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 struct address_space *faults_disabled_mapping(void)
  39 {
  40         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  41 }
  42
  43 static inline void set_fdm_dropped_locks(void)
  44 {
  45         current->faults_disabled_mapping =
  46                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  47 }
  48
  49 static inline bool fdm_dropped_locks(void)
  50 {
  51         return ((unsigned long) current->faults_disabled_mapping) & 1;
  52 }
  53
  54 struct quota_res {
  55         u64                             sectors;
  56 };
  57
  58 struct bch_writepage_io {
  59         struct closure                  cl;
  60         struct bch_inode_info           *inode;
  61
  62         /* must be last: */
  63         struct bch_write_op             op;
  64 };
  65
  66 struct dio_write {
  67         struct completion               done;
  68         struct kiocb                    *req;
  69         struct mm_struct                *mm;
  70         unsigned                        loop:1,
  71                                         sync:1,
  72                                         free_iov:1;
  73         struct quota_res                quota_res;
  74         u64                             written;
  75
  76         struct iov_iter                 iter;
  77         struct iovec                    inline_vecs[2];
  78
  79         /* must be last: */
  80         struct bch_write_op             op;
  81 };
  82
  83 struct dio_read {
  84         struct closure                  cl;
  85         struct kiocb                    *req;
  86         long                            ret;
  87         bool                            should_dirty;
  88         struct bch_read_bio             rbio;
  89 };
  90
  91 /* pagecache_block must be held */
  92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  93                                               loff_t start, loff_t end)
  94 {
  95         int ret;
  96
  97         /*
  98          * XXX: the way this is currently implemented, we can spin if a process
  99          * is continually redirtying a specific page
 100          */
 101         do {
 102                 if (!mapping->nrpages)
 103                         return 0;
 104
 105                 ret = filemap_write_and_wait_range(mapping, start, end);
 106                 if (ret)
 107                         break;
 108
 109                 if (!mapping->nrpages)
 110                         return 0;
 111
 112                 ret = invalidate_inode_pages2_range(mapping,
 113                                 start >> PAGE_SHIFT,
 114                                 end >> PAGE_SHIFT);
 115         } while (ret == -EBUSY);
 116
 117         return ret;
 118 }
 119
 120 /* quotas */
 121
 122 #ifdef CONFIG_BCACHEFS_QUOTA
 123
 124 static void bch2_quota_reservation_put(struct bch_fs *c,
 125                                        struct bch_inode_info *inode,
 126                                        struct quota_res *res)
 127 {
 128         if (!res->sectors)
 129                 return;
 130
 131         mutex_lock(&inode->ei_quota_lock);
 132         BUG_ON(res->sectors > inode->ei_quota_reserved);
 133
 134         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 135                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 136         inode->ei_quota_reserved -= res->sectors;
 137         mutex_unlock(&inode->ei_quota_lock);
 138
 139         res->sectors = 0;
 140 }
 141
 142 static int bch2_quota_reservation_add(struct bch_fs *c,
 143                                       struct bch_inode_info *inode,
 144                                       struct quota_res *res,
 145                                       unsigned sectors,
 146                                       bool check_enospc)
 147 {
 148         int ret;
 149
 150         mutex_lock(&inode->ei_quota_lock);
 151         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 152                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 153         if (likely(!ret)) {
 154                 inode->ei_quota_reserved += sectors;
 155                 res->sectors += sectors;
 156         }
 157         mutex_unlock(&inode->ei_quota_lock);
 158
 159         return ret;
 160 }
 161
 162 #else
 163
 164 static void bch2_quota_reservation_put(struct bch_fs *c,
 165                                        struct bch_inode_info *inode,
 166                                        struct quota_res *res)
 167 {
 168 }
 169
 170 static int bch2_quota_reservation_add(struct bch_fs *c,
 171                                       struct bch_inode_info *inode,
 172                                       struct quota_res *res,
 173                                       unsigned sectors,
 174                                       bool check_enospc)
 175 {
 176         return 0;
 177 }
 178
 179 #endif
 180
 181 /* i_size updates: */
 182
 183 struct inode_new_size {
 184         loff_t          new_size;
 185         u64             now;
 186         unsigned        fields;
 187 };
 188
 189 static int inode_set_size(struct bch_inode_info *inode,
 190                           struct bch_inode_unpacked *bi,
 191                           void *p)
 192 {
 193         struct inode_new_size *s = p;
 194
 195         bi->bi_size = s->new_size;
 196         if (s->fields & ATTR_ATIME)
 197                 bi->bi_atime = s->now;
 198         if (s->fields & ATTR_MTIME)
 199                 bi->bi_mtime = s->now;
 200         if (s->fields & ATTR_CTIME)
 201                 bi->bi_ctime = s->now;
 202
 203         return 0;
 204 }
 205
 206 int __must_check bch2_write_inode_size(struct bch_fs *c,
 207                                        struct bch_inode_info *inode,
 208                                        loff_t new_size, unsigned fields)
 209 {
 210         struct inode_new_size s = {
 211                 .new_size       = new_size,
 212                 .now            = bch2_current_time(c),
 213                 .fields         = fields,
 214         };
 215
 216         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 217 }
 218
 219 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 220                            struct quota_res *quota_res, s64 sectors)
 221 {
 222         if (!sectors)
 223                 return;
 224
 225         mutex_lock(&inode->ei_quota_lock);
 226         BUG_ON((s64) inode->v.i_blocks + sectors < 0);
 227         inode->v.i_blocks += sectors;
 228
 229 #ifdef CONFIG_BCACHEFS_QUOTA
 230         if (quota_res && sectors > 0) {
 231                 BUG_ON(sectors > quota_res->sectors);
 232                 BUG_ON(sectors > inode->ei_quota_reserved);
 233
 234                 quota_res->sectors -= sectors;
 235                 inode->ei_quota_reserved -= sectors;
 236         } else {
 237                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 238         }
 239 #endif
 240         mutex_unlock(&inode->ei_quota_lock);
 241 }
 242
 243 /* page state: */
 244
 245 /* stored in page->private: */
 246
 247 struct bch_page_sector {
 248         /* Uncompressed, fully allocated replicas (or on disk reservation): */
 249         unsigned                nr_replicas:4;
 250
 251         /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
 252         unsigned                replicas_reserved:4;
 253
 254         /* i_sectors: */
 255         enum {
 256                 SECTOR_UNALLOCATED,
 257                 SECTOR_RESERVED,
 258                 SECTOR_DIRTY,
 259                 SECTOR_DIRTY_RESERVED,
 260                 SECTOR_ALLOCATED,
 261         }                       state:8;
 262 };
 263
 264 struct bch_page_state {
 265         spinlock_t              lock;
 266         atomic_t                write_count;
 267         bool                    uptodate;
 268         struct bch_page_sector  s[PAGE_SECTORS];
 269 };
 270
 271 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 272 {
 273         return page_has_private(page)
 274                 ? (struct bch_page_state *) page_private(page)
 275                 : NULL;
 276 }
 277
 278 static inline struct bch_page_state *bch2_page_state(struct page *page)
 279 {
 280         EBUG_ON(!PageLocked(page));
 281
 282         return __bch2_page_state(page);
 283 }
 284
 285 /* for newly allocated pages: */
 286 static void __bch2_page_state_release(struct page *page)
 287 {
 288         kfree(detach_page_private(page));
 289 }
 290
 291 static void bch2_page_state_release(struct page *page)
 292 {
 293         EBUG_ON(!PageLocked(page));
 294         __bch2_page_state_release(page);
 295 }
 296
 297 /* for newly allocated pages: */
 298 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 299                                                        gfp_t gfp)
 300 {
 301         struct bch_page_state *s;
 302
 303         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 304         if (!s)
 305                 return NULL;
 306
 307         spin_lock_init(&s->lock);
 308         attach_page_private(page, s);
 309         return s;
 310 }
 311
 312 static struct bch_page_state *bch2_page_state_create(struct page *page,
 313                                                      gfp_t gfp)
 314 {
 315         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 316 }
 317
 318 static unsigned bkey_to_sector_state(const struct bkey *k)
 319 {
 320         if (k->type == KEY_TYPE_reservation)
 321                 return SECTOR_RESERVED;
 322         if (bkey_extent_is_allocation(k))
 323                 return SECTOR_ALLOCATED;
 324         return SECTOR_UNALLOCATED;
 325 }
 326
 327 static void __bch2_page_state_set(struct page *page,
 328                                   unsigned pg_offset, unsigned pg_len,
 329                                   unsigned nr_ptrs, unsigned state)
 330 {
 331         struct bch_page_state *s = bch2_page_state_create(page, __GFP_NOFAIL);
 332         unsigned i;
 333
 334         BUG_ON(pg_offset >= PAGE_SECTORS);
 335         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 336
 337         spin_lock(&s->lock);
 338
 339         for (i = pg_offset; i < pg_offset + pg_len; i++) {
 340                 s->s[i].nr_replicas = nr_ptrs;
 341                 s->s[i].state = state;
 342         }
 343
 344         if (i == PAGE_SECTORS)
 345                 s->uptodate = true;
 346
 347         spin_unlock(&s->lock);
 348 }
 349
 350 static int bch2_page_state_set(struct bch_fs *c, subvol_inum inum,
 351                                struct page **pages, unsigned nr_pages)
 352 {
 353         struct btree_trans trans;
 354         struct btree_iter iter;
 355         struct bkey_s_c k;
 356         u64 offset = pages[0]->index << PAGE_SECTORS_SHIFT;
 357         unsigned pg_idx = 0;
 358         u32 snapshot;
 359         int ret;
 360
 361         bch2_trans_init(&trans, c, 0, 0);
 362 retry:
 363         bch2_trans_begin(&trans);
 364
 365         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 366         if (ret)
 367                 goto err;
 368
 369         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
 370                            SPOS(inum.inum, offset, snapshot),
 371                            BTREE_ITER_SLOTS, k, ret) {
 372                 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
 373                 unsigned state = bkey_to_sector_state(k.k);
 374
 375                 while (pg_idx < nr_pages) {
 376                         struct page *page = pages[pg_idx];
 377                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 378                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 379                         unsigned pg_offset = max(bkey_start_offset(k.k), pg_start) - pg_start;
 380                         unsigned pg_len = min(k.k->p.offset, pg_end) - pg_offset - pg_start;
 381
 382                         BUG_ON(k.k->p.offset < pg_start);
 383                         BUG_ON(bkey_start_offset(k.k) > pg_end);
 384
 385                         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate)
 386                                 __bch2_page_state_set(page, pg_offset, pg_len, nr_ptrs, state);
 387
 388                         if (k.k->p.offset < pg_end)
 389                                 break;
 390                         pg_idx++;
 391                 }
 392
 393                 if (pg_idx == nr_pages)
 394                         break;
 395         }
 396
 397         offset = iter.pos.offset;
 398         bch2_trans_iter_exit(&trans, &iter);
 399 err:
 400         if (ret == -EINTR)
 401                 goto retry;
 402         bch2_trans_exit(&trans);
 403
 404         return ret;
 405 }
 406
 407 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
 408 {
 409         struct bvec_iter iter;
 410         struct bio_vec bv;
 411         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 412                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 413         unsigned state = bkey_to_sector_state(k.k);
 414
 415         bio_for_each_segment(bv, bio, iter)
 416                 __bch2_page_state_set(bv.bv_page, bv.bv_offset >> 9,
 417                                       bv.bv_len >> 9, nr_ptrs, state);
 418 }
 419
 420 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
 421                                        u64 start, u64 end)
 422 {
 423         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 424         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 425         struct pagevec pvec;
 426
 427         if (end <= start)
 428                 return;
 429
 430         pagevec_init(&pvec);
 431
 432         do {
 433                 unsigned nr_pages, i, j;
 434
 435                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
 436                                                 &index, end_index);
 437                 for (i = 0; i < nr_pages; i++) {
 438                         struct page *page = pvec.pages[i];
 439                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 440                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 441                         unsigned pg_offset = max(start, pg_start) - pg_start;
 442                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 443                         struct bch_page_state *s;
 444
 445                         BUG_ON(end <= pg_start);
 446                         BUG_ON(pg_offset >= PAGE_SECTORS);
 447                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 448
 449                         lock_page(page);
 450                         s = bch2_page_state(page);
 451
 452                         if (s) {
 453                                 spin_lock(&s->lock);
 454                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 455                                         s->s[j].nr_replicas = 0;
 456                                 spin_unlock(&s->lock);
 457                         }
 458
 459                         unlock_page(page);
 460                 }
 461                 pagevec_release(&pvec);
 462         } while (index <= end_index);
 463 }
 464
 465 static void mark_pagecache_reserved(struct bch_inode_info *inode,
 466                                     u64 start, u64 end)
 467 {
 468         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 469         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 470         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 471         struct pagevec pvec;
 472         s64 i_sectors_delta = 0;
 473
 474         if (end <= start)
 475                 return;
 476
 477         pagevec_init(&pvec);
 478
 479         do {
 480                 unsigned nr_pages, i, j;
 481
 482                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
 483                                                 &index, end_index);
 484                 for (i = 0; i < nr_pages; i++) {
 485                         struct page *page = pvec.pages[i];
 486                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 487                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 488                         unsigned pg_offset = max(start, pg_start) - pg_start;
 489                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 490                         struct bch_page_state *s;
 491
 492                         BUG_ON(end <= pg_start);
 493                         BUG_ON(pg_offset >= PAGE_SECTORS);
 494                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 495
 496                         lock_page(page);
 497                         s = bch2_page_state(page);
 498
 499                         if (s) {
 500                                 spin_lock(&s->lock);
 501                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 502                                         switch (s->s[j].state) {
 503                                         case SECTOR_UNALLOCATED:
 504                                                 s->s[j].state = SECTOR_RESERVED;
 505                                                 break;
 506                                         case SECTOR_DIRTY:
 507                                                 s->s[j].state = SECTOR_DIRTY_RESERVED;
 508                                                 i_sectors_delta--;
 509                                                 break;
 510                                         default:
 511                                                 break;
 512                                         }
 513                                 spin_unlock(&s->lock);
 514                         }
 515
 516                         unlock_page(page);
 517                 }
 518                 pagevec_release(&pvec);
 519         } while (index <= end_index);
 520
 521         i_sectors_acct(c, inode, NULL, i_sectors_delta);
 522 }
 523
 524 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 525 {
 526         /* XXX: this should not be open coded */
 527         return inode->ei_inode.bi_data_replicas
 528                 ? inode->ei_inode.bi_data_replicas - 1
 529                 : c->opts.data_replicas;
 530 }
 531
 532 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 533                                                   unsigned nr_replicas)
 534 {
 535         return max(0, (int) nr_replicas -
 536                    s->nr_replicas -
 537                    s->replicas_reserved);
 538 }
 539
 540 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 541                                 struct bch_inode_info *inode,
 542                                 struct page *page, bool check_enospc)
 543 {
 544         struct bch_page_state *s = bch2_page_state_create(page, 0);
 545         unsigned nr_replicas = inode_nr_replicas(c, inode);
 546         struct disk_reservation disk_res = { 0 };
 547         unsigned i, disk_res_sectors = 0;
 548         int ret;
 549
 550         if (!s)
 551                 return -ENOMEM;
 552
 553         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 554                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 555
 556         if (!disk_res_sectors)
 557                 return 0;
 558
 559         ret = bch2_disk_reservation_get(c, &disk_res,
 560                                         disk_res_sectors, 1,
 561                                         !check_enospc
 562                                         ? BCH_DISK_RESERVATION_NOFAIL
 563                                         : 0);
 564         if (unlikely(ret))
 565                 return ret;
 566
 567         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 568                 s->s[i].replicas_reserved +=
 569                         sectors_to_reserve(&s->s[i], nr_replicas);
 570
 571         return 0;
 572 }
 573
 574 struct bch2_page_reservation {
 575         struct disk_reservation disk;
 576         struct quota_res        quota;
 577 };
 578
 579 static void bch2_page_reservation_init(struct bch_fs *c,
 580                         struct bch_inode_info *inode,
 581                         struct bch2_page_reservation *res)
 582 {
 583         memset(res, 0, sizeof(*res));
 584
 585         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 586 }
 587
 588 static void bch2_page_reservation_put(struct bch_fs *c,
 589                         struct bch_inode_info *inode,
 590                         struct bch2_page_reservation *res)
 591 {
 592         bch2_disk_reservation_put(c, &res->disk);
 593         bch2_quota_reservation_put(c, inode, &res->quota);
 594 }
 595
 596 static int bch2_page_reservation_get(struct bch_fs *c,
 597                         struct bch_inode_info *inode, struct page *page,
 598                         struct bch2_page_reservation *res,
 599                         unsigned offset, unsigned len, bool check_enospc)
 600 {
 601         struct bch_page_state *s = bch2_page_state_create(page, 0);
 602         unsigned i, disk_sectors = 0, quota_sectors = 0;
 603         int ret;
 604
 605         if (!s)
 606                 return -ENOMEM;
 607
 608         BUG_ON(!s->uptodate);
 609
 610         for (i = round_down(offset, block_bytes(c)) >> 9;
 611              i < round_up(offset + len, block_bytes(c)) >> 9;
 612              i++) {
 613                 disk_sectors += sectors_to_reserve(&s->s[i],
 614                                                 res->disk.nr_replicas);
 615                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 616         }
 617
 618         if (disk_sectors) {
 619                 ret = bch2_disk_reservation_add(c, &res->disk,
 620                                                 disk_sectors,
 621                                                 !check_enospc
 622                                                 ? BCH_DISK_RESERVATION_NOFAIL
 623                                                 : 0);
 624                 if (unlikely(ret))
 625                         return ret;
 626         }
 627
 628         if (quota_sectors) {
 629                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 630                                                  quota_sectors,
 631                                                  check_enospc);
 632                 if (unlikely(ret)) {
 633                         struct disk_reservation tmp = {
 634                                 .sectors = disk_sectors
 635                         };
 636
 637                         bch2_disk_reservation_put(c, &tmp);
 638                         res->disk.sectors -= disk_sectors;
 639                         return ret;
 640                 }
 641         }
 642
 643         return 0;
 644 }
 645
 646 static void bch2_clear_page_bits(struct page *page)
 647 {
 648         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 649         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 650         struct bch_page_state *s = bch2_page_state(page);
 651         struct disk_reservation disk_res = { 0 };
 652         int i, dirty_sectors = 0;
 653
 654         if (!s)
 655                 return;
 656
 657         EBUG_ON(!PageLocked(page));
 658         EBUG_ON(PageWriteback(page));
 659
 660         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 661                 disk_res.sectors += s->s[i].replicas_reserved;
 662                 s->s[i].replicas_reserved = 0;
 663
 664                 switch (s->s[i].state) {
 665                 case SECTOR_DIRTY:
 666                         s->s[i].state = SECTOR_UNALLOCATED;
 667                         --dirty_sectors;
 668                         break;
 669                 case SECTOR_DIRTY_RESERVED:
 670                         s->s[i].state = SECTOR_RESERVED;
 671                         break;
 672                 default:
 673                         break;
 674                 }
 675         }
 676
 677         bch2_disk_reservation_put(c, &disk_res);
 678
 679         i_sectors_acct(c, inode, NULL, dirty_sectors);
 680
 681         bch2_page_state_release(page);
 682 }
 683
 684 static void bch2_set_page_dirty(struct bch_fs *c,
 685                         struct bch_inode_info *inode, struct page *page,
 686                         struct bch2_page_reservation *res,
 687                         unsigned offset, unsigned len)
 688 {
 689         struct bch_page_state *s = bch2_page_state(page);
 690         unsigned i, dirty_sectors = 0;
 691
 692         WARN_ON((u64) page_offset(page) + offset + len >
 693                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 694
 695         spin_lock(&s->lock);
 696
 697         for (i = round_down(offset, block_bytes(c)) >> 9;
 698              i < round_up(offset + len, block_bytes(c)) >> 9;
 699              i++) {
 700                 unsigned sectors = sectors_to_reserve(&s->s[i],
 701                                                 res->disk.nr_replicas);
 702
 703                 /*
 704                  * This can happen if we race with the error path in
 705                  * bch2_writepage_io_done():
 706                  */
 707                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 708
 709                 s->s[i].replicas_reserved += sectors;
 710                 res->disk.sectors -= sectors;
 711
 712                 switch (s->s[i].state) {
 713                 case SECTOR_UNALLOCATED:
 714                         s->s[i].state = SECTOR_DIRTY;
 715                         dirty_sectors++;
 716                         break;
 717                 case SECTOR_RESERVED:
 718                         s->s[i].state = SECTOR_DIRTY_RESERVED;
 719                         break;
 720                 default:
 721                         break;
 722                 }
 723         }
 724
 725         spin_unlock(&s->lock);
 726
 727         i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 728
 729         if (!PageDirty(page))
 730                 __set_page_dirty_nobuffers(page);
 731 }
 732
 733 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 734 {
 735         struct file *file = vmf->vma->vm_file;
 736         struct address_space *mapping = file->f_mapping;
 737         struct address_space *fdm = faults_disabled_mapping();
 738         struct bch_inode_info *inode = file_bch_inode(file);
 739         int ret;
 740
 741         if (fdm == mapping)
 742                 return VM_FAULT_SIGBUS;
 743
 744         /* Lock ordering: */
 745         if (fdm > mapping) {
 746                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 747
 748                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 749                         goto got_lock;
 750
 751                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 752
 753                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 754                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 755
 756                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 757
 758                 /* Signal that lock has been dropped: */
 759                 set_fdm_dropped_locks();
 760                 return VM_FAULT_SIGBUS;
 761         }
 762
 763         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 764 got_lock:
 765         ret = filemap_fault(vmf);
 766         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 767
 768         return ret;
 769 }
 770
 771 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 772 {
 773         struct page *page = vmf->page;
 774         struct file *file = vmf->vma->vm_file;
 775         struct bch_inode_info *inode = file_bch_inode(file);
 776         struct address_space *mapping = file->f_mapping;
 777         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 778         struct bch2_page_reservation res;
 779         unsigned len;
 780         loff_t isize;
 781         int ret;
 782
 783         bch2_page_reservation_init(c, inode, &res);
 784
 785         sb_start_pagefault(inode->v.i_sb);
 786         file_update_time(file);
 787
 788         /*
 789          * Not strictly necessary, but helps avoid dio writes livelocking in
 790          * write_invalidate_inode_pages_range() - can drop this if/when we get
 791          * a write_invalidate_inode_pages_range() that works without dropping
 792          * page lock before invalidating page
 793          */
 794         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 795
 796         lock_page(page);
 797         isize = i_size_read(&inode->v);
 798
 799         if (page->mapping != mapping || page_offset(page) >= isize) {
 800                 unlock_page(page);
 801                 ret = VM_FAULT_NOPAGE;
 802                 goto out;
 803         }
 804
 805         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 806
 807         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
 808                 if (bch2_page_state_set(c, inode_inum(inode), &page, 1)) {
 809                         unlock_page(page);
 810                         ret = VM_FAULT_SIGBUS;
 811                         goto out;
 812                 }
 813         }
 814
 815         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 816                 unlock_page(page);
 817                 ret = VM_FAULT_SIGBUS;
 818                 goto out;
 819         }
 820
 821         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 822         bch2_page_reservation_put(c, inode, &res);
 823
 824         wait_for_stable_page(page);
 825         ret = VM_FAULT_LOCKED;
 826 out:
 827         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 828         sb_end_pagefault(inode->v.i_sb);
 829
 830         return ret;
 831 }
 832
 833 void bch2_invalidatepage(struct page *page, unsigned int offset,
 834                          unsigned int length)
 835 {
 836         if (offset || length < PAGE_SIZE)
 837                 return;
 838
 839         bch2_clear_page_bits(page);
 840 }
 841
 842 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 843 {
 844         if (PageDirty(page))
 845                 return 0;
 846
 847         bch2_clear_page_bits(page);
 848         return 1;
 849 }
 850
 851 #ifdef CONFIG_MIGRATION
 852 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 853                       struct page *page, enum migrate_mode mode)
 854 {
 855         int ret;
 856
 857         EBUG_ON(!PageLocked(page));
 858         EBUG_ON(!PageLocked(newpage));
 859
 860         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 861         if (ret != MIGRATEPAGE_SUCCESS)
 862                 return ret;
 863
 864         if (PagePrivate(page))
 865                 attach_page_private(newpage, detach_page_private(page));
 866
 867         if (mode != MIGRATE_SYNC_NO_COPY)
 868                 migrate_page_copy(newpage, page);
 869         else
 870                 migrate_page_states(newpage, page);
 871         return MIGRATEPAGE_SUCCESS;
 872 }
 873 #endif
 874
 875 /* readpage(s): */
 876
 877 static void bch2_readpages_end_io(struct bio *bio)
 878 {
 879         struct bvec_iter_all iter;
 880         struct bio_vec *bv;
 881
 882         bio_for_each_segment_all(bv, bio, iter) {
 883                 struct page *page = bv->bv_page;
 884
 885                 if (!bio->bi_status) {
 886                         SetPageUptodate(page);
 887                 } else {
 888                         ClearPageUptodate(page);
 889                         SetPageError(page);
 890                 }
 891                 unlock_page(page);
 892         }
 893
 894         bio_put(bio);
 895 }
 896
 897 struct readpages_iter {
 898         struct address_space    *mapping;
 899         struct page             **pages;
 900         unsigned                nr_pages;
 901         unsigned                idx;
 902         pgoff_t                 offset;
 903 };
 904
 905 static int readpages_iter_init(struct readpages_iter *iter,
 906                                struct readahead_control *ractl)
 907 {
 908         unsigned i, nr_pages = readahead_count(ractl);
 909
 910         memset(iter, 0, sizeof(*iter));
 911
 912         iter->mapping   = ractl->mapping;
 913         iter->offset    = readahead_index(ractl);
 914         iter->nr_pages  = nr_pages;
 915
 916         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 917         if (!iter->pages)
 918                 return -ENOMEM;
 919
 920         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 921         for (i = 0; i < nr_pages; i++) {
 922                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 923                 put_page(iter->pages[i]);
 924         }
 925
 926         return 0;
 927 }
 928
 929 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 930 {
 931         if (iter->idx >= iter->nr_pages)
 932                 return NULL;
 933
 934         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 935
 936         return iter->pages[iter->idx];
 937 }
 938
 939 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 940 {
 941         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 942         struct bch_extent_crc_unpacked crc;
 943         const union bch_extent_entry *i;
 944
 945         bkey_for_each_crc(k.k, ptrs, crc, i)
 946                 if (crc.csum_type || crc.compression_type)
 947                         return true;
 948         return false;
 949 }
 950
 951 static void readpage_bio_extend(struct readpages_iter *iter,
 952                                 struct bio *bio,
 953                                 unsigned sectors_this_extent,
 954                                 bool get_more)
 955 {
 956         while (bio_sectors(bio) < sectors_this_extent &&
 957                bio->bi_vcnt < bio->bi_max_vecs) {
 958                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
 959                 struct page *page = readpage_iter_next(iter);
 960                 int ret;
 961
 962                 if (page) {
 963                         if (iter->offset + iter->idx != page_offset)
 964                                 break;
 965
 966                         iter->idx++;
 967                 } else {
 968                         if (!get_more)
 969                                 break;
 970
 971                         page = xa_load(&iter->mapping->i_pages, page_offset);
 972                         if (page && !xa_is_value(page))
 973                                 break;
 974
 975                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 976                         if (!page)
 977                                 break;
 978
 979                         if (!__bch2_page_state_create(page, 0)) {
 980                                 put_page(page);
 981                                 break;
 982                         }
 983
 984                         ret = add_to_page_cache_lru(page, iter->mapping,
 985                                                     page_offset, GFP_NOFS);
 986                         if (ret) {
 987                                 __bch2_page_state_release(page);
 988                                 put_page(page);
 989                                 break;
 990                         }
 991
 992                         put_page(page);
 993                 }
 994
 995                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 996         }
 997 }
 998
 999 static void bchfs_read(struct btree_trans *trans,
1000                        struct bch_read_bio *rbio,
1001                        subvol_inum inum,
1002                        struct readpages_iter *readpages_iter)
1003 {
1004         struct bch_fs *c = trans->c;
1005         struct btree_iter iter;
1006         struct bkey_buf sk;
1007         int flags = BCH_READ_RETRY_IF_STALE|
1008                 BCH_READ_MAY_PROMOTE;
1009         u32 snapshot;
1010         int ret = 0;
1011
1012         rbio->c = c;
1013         rbio->start_time = local_clock();
1014         rbio->subvol = inum.subvol;
1015
1016         bch2_bkey_buf_init(&sk);
1017 retry:
1018         bch2_trans_begin(trans);
1019         iter = (struct btree_iter) { NULL };
1020
1021         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1022         if (ret)
1023                 goto err;
1024
1025         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1026                              SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1027                              BTREE_ITER_SLOTS);
1028         while (1) {
1029                 struct bkey_s_c k;
1030                 unsigned bytes, sectors, offset_into_extent;
1031                 enum btree_id data_btree = BTREE_ID_extents;
1032
1033                 /*
1034                  * read_extent -> io_time_reset may cause a transaction restart
1035                  * without returning an error, we need to check for that here:
1036                  */
1037                 if (!bch2_trans_relock(trans)) {
1038                         ret = -EINTR;
1039                         break;
1040                 }
1041
1042                 bch2_btree_iter_set_pos(&iter,
1043                                 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1044
1045                 k = bch2_btree_iter_peek_slot(&iter);
1046                 ret = bkey_err(k);
1047                 if (ret)
1048                         break;
1049
1050                 offset_into_extent = iter.pos.offset -
1051                         bkey_start_offset(k.k);
1052                 sectors = k.k->size - offset_into_extent;
1053
1054                 bch2_bkey_buf_reassemble(&sk, c, k);
1055
1056                 ret = bch2_read_indirect_extent(trans, &data_btree,
1057                                         &offset_into_extent, &sk);
1058                 if (ret)
1059                         break;
1060
1061                 k = bkey_i_to_s_c(sk.k);
1062
1063                 sectors = min(sectors, k.k->size - offset_into_extent);
1064
1065                 if (readpages_iter)
1066                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
1067                                             extent_partial_reads_expensive(k));
1068
1069                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1070                 swap(rbio->bio.bi_iter.bi_size, bytes);
1071
1072                 if (rbio->bio.bi_iter.bi_size == bytes)
1073                         flags |= BCH_READ_LAST_FRAGMENT;
1074
1075                 bch2_bio_page_state_set(&rbio->bio, k);
1076
1077                 bch2_read_extent(trans, rbio, iter.pos,
1078                                  data_btree, k, offset_into_extent, flags);
1079
1080                 if (flags & BCH_READ_LAST_FRAGMENT)
1081                         break;
1082
1083                 swap(rbio->bio.bi_iter.bi_size, bytes);
1084                 bio_advance(&rbio->bio, bytes);
1085
1086                 ret = btree_trans_too_many_iters(trans);
1087                 if (ret)
1088                         break;
1089         }
1090 err:
1091         bch2_trans_iter_exit(trans, &iter);
1092
1093         if (ret == -EINTR)
1094                 goto retry;
1095
1096         if (ret) {
1097                 bch_err_inum_ratelimited(c, inum.inum,
1098                                 "read error %i from btree lookup", ret);
1099                 rbio->bio.bi_status = BLK_STS_IOERR;
1100                 bio_endio(&rbio->bio);
1101         }
1102
1103         bch2_bkey_buf_exit(&sk, c);
1104 }
1105
1106 void bch2_readahead(struct readahead_control *ractl)
1107 {
1108         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1109         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1110         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1111         struct btree_trans trans;
1112         struct page *page;
1113         struct readpages_iter readpages_iter;
1114         int ret;
1115
1116         ret = readpages_iter_init(&readpages_iter, ractl);
1117         BUG_ON(ret);
1118
1119         bch2_trans_init(&trans, c, 0, 0);
1120
1121         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1122
1123         while ((page = readpage_iter_next(&readpages_iter))) {
1124                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1125                 unsigned n = min_t(unsigned,
1126                                    readpages_iter.nr_pages -
1127                                    readpages_iter.idx,
1128                                    BIO_MAX_VECS);
1129                 struct bch_read_bio *rbio =
1130                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
1131                                   opts);
1132
1133                 readpages_iter.idx++;
1134
1135                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
1136                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTORS_SHIFT;
1137                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1138                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1139
1140                 bchfs_read(&trans, rbio, inode_inum(inode),
1141                            &readpages_iter);
1142         }
1143
1144         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1145
1146         bch2_trans_exit(&trans);
1147         kfree(readpages_iter.pages);
1148 }
1149
1150 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1151                              subvol_inum inum, struct page *page)
1152 {
1153         struct btree_trans trans;
1154
1155         bch2_page_state_create(page, __GFP_NOFAIL);
1156
1157         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1158         rbio->bio.bi_iter.bi_sector =
1159                 (sector_t) page->index << PAGE_SECTORS_SHIFT;
1160         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1161
1162         bch2_trans_init(&trans, c, 0, 0);
1163         bchfs_read(&trans, rbio, inum, NULL);
1164         bch2_trans_exit(&trans);
1165 }
1166
1167 int bch2_readpage(struct file *file, struct page *page)
1168 {
1169         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1170         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1171         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1172         struct bch_read_bio *rbio;
1173
1174         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
1175         rbio->bio.bi_end_io = bch2_readpages_end_io;
1176
1177         __bchfs_readpage(c, rbio, inode_inum(inode), page);
1178         return 0;
1179 }
1180
1181 static void bch2_read_single_page_end_io(struct bio *bio)
1182 {
1183         complete(bio->bi_private);
1184 }
1185
1186 static int bch2_read_single_page(struct page *page,
1187                                  struct address_space *mapping)
1188 {
1189         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1190         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1191         struct bch_read_bio *rbio;
1192         int ret;
1193         DECLARE_COMPLETION_ONSTACK(done);
1194
1195         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1196                          io_opts(c, &inode->ei_inode));
1197         rbio->bio.bi_private = &done;
1198         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1199
1200         __bchfs_readpage(c, rbio, inode_inum(inode), page);
1201         wait_for_completion(&done);
1202
1203         ret = blk_status_to_errno(rbio->bio.bi_status);
1204         bio_put(&rbio->bio);
1205
1206         if (ret < 0)
1207                 return ret;
1208
1209         SetPageUptodate(page);
1210         return 0;
1211 }
1212
1213 /* writepages: */
1214
1215 struct bch_writepage_state {
1216         struct bch_writepage_io *io;
1217         struct bch_io_opts      opts;
1218 };
1219
1220 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1221                                                                   struct bch_inode_info *inode)
1222 {
1223         return (struct bch_writepage_state) {
1224                 .opts = io_opts(c, &inode->ei_inode)
1225         };
1226 }
1227
1228 static void bch2_writepage_io_free(struct closure *cl)
1229 {
1230         struct bch_writepage_io *io = container_of(cl,
1231                                         struct bch_writepage_io, cl);
1232
1233         bio_put(&io->op.wbio.bio);
1234 }
1235
1236 static void bch2_writepage_io_done(struct closure *cl)
1237 {
1238         struct bch_writepage_io *io = container_of(cl,
1239                                         struct bch_writepage_io, cl);
1240         struct bch_fs *c = io->op.c;
1241         struct bio *bio = &io->op.wbio.bio;
1242         struct bvec_iter_all iter;
1243         struct bio_vec *bvec;
1244         unsigned i;
1245
1246         up(&io->op.c->io_in_flight);
1247
1248         if (io->op.error) {
1249                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1250
1251                 bio_for_each_segment_all(bvec, bio, iter) {
1252                         struct bch_page_state *s;
1253
1254                         SetPageError(bvec->bv_page);
1255                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1256
1257                         s = __bch2_page_state(bvec->bv_page);
1258                         spin_lock(&s->lock);
1259                         for (i = 0; i < PAGE_SECTORS; i++)
1260                                 s->s[i].nr_replicas = 0;
1261                         spin_unlock(&s->lock);
1262                 }
1263         }
1264
1265         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1266                 bio_for_each_segment_all(bvec, bio, iter) {
1267                         struct bch_page_state *s;
1268
1269                         s = __bch2_page_state(bvec->bv_page);
1270                         spin_lock(&s->lock);
1271                         for (i = 0; i < PAGE_SECTORS; i++)
1272                                 s->s[i].nr_replicas = 0;
1273                         spin_unlock(&s->lock);
1274                 }
1275         }
1276
1277         /*
1278          * racing with fallocate can cause us to add fewer sectors than
1279          * expected - but we shouldn't add more sectors than expected:
1280          */
1281         WARN_ON(io->op.i_sectors_delta > 0);
1282
1283         /*
1284          * (error (due to going RO) halfway through a page can screw that up
1285          * slightly)
1286          * XXX wtf?
1287            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1288          */
1289
1290         /*
1291          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1292          * before calling end_page_writeback:
1293          */
1294         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1295
1296         bio_for_each_segment_all(bvec, bio, iter) {
1297                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1298
1299                 if (atomic_dec_and_test(&s->write_count))
1300                         end_page_writeback(bvec->bv_page);
1301         }
1302
1303         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1304 }
1305
1306 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1307 {
1308         struct bch_writepage_io *io = w->io;
1309
1310         down(&io->op.c->io_in_flight);
1311
1312         w->io = NULL;
1313         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1314         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1315 }
1316
1317 /*
1318  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1319  * possible, else allocating a new one:
1320  */
1321 static void bch2_writepage_io_alloc(struct bch_fs *c,
1322                                     struct writeback_control *wbc,
1323                                     struct bch_writepage_state *w,
1324                                     struct bch_inode_info *inode,
1325                                     u64 sector,
1326                                     unsigned nr_replicas)
1327 {
1328         struct bch_write_op *op;
1329
1330         w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1331                                               &c->writepage_bioset),
1332                              struct bch_writepage_io, op.wbio.bio);
1333
1334         closure_init(&w->io->cl, NULL);
1335         w->io->inode            = inode;
1336
1337         op                      = &w->io->op;
1338         bch2_write_op_init(op, c, w->opts);
1339         op->target              = w->opts.foreground_target;
1340         op->nr_replicas         = nr_replicas;
1341         op->res.nr_replicas     = nr_replicas;
1342         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1343         op->subvol              = inode->ei_subvol;
1344         op->pos                 = POS(inode->v.i_ino, sector);
1345         op->wbio.bio.bi_iter.bi_sector = sector;
1346         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1347 }
1348
1349 static int __bch2_writepage(struct page *page,
1350                             struct writeback_control *wbc,
1351                             void *data)
1352 {
1353         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1354         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1355         struct bch_writepage_state *w = data;
1356         struct bch_page_state *s, orig;
1357         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1358         loff_t i_size = i_size_read(&inode->v);
1359         pgoff_t end_index = i_size >> PAGE_SHIFT;
1360         int ret;
1361
1362         EBUG_ON(!PageUptodate(page));
1363
1364         /* Is the page fully inside i_size? */
1365         if (page->index < end_index)
1366                 goto do_io;
1367
1368         /* Is the page fully outside i_size? (truncate in progress) */
1369         offset = i_size & (PAGE_SIZE - 1);
1370         if (page->index > end_index || !offset) {
1371                 unlock_page(page);
1372                 return 0;
1373         }
1374
1375         /*
1376          * The page straddles i_size.  It must be zeroed out on each and every
1377          * writepage invocation because it may be mmapped.  "A file is mapped
1378          * in multiples of the page size.  For a file that is not a multiple of
1379          * the  page size, the remaining memory is zeroed when mapped, and
1380          * writes to that region are not written out to the file."
1381          */
1382         zero_user_segment(page, offset, PAGE_SIZE);
1383 do_io:
1384         s = bch2_page_state_create(page, __GFP_NOFAIL);
1385
1386         /*
1387          * Things get really hairy with errors during writeback:
1388          */
1389         ret = bch2_get_page_disk_reservation(c, inode, page, false);
1390         BUG_ON(ret);
1391
1392         /* Before unlocking the page, get copy of reservations: */
1393         spin_lock(&s->lock);
1394         orig = *s;
1395         spin_unlock(&s->lock);
1396
1397         for (i = 0; i < PAGE_SECTORS; i++) {
1398                 if (s->s[i].state < SECTOR_DIRTY)
1399                         continue;
1400
1401                 nr_replicas_this_write =
1402                         min_t(unsigned, nr_replicas_this_write,
1403                               s->s[i].nr_replicas +
1404                               s->s[i].replicas_reserved);
1405         }
1406
1407         for (i = 0; i < PAGE_SECTORS; i++) {
1408                 if (s->s[i].state < SECTOR_DIRTY)
1409                         continue;
1410
1411                 s->s[i].nr_replicas = w->opts.compression
1412                         ? 0 : nr_replicas_this_write;
1413
1414                 s->s[i].replicas_reserved = 0;
1415                 s->s[i].state = SECTOR_ALLOCATED;
1416         }
1417
1418         BUG_ON(atomic_read(&s->write_count));
1419         atomic_set(&s->write_count, 1);
1420
1421         BUG_ON(PageWriteback(page));
1422         set_page_writeback(page);
1423
1424         unlock_page(page);
1425
1426         offset = 0;
1427         while (1) {
1428                 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1429                 u64 sector;
1430
1431                 while (offset < PAGE_SECTORS &&
1432                        orig.s[offset].state < SECTOR_DIRTY)
1433                         offset++;
1434
1435                 if (offset == PAGE_SECTORS)
1436                         break;
1437
1438                 while (offset + sectors < PAGE_SECTORS &&
1439                        orig.s[offset + sectors].state >= SECTOR_DIRTY) {
1440                         reserved_sectors += orig.s[offset + sectors].replicas_reserved;
1441                         dirty_sectors += orig.s[offset + sectors].state == SECTOR_DIRTY;
1442                         sectors++;
1443                 }
1444                 BUG_ON(!sectors);
1445
1446                 sector = ((u64) page->index << PAGE_SECTORS_SHIFT) + offset;
1447
1448                 if (w->io &&
1449                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1450                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1451                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1452                      (BIO_MAX_VECS * PAGE_SIZE) ||
1453                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1454                         bch2_writepage_do_io(w);
1455
1456                 if (!w->io)
1457                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1458                                                 nr_replicas_this_write);
1459
1460                 atomic_inc(&s->write_count);
1461
1462                 BUG_ON(inode != w->io->inode);
1463                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1464                                      sectors << 9, offset << 9));
1465
1466                 /* Check for writing past i_size: */
1467                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1468                         round_up(i_size, block_bytes(c)));
1469
1470                 w->io->op.res.sectors += reserved_sectors;
1471                 w->io->op.i_sectors_delta -= dirty_sectors;
1472                 w->io->op.new_i_size = i_size;
1473
1474                 offset += sectors;
1475         }
1476
1477         if (atomic_dec_and_test(&s->write_count))
1478                 end_page_writeback(page);
1479
1480         return 0;
1481 }
1482
1483 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1484 {
1485         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1486         struct bch_writepage_state w =
1487                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1488         struct blk_plug plug;
1489         int ret;
1490
1491         blk_start_plug(&plug);
1492         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1493         if (w.io)
1494                 bch2_writepage_do_io(&w);
1495         blk_finish_plug(&plug);
1496         return ret;
1497 }
1498
1499 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1500 {
1501         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1502         struct bch_writepage_state w =
1503                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1504         int ret;
1505
1506         ret = __bch2_writepage(page, wbc, &w);
1507         if (w.io)
1508                 bch2_writepage_do_io(&w);
1509
1510         return ret;
1511 }
1512
1513 /* buffered writes: */
1514
1515 int bch2_write_begin(struct file *file, struct address_space *mapping,
1516                      loff_t pos, unsigned len, unsigned flags,
1517                      struct page **pagep, void **fsdata)
1518 {
1519         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1520         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1521         struct bch2_page_reservation *res;
1522         pgoff_t index = pos >> PAGE_SHIFT;
1523         unsigned offset = pos & (PAGE_SIZE - 1);
1524         struct page *page;
1525         int ret = -ENOMEM;
1526
1527         res = kmalloc(sizeof(*res), GFP_KERNEL);
1528         if (!res)
1529                 return -ENOMEM;
1530
1531         bch2_page_reservation_init(c, inode, res);
1532         *fsdata = res;
1533
1534         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1535
1536         page = grab_cache_page_write_begin(mapping, index, flags);
1537         if (!page)
1538                 goto err_unlock;
1539
1540         if (PageUptodate(page))
1541                 goto out;
1542
1543         /* If we're writing entire page, don't need to read it in first: */
1544         if (len == PAGE_SIZE)
1545                 goto out;
1546
1547         if (!offset && pos + len >= inode->v.i_size) {
1548                 zero_user_segment(page, len, PAGE_SIZE);
1549                 flush_dcache_page(page);
1550                 goto out;
1551         }
1552
1553         if (index > inode->v.i_size >> PAGE_SHIFT) {
1554                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1555                 flush_dcache_page(page);
1556                 goto out;
1557         }
1558 readpage:
1559         ret = bch2_read_single_page(page, mapping);
1560         if (ret)
1561                 goto err;
1562 out:
1563         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1564                 ret = bch2_page_state_set(c, inode_inum(inode), &page, 1);
1565                 if (ret)
1566                         goto out;
1567         }
1568
1569         ret = bch2_page_reservation_get(c, inode, page, res,
1570                                         offset, len, true);
1571         if (ret) {
1572                 if (!PageUptodate(page)) {
1573                         /*
1574                          * If the page hasn't been read in, we won't know if we
1575                          * actually need a reservation - we don't actually need
1576                          * to read here, we just need to check if the page is
1577                          * fully backed by uncompressed data:
1578                          */
1579                         goto readpage;
1580                 }
1581
1582                 goto err;
1583         }
1584
1585         *pagep = page;
1586         return 0;
1587 err:
1588         unlock_page(page);
1589         put_page(page);
1590         *pagep = NULL;
1591 err_unlock:
1592         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1593         kfree(res);
1594         *fsdata = NULL;
1595         return ret;
1596 }
1597
1598 int bch2_write_end(struct file *file, struct address_space *mapping,
1599                    loff_t pos, unsigned len, unsigned copied,
1600                    struct page *page, void *fsdata)
1601 {
1602         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1603         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1604         struct bch2_page_reservation *res = fsdata;
1605         unsigned offset = pos & (PAGE_SIZE - 1);
1606
1607         lockdep_assert_held(&inode->v.i_rwsem);
1608
1609         if (unlikely(copied < len && !PageUptodate(page))) {
1610                 /*
1611                  * The page needs to be read in, but that would destroy
1612                  * our partial write - simplest thing is to just force
1613                  * userspace to redo the write:
1614                  */
1615                 zero_user(page, 0, PAGE_SIZE);
1616                 flush_dcache_page(page);
1617                 copied = 0;
1618         }
1619
1620         spin_lock(&inode->v.i_lock);
1621         if (pos + copied > inode->v.i_size)
1622                 i_size_write(&inode->v, pos + copied);
1623         spin_unlock(&inode->v.i_lock);
1624
1625         if (copied) {
1626                 if (!PageUptodate(page))
1627                         SetPageUptodate(page);
1628
1629                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1630
1631                 inode->ei_last_dirtied = (unsigned long) current;
1632         }
1633
1634         unlock_page(page);
1635         put_page(page);
1636         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1637
1638         bch2_page_reservation_put(c, inode, res);
1639         kfree(res);
1640
1641         return copied;
1642 }
1643
1644 #define WRITE_BATCH_PAGES       32
1645
1646 static int __bch2_buffered_write(struct bch_inode_info *inode,
1647                                  struct address_space *mapping,
1648                                  struct iov_iter *iter,
1649                                  loff_t pos, unsigned len)
1650 {
1651         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1652         struct page *pages[WRITE_BATCH_PAGES];
1653         struct bch2_page_reservation res;
1654         unsigned long index = pos >> PAGE_SHIFT;
1655         unsigned offset = pos & (PAGE_SIZE - 1);
1656         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1657         unsigned i, reserved = 0, set_dirty = 0;
1658         unsigned copied = 0, nr_pages_copied = 0;
1659         int ret = 0;
1660
1661         BUG_ON(!len);
1662         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1663
1664         bch2_page_reservation_init(c, inode, &res);
1665
1666         for (i = 0; i < nr_pages; i++) {
1667                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1668                 if (!pages[i]) {
1669                         nr_pages = i;
1670                         if (!i) {
1671                                 ret = -ENOMEM;
1672                                 goto out;
1673                         }
1674                         len = min_t(unsigned, len,
1675                                     nr_pages * PAGE_SIZE - offset);
1676                         break;
1677                 }
1678         }
1679
1680         if (offset && !PageUptodate(pages[0])) {
1681                 ret = bch2_read_single_page(pages[0], mapping);
1682                 if (ret)
1683                         goto out;
1684         }
1685
1686         if ((pos + len) & (PAGE_SIZE - 1) &&
1687             !PageUptodate(pages[nr_pages - 1])) {
1688                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1689                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1690                 } else {
1691                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1692                         if (ret)
1693                                 goto out;
1694                 }
1695         }
1696
1697         while (reserved < len) {
1698                 unsigned i = (offset + reserved) >> PAGE_SHIFT;
1699                 struct page *page = pages[i];
1700                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1701                 unsigned pg_len = min_t(unsigned, len - reserved,
1702                                         PAGE_SIZE - pg_offset);
1703
1704                 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1705                         ret = bch2_page_state_set(c, inode_inum(inode),
1706                                                   pages + i, nr_pages - i);
1707                         if (ret)
1708                                 goto out;
1709                 }
1710
1711                 ret = bch2_page_reservation_get(c, inode, page, &res,
1712                                                 pg_offset, pg_len, true);
1713                 if (ret)
1714                         goto out;
1715
1716                 reserved += pg_len;
1717         }
1718
1719         if (mapping_writably_mapped(mapping))
1720                 for (i = 0; i < nr_pages; i++)
1721                         flush_dcache_page(pages[i]);
1722
1723         while (copied < len) {
1724                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1725                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1726                 unsigned pg_len = min_t(unsigned, len - copied,
1727                                         PAGE_SIZE - pg_offset);
1728                 unsigned pg_copied = copy_page_from_iter_atomic(page,
1729                                                 pg_offset, pg_len,iter);
1730
1731                 if (!pg_copied)
1732                         break;
1733
1734                 if (!PageUptodate(page) &&
1735                     pg_copied != PAGE_SIZE &&
1736                     pos + copied + pg_copied < inode->v.i_size) {
1737                         zero_user(page, 0, PAGE_SIZE);
1738                         break;
1739                 }
1740
1741                 flush_dcache_page(page);
1742                 copied += pg_copied;
1743
1744                 if (pg_copied != pg_len)
1745                         break;
1746         }
1747
1748         if (!copied)
1749                 goto out;
1750
1751         spin_lock(&inode->v.i_lock);
1752         if (pos + copied > inode->v.i_size)
1753                 i_size_write(&inode->v, pos + copied);
1754         spin_unlock(&inode->v.i_lock);
1755
1756         while (set_dirty < copied) {
1757                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1758                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1759                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1760                                         PAGE_SIZE - pg_offset);
1761
1762                 if (!PageUptodate(page))
1763                         SetPageUptodate(page);
1764
1765                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1766                 unlock_page(page);
1767                 put_page(page);
1768
1769                 set_dirty += pg_len;
1770         }
1771
1772         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1773         inode->ei_last_dirtied = (unsigned long) current;
1774 out:
1775         for (i = nr_pages_copied; i < nr_pages; i++) {
1776                 unlock_page(pages[i]);
1777                 put_page(pages[i]);
1778         }
1779
1780         bch2_page_reservation_put(c, inode, &res);
1781
1782         return copied ?: ret;
1783 }
1784
1785 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1786 {
1787         struct file *file = iocb->ki_filp;
1788         struct address_space *mapping = file->f_mapping;
1789         struct bch_inode_info *inode = file_bch_inode(file);
1790         loff_t pos = iocb->ki_pos;
1791         ssize_t written = 0;
1792         int ret = 0;
1793
1794         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1795
1796         do {
1797                 unsigned offset = pos & (PAGE_SIZE - 1);
1798                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1799                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1800 again:
1801                 /*
1802                  * Bring in the user page that we will copy from _first_.
1803                  * Otherwise there's a nasty deadlock on copying from the
1804                  * same page as we're writing to, without it being marked
1805                  * up-to-date.
1806                  *
1807                  * Not only is this an optimisation, but it is also required
1808                  * to check that the address is actually valid, when atomic
1809                  * usercopies are used, below.
1810                  */
1811                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1812                         bytes = min_t(unsigned long, iov_iter_count(iter),
1813                                       PAGE_SIZE - offset);
1814
1815                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1816                                 ret = -EFAULT;
1817                                 break;
1818                         }
1819                 }
1820
1821                 if (unlikely(fatal_signal_pending(current))) {
1822                         ret = -EINTR;
1823                         break;
1824                 }
1825
1826                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1827                 if (unlikely(ret < 0))
1828                         break;
1829
1830                 cond_resched();
1831
1832                 if (unlikely(ret == 0)) {
1833                         /*
1834                          * If we were unable to copy any data at all, we must
1835                          * fall back to a single segment length write.
1836                          *
1837                          * If we didn't fallback here, we could livelock
1838                          * because not all segments in the iov can be copied at
1839                          * once without a pagefault.
1840                          */
1841                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1842                                       iov_iter_single_seg_count(iter));
1843                         goto again;
1844                 }
1845                 pos += ret;
1846                 written += ret;
1847                 ret = 0;
1848
1849                 balance_dirty_pages_ratelimited(mapping);
1850         } while (iov_iter_count(iter));
1851
1852         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1853
1854         return written ? written : ret;
1855 }
1856
1857 /* O_DIRECT reads */
1858
1859 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1860 {
1861         if (check_dirty) {
1862                 bio_check_pages_dirty(bio);
1863         } else {
1864                 bio_release_pages(bio, false);
1865                 bio_put(bio);
1866         }
1867 }
1868
1869 static void bch2_dio_read_complete(struct closure *cl)
1870 {
1871         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1872
1873         dio->req->ki_complete(dio->req, dio->ret, 0);
1874         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1875 }
1876
1877 static void bch2_direct_IO_read_endio(struct bio *bio)
1878 {
1879         struct dio_read *dio = bio->bi_private;
1880
1881         if (bio->bi_status)
1882                 dio->ret = blk_status_to_errno(bio->bi_status);
1883
1884         closure_put(&dio->cl);
1885 }
1886
1887 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1888 {
1889         struct dio_read *dio = bio->bi_private;
1890         bool should_dirty = dio->should_dirty;
1891
1892         bch2_direct_IO_read_endio(bio);
1893         bio_check_or_release(bio, should_dirty);
1894 }
1895
1896 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1897 {
1898         struct file *file = req->ki_filp;
1899         struct bch_inode_info *inode = file_bch_inode(file);
1900         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1901         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1902         struct dio_read *dio;
1903         struct bio *bio;
1904         loff_t offset = req->ki_pos;
1905         bool sync = is_sync_kiocb(req);
1906         size_t shorten;
1907         ssize_t ret;
1908
1909         if ((offset|iter->count) & (block_bytes(c) - 1))
1910                 return -EINVAL;
1911
1912         ret = min_t(loff_t, iter->count,
1913                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1914
1915         if (!ret)
1916                 return ret;
1917
1918         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1919         iter->count -= shorten;
1920
1921         bio = bio_alloc_bioset(GFP_KERNEL,
1922                                iov_iter_npages(iter, BIO_MAX_VECS),
1923                                &c->dio_read_bioset);
1924
1925         bio->bi_end_io = bch2_direct_IO_read_endio;
1926
1927         dio = container_of(bio, struct dio_read, rbio.bio);
1928         closure_init(&dio->cl, NULL);
1929
1930         /*
1931          * this is a _really_ horrible hack just to avoid an atomic sub at the
1932          * end:
1933          */
1934         if (!sync) {
1935                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1936                 atomic_set(&dio->cl.remaining,
1937                            CLOSURE_REMAINING_INITIALIZER -
1938                            CLOSURE_RUNNING +
1939                            CLOSURE_DESTRUCTOR);
1940         } else {
1941                 atomic_set(&dio->cl.remaining,
1942                            CLOSURE_REMAINING_INITIALIZER + 1);
1943         }
1944
1945         dio->req        = req;
1946         dio->ret        = ret;
1947         /*
1948          * This is one of the sketchier things I've encountered: we have to skip
1949          * the dirtying of requests that are internal from the kernel (i.e. from
1950          * loopback), because we'll deadlock on page_lock.
1951          */
1952         dio->should_dirty = iter_is_iovec(iter);
1953
1954         goto start;
1955         while (iter->count) {
1956                 bio = bio_alloc_bioset(GFP_KERNEL,
1957                                        iov_iter_npages(iter, BIO_MAX_VECS),
1958                                        &c->bio_read);
1959                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1960 start:
1961                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1962                 bio->bi_iter.bi_sector  = offset >> 9;
1963                 bio->bi_private         = dio;
1964
1965                 ret = bio_iov_iter_get_pages(bio, iter);
1966                 if (ret < 0) {
1967                         /* XXX: fault inject this path */
1968                         bio->bi_status = BLK_STS_RESOURCE;
1969                         bio_endio(bio);
1970                         break;
1971                 }
1972
1973                 offset += bio->bi_iter.bi_size;
1974
1975                 if (dio->should_dirty)
1976                         bio_set_pages_dirty(bio);
1977
1978                 if (iter->count)
1979                         closure_get(&dio->cl);
1980
1981                 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
1982         }
1983
1984         iter->count += shorten;
1985
1986         if (sync) {
1987                 closure_sync(&dio->cl);
1988                 closure_debug_destroy(&dio->cl);
1989                 ret = dio->ret;
1990                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1991                 return ret;
1992         } else {
1993                 return -EIOCBQUEUED;
1994         }
1995 }
1996
1997 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1998 {
1999         struct file *file = iocb->ki_filp;
2000         struct bch_inode_info *inode = file_bch_inode(file);
2001         struct address_space *mapping = file->f_mapping;
2002         size_t count = iov_iter_count(iter);
2003         ssize_t ret;
2004
2005         if (!count)
2006                 return 0; /* skip atime */
2007
2008         if (iocb->ki_flags & IOCB_DIRECT) {
2009                 struct blk_plug plug;
2010
2011                 ret = filemap_write_and_wait_range(mapping,
2012                                         iocb->ki_pos,
2013                                         iocb->ki_pos + count - 1);
2014                 if (ret < 0)
2015                         return ret;
2016
2017                 file_accessed(file);
2018
2019                 blk_start_plug(&plug);
2020                 ret = bch2_direct_IO_read(iocb, iter);
2021                 blk_finish_plug(&plug);
2022
2023                 if (ret >= 0)
2024                         iocb->ki_pos += ret;
2025         } else {
2026                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
2027                 ret = generic_file_read_iter(iocb, iter);
2028                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
2029         }
2030
2031         return ret;
2032 }
2033
2034 /* O_DIRECT writes */
2035
2036 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2037                                        u64 offset, u64 size,
2038                                        unsigned nr_replicas, bool compressed)
2039 {
2040         struct btree_trans trans;
2041         struct btree_iter iter;
2042         struct bkey_s_c k;
2043         u64 end = offset + size;
2044         u32 snapshot;
2045         bool ret = true;
2046         int err;
2047
2048         bch2_trans_init(&trans, c, 0, 0);
2049 retry:
2050         bch2_trans_begin(&trans);
2051
2052         err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2053         if (err)
2054                 goto err;
2055
2056         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2057                            SPOS(inum.inum, offset, snapshot),
2058                            BTREE_ITER_SLOTS, k, err) {
2059                 if (bkey_cmp(bkey_start_pos(k.k), POS(inum.inum, end)) >= 0)
2060                         break;
2061
2062                 if (k.k->p.snapshot != snapshot ||
2063                     nr_replicas > bch2_bkey_replicas(c, k) ||
2064                     (!compressed && bch2_bkey_sectors_compressed(k))) {
2065                         ret = false;
2066                         break;
2067                 }
2068         }
2069
2070         offset = iter.pos.offset;
2071         bch2_trans_iter_exit(&trans, &iter);
2072 err:
2073         if (err == -EINTR)
2074                 goto retry;
2075         bch2_trans_exit(&trans);
2076
2077         return err ? false : ret;
2078 }
2079
2080 static void bch2_dio_write_loop_async(struct bch_write_op *);
2081
2082 static long bch2_dio_write_loop(struct dio_write *dio)
2083 {
2084         bool kthread = (current->flags & PF_KTHREAD) != 0;
2085         struct kiocb *req = dio->req;
2086         struct address_space *mapping = req->ki_filp->f_mapping;
2087         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
2088         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2089         struct bio *bio = &dio->op.wbio.bio;
2090         struct bvec_iter_all iter;
2091         struct bio_vec *bv;
2092         unsigned unaligned, iter_count;
2093         bool sync = dio->sync, dropped_locks;
2094         long ret;
2095
2096         if (dio->loop)
2097                 goto loop;
2098
2099         down(&c->io_in_flight);
2100
2101         while (1) {
2102                 iter_count = dio->iter.count;
2103
2104                 if (kthread)
2105                         kthread_use_mm(dio->mm);
2106                 BUG_ON(current->faults_disabled_mapping);
2107                 current->faults_disabled_mapping = mapping;
2108
2109                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2110
2111                 dropped_locks = fdm_dropped_locks();
2112
2113                 current->faults_disabled_mapping = NULL;
2114                 if (kthread)
2115                         kthread_unuse_mm(dio->mm);
2116
2117                 /*
2118                  * If the fault handler returned an error but also signalled
2119                  * that it dropped & retook ei_pagecache_lock, we just need to
2120                  * re-shoot down the page cache and retry:
2121                  */
2122                 if (dropped_locks && ret)
2123                         ret = 0;
2124
2125                 if (unlikely(ret < 0))
2126                         goto err;
2127
2128                 if (unlikely(dropped_locks)) {
2129                         ret = write_invalidate_inode_pages_range(mapping,
2130                                         req->ki_pos,
2131                                         req->ki_pos + iter_count - 1);
2132                         if (unlikely(ret))
2133                                 goto err;
2134
2135                         if (!bio->bi_iter.bi_size)
2136                                 continue;
2137                 }
2138
2139                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2140                 bio->bi_iter.bi_size -= unaligned;
2141                 iov_iter_revert(&dio->iter, unaligned);
2142
2143                 if (!bio->bi_iter.bi_size) {
2144                         /*
2145                          * bio_iov_iter_get_pages was only able to get <
2146                          * blocksize worth of pages:
2147                          */
2148                         ret = -EFAULT;
2149                         goto err;
2150                 }
2151
2152                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
2153                 dio->op.end_io          = bch2_dio_write_loop_async;
2154                 dio->op.target          = dio->op.opts.foreground_target;
2155                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
2156                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
2157                 dio->op.subvol          = inode->ei_subvol;
2158                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2159
2160                 if ((req->ki_flags & IOCB_DSYNC) &&
2161                     !c->opts.journal_flush_disabled)
2162                         dio->op.flags |= BCH_WRITE_FLUSH;
2163                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2164
2165                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2166                                                 dio->op.opts.data_replicas, 0);
2167                 if (unlikely(ret) &&
2168                     !bch2_check_range_allocated(c, inode_inum(inode),
2169                                 dio->op.pos.offset, bio_sectors(bio),
2170                                 dio->op.opts.data_replicas,
2171                                 dio->op.opts.compression != 0))
2172                         goto err;
2173
2174                 task_io_account_write(bio->bi_iter.bi_size);
2175
2176                 if (!dio->sync && !dio->loop && dio->iter.count) {
2177                         struct iovec *iov = dio->inline_vecs;
2178
2179                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2180                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
2181                                               GFP_KERNEL);
2182                                 if (unlikely(!iov)) {
2183                                         dio->sync = sync = true;
2184                                         goto do_io;
2185                                 }
2186
2187                                 dio->free_iov = true;
2188                         }
2189
2190                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2191                         dio->iter.iov = iov;
2192                 }
2193 do_io:
2194                 dio->loop = true;
2195                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2196
2197                 if (sync)
2198                         wait_for_completion(&dio->done);
2199                 else
2200                         return -EIOCBQUEUED;
2201 loop:
2202                 i_sectors_acct(c, inode, &dio->quota_res,
2203                                dio->op.i_sectors_delta);
2204                 req->ki_pos += (u64) dio->op.written << 9;
2205                 dio->written += dio->op.written;
2206
2207                 spin_lock(&inode->v.i_lock);
2208                 if (req->ki_pos > inode->v.i_size)
2209                         i_size_write(&inode->v, req->ki_pos);
2210                 spin_unlock(&inode->v.i_lock);
2211
2212                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2213                         bio_for_each_segment_all(bv, bio, iter)
2214                                 put_page(bv->bv_page);
2215                 bio->bi_vcnt = 0;
2216
2217                 if (dio->op.error) {
2218                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
2219                         break;
2220                 }
2221
2222                 if (!dio->iter.count)
2223                         break;
2224
2225                 bio_reset(bio);
2226                 reinit_completion(&dio->done);
2227         }
2228
2229         ret = dio->op.error ?: ((long) dio->written << 9);
2230 err:
2231         up(&c->io_in_flight);
2232         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2233         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2234
2235         if (dio->free_iov)
2236                 kfree(dio->iter.iov);
2237
2238         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2239                 bio_for_each_segment_all(bv, bio, iter)
2240                         put_page(bv->bv_page);
2241         bio_put(bio);
2242
2243         /* inode->i_dio_count is our ref on inode and thus bch_fs */
2244         inode_dio_end(&inode->v);
2245
2246         if (!sync) {
2247                 req->ki_complete(req, ret, 0);
2248                 ret = -EIOCBQUEUED;
2249         }
2250         return ret;
2251 }
2252
2253 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2254 {
2255         struct dio_write *dio = container_of(op, struct dio_write, op);
2256
2257         if (dio->sync)
2258                 complete(&dio->done);
2259         else
2260                 bch2_dio_write_loop(dio);
2261 }
2262
2263 static noinline
2264 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2265 {
2266         struct file *file = req->ki_filp;
2267         struct address_space *mapping = file->f_mapping;
2268         struct bch_inode_info *inode = file_bch_inode(file);
2269         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2270         struct dio_write *dio;
2271         struct bio *bio;
2272         bool locked = true, extending;
2273         ssize_t ret;
2274
2275         prefetch(&c->opts);
2276         prefetch((void *) &c->opts + 64);
2277         prefetch(&inode->ei_inode);
2278         prefetch((void *) &inode->ei_inode + 64);
2279
2280         inode_lock(&inode->v);
2281
2282         ret = generic_write_checks(req, iter);
2283         if (unlikely(ret <= 0))
2284                 goto err;
2285
2286         ret = file_remove_privs(file);
2287         if (unlikely(ret))
2288                 goto err;
2289
2290         ret = file_update_time(file);
2291         if (unlikely(ret))
2292                 goto err;
2293
2294         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2295                 goto err;
2296
2297         inode_dio_begin(&inode->v);
2298         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2299
2300         extending = req->ki_pos + iter->count > inode->v.i_size;
2301         if (!extending) {
2302                 inode_unlock(&inode->v);
2303                 locked = false;
2304         }
2305
2306         bio = bio_alloc_bioset(GFP_KERNEL,
2307                                iov_iter_is_bvec(iter)
2308                                ? 0
2309                                : iov_iter_npages(iter, BIO_MAX_VECS),
2310                                &c->dio_write_bioset);
2311         dio = container_of(bio, struct dio_write, op.wbio.bio);
2312         init_completion(&dio->done);
2313         dio->req                = req;
2314         dio->mm                 = current->mm;
2315         dio->loop               = false;
2316         dio->sync               = is_sync_kiocb(req) || extending;
2317         dio->free_iov           = false;
2318         dio->quota_res.sectors  = 0;
2319         dio->written            = 0;
2320         dio->iter               = *iter;
2321
2322         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2323                                          iter->count >> 9, true);
2324         if (unlikely(ret))
2325                 goto err_put_bio;
2326
2327         ret = write_invalidate_inode_pages_range(mapping,
2328                                         req->ki_pos,
2329                                         req->ki_pos + iter->count - 1);
2330         if (unlikely(ret))
2331                 goto err_put_bio;
2332
2333         ret = bch2_dio_write_loop(dio);
2334 err:
2335         if (locked)
2336                 inode_unlock(&inode->v);
2337         return ret;
2338 err_put_bio:
2339         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2340         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2341         bio_put(bio);
2342         inode_dio_end(&inode->v);
2343         goto err;
2344 }
2345
2346 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2347 {
2348         struct file *file = iocb->ki_filp;
2349         struct bch_inode_info *inode = file_bch_inode(file);
2350         ssize_t ret;
2351
2352         if (iocb->ki_flags & IOCB_DIRECT)
2353                 return bch2_direct_write(iocb, from);
2354
2355         /* We can write back this queue in page reclaim */
2356         current->backing_dev_info = inode_to_bdi(&inode->v);
2357         inode_lock(&inode->v);
2358
2359         ret = generic_write_checks(iocb, from);
2360         if (ret <= 0)
2361                 goto unlock;
2362
2363         ret = file_remove_privs(file);
2364         if (ret)
2365                 goto unlock;
2366
2367         ret = file_update_time(file);
2368         if (ret)
2369                 goto unlock;
2370
2371         ret = bch2_buffered_write(iocb, from);
2372         if (likely(ret > 0))
2373                 iocb->ki_pos += ret;
2374 unlock:
2375         inode_unlock(&inode->v);
2376         current->backing_dev_info = NULL;
2377
2378         if (ret > 0)
2379                 ret = generic_write_sync(iocb, ret);
2380
2381         return ret;
2382 }
2383
2384 /* fsync: */
2385
2386 /*
2387  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2388  * insert trigger: look up the btree inode instead
2389  */
2390 static int bch2_flush_inode(struct bch_fs *c, subvol_inum inum)
2391 {
2392         struct bch_inode_unpacked inode;
2393         int ret;
2394
2395         if (c->opts.journal_flush_disabled)
2396                 return 0;
2397
2398         ret = bch2_inode_find_by_inum(c, inum, &inode);
2399         if (ret)
2400                 return ret;
2401
2402         return bch2_journal_flush_seq(&c->journal, inode.bi_journal_seq);
2403 }
2404
2405 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2406 {
2407         struct bch_inode_info *inode = file_bch_inode(file);
2408         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2409         int ret, ret2, ret3;
2410
2411         ret = file_write_and_wait_range(file, start, end);
2412         ret2 = sync_inode_metadata(&inode->v, 1);
2413         ret3 = bch2_flush_inode(c, inode_inum(inode));
2414
2415         return ret ?: ret2 ?: ret3;
2416 }
2417
2418 /* truncate: */
2419
2420 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2421                                  struct bpos start,
2422                                  struct bpos end)
2423 {
2424         struct btree_trans trans;
2425         struct btree_iter iter;
2426         struct bkey_s_c k;
2427         int ret = 0;
2428
2429         bch2_trans_init(&trans, c, 0, 0);
2430 retry:
2431         bch2_trans_begin(&trans);
2432
2433         ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2434         if (ret)
2435                 goto err;
2436
2437         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2438                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2439                         break;
2440
2441                 if (bkey_extent_is_data(k.k)) {
2442                         ret = 1;
2443                         break;
2444                 }
2445         }
2446         start = iter.pos;
2447         bch2_trans_iter_exit(&trans, &iter);
2448 err:
2449         if (ret == -EINTR)
2450                 goto retry;
2451
2452         bch2_trans_exit(&trans);
2453         return ret;
2454 }
2455
2456 static int __bch2_truncate_page(struct bch_inode_info *inode,
2457                                 pgoff_t index, loff_t start, loff_t end)
2458 {
2459         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2460         struct address_space *mapping = inode->v.i_mapping;
2461         struct bch_page_state *s;
2462         unsigned start_offset = start & (PAGE_SIZE - 1);
2463         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2464         unsigned i;
2465         struct page *page;
2466         s64 i_sectors_delta = 0;
2467         int ret = 0;
2468
2469         /* Page boundary? Nothing to do */
2470         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2471               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2472                 return 0;
2473
2474         /* Above i_size? */
2475         if (index << PAGE_SHIFT >= inode->v.i_size)
2476                 return 0;
2477
2478         page = find_lock_page(mapping, index);
2479         if (!page) {
2480                 /*
2481                  * XXX: we're doing two index lookups when we end up reading the
2482                  * page
2483                  */
2484                 ret = range_has_data(c, inode->ei_subvol,
2485                                 POS(inode->v.i_ino, index << PAGE_SECTORS_SHIFT),
2486                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTORS_SHIFT));
2487                 if (ret <= 0)
2488                         return ret;
2489
2490                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2491                 if (unlikely(!page)) {
2492                         ret = -ENOMEM;
2493                         goto out;
2494                 }
2495         }
2496
2497         s = bch2_page_state_create(page, 0);
2498         if (!s) {
2499                 ret = -ENOMEM;
2500                 goto unlock;
2501         }
2502
2503         if (!PageUptodate(page)) {
2504                 ret = bch2_read_single_page(page, mapping);
2505                 if (ret)
2506                         goto unlock;
2507         }
2508
2509         if (index != start >> PAGE_SHIFT)
2510                 start_offset = 0;
2511         if (index != end >> PAGE_SHIFT)
2512                 end_offset = PAGE_SIZE;
2513
2514         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2515              i < round_down(end_offset, block_bytes(c)) >> 9;
2516              i++) {
2517                 s->s[i].nr_replicas     = 0;
2518                 if (s->s[i].state == SECTOR_DIRTY)
2519                         i_sectors_delta--;
2520                 s->s[i].state           = SECTOR_UNALLOCATED;
2521         }
2522
2523         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2524
2525         /*
2526          * Caller needs to know whether this page will be written out by
2527          * writeback - doing an i_size update if necessary - or whether it will
2528          * be responsible for the i_size update:
2529          */
2530         ret = s->s[(min_t(u64, inode->v.i_size - (index << PAGE_SHIFT),
2531                           PAGE_SIZE) - 1) >> 9].state >= SECTOR_DIRTY;
2532
2533         zero_user_segment(page, start_offset, end_offset);
2534
2535         /*
2536          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2537          *
2538          * XXX: because we aren't currently tracking whether the page has actual
2539          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2540          */
2541         BUG_ON(bch2_get_page_disk_reservation(c, inode, page, false));
2542
2543         /*
2544          * This removes any writeable userspace mappings; we need to force
2545          * .page_mkwrite to be called again before any mmapped writes, to
2546          * redirty the full page:
2547          */
2548         page_mkclean(page);
2549         __set_page_dirty_nobuffers(page);
2550 unlock:
2551         unlock_page(page);
2552         put_page(page);
2553 out:
2554         return ret;
2555 }
2556
2557 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2558 {
2559         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2560                                     from, round_up(from, PAGE_SIZE));
2561 }
2562
2563 static int bch2_truncate_pages(struct bch_inode_info *inode,
2564                                loff_t start, loff_t end)
2565 {
2566         int ret = __bch2_truncate_page(inode, start >> PAGE_SHIFT,
2567                                        start, end);
2568
2569         if (ret >= 0 &&
2570             start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2571                 ret = __bch2_truncate_page(inode,
2572                                            end >> PAGE_SHIFT,
2573                                            start, end);
2574         return ret;
2575 }
2576
2577 static int bch2_extend(struct user_namespace *mnt_userns,
2578                        struct bch_inode_info *inode,
2579                        struct bch_inode_unpacked *inode_u,
2580                        struct iattr *iattr)
2581 {
2582         struct address_space *mapping = inode->v.i_mapping;
2583         int ret;
2584
2585         /*
2586          * sync appends:
2587          *
2588          * this has to be done _before_ extending i_size:
2589          */
2590         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2591         if (ret)
2592                 return ret;
2593
2594         truncate_setsize(&inode->v, iattr->ia_size);
2595
2596         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2597 }
2598
2599 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2600                                    struct bch_inode_unpacked *bi,
2601                                    void *p)
2602 {
2603         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2604         return 0;
2605 }
2606
2607 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2608                                   struct bch_inode_unpacked *bi, void *p)
2609 {
2610         u64 *new_i_size = p;
2611
2612         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2613         bi->bi_size = *new_i_size;
2614         return 0;
2615 }
2616
2617 int bch2_truncate(struct user_namespace *mnt_userns,
2618                   struct bch_inode_info *inode, struct iattr *iattr)
2619 {
2620         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2621         struct address_space *mapping = inode->v.i_mapping;
2622         struct bch_inode_unpacked inode_u;
2623         u64 new_i_size = iattr->ia_size;
2624         s64 i_sectors_delta = 0;
2625         int ret = 0;
2626
2627         /*
2628          * If the truncate call with change the size of the file, the
2629          * cmtimes should be updated. If the size will not change, we
2630          * do not need to update the cmtimes.
2631          */
2632         if (iattr->ia_size != inode->v.i_size) {
2633                 if (!(iattr->ia_valid & ATTR_MTIME))
2634                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2635                 if (!(iattr->ia_valid & ATTR_CTIME))
2636                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2637                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2638         }
2639
2640         inode_dio_wait(&inode->v);
2641         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2642
2643         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
2644         if (ret)
2645                 goto err;
2646
2647         /*
2648          * check this before next assertion; on filesystem error our normal
2649          * invariants are a bit broken (truncate has to truncate the page cache
2650          * before the inode).
2651          */
2652         ret = bch2_journal_error(&c->journal);
2653         if (ret)
2654                 goto err;
2655
2656         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2657                 inode->v.i_size < inode_u.bi_size);
2658
2659         if (iattr->ia_size > inode->v.i_size) {
2660                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2661                 goto err;
2662         }
2663
2664         iattr->ia_valid &= ~ATTR_SIZE;
2665
2666         ret = bch2_truncate_page(inode, iattr->ia_size);
2667         if (unlikely(ret < 0))
2668                 goto err;
2669
2670         /*
2671          * When extending, we're going to write the new i_size to disk
2672          * immediately so we need to flush anything above the current on disk
2673          * i_size first:
2674          *
2675          * Also, when extending we need to flush the page that i_size currently
2676          * straddles - if it's mapped to userspace, we need to ensure that
2677          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2678          * again to allocate the part of the page that was extended.
2679          */
2680         if (iattr->ia_size > inode_u.bi_size)
2681                 ret = filemap_write_and_wait_range(mapping,
2682                                 inode_u.bi_size,
2683                                 iattr->ia_size - 1);
2684         else if (iattr->ia_size & (PAGE_SIZE - 1))
2685                 ret = filemap_write_and_wait_range(mapping,
2686                                 round_down(iattr->ia_size, PAGE_SIZE),
2687                                 iattr->ia_size - 1);
2688         if (ret)
2689                 goto err;
2690
2691         mutex_lock(&inode->ei_update_lock);
2692         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2693                                &new_i_size, 0);
2694         mutex_unlock(&inode->ei_update_lock);
2695
2696         if (unlikely(ret))
2697                 goto err;
2698
2699         truncate_setsize(&inode->v, iattr->ia_size);
2700
2701         ret = bch2_fpunch(c, inode_inum(inode),
2702                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2703                         U64_MAX, &i_sectors_delta);
2704         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2705
2706         WARN_ON(!inode->v.i_size && inode->v.i_blocks &&
2707                 !bch2_journal_error(&c->journal));
2708
2709         if (unlikely(ret))
2710                 goto err;
2711
2712         mutex_lock(&inode->ei_update_lock);
2713         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2714         mutex_unlock(&inode->ei_update_lock);
2715
2716         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2717 err:
2718         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2719         return ret;
2720 }
2721
2722 /* fallocate: */
2723
2724 static int inode_update_times_fn(struct bch_inode_info *inode,
2725                                  struct bch_inode_unpacked *bi, void *p)
2726 {
2727         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2728
2729         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2730         return 0;
2731 }
2732
2733 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2734 {
2735         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2736         u64 end         = offset + len;
2737         u64 block_start = round_up(offset, block_bytes(c));
2738         u64 block_end   = round_down(end, block_bytes(c));
2739         bool truncated_last_page;
2740         int ret = 0;
2741
2742         ret = bch2_truncate_pages(inode, offset, end);
2743         if (unlikely(ret < 0))
2744                 goto err;
2745
2746         truncated_last_page = ret;
2747
2748         truncate_pagecache_range(&inode->v, offset, end - 1);
2749
2750         if (block_start < block_end ) {
2751                 s64 i_sectors_delta = 0;
2752
2753                 ret = bch2_fpunch(c, inode_inum(inode),
2754                                   block_start >> 9, block_end >> 9,
2755                                   &i_sectors_delta);
2756                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2757         }
2758
2759         mutex_lock(&inode->ei_update_lock);
2760         if (end >= inode->v.i_size && !truncated_last_page) {
2761                 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2762                                             ATTR_MTIME|ATTR_CTIME);
2763         } else {
2764                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2765                                        ATTR_MTIME|ATTR_CTIME);
2766         }
2767         mutex_unlock(&inode->ei_update_lock);
2768 err:
2769         return ret;
2770 }
2771
2772 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2773                                    loff_t offset, loff_t len,
2774                                    bool insert)
2775 {
2776         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2777         struct address_space *mapping = inode->v.i_mapping;
2778         struct bkey_buf copy;
2779         struct btree_trans trans;
2780         struct btree_iter src, dst, del;
2781         loff_t shift, new_size;
2782         u64 src_start;
2783         int ret = 0;
2784
2785         if ((offset | len) & (block_bytes(c) - 1))
2786                 return -EINVAL;
2787
2788         if (insert) {
2789                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2790                         return -EFBIG;
2791
2792                 if (offset >= inode->v.i_size)
2793                         return -EINVAL;
2794
2795                 src_start       = U64_MAX;
2796                 shift           = len;
2797         } else {
2798                 if (offset + len >= inode->v.i_size)
2799                         return -EINVAL;
2800
2801                 src_start       = offset + len;
2802                 shift           = -len;
2803         }
2804
2805         new_size = inode->v.i_size + shift;
2806
2807         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2808         if (ret)
2809                 return ret;
2810
2811         if (insert) {
2812                 i_size_write(&inode->v, new_size);
2813                 mutex_lock(&inode->ei_update_lock);
2814                 ret = bch2_write_inode_size(c, inode, new_size,
2815                                             ATTR_MTIME|ATTR_CTIME);
2816                 mutex_unlock(&inode->ei_update_lock);
2817         } else {
2818                 s64 i_sectors_delta = 0;
2819
2820                 ret = bch2_fpunch(c, inode_inum(inode),
2821                                   offset >> 9, (offset + len) >> 9,
2822                                   &i_sectors_delta);
2823                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2824
2825                 if (ret)
2826                         return ret;
2827         }
2828
2829         bch2_bkey_buf_init(&copy);
2830         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2831         bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
2832                         POS(inode->v.i_ino, src_start >> 9),
2833                         BTREE_ITER_INTENT);
2834         bch2_trans_copy_iter(&dst, &src);
2835         bch2_trans_copy_iter(&del, &src);
2836
2837         while (ret == 0 || ret == -EINTR) {
2838                 struct disk_reservation disk_res =
2839                         bch2_disk_reservation_init(c, 0);
2840                 struct bkey_i delete;
2841                 struct bkey_s_c k;
2842                 struct bpos next_pos;
2843                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2844                 struct bpos atomic_end;
2845                 unsigned trigger_flags = 0;
2846                 u32 snapshot;
2847
2848                 bch2_trans_begin(&trans);
2849
2850                 ret = bch2_subvolume_get_snapshot(&trans,
2851                                         inode->ei_subvol, &snapshot);
2852                 if (ret)
2853                         continue;
2854
2855                 bch2_btree_iter_set_snapshot(&src, snapshot);
2856                 bch2_btree_iter_set_snapshot(&dst, snapshot);
2857                 bch2_btree_iter_set_snapshot(&del, snapshot);
2858
2859                 bch2_trans_begin(&trans);
2860
2861                 k = insert
2862                         ? bch2_btree_iter_peek_prev(&src)
2863                         : bch2_btree_iter_peek(&src);
2864                 if ((ret = bkey_err(k)))
2865                         continue;
2866
2867                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2868                         break;
2869
2870                 if (insert &&
2871                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2872                         break;
2873 reassemble:
2874                 bch2_bkey_buf_reassemble(&copy, c, k);
2875
2876                 if (insert &&
2877                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2878                         bch2_cut_front(move_pos, copy.k);
2879
2880                 copy.k->k.p.offset += shift >> 9;
2881                 bch2_btree_iter_set_pos(&dst, bkey_start_pos(&copy.k->k));
2882
2883                 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
2884                 if (ret)
2885                         continue;
2886
2887                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2888                         if (insert) {
2889                                 move_pos = atomic_end;
2890                                 move_pos.offset -= shift >> 9;
2891                                 goto reassemble;
2892                         } else {
2893                                 bch2_cut_back(atomic_end, copy.k);
2894                         }
2895                 }
2896
2897                 bkey_init(&delete.k);
2898                 delete.k.p = copy.k->k.p;
2899                 delete.k.size = copy.k->k.size;
2900                 delete.k.p.offset -= shift >> 9;
2901                 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
2902
2903                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2904
2905                 if (copy.k->k.size == k.k->size) {
2906                         /*
2907                          * If we're moving the entire extent, we can skip
2908                          * running triggers:
2909                          */
2910                         trigger_flags |= BTREE_TRIGGER_NORUN;
2911                 } else {
2912                         /* We might end up splitting compressed extents: */
2913                         unsigned nr_ptrs =
2914                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2915
2916                         ret = bch2_disk_reservation_get(c, &disk_res,
2917                                         copy.k->k.size, nr_ptrs,
2918                                         BCH_DISK_RESERVATION_NOFAIL);
2919                         BUG_ON(ret);
2920                 }
2921
2922                 ret =   bch2_btree_iter_traverse(&del) ?:
2923                         bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
2924                         bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
2925                         bch2_trans_commit(&trans, &disk_res, NULL,
2926                                           BTREE_INSERT_NOFAIL);
2927                 bch2_disk_reservation_put(c, &disk_res);
2928
2929                 if (!ret)
2930                         bch2_btree_iter_set_pos(&src, next_pos);
2931         }
2932         bch2_trans_iter_exit(&trans, &del);
2933         bch2_trans_iter_exit(&trans, &dst);
2934         bch2_trans_iter_exit(&trans, &src);
2935         bch2_trans_exit(&trans);
2936         bch2_bkey_buf_exit(&copy, c);
2937
2938         if (ret)
2939                 return ret;
2940
2941         mutex_lock(&inode->ei_update_lock);
2942         if (!insert) {
2943                 i_size_write(&inode->v, new_size);
2944                 ret = bch2_write_inode_size(c, inode, new_size,
2945                                             ATTR_MTIME|ATTR_CTIME);
2946         } else {
2947                 /* We need an inode update to update bi_journal_seq for fsync: */
2948                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2949                                        ATTR_MTIME|ATTR_CTIME);
2950         }
2951         mutex_unlock(&inode->ei_update_lock);
2952         return ret;
2953 }
2954
2955 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2956                              u64 start_sector, u64 end_sector)
2957 {
2958         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2959         struct btree_trans trans;
2960         struct btree_iter iter;
2961         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2962         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2963         int ret = 0;
2964
2965         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2966
2967         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2968                         POS(inode->v.i_ino, start_sector),
2969                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2970
2971         while (!ret && bkey_cmp(iter.pos, end_pos) < 0) {
2972                 s64 i_sectors_delta = 0;
2973                 struct disk_reservation disk_res = { 0 };
2974                 struct quota_res quota_res = { 0 };
2975                 struct bkey_i_reservation reservation;
2976                 struct bkey_s_c k;
2977                 unsigned sectors;
2978                 u32 snapshot;
2979
2980                 bch2_trans_begin(&trans);
2981
2982                 ret = bch2_subvolume_get_snapshot(&trans,
2983                                         inode->ei_subvol, &snapshot);
2984                 if (ret)
2985                         goto bkey_err;
2986
2987                 bch2_btree_iter_set_snapshot(&iter, snapshot);
2988
2989                 k = bch2_btree_iter_peek_slot(&iter);
2990                 if ((ret = bkey_err(k)))
2991                         goto bkey_err;
2992
2993                 /* already reserved */
2994                 if (k.k->type == KEY_TYPE_reservation &&
2995                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2996                         bch2_btree_iter_advance(&iter);
2997                         continue;
2998                 }
2999
3000                 if (bkey_extent_is_data(k.k) &&
3001                     !(mode & FALLOC_FL_ZERO_RANGE)) {
3002                         bch2_btree_iter_advance(&iter);
3003                         continue;
3004                 }
3005
3006                 bkey_reservation_init(&reservation.k_i);
3007                 reservation.k.type      = KEY_TYPE_reservation;
3008                 reservation.k.p         = k.k->p;
3009                 reservation.k.size      = k.k->size;
3010
3011                 bch2_cut_front(iter.pos,        &reservation.k_i);
3012                 bch2_cut_back(end_pos,          &reservation.k_i);
3013
3014                 sectors = reservation.k.size;
3015                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
3016
3017                 if (!bkey_extent_is_allocation(k.k)) {
3018                         ret = bch2_quota_reservation_add(c, inode,
3019                                         &quota_res,
3020                                         sectors, true);
3021                         if (unlikely(ret))
3022                                 goto bkey_err;
3023                 }
3024
3025                 if (reservation.v.nr_replicas < replicas ||
3026                     bch2_bkey_sectors_compressed(k)) {
3027                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
3028                                                         replicas, 0);
3029                         if (unlikely(ret))
3030                                 goto bkey_err;
3031
3032                         reservation.v.nr_replicas = disk_res.nr_replicas;
3033                 }
3034
3035                 ret = bch2_extent_update(&trans, inode_inum(inode), &iter,
3036                                          &reservation.k_i,
3037                                 &disk_res, NULL,
3038                                 0, &i_sectors_delta, true);
3039                 if (ret)
3040                         goto bkey_err;
3041                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3042 bkey_err:
3043                 bch2_quota_reservation_put(c, inode, &quota_res);
3044                 bch2_disk_reservation_put(c, &disk_res);
3045                 if (ret == -EINTR)
3046                         ret = 0;
3047         }
3048
3049         bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3050         mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3051
3052         if (ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)) {
3053                 struct quota_res quota_res = { 0 };
3054                 s64 i_sectors_delta = 0;
3055
3056                 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3057                                end_sector, &i_sectors_delta);
3058                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3059                 bch2_quota_reservation_put(c, inode, &quota_res);
3060         }
3061
3062         bch2_trans_iter_exit(&trans, &iter);
3063         bch2_trans_exit(&trans);
3064         return ret;
3065 }
3066
3067 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3068                             loff_t offset, loff_t len)
3069 {
3070         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3071         u64 end         = offset + len;
3072         u64 block_start = round_down(offset,    block_bytes(c));
3073         u64 block_end   = round_up(end,         block_bytes(c));
3074         bool truncated_last_page = false;
3075         int ret, ret2 = 0;
3076
3077         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3078                 ret = inode_newsize_ok(&inode->v, end);
3079                 if (ret)
3080                         return ret;
3081         }
3082
3083         if (mode & FALLOC_FL_ZERO_RANGE) {
3084                 ret = bch2_truncate_pages(inode, offset, end);
3085                 if (unlikely(ret < 0))
3086                         return ret;
3087
3088                 truncated_last_page = ret;
3089
3090                 truncate_pagecache_range(&inode->v, offset, end - 1);
3091
3092                 block_start     = round_up(offset,      block_bytes(c));
3093                 block_end       = round_down(end,       block_bytes(c));
3094         }
3095
3096         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3097
3098         /*
3099          * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3100          * so that the VFS cache i_size is consistent with the btree i_size:
3101          */
3102         if (ret &&
3103             !(ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)))
3104                 return ret;
3105
3106         if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3107                 end = inode->v.i_size;
3108
3109         if (end >= inode->v.i_size &&
3110             (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3111              !(mode & FALLOC_FL_KEEP_SIZE))) {
3112                 spin_lock(&inode->v.i_lock);
3113                 i_size_write(&inode->v, end);
3114                 spin_unlock(&inode->v.i_lock);
3115
3116                 mutex_lock(&inode->ei_update_lock);
3117                 ret2 = bch2_write_inode_size(c, inode, end, 0);
3118                 mutex_unlock(&inode->ei_update_lock);
3119         }
3120
3121         return ret ?: ret2;
3122 }
3123
3124 long bch2_fallocate_dispatch(struct file *file, int mode,
3125                              loff_t offset, loff_t len)
3126 {
3127         struct bch_inode_info *inode = file_bch_inode(file);
3128         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3129         long ret;
3130
3131         if (!percpu_ref_tryget(&c->writes))
3132                 return -EROFS;
3133
3134         inode_lock(&inode->v);
3135         inode_dio_wait(&inode->v);
3136         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
3137
3138         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3139                 ret = bchfs_fallocate(inode, mode, offset, len);
3140         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3141                 ret = bchfs_fpunch(inode, offset, len);
3142         else if (mode == FALLOC_FL_INSERT_RANGE)
3143                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3144         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3145                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3146         else
3147                 ret = -EOPNOTSUPP;
3148
3149
3150         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
3151         inode_unlock(&inode->v);
3152         percpu_ref_put(&c->writes);
3153
3154         return ret;
3155 }
3156
3157 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3158                              struct file *file_dst, loff_t pos_dst,
3159                              loff_t len, unsigned remap_flags)
3160 {
3161         struct bch_inode_info *src = file_bch_inode(file_src);
3162         struct bch_inode_info *dst = file_bch_inode(file_dst);
3163         struct bch_fs *c = src->v.i_sb->s_fs_info;
3164         s64 i_sectors_delta = 0;
3165         u64 aligned_len;
3166         loff_t ret = 0;
3167
3168         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3169                 return -EINVAL;
3170
3171         if (remap_flags & REMAP_FILE_DEDUP)
3172                 return -EOPNOTSUPP;
3173
3174         if ((pos_src & (block_bytes(c) - 1)) ||
3175             (pos_dst & (block_bytes(c) - 1)))
3176                 return -EINVAL;
3177
3178         if (src == dst &&
3179             abs(pos_src - pos_dst) < len)
3180                 return -EINVAL;
3181
3182         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3183
3184         file_update_time(file_dst);
3185
3186         inode_dio_wait(&src->v);
3187         inode_dio_wait(&dst->v);
3188
3189         ret = generic_remap_file_range_prep(file_src, pos_src,
3190                                             file_dst, pos_dst,
3191                                             &len, remap_flags);
3192         if (ret < 0 || len == 0)
3193                 goto err;
3194
3195         aligned_len = round_up((u64) len, block_bytes(c));
3196
3197         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3198                                 pos_dst, pos_dst + len - 1);
3199         if (ret)
3200                 goto err;
3201
3202         mark_pagecache_unallocated(src, pos_src >> 9,
3203                                    (pos_src + aligned_len) >> 9);
3204
3205         ret = bch2_remap_range(c,
3206                                inode_inum(dst), pos_dst >> 9,
3207                                inode_inum(src), pos_src >> 9,
3208                                aligned_len >> 9,
3209                                pos_dst + len, &i_sectors_delta);
3210         if (ret < 0)
3211                 goto err;
3212
3213         /*
3214          * due to alignment, we might have remapped slightly more than requsted
3215          */
3216         ret = min((u64) ret << 9, (u64) len);
3217
3218         /* XXX get a quota reservation */
3219         i_sectors_acct(c, dst, NULL, i_sectors_delta);
3220
3221         spin_lock(&dst->v.i_lock);
3222         if (pos_dst + ret > dst->v.i_size)
3223                 i_size_write(&dst->v, pos_dst + ret);
3224         spin_unlock(&dst->v.i_lock);
3225
3226         if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3227             IS_SYNC(file_inode(file_dst)))
3228                 ret = bch2_flush_inode(c, inode_inum(dst));
3229 err:
3230         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3231
3232         return ret;
3233 }
3234
3235 /* fseek: */
3236
3237 static int page_data_offset(struct page *page, unsigned offset)
3238 {
3239         struct bch_page_state *s = bch2_page_state(page);
3240         unsigned i;
3241
3242         if (s)
3243                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3244                         if (s->s[i].state >= SECTOR_DIRTY)
3245                                 return i << 9;
3246
3247         return -1;
3248 }
3249
3250 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3251                                        loff_t start_offset,
3252                                        loff_t end_offset)
3253 {
3254         struct address_space *mapping = vinode->i_mapping;
3255         struct page *page;
3256         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
3257         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
3258         pgoff_t index           = start_index;
3259         loff_t ret;
3260         int offset;
3261
3262         while (index <= end_index) {
3263                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
3264                         lock_page(page);
3265
3266                         offset = page_data_offset(page,
3267                                         page->index == start_index
3268                                         ? start_offset & (PAGE_SIZE - 1)
3269                                         : 0);
3270                         if (offset >= 0) {
3271                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
3272                                             offset,
3273                                             start_offset, end_offset);
3274                                 unlock_page(page);
3275                                 put_page(page);
3276                                 return ret;
3277                         }
3278
3279                         unlock_page(page);
3280                         put_page(page);
3281                 } else {
3282                         break;
3283                 }
3284         }
3285
3286         return end_offset;
3287 }
3288
3289 static loff_t bch2_seek_data(struct file *file, u64 offset)
3290 {
3291         struct bch_inode_info *inode = file_bch_inode(file);
3292         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3293         struct btree_trans trans;
3294         struct btree_iter iter;
3295         struct bkey_s_c k;
3296         subvol_inum inum = inode_inum(inode);
3297         u64 isize, next_data = MAX_LFS_FILESIZE;
3298         u32 snapshot;
3299         int ret;
3300
3301         isize = i_size_read(&inode->v);
3302         if (offset >= isize)
3303                 return -ENXIO;
3304
3305         bch2_trans_init(&trans, c, 0, 0);
3306 retry:
3307         bch2_trans_begin(&trans);
3308
3309         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3310         if (ret)
3311                 goto err;
3312
3313         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3314                            SPOS(inode->v.i_ino, offset >> 9, snapshot), 0, k, ret) {
3315                 if (k.k->p.inode != inode->v.i_ino) {
3316                         break;
3317                 } else if (bkey_extent_is_data(k.k)) {
3318                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3319                         break;
3320                 } else if (k.k->p.offset >> 9 > isize)
3321                         break;
3322         }
3323         bch2_trans_iter_exit(&trans, &iter);
3324 err:
3325         if (ret == -EINTR)
3326                 goto retry;
3327
3328         bch2_trans_exit(&trans);
3329         if (ret)
3330                 return ret;
3331
3332         if (next_data > offset)
3333                 next_data = bch2_seek_pagecache_data(&inode->v,
3334                                                      offset, next_data);
3335
3336         if (next_data >= isize)
3337                 return -ENXIO;
3338
3339         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3340 }
3341
3342 static int __page_hole_offset(struct page *page, unsigned offset)
3343 {
3344         struct bch_page_state *s = bch2_page_state(page);
3345         unsigned i;
3346
3347         if (!s)
3348                 return 0;
3349
3350         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3351                 if (s->s[i].state < SECTOR_DIRTY)
3352                         return i << 9;
3353
3354         return -1;
3355 }
3356
3357 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3358 {
3359         pgoff_t index = offset >> PAGE_SHIFT;
3360         struct page *page;
3361         int pg_offset;
3362         loff_t ret = -1;
3363
3364         page = find_lock_page(mapping, index);
3365         if (!page)
3366                 return offset;
3367
3368         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3369         if (pg_offset >= 0)
3370                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3371
3372         unlock_page(page);
3373
3374         return ret;
3375 }
3376
3377 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3378                                        loff_t start_offset,
3379                                        loff_t end_offset)
3380 {
3381         struct address_space *mapping = vinode->i_mapping;
3382         loff_t offset = start_offset, hole;
3383
3384         while (offset < end_offset) {
3385                 hole = page_hole_offset(mapping, offset);
3386                 if (hole >= 0 && hole <= end_offset)
3387                         return max(start_offset, hole);
3388
3389                 offset += PAGE_SIZE;
3390                 offset &= PAGE_MASK;
3391         }
3392
3393         return end_offset;
3394 }
3395
3396 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3397 {
3398         struct bch_inode_info *inode = file_bch_inode(file);
3399         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3400         struct btree_trans trans;
3401         struct btree_iter iter;
3402         struct bkey_s_c k;
3403         subvol_inum inum = inode_inum(inode);
3404         u64 isize, next_hole = MAX_LFS_FILESIZE;
3405         u32 snapshot;
3406         int ret;
3407
3408         isize = i_size_read(&inode->v);
3409         if (offset >= isize)
3410                 return -ENXIO;
3411
3412         bch2_trans_init(&trans, c, 0, 0);
3413 retry:
3414         bch2_trans_begin(&trans);
3415
3416         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3417         if (ret)
3418                 goto err;
3419
3420         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3421                            SPOS(inode->v.i_ino, offset >> 9, snapshot),
3422                            BTREE_ITER_SLOTS, k, ret) {
3423                 if (k.k->p.inode != inode->v.i_ino) {
3424                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3425                                         offset, MAX_LFS_FILESIZE);
3426                         break;
3427                 } else if (!bkey_extent_is_data(k.k)) {
3428                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3429                                         max(offset, bkey_start_offset(k.k) << 9),
3430                                         k.k->p.offset << 9);
3431
3432                         if (next_hole < k.k->p.offset << 9)
3433                                 break;
3434                 } else {
3435                         offset = max(offset, bkey_start_offset(k.k) << 9);
3436                 }
3437         }
3438         bch2_trans_iter_exit(&trans, &iter);
3439 err:
3440         if (ret == -EINTR)
3441                 goto retry;
3442
3443         bch2_trans_exit(&trans);
3444         if (ret)
3445                 return ret;
3446
3447         if (next_hole > isize)
3448                 next_hole = isize;
3449
3450         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3451 }
3452
3453 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3454 {
3455         switch (whence) {
3456         case SEEK_SET:
3457         case SEEK_CUR:
3458         case SEEK_END:
3459                 return generic_file_llseek(file, offset, whence);
3460         case SEEK_DATA:
3461                 return bch2_seek_data(file, offset);
3462         case SEEK_HOLE:
3463                 return bch2_seek_hole(file, offset);
3464         }
3465
3466         return -EINVAL;
3467 }
3468
3469 void bch2_fs_fsio_exit(struct bch_fs *c)
3470 {
3471         bioset_exit(&c->dio_write_bioset);
3472         bioset_exit(&c->dio_read_bioset);
3473         bioset_exit(&c->writepage_bioset);
3474 }
3475
3476 int bch2_fs_fsio_init(struct bch_fs *c)
3477 {
3478         int ret = 0;
3479
3480         pr_verbose_init(c->opts, "");
3481
3482         if (bioset_init(&c->writepage_bioset,
3483                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3484                         BIOSET_NEED_BVECS) ||
3485             bioset_init(&c->dio_read_bioset,
3486                         4, offsetof(struct dio_read, rbio.bio),
3487                         BIOSET_NEED_BVECS) ||
3488             bioset_init(&c->dio_write_bioset,
3489                         4, offsetof(struct dio_write, op.wbio.bio),
3490                         BIOSET_NEED_BVECS))
3491                 ret = -ENOMEM;
3492
3493         pr_verbose_init(c->opts, "ret %i", ret);
3494         return ret;
3495 }
3496
3497 #endif /* NO_BCACHEFS_FS */