&trans)) &&
3146                (k = bch2_btree_iter_peek_upto(&iter, POS(inode->v.i_ino, end - 1))).k &&
3147                !(ret = bkey_err(k))) {
3148                 if (bkey_extent_is_allocation(k.k)) {
3149                         u64 s = min(end, k.k->p.offset) -
3150                                 max(start, bkey_start_offset(k.k));
3151                         BUG_ON(s > sectors);
3152                         sectors -= s;
3153                 }
3154                 bch2_btree_iter_advance(&iter);
3155         }
3156         pos = iter.pos.offset;
3157         bch2_trans_iter_exit(&trans, &iter);
3158 err:
3159         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3160                 goto retry;
3161
3162         bch2_trans_exit(&trans);
3163
3164         if (ret)
3165                 return ret;
3166
3167         return bch2_quota_reservation_add(c, inode, res, sectors, true);
3168 }
3169
3170 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3171                              struct file *file_dst, loff_t pos_dst,
3172                              loff_t len, unsigned remap_flags)
3173 {
3174         struct bch_inode_info *src = file_bch_inode(file_src);
3175         struct bch_inode_info *dst = file_bch_inode(file_dst);
3176         struct bch_fs *c = src->v.i_sb->s_fs_info;
3177         struct quota_res quota_res = { 0 };
3178         s64 i_sectors_delta = 0;
3179         u64 aligned_len;
3180         loff_t ret = 0;
3181
3182         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3183                 return -EINVAL;
3184
3185         if (remap_flags & REMAP_FILE_DEDUP)
3186                 return -EOPNOTSUPP;
3187
3188         if ((pos_src & (block_bytes(c) - 1)) ||
3189             (pos_dst & (block_bytes(c) - 1)))
3190                 return -EINVAL;
3191
3192         if (src == dst &&
3193             abs(pos_src - pos_dst) < len)
3194                 return -EINVAL;
3195
3196         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3197
3198         inode_dio_wait(&src->v);
3199         inode_dio_wait(&dst->v);
3200
3201         ret = generic_remap_file_range_prep(file_src, pos_src,
3202                                             file_dst, pos_dst,
3203                                             &len, remap_flags);
3204         if (ret < 0 || len == 0)
3205                 goto err;
3206
3207         aligned_len = round_up((u64) len, block_bytes(c));
3208
3209         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3210                                 pos_dst, pos_dst + len - 1);
3211         if (ret)
3212                 goto err;
3213
3214         ret = quota_reserve_range(dst, &quota_res, pos_dst >> 9,
3215                                   (pos_dst + aligned_len) >> 9);
3216         if (ret)
3217                 goto err;
3218
3219         file_update_time(file_dst);
3220
3221         mark_pagecache_unallocated(src, pos_src >> 9,
3222                                    (pos_src + aligned_len) >> 9);
3223
3224         ret = bch2_remap_range(c,
3225                                inode_inum(dst), pos_dst >> 9,
3226                                inode_inum(src), pos_src >> 9,
3227                                aligned_len >> 9,
3228                                pos_dst + len, &i_sectors_delta);
3229         if (ret < 0)
3230                 goto err;
3231
3232         /*
3233          * due to alignment, we might have remapped slightly more than requsted
3234          */
3235         ret = min((u64) ret << 9, (u64) len);
3236
3237         i_sectors_acct(c, dst, &quota_res, i_sectors_delta);
3238
3239         spin_lock(&dst->v.i_lock);
3240         if (pos_dst + ret > dst->v.i_size)
3241                 i_size_write(&dst->v, pos_dst + ret);
3242         spin_unlock(&dst->v.i_lock);
3243
3244         if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3245             IS_SYNC(file_inode(file_dst)))
3246                 ret = bch2_flush_inode(c, inode_inum(dst));
3247 err:
3248         bch2_quota_reservation_put(c, dst, &quota_res);
3249         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3250
3251         return bch2_err_class(ret);
3252 }
3253
3254 /* fseek: */
3255
3256 static int page_data_offset(struct page *page, unsigned offset)
3257 {
3258         struct bch_page_state *s = bch2_page_state(page);
3259         unsigned i;
3260
3261         if (s)
3262                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3263                         if (s->s[i].state >= SECTOR_DIRTY)
3264                                 return i << 9;
3265
3266         return -1;
3267 }
3268
3269 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3270                                        loff_t start_offset,
3271                                        loff_t end_offset)
3272 {
3273         struct folio_batch fbatch;
3274         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
3275         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
3276         pgoff_t index           = start_index;
3277         unsigned i;
3278         loff_t ret;
3279         int offset;
3280
3281         folio_batch_init(&fbatch);
3282
3283         while (filemap_get_folios(vinode->i_mapping,
3284                                   &index, end_index, &fbatch)) {
3285                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
3286                         struct folio *folio = fbatch.folios[i];
3287
3288                         folio_lock(folio);
3289
3290                         offset = page_data_offset(&folio->page,
3291                                         folio->index == start_index
3292                                         ? start_offset & (PAGE_SIZE - 1)
3293                                         : 0);
3294                         if (offset >= 0) {
3295                                 ret = clamp(((loff_t) folio->index << PAGE_SHIFT) +
3296                                             offset,
3297                                             start_offset, end_offset);
3298                                 folio_unlock(folio);
3299                                 folio_batch_release(&fbatch);
3300                                 return ret;
3301                         }
3302
3303                         folio_unlock(folio);
3304                 }
3305                 folio_batch_release(&fbatch);
3306                 cond_resched();
3307         }
3308
3309         return end_offset;
3310 }
3311
3312 static loff_t bch2_seek_data(struct file *file, u64 offset)
3313 {
3314         struct bch_inode_info *inode = file_bch_inode(file);
3315         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3316         struct btree_trans trans;
3317         struct btree_iter iter;
3318         struct bkey_s_c k;
3319         subvol_inum inum = inode_inum(inode);
3320         u64 isize, next_data = MAX_LFS_FILESIZE;
3321         u32 snapshot;
3322         int ret;
3323
3324         isize = i_size_read(&inode->v);
3325         if (offset >= isize)
3326                 return -ENXIO;
3327
3328         bch2_trans_init(&trans, c, 0, 0);
3329 retry:
3330         bch2_trans_begin(&trans);
3331
3332         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3333         if (ret)
3334                 goto err;
3335
3336         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3337                            SPOS(inode->v.i_ino, offset >> 9, snapshot), 0, k, ret) {
3338                 if (k.k->p.inode != inode->v.i_ino) {
3339                         break;
3340                 } else if (bkey_extent_is_data(k.k)) {
3341                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3342                         break;
3343                 } else if (k.k->p.offset >> 9 > isize)
3344                         break;
3345         }
3346         bch2_trans_iter_exit(&trans, &iter);
3347 err:
3348         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3349                 goto retry;
3350
3351         bch2_trans_exit(&trans);
3352         if (ret)
3353                 return ret;
3354
3355         if (next_data > offset)
3356                 next_data = bch2_seek_pagecache_data(&inode->v,
3357                                                      offset, next_data);
3358
3359         if (next_data >= isize)
3360                 return -ENXIO;
3361
3362         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3363 }
3364
3365 static int __page_hole_offset(struct page *page, unsigned offset)
3366 {
3367         struct bch_page_state *s = bch2_page_state(page);
3368         unsigned i;
3369
3370         if (!s)
3371                 return 0;
3372
3373         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3374                 if (s->s[i].state < SECTOR_DIRTY)
3375                         return i << 9;
3376
3377         return -1;
3378 }
3379
3380 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3381 {
3382         pgoff_t index = offset >> PAGE_SHIFT;
3383         struct page *page;
3384         int pg_offset;
3385         loff_t ret = -1;
3386
3387         page = find_lock_page(mapping, index);
3388         if (!page)
3389                 return offset;
3390
3391         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3392         if (pg_offset >= 0)
3393                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3394
3395         unlock_page(page);
3396
3397         return ret;
3398 }
3399
3400 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3401                                        loff_t start_offset,
3402                                        loff_t end_offset)
3403 {
3404         struct address_space *mapping = vinode->i_mapping;
3405         loff_t offset = start_offset, hole;
3406
3407         while (offset < end_offset) {
3408                 hole = page_hole_offset(mapping, offset);
3409                 if (hole >= 0 && hole <= end_offset)
3410                         return max(start_offset, hole);
3411
3412                 offset += PAGE_SIZE;
3413                 offset &= PAGE_MASK;
3414         }
3415
3416         return end_offset;
3417 }
3418
3419 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3420 {
3421         struct bch_inode_info *inode = file_bch_inode(file);
3422         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3423         struct btree_trans trans;
3424         struct btree_iter iter;
3425         struct bkey_s_c k;
3426         subvol_inum inum = inode_inum(inode);
3427         u64 isize, next_hole = MAX_LFS_FILESIZE;
3428         u32 snapshot;
3429         int ret;
3430
3431         isize = i_size_read(&inode->v);
3432         if (offset >= isize)
3433                 return -ENXIO;
3434
3435         bch2_trans_init(&trans, c, 0, 0);
3436 retry:
3437         bch2_trans_begin(&trans);
3438
3439         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3440         if (ret)
3441                 goto err;
3442
3443         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3444                            SPOS(inode->v.i_ino, offset >> 9, snapshot),
3445                            BTREE_ITER_SLOTS, k, ret) {
3446                 if (k.k->p.inode != inode->v.i_ino) {
3447                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3448                                         offset, MAX_LFS_FILESIZE);
3449                         break;
3450                 } else if (!bkey_extent_is_data(k.k)) {
3451                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3452                                         max(offset, bkey_start_offset(k.k) << 9),
3453                                         k.k->p.offset << 9);
3454
3455                         if (next_hole < k.k->p.offset << 9)
3456                                 break;
3457                 } else {
3458                         offset = max(offset, bkey_start_offset(k.k) << 9);
3459                 }
3460         }
3461         bch2_trans_iter_exit(&trans, &iter);
3462 err:
3463         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3464                 goto retry;
3465
3466         bch2_trans_exit(&trans);
3467         if (ret)
3468                 return ret;
3469
3470         if (next_hole > isize)
3471                 next_hole = isize;
3472
3473         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3474 }
3475
3476 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3477 {
3478         loff_t ret;
3479
3480         switch (whence) {
3481         case SEEK_SET:
3482         case SEEK_CUR:
3483         case SEEK_END:
3484                 ret = generic_file_llseek(file, offset, whence);
3485                 break;
3486         case SEEK_DATA:
3487                 ret = bch2_seek_data(file, offset);
3488                 break;
3489         case SEEK_HOLE:
3490                 ret = bch2_seek_hole(file, offset);
3491                 break;
3492         default:
3493                 ret = -EINVAL;
3494                 break;
3495         }
3496
3497         return bch2_err_class(ret);
3498 }
3499
3500 void bch2_fs_fsio_exit(struct bch_fs *c)
3501 {
3502         bioset_exit(&c->dio_write_bioset);
3503         bioset_exit(&c->dio_read_bioset);
3504         bioset_exit(&c->writepage_bioset);
3505 }
3506
3507 int bch2_fs_fsio_init(struct bch_fs *c)
3508 {
3509         int ret = 0;
3510
3511         pr_verbose_init(c->opts, "");
3512
3513         if (bioset_init(&c->writepage_bioset,
3514                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3515                         BIOSET_NEED_BVECS) ||
3516             bioset_init(&c->dio_read_bioset,
3517                         4, offsetof(struct dio_read, rbio.bio),
3518                         BIOSET_NEED_BVECS) ||
3519             bioset_init(&c->dio_write_bioset,
3520                         4, offsetof(struct dio_write, op.wbio.bio),
3521                         BIOSET_NEED_BVECS))
3522                 ret = -ENOMEM;
3523
3524         pr_verbose_init(c->opts, "ret %i", ret);
3525         return ret;
3526 }
3527
3528 #endif /* NO_BCACHEFS_FS */