git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/fs-io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 static inline bool bio_full(struct bio *bio, unsigned len)
  39 {
  40         if (bio->bi_vcnt >= bio->bi_max_vecs)
  41                 return true;
  42         if (bio->bi_iter.bi_size > UINT_MAX - len)
  43                 return true;
  44         return false;
  45 }
  46
  47 static inline struct address_space *faults_disabled_mapping(void)
  48 {
  49         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  50 }
  51
  52 static inline void set_fdm_dropped_locks(void)
  53 {
  54         current->faults_disabled_mapping =
  55                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  56 }
  57
  58 static inline bool fdm_dropped_locks(void)
  59 {
  60         return ((unsigned long) current->faults_disabled_mapping) & 1;
  61 }
  62
  63 struct quota_res {
  64         u64                             sectors;
  65 };
  66
  67 struct bch_writepage_io {
  68         struct closure                  cl;
  69         struct bch_inode_info           *inode;
  70
  71         /* must be last: */
  72         struct bch_write_op             op;
  73 };
  74
  75 struct dio_write {
  76         struct completion               done;
  77         struct kiocb                    *req;
  78         struct mm_struct                *mm;
  79         unsigned                        loop:1,
  80                                         sync:1,
  81                                         free_iov:1;
  82         struct quota_res                quota_res;
  83         u64                             written;
  84
  85         struct iov_iter                 iter;
  86         struct iovec                    inline_vecs[2];
  87
  88         /* must be last: */
  89         struct bch_write_op             op;
  90 };
  91
  92 struct dio_read {
  93         struct closure                  cl;
  94         struct kiocb                    *req;
  95         long                            ret;
  96         bool                            should_dirty;
  97         struct bch_read_bio             rbio;
  98 };
  99
 100 /* pagecache_block must be held */
 101 static int write_invalidate_inode_pages_range(struct address_space *mapping,
 102                                               loff_t start, loff_t end)
 103 {
 104         int ret;
 105
 106         /*
 107          * XXX: the way this is currently implemented, we can spin if a process
 108          * is continually redirtying a specific page
 109          */
 110         do {
 111                 if (!mapping->nrpages)
 112                         return 0;
 113
 114                 ret = filemap_write_and_wait_range(mapping, start, end);
 115                 if (ret)
 116                         break;
 117
 118                 if (!mapping->nrpages)
 119                         return 0;
 120
 121                 ret = invalidate_inode_pages2_range(mapping,
 122                                 start >> PAGE_SHIFT,
 123                                 end >> PAGE_SHIFT);
 124         } while (ret == -EBUSY);
 125
 126         return ret;
 127 }
 128
 129 /* quotas */
 130
 131 #ifdef CONFIG_BCACHEFS_QUOTA
 132
 133 static void bch2_quota_reservation_put(struct bch_fs *c,
 134                                        struct bch_inode_info *inode,
 135                                        struct quota_res *res)
 136 {
 137         if (!res->sectors)
 138                 return;
 139
 140         mutex_lock(&inode->ei_quota_lock);
 141         BUG_ON(res->sectors > inode->ei_quota_reserved);
 142
 143         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 144                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 145         inode->ei_quota_reserved -= res->sectors;
 146         mutex_unlock(&inode->ei_quota_lock);
 147
 148         res->sectors = 0;
 149 }
 150
 151 static int bch2_quota_reservation_add(struct bch_fs *c,
 152                                       struct bch_inode_info *inode,
 153                                       struct quota_res *res,
 154                                       unsigned sectors,
 155                                       bool check_enospc)
 156 {
 157         int ret;
 158
 159         mutex_lock(&inode->ei_quota_lock);
 160         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 161                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 162         if (likely(!ret)) {
 163                 inode->ei_quota_reserved += sectors;
 164                 res->sectors += sectors;
 165         }
 166         mutex_unlock(&inode->ei_quota_lock);
 167
 168         return ret;
 169 }
 170
 171 #else
 172
 173 static void bch2_quota_reservation_put(struct bch_fs *c,
 174                                        struct bch_inode_info *inode,
 175                                        struct quota_res *res)
 176 {
 177 }
 178
 179 static int bch2_quota_reservation_add(struct bch_fs *c,
 180                                       struct bch_inode_info *inode,
 181                                       struct quota_res *res,
 182                                       unsigned sectors,
 183                                       bool check_enospc)
 184 {
 185         return 0;
 186 }
 187
 188 #endif
 189
 190 /* i_size updates: */
 191
 192 struct inode_new_size {
 193         loff_t          new_size;
 194         u64             now;
 195         unsigned        fields;
 196 };
 197
 198 static int inode_set_size(struct bch_inode_info *inode,
 199                           struct bch_inode_unpacked *bi,
 200                           void *p)
 201 {
 202         struct inode_new_size *s = p;
 203
 204         bi->bi_size = s->new_size;
 205         if (s->fields & ATTR_ATIME)
 206                 bi->bi_atime = s->now;
 207         if (s->fields & ATTR_MTIME)
 208                 bi->bi_mtime = s->now;
 209         if (s->fields & ATTR_CTIME)
 210                 bi->bi_ctime = s->now;
 211
 212         return 0;
 213 }
 214
 215 int __must_check bch2_write_inode_size(struct bch_fs *c,
 216                                        struct bch_inode_info *inode,
 217                                        loff_t new_size, unsigned fields)
 218 {
 219         struct inode_new_size s = {
 220                 .new_size       = new_size,
 221                 .now            = bch2_current_time(c),
 222                 .fields         = fields,
 223         };
 224
 225         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 226 }
 227
 228 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 229                            struct quota_res *quota_res, s64 sectors)
 230 {
 231         if (!sectors)
 232                 return;
 233
 234         mutex_lock(&inode->ei_quota_lock);
 235         bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
 236                                 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
 237                                 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
 238                                 inode->ei_inode.bi_sectors);
 239         inode->v.i_blocks += sectors;
 240
 241 #ifdef CONFIG_BCACHEFS_QUOTA
 242         if (quota_res && sectors > 0) {
 243                 BUG_ON(sectors > quota_res->sectors);
 244                 BUG_ON(sectors > inode->ei_quota_reserved);
 245
 246                 quota_res->sectors -= sectors;
 247                 inode->ei_quota_reserved -= sectors;
 248         } else {
 249                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 250         }
 251 #endif
 252         mutex_unlock(&inode->ei_quota_lock);
 253 }
 254
 255 /* page state: */
 256
 257 /* stored in page->private: */
 258
 259 struct bch_page_sector {
 260         /* Uncompressed, fully allocated replicas (or on disk reservation): */
 261         unsigned                nr_replicas:4;
 262
 263         /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
 264         unsigned                replicas_reserved:4;
 265
 266         /* i_sectors: */
 267         enum {
 268                 SECTOR_UNALLOCATED,
 269                 SECTOR_RESERVED,
 270                 SECTOR_DIRTY,
 271                 SECTOR_DIRTY_RESERVED,
 272                 SECTOR_ALLOCATED,
 273         }                       state:8;
 274 };
 275
 276 struct bch_page_state {
 277         spinlock_t              lock;
 278         atomic_t                write_count;
 279         bool                    uptodate;
 280         struct bch_page_sector  s[PAGE_SECTORS];
 281 };
 282
 283 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 284 {
 285         return page_has_private(page)
 286                 ? (struct bch_page_state *) page_private(page)
 287                 : NULL;
 288 }
 289
 290 static inline struct bch_page_state *bch2_page_state(struct page *page)
 291 {
 292         EBUG_ON(!PageLocked(page));
 293
 294         return __bch2_page_state(page);
 295 }
 296
 297 /* for newly allocated pages: */
 298 static void __bch2_page_state_release(struct page *page)
 299 {
 300         kfree(detach_page_private(page));
 301 }
 302
 303 static void bch2_page_state_release(struct page *page)
 304 {
 305         EBUG_ON(!PageLocked(page));
 306         __bch2_page_state_release(page);
 307 }
 308
 309 /* for newly allocated pages: */
 310 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 311                                                        gfp_t gfp)
 312 {
 313         struct bch_page_state *s;
 314
 315         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 316         if (!s)
 317                 return NULL;
 318
 319         spin_lock_init(&s->lock);
 320         attach_page_private(page, s);
 321         return s;
 322 }
 323
 324 static struct bch_page_state *bch2_page_state_create(struct page *page,
 325                                                      gfp_t gfp)
 326 {
 327         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 328 }
 329
 330 static unsigned bkey_to_sector_state(const struct bkey *k)
 331 {
 332         if (k->type == KEY_TYPE_reservation)
 333                 return SECTOR_RESERVED;
 334         if (bkey_extent_is_allocation(k))
 335                 return SECTOR_ALLOCATED;
 336         return SECTOR_UNALLOCATED;
 337 }
 338
 339 static void __bch2_page_state_set(struct page *page,
 340                                   unsigned pg_offset, unsigned pg_len,
 341                                   unsigned nr_ptrs, unsigned state)
 342 {
 343         struct bch_page_state *s = bch2_page_state_create(page, __GFP_NOFAIL);
 344         unsigned i;
 345
 346         BUG_ON(pg_offset >= PAGE_SECTORS);
 347         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 348
 349         spin_lock(&s->lock);
 350
 351         for (i = pg_offset; i < pg_offset + pg_len; i++) {
 352                 s->s[i].nr_replicas = nr_ptrs;
 353                 s->s[i].state = state;
 354         }
 355
 356         if (i == PAGE_SECTORS)
 357                 s->uptodate = true;
 358
 359         spin_unlock(&s->lock);
 360 }
 361
 362 static int bch2_page_state_set(struct bch_fs *c, subvol_inum inum,
 363                                struct page **pages, unsigned nr_pages)
 364 {
 365         struct btree_trans trans;
 366         struct btree_iter iter;
 367         struct bkey_s_c k;
 368         u64 offset = pages[0]->index << PAGE_SECTORS_SHIFT;
 369         unsigned pg_idx = 0;
 370         u32 snapshot;
 371         int ret;
 372
 373         bch2_trans_init(&trans, c, 0, 0);
 374 retry:
 375         bch2_trans_begin(&trans);
 376
 377         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 378         if (ret)
 379                 goto err;
 380
 381         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
 382                            SPOS(inum.inum, offset, snapshot),
 383                            BTREE_ITER_SLOTS, k, ret) {
 384                 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
 385                 unsigned state = bkey_to_sector_state(k.k);
 386
 387                 while (pg_idx < nr_pages) {
 388                         struct page *page = pages[pg_idx];
 389                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 390                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 391                         unsigned pg_offset = max(bkey_start_offset(k.k), pg_start) - pg_start;
 392                         unsigned pg_len = min(k.k->p.offset, pg_end) - pg_offset - pg_start;
 393
 394                         BUG_ON(k.k->p.offset < pg_start);
 395                         BUG_ON(bkey_start_offset(k.k) > pg_end);
 396
 397                         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate)
 398                                 __bch2_page_state_set(page, pg_offset, pg_len, nr_ptrs, state);
 399
 400                         if (k.k->p.offset < pg_end)
 401                                 break;
 402                         pg_idx++;
 403                 }
 404
 405                 if (pg_idx == nr_pages)
 406                         break;
 407         }
 408
 409         offset = iter.pos.offset;
 410         bch2_trans_iter_exit(&trans, &iter);
 411 err:
 412         if (ret == -EINTR)
 413                 goto retry;
 414         bch2_trans_exit(&trans);
 415
 416         return ret;
 417 }
 418
 419 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
 420 {
 421         struct bvec_iter iter;
 422         struct bio_vec bv;
 423         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 424                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 425         unsigned state = bkey_to_sector_state(k.k);
 426
 427         bio_for_each_segment(bv, bio, iter)
 428                 __bch2_page_state_set(bv.bv_page, bv.bv_offset >> 9,
 429                                       bv.bv_len >> 9, nr_ptrs, state);
 430 }
 431
 432 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
 433                                        u64 start, u64 end)
 434 {
 435         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 436         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 437         struct pagevec pvec;
 438
 439         if (end <= start)
 440                 return;
 441
 442         pagevec_init(&pvec);
 443
 444         do {
 445                 unsigned nr_pages, i, j;
 446
 447                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
 448                                                 &index, end_index);
 449                 for (i = 0; i < nr_pages; i++) {
 450                         struct page *page = pvec.pages[i];
 451                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 452                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 453                         unsigned pg_offset = max(start, pg_start) - pg_start;
 454                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 455                         struct bch_page_state *s;
 456
 457                         BUG_ON(end <= pg_start);
 458                         BUG_ON(pg_offset >= PAGE_SECTORS);
 459                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 460
 461                         lock_page(page);
 462                         s = bch2_page_state(page);
 463
 464                         if (s) {
 465                                 spin_lock(&s->lock);
 466                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 467                                         s->s[j].nr_replicas = 0;
 468                                 spin_unlock(&s->lock);
 469                         }
 470
 471                         unlock_page(page);
 472                 }
 473                 pagevec_release(&pvec);
 474         } while (index <= end_index);
 475 }
 476
 477 static void mark_pagecache_reserved(struct bch_inode_info *inode,
 478                                     u64 start, u64 end)
 479 {
 480         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 481         pgoff_t index = start >> PAGE_SECTORS_SHIFT;
 482         pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
 483         struct pagevec pvec;
 484         s64 i_sectors_delta = 0;
 485
 486         if (end <= start)
 487                 return;
 488
 489         pagevec_init(&pvec);
 490
 491         do {
 492                 unsigned nr_pages, i, j;
 493
 494                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
 495                                                 &index, end_index);
 496                 for (i = 0; i < nr_pages; i++) {
 497                         struct page *page = pvec.pages[i];
 498                         u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
 499                         u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
 500                         unsigned pg_offset = max(start, pg_start) - pg_start;
 501                         unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
 502                         struct bch_page_state *s;
 503
 504                         BUG_ON(end <= pg_start);
 505                         BUG_ON(pg_offset >= PAGE_SECTORS);
 506                         BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
 507
 508                         lock_page(page);
 509                         s = bch2_page_state(page);
 510
 511                         if (s) {
 512                                 spin_lock(&s->lock);
 513                                 for (j = pg_offset; j < pg_offset + pg_len; j++)
 514                                         switch (s->s[j].state) {
 515                                         case SECTOR_UNALLOCATED:
 516                                                 s->s[j].state = SECTOR_RESERVED;
 517                                                 break;
 518                                         case SECTOR_DIRTY:
 519                                                 s->s[j].state = SECTOR_DIRTY_RESERVED;
 520                                                 i_sectors_delta--;
 521                                                 break;
 522                                         default:
 523                                                 break;
 524                                         }
 525                                 spin_unlock(&s->lock);
 526                         }
 527
 528                         unlock_page(page);
 529                 }
 530                 pagevec_release(&pvec);
 531         } while (index <= end_index);
 532
 533         i_sectors_acct(c, inode, NULL, i_sectors_delta);
 534 }
 535
 536 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 537 {
 538         /* XXX: this should not be open coded */
 539         return inode->ei_inode.bi_data_replicas
 540                 ? inode->ei_inode.bi_data_replicas - 1
 541                 : c->opts.data_replicas;
 542 }
 543
 544 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 545                                                   unsigned nr_replicas)
 546 {
 547         return max(0, (int) nr_replicas -
 548                    s->nr_replicas -
 549                    s->replicas_reserved);
 550 }
 551
 552 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 553                                 struct bch_inode_info *inode,
 554                                 struct page *page, bool check_enospc)
 555 {
 556         struct bch_page_state *s = bch2_page_state_create(page, 0);
 557         unsigned nr_replicas = inode_nr_replicas(c, inode);
 558         struct disk_reservation disk_res = { 0 };
 559         unsigned i, disk_res_sectors = 0;
 560         int ret;
 561
 562         if (!s)
 563                 return -ENOMEM;
 564
 565         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 566                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 567
 568         if (!disk_res_sectors)
 569                 return 0;
 570
 571         ret = bch2_disk_reservation_get(c, &disk_res,
 572                                         disk_res_sectors, 1,
 573                                         !check_enospc
 574                                         ? BCH_DISK_RESERVATION_NOFAIL
 575                                         : 0);
 576         if (unlikely(ret))
 577                 return ret;
 578
 579         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 580                 s->s[i].replicas_reserved +=
 581                         sectors_to_reserve(&s->s[i], nr_replicas);
 582
 583         return 0;
 584 }
 585
 586 struct bch2_page_reservation {
 587         struct disk_reservation disk;
 588         struct quota_res        quota;
 589 };
 590
 591 static void bch2_page_reservation_init(struct bch_fs *c,
 592                         struct bch_inode_info *inode,
 593                         struct bch2_page_reservation *res)
 594 {
 595         memset(res, 0, sizeof(*res));
 596
 597         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 598 }
 599
 600 static void bch2_page_reservation_put(struct bch_fs *c,
 601                         struct bch_inode_info *inode,
 602                         struct bch2_page_reservation *res)
 603 {
 604         bch2_disk_reservation_put(c, &res->disk);
 605         bch2_quota_reservation_put(c, inode, &res->quota);
 606 }
 607
 608 static int bch2_page_reservation_get(struct bch_fs *c,
 609                         struct bch_inode_info *inode, struct page *page,
 610                         struct bch2_page_reservation *res,
 611                         unsigned offset, unsigned len, bool check_enospc)
 612 {
 613         struct bch_page_state *s = bch2_page_state_create(page, 0);
 614         unsigned i, disk_sectors = 0, quota_sectors = 0;
 615         int ret;
 616
 617         if (!s)
 618                 return -ENOMEM;
 619
 620         BUG_ON(!s->uptodate);
 621
 622         for (i = round_down(offset, block_bytes(c)) >> 9;
 623              i < round_up(offset + len, block_bytes(c)) >> 9;
 624              i++) {
 625                 disk_sectors += sectors_to_reserve(&s->s[i],
 626                                                 res->disk.nr_replicas);
 627                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 628         }
 629
 630         if (disk_sectors) {
 631                 ret = bch2_disk_reservation_add(c, &res->disk,
 632                                                 disk_sectors,
 633                                                 !check_enospc
 634                                                 ? BCH_DISK_RESERVATION_NOFAIL
 635                                                 : 0);
 636                 if (unlikely(ret))
 637                         return ret;
 638         }
 639
 640         if (quota_sectors) {
 641                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 642                                                  quota_sectors,
 643                                                  check_enospc);
 644                 if (unlikely(ret)) {
 645                         struct disk_reservation tmp = {
 646                                 .sectors = disk_sectors
 647                         };
 648
 649                         bch2_disk_reservation_put(c, &tmp);
 650                         res->disk.sectors -= disk_sectors;
 651                         return ret;
 652                 }
 653         }
 654
 655         return 0;
 656 }
 657
 658 static void bch2_clear_page_bits(struct page *page)
 659 {
 660         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 661         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 662         struct bch_page_state *s = bch2_page_state(page);
 663         struct disk_reservation disk_res = { 0 };
 664         int i, dirty_sectors = 0;
 665
 666         if (!s)
 667                 return;
 668
 669         EBUG_ON(!PageLocked(page));
 670         EBUG_ON(PageWriteback(page));
 671
 672         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 673                 disk_res.sectors += s->s[i].replicas_reserved;
 674                 s->s[i].replicas_reserved = 0;
 675
 676                 switch (s->s[i].state) {
 677                 case SECTOR_DIRTY:
 678                         s->s[i].state = SECTOR_UNALLOCATED;
 679                         --dirty_sectors;
 680                         break;
 681                 case SECTOR_DIRTY_RESERVED:
 682                         s->s[i].state = SECTOR_RESERVED;
 683                         break;
 684                 default:
 685                         break;
 686                 }
 687         }
 688
 689         bch2_disk_reservation_put(c, &disk_res);
 690
 691         i_sectors_acct(c, inode, NULL, dirty_sectors);
 692
 693         bch2_page_state_release(page);
 694 }
 695
 696 static void bch2_set_page_dirty(struct bch_fs *c,
 697                         struct bch_inode_info *inode, struct page *page,
 698                         struct bch2_page_reservation *res,
 699                         unsigned offset, unsigned len)
 700 {
 701         struct bch_page_state *s = bch2_page_state(page);
 702         unsigned i, dirty_sectors = 0;
 703
 704         WARN_ON((u64) page_offset(page) + offset + len >
 705                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 706
 707         spin_lock(&s->lock);
 708
 709         for (i = round_down(offset, block_bytes(c)) >> 9;
 710              i < round_up(offset + len, block_bytes(c)) >> 9;
 711              i++) {
 712                 unsigned sectors = sectors_to_reserve(&s->s[i],
 713                                                 res->disk.nr_replicas);
 714
 715                 /*
 716                  * This can happen if we race with the error path in
 717                  * bch2_writepage_io_done():
 718                  */
 719                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 720
 721                 s->s[i].replicas_reserved += sectors;
 722                 res->disk.sectors -= sectors;
 723
 724                 switch (s->s[i].state) {
 725                 case SECTOR_UNALLOCATED:
 726                         s->s[i].state = SECTOR_DIRTY;
 727                         dirty_sectors++;
 728                         break;
 729                 case SECTOR_RESERVED:
 730                         s->s[i].state = SECTOR_DIRTY_RESERVED;
 731                         break;
 732                 default:
 733                         break;
 734                 }
 735         }
 736
 737         spin_unlock(&s->lock);
 738
 739         i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 740
 741         if (!PageDirty(page))
 742                 __set_page_dirty_nobuffers(page);
 743 }
 744
 745 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 746 {
 747         struct file *file = vmf->vma->vm_file;
 748         struct address_space *mapping = file->f_mapping;
 749         struct address_space *fdm = faults_disabled_mapping();
 750         struct bch_inode_info *inode = file_bch_inode(file);
 751         int ret;
 752
 753         if (fdm == mapping)
 754                 return VM_FAULT_SIGBUS;
 755
 756         /* Lock ordering: */
 757         if (fdm > mapping) {
 758                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 759
 760                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 761                         goto got_lock;
 762
 763                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 764
 765                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 766                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 767
 768                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 769
 770                 /* Signal that lock has been dropped: */
 771                 set_fdm_dropped_locks();
 772                 return VM_FAULT_SIGBUS;
 773         }
 774
 775         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 776 got_lock:
 777         ret = filemap_fault(vmf);
 778         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 779
 780         return ret;
 781 }
 782
 783 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 784 {
 785         struct page *page = vmf->page;
 786         struct file *file = vmf->vma->vm_file;
 787         struct bch_inode_info *inode = file_bch_inode(file);
 788         struct address_space *mapping = file->f_mapping;
 789         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 790         struct bch2_page_reservation res;
 791         unsigned len;
 792         loff_t isize;
 793         int ret;
 794
 795         bch2_page_reservation_init(c, inode, &res);
 796
 797         sb_start_pagefault(inode->v.i_sb);
 798         file_update_time(file);
 799
 800         /*
 801          * Not strictly necessary, but helps avoid dio writes livelocking in
 802          * write_invalidate_inode_pages_range() - can drop this if/when we get
 803          * a write_invalidate_inode_pages_range() that works without dropping
 804          * page lock before invalidating page
 805          */
 806         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 807
 808         lock_page(page);
 809         isize = i_size_read(&inode->v);
 810
 811         if (page->mapping != mapping || page_offset(page) >= isize) {
 812                 unlock_page(page);
 813                 ret = VM_FAULT_NOPAGE;
 814                 goto out;
 815         }
 816
 817         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 818
 819         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
 820                 if (bch2_page_state_set(c, inode_inum(inode), &page, 1)) {
 821                         unlock_page(page);
 822                         ret = VM_FAULT_SIGBUS;
 823                         goto out;
 824                 }
 825         }
 826
 827         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 828                 unlock_page(page);
 829                 ret = VM_FAULT_SIGBUS;
 830                 goto out;
 831         }
 832
 833         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 834         bch2_page_reservation_put(c, inode, &res);
 835
 836         wait_for_stable_page(page);
 837         ret = VM_FAULT_LOCKED;
 838 out:
 839         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 840         sb_end_pagefault(inode->v.i_sb);
 841
 842         return ret;
 843 }
 844
 845 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
 846 {
 847         if (offset || length < folio_size(folio))
 848                 return;
 849
 850         bch2_clear_page_bits(&folio->page);
 851 }
 852
 853 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 854 {
 855         if (PageDirty(page))
 856                 return 0;
 857
 858         bch2_clear_page_bits(page);
 859         return 1;
 860 }
 861
 862 #ifdef CONFIG_MIGRATION
 863 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 864                       struct page *page, enum migrate_mode mode)
 865 {
 866         int ret;
 867
 868         EBUG_ON(!PageLocked(page));
 869         EBUG_ON(!PageLocked(newpage));
 870
 871         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 872         if (ret != MIGRATEPAGE_SUCCESS)
 873                 return ret;
 874
 875         if (PagePrivate(page))
 876                 attach_page_private(newpage, detach_page_private(page));
 877
 878         if (mode != MIGRATE_SYNC_NO_COPY)
 879                 migrate_page_copy(newpage, page);
 880         else
 881                 migrate_page_states(newpage, page);
 882         return MIGRATEPAGE_SUCCESS;
 883 }
 884 #endif
 885
 886 /* readpage(s): */
 887
 888 static void bch2_readpages_end_io(struct bio *bio)
 889 {
 890         struct bvec_iter_all iter;
 891         struct bio_vec *bv;
 892
 893         bio_for_each_segment_all(bv, bio, iter) {
 894                 struct page *page = bv->bv_page;
 895
 896                 if (!bio->bi_status) {
 897                         SetPageUptodate(page);
 898                 } else {
 899                         ClearPageUptodate(page);
 900                         SetPageError(page);
 901                 }
 902                 unlock_page(page);
 903         }
 904
 905         bio_put(bio);
 906 }
 907
 908 struct readpages_iter {
 909         struct address_space    *mapping;
 910         struct page             **pages;
 911         unsigned                nr_pages;
 912         unsigned                idx;
 913         pgoff_t                 offset;
 914 };
 915
 916 static int readpages_iter_init(struct readpages_iter *iter,
 917                                struct readahead_control *ractl)
 918 {
 919         unsigned i, nr_pages = readahead_count(ractl);
 920
 921         memset(iter, 0, sizeof(*iter));
 922
 923         iter->mapping   = ractl->mapping;
 924         iter->offset    = readahead_index(ractl);
 925         iter->nr_pages  = nr_pages;
 926
 927         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 928         if (!iter->pages)
 929                 return -ENOMEM;
 930
 931         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 932         for (i = 0; i < nr_pages; i++) {
 933                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 934                 put_page(iter->pages[i]);
 935         }
 936
 937         return 0;
 938 }
 939
 940 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 941 {
 942         if (iter->idx >= iter->nr_pages)
 943                 return NULL;
 944
 945         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 946
 947         return iter->pages[iter->idx];
 948 }
 949
 950 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 951 {
 952         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 953         struct bch_extent_crc_unpacked crc;
 954         const union bch_extent_entry *i;
 955
 956         bkey_for_each_crc(k.k, ptrs, crc, i)
 957                 if (crc.csum_type || crc.compression_type)
 958                         return true;
 959         return false;
 960 }
 961
 962 static void readpage_bio_extend(struct readpages_iter *iter,
 963                                 struct bio *bio,
 964                                 unsigned sectors_this_extent,
 965                                 bool get_more)
 966 {
 967         while (bio_sectors(bio) < sectors_this_extent &&
 968                bio->bi_vcnt < bio->bi_max_vecs) {
 969                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
 970                 struct page *page = readpage_iter_next(iter);
 971                 int ret;
 972
 973                 if (page) {
 974                         if (iter->offset + iter->idx != page_offset)
 975                                 break;
 976
 977                         iter->idx++;
 978                 } else {
 979                         if (!get_more)
 980                                 break;
 981
 982                         page = xa_load(&iter->mapping->i_pages, page_offset);
 983                         if (page && !xa_is_value(page))
 984                                 break;
 985
 986                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 987                         if (!page)
 988                                 break;
 989
 990                         if (!__bch2_page_state_create(page, 0)) {
 991                                 put_page(page);
 992                                 break;
 993                         }
 994
 995                         ret = add_to_page_cache_lru(page, iter->mapping,
 996                                                     page_offset, GFP_NOFS);
 997                         if (ret) {
 998                                 __bch2_page_state_release(page);
 999                                 put_page(page);
1000                                 break;
1001                         }
1002
1003                         put_page(page);
1004                 }
1005
1006                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
1007         }
1008 }
1009
1010 static void bchfs_read(struct btree_trans *trans,
1011                        struct bch_read_bio *rbio,
1012                        subvol_inum inum,
1013                        struct readpages_iter *readpages_iter)
1014 {
1015         struct bch_fs *c = trans->c;
1016         struct btree_iter iter;
1017         struct bkey_buf sk;
1018         int flags = BCH_READ_RETRY_IF_STALE|
1019                 BCH_READ_MAY_PROMOTE;
1020         u32 snapshot;
1021         int ret = 0;
1022
1023         rbio->c = c;
1024         rbio->start_time = local_clock();
1025         rbio->subvol = inum.subvol;
1026
1027         bch2_bkey_buf_init(&sk);
1028 retry:
1029         bch2_trans_begin(trans);
1030         iter = (struct btree_iter) { NULL };
1031
1032         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1033         if (ret)
1034                 goto err;
1035
1036         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1037                              SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1038                              BTREE_ITER_SLOTS);
1039         while (1) {
1040                 struct bkey_s_c k;
1041                 unsigned bytes, sectors, offset_into_extent;
1042                 enum btree_id data_btree = BTREE_ID_extents;
1043
1044                 /*
1045                  * read_extent -> io_time_reset may cause a transaction restart
1046                  * without returning an error, we need to check for that here:
1047                  */
1048                 if (!bch2_trans_relock(trans)) {
1049                         ret = -EINTR;
1050                         break;
1051                 }
1052
1053                 bch2_btree_iter_set_pos(&iter,
1054                                 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1055
1056                 k = bch2_btree_iter_peek_slot(&iter);
1057                 ret = bkey_err(k);
1058                 if (ret)
1059                         break;
1060
1061                 offset_into_extent = iter.pos.offset -
1062                         bkey_start_offset(k.k);
1063                 sectors = k.k->size - offset_into_extent;
1064
1065                 bch2_bkey_buf_reassemble(&sk, c, k);
1066
1067                 ret = bch2_read_indirect_extent(trans, &data_btree,
1068                                         &offset_into_extent, &sk);
1069                 if (ret)
1070                         break;
1071
1072                 k = bkey_i_to_s_c(sk.k);
1073
1074                 sectors = min(sectors, k.k->size - offset_into_extent);
1075
1076                 if (readpages_iter)
1077                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
1078                                             extent_partial_reads_expensive(k));
1079
1080                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1081                 swap(rbio->bio.bi_iter.bi_size, bytes);
1082
1083                 if (rbio->bio.bi_iter.bi_size == bytes)
1084                         flags |= BCH_READ_LAST_FRAGMENT;
1085
1086                 bch2_bio_page_state_set(&rbio->bio, k);
1087
1088                 bch2_read_extent(trans, rbio, iter.pos,
1089                                  data_btree, k, offset_into_extent, flags);
1090
1091                 if (flags & BCH_READ_LAST_FRAGMENT)
1092                         break;
1093
1094                 swap(rbio->bio.bi_iter.bi_size, bytes);
1095                 bio_advance(&rbio->bio, bytes);
1096
1097                 ret = btree_trans_too_many_iters(trans);
1098                 if (ret)
1099                         break;
1100         }
1101 err:
1102         bch2_trans_iter_exit(trans, &iter);
1103
1104         if (ret == -EINTR)
1105                 goto retry;
1106
1107         if (ret) {
1108                 bch_err_inum_ratelimited(c, inum.inum,
1109                                 "read error %i from btree lookup", ret);
1110                 rbio->bio.bi_status = BLK_STS_IOERR;
1111                 bio_endio(&rbio->bio);
1112         }
1113
1114         bch2_bkey_buf_exit(&sk, c);
1115 }
1116
1117 void bch2_readahead(struct readahead_control *ractl)
1118 {
1119         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1120         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1121         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1122         struct btree_trans trans;
1123         struct page *page;
1124         struct readpages_iter readpages_iter;
1125         int ret;
1126
1127         ret = readpages_iter_init(&readpages_iter, ractl);
1128         BUG_ON(ret);
1129
1130         bch2_trans_init(&trans, c, 0, 0);
1131
1132         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1133
1134         while ((page = readpage_iter_next(&readpages_iter))) {
1135                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1136                 unsigned n = min_t(unsigned,
1137                                    readpages_iter.nr_pages -
1138                                    readpages_iter.idx,
1139                                    BIO_MAX_VECS);
1140                 struct bch_read_bio *rbio =
1141                         rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
1142                                                    GFP_NOFS, &c->bio_read),
1143                                   opts);
1144
1145                 readpages_iter.idx++;
1146
1147                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTORS_SHIFT;
1148                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1149                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1150
1151                 bchfs_read(&trans, rbio, inode_inum(inode),
1152                            &readpages_iter);
1153         }
1154
1155         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1156
1157         bch2_trans_exit(&trans);
1158         kfree(readpages_iter.pages);
1159 }
1160
1161 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1162                              subvol_inum inum, struct page *page)
1163 {
1164         struct btree_trans trans;
1165
1166         bch2_page_state_create(page, __GFP_NOFAIL);
1167
1168         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1169         rbio->bio.bi_iter.bi_sector =
1170                 (sector_t) page->index << PAGE_SECTORS_SHIFT;
1171         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1172
1173         bch2_trans_init(&trans, c, 0, 0);
1174         bchfs_read(&trans, rbio, inum, NULL);
1175         bch2_trans_exit(&trans);
1176 }
1177
1178 int bch2_readpage(struct file *file, struct page *page)
1179 {
1180         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1181         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1182         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1183         struct bch_read_bio *rbio;
1184
1185         rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS, &c->bio_read), opts);
1186         rbio->bio.bi_end_io = bch2_readpages_end_io;
1187
1188         __bchfs_readpage(c, rbio, inode_inum(inode), page);
1189         return 0;
1190 }
1191
1192 static void bch2_read_single_page_end_io(struct bio *bio)
1193 {
1194         complete(bio->bi_private);
1195 }
1196
1197 static int bch2_read_single_page(struct page *page,
1198                                  struct address_space *mapping)
1199 {
1200         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1201         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1202         struct bch_read_bio *rbio;
1203         int ret;
1204         DECLARE_COMPLETION_ONSTACK(done);
1205
1206         rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS, &c->bio_read),
1207                          io_opts(c, &inode->ei_inode));
1208         rbio->bio.bi_private = &done;
1209         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1210
1211         __bchfs_readpage(c, rbio, inode_inum(inode), page);
1212         wait_for_completion(&done);
1213
1214         ret = blk_status_to_errno(rbio->bio.bi_status);
1215         bio_put(&rbio->bio);
1216
1217         if (ret < 0)
1218                 return ret;
1219
1220         SetPageUptodate(page);
1221         return 0;
1222 }
1223
1224 /* writepages: */
1225
1226 struct bch_writepage_state {
1227         struct bch_writepage_io *io;
1228         struct bch_io_opts      opts;
1229 };
1230
1231 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1232                                                                   struct bch_inode_info *inode)
1233 {
1234         return (struct bch_writepage_state) {
1235                 .opts = io_opts(c, &inode->ei_inode)
1236         };
1237 }
1238
1239 static void bch2_writepage_io_free(struct closure *cl)
1240 {
1241         struct bch_writepage_io *io = container_of(cl,
1242                                         struct bch_writepage_io, cl);
1243
1244         bio_put(&io->op.wbio.bio);
1245 }
1246
1247 static void bch2_writepage_io_done(struct closure *cl)
1248 {
1249         struct bch_writepage_io *io = container_of(cl,
1250                                         struct bch_writepage_io, cl);
1251         struct bch_fs *c = io->op.c;
1252         struct bio *bio = &io->op.wbio.bio;
1253         struct bvec_iter_all iter;
1254         struct bio_vec *bvec;
1255         unsigned i;
1256
1257         up(&io->op.c->io_in_flight);
1258
1259         if (io->op.error) {
1260                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1261
1262                 bio_for_each_segment_all(bvec, bio, iter) {
1263                         struct bch_page_state *s;
1264
1265                         SetPageError(bvec->bv_page);
1266                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1267
1268                         s = __bch2_page_state(bvec->bv_page);
1269                         spin_lock(&s->lock);
1270                         for (i = 0; i < PAGE_SECTORS; i++)
1271                                 s->s[i].nr_replicas = 0;
1272                         spin_unlock(&s->lock);
1273                 }
1274         }
1275
1276         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1277                 bio_for_each_segment_all(bvec, bio, iter) {
1278                         struct bch_page_state *s;
1279
1280                         s = __bch2_page_state(bvec->bv_page);
1281                         spin_lock(&s->lock);
1282                         for (i = 0; i < PAGE_SECTORS; i++)
1283                                 s->s[i].nr_replicas = 0;
1284                         spin_unlock(&s->lock);
1285                 }
1286         }
1287
1288         /*
1289          * racing with fallocate can cause us to add fewer sectors than
1290          * expected - but we shouldn't add more sectors than expected:
1291          */
1292         WARN_ON_ONCE(io->op.i_sectors_delta > 0);
1293
1294         /*
1295          * (error (due to going RO) halfway through a page can screw that up
1296          * slightly)
1297          * XXX wtf?
1298            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1299          */
1300
1301         /*
1302          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1303          * before calling end_page_writeback:
1304          */
1305         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1306
1307         bio_for_each_segment_all(bvec, bio, iter) {
1308                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1309
1310                 if (atomic_dec_and_test(&s->write_count))
1311                         end_page_writeback(bvec->bv_page);
1312         }
1313
1314         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1315 }
1316
1317 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1318 {
1319         struct bch_writepage_io *io = w->io;
1320
1321         down(&io->op.c->io_in_flight);
1322
1323         w->io = NULL;
1324         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1325         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1326 }
1327
1328 /*
1329  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1330  * possible, else allocating a new one:
1331  */
1332 static void bch2_writepage_io_alloc(struct bch_fs *c,
1333                                     struct writeback_control *wbc,
1334                                     struct bch_writepage_state *w,
1335                                     struct bch_inode_info *inode,
1336                                     u64 sector,
1337                                     unsigned nr_replicas)
1338 {
1339         struct bch_write_op *op;
1340
1341         w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1342                                               REQ_OP_WRITE,
1343                                               GFP_NOFS,
1344                                               &c->writepage_bioset),
1345                              struct bch_writepage_io, op.wbio.bio);
1346
1347         closure_init(&w->io->cl, NULL);
1348         w->io->inode            = inode;
1349
1350         op                      = &w->io->op;
1351         bch2_write_op_init(op, c, w->opts);
1352         op->target              = w->opts.foreground_target;
1353         op->nr_replicas         = nr_replicas;
1354         op->res.nr_replicas     = nr_replicas;
1355         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1356         op->subvol              = inode->ei_subvol;
1357         op->pos                 = POS(inode->v.i_ino, sector);
1358         op->wbio.bio.bi_iter.bi_sector = sector;
1359         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1360 }
1361
1362 static int __bch2_writepage(struct page *page,
1363                             struct writeback_control *wbc,
1364                             void *data)
1365 {
1366         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1367         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1368         struct bch_writepage_state *w = data;
1369         struct bch_page_state *s, orig;
1370         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1371         loff_t i_size = i_size_read(&inode->v);
1372         pgoff_t end_index = i_size >> PAGE_SHIFT;
1373         int ret;
1374
1375         EBUG_ON(!PageUptodate(page));
1376
1377         /* Is the page fully inside i_size? */
1378         if (page->index < end_index)
1379                 goto do_io;
1380
1381         /* Is the page fully outside i_size? (truncate in progress) */
1382         offset = i_size & (PAGE_SIZE - 1);
1383         if (page->index > end_index || !offset) {
1384                 unlock_page(page);
1385                 return 0;
1386         }
1387
1388         /*
1389          * The page straddles i_size.  It must be zeroed out on each and every
1390          * writepage invocation because it may be mmapped.  "A file is mapped
1391          * in multiples of the page size.  For a file that is not a multiple of
1392          * the  page size, the remaining memory is zeroed when mapped, and
1393          * writes to that region are not written out to the file."
1394          */
1395         zero_user_segment(page, offset, PAGE_SIZE);
1396 do_io:
1397         s = bch2_page_state_create(page, __GFP_NOFAIL);
1398
1399         /*
1400          * Things get really hairy with errors during writeback:
1401          */
1402         ret = bch2_get_page_disk_reservation(c, inode, page, false);
1403         BUG_ON(ret);
1404
1405         /* Before unlocking the page, get copy of reservations: */
1406         spin_lock(&s->lock);
1407         orig = *s;
1408         spin_unlock(&s->lock);
1409
1410         for (i = 0; i < PAGE_SECTORS; i++) {
1411                 if (s->s[i].state < SECTOR_DIRTY)
1412                         continue;
1413
1414                 nr_replicas_this_write =
1415                         min_t(unsigned, nr_replicas_this_write,
1416                               s->s[i].nr_replicas +
1417                               s->s[i].replicas_reserved);
1418         }
1419
1420         for (i = 0; i < PAGE_SECTORS; i++) {
1421                 if (s->s[i].state < SECTOR_DIRTY)
1422                         continue;
1423
1424                 s->s[i].nr_replicas = w->opts.compression
1425                         ? 0 : nr_replicas_this_write;
1426
1427                 s->s[i].replicas_reserved = 0;
1428                 s->s[i].state = SECTOR_ALLOCATED;
1429         }
1430
1431         BUG_ON(atomic_read(&s->write_count));
1432         atomic_set(&s->write_count, 1);
1433
1434         BUG_ON(PageWriteback(page));
1435         set_page_writeback(page);
1436
1437         unlock_page(page);
1438
1439         offset = 0;
1440         while (1) {
1441                 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1442                 u64 sector;
1443
1444                 while (offset < PAGE_SECTORS &&
1445                        orig.s[offset].state < SECTOR_DIRTY)
1446                         offset++;
1447
1448                 if (offset == PAGE_SECTORS)
1449                         break;
1450
1451                 while (offset + sectors < PAGE_SECTORS &&
1452                        orig.s[offset + sectors].state >= SECTOR_DIRTY) {
1453                         reserved_sectors += orig.s[offset + sectors].replicas_reserved;
1454                         dirty_sectors += orig.s[offset + sectors].state == SECTOR_DIRTY;
1455                         sectors++;
1456                 }
1457                 BUG_ON(!sectors);
1458
1459                 sector = ((u64) page->index << PAGE_SECTORS_SHIFT) + offset;
1460
1461                 if (w->io &&
1462                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1463                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1464                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1465                      (BIO_MAX_VECS * PAGE_SIZE) ||
1466                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1467                         bch2_writepage_do_io(w);
1468
1469                 if (!w->io)
1470                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1471                                                 nr_replicas_this_write);
1472
1473                 atomic_inc(&s->write_count);
1474
1475                 BUG_ON(inode != w->io->inode);
1476                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1477                                      sectors << 9, offset << 9));
1478
1479                 /* Check for writing past i_size: */
1480                 WARN_ON_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1481                              round_up(i_size, block_bytes(c)));
1482
1483                 w->io->op.res.sectors += reserved_sectors;
1484                 w->io->op.i_sectors_delta -= dirty_sectors;
1485                 w->io->op.new_i_size = i_size;
1486
1487                 offset += sectors;
1488         }
1489
1490         if (atomic_dec_and_test(&s->write_count))
1491                 end_page_writeback(page);
1492
1493         return 0;
1494 }
1495
1496 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1497 {
1498         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1499         struct bch_writepage_state w =
1500                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1501         struct blk_plug plug;
1502         int ret;
1503
1504         blk_start_plug(&plug);
1505         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1506         if (w.io)
1507                 bch2_writepage_do_io(&w);
1508         blk_finish_plug(&plug);
1509         return ret;
1510 }
1511
1512 /* buffered writes: */
1513
1514 int bch2_write_begin(struct file *file, struct address_space *mapping,
1515                      loff_t pos, unsigned len, unsigned flags,
1516                      struct page **pagep, void **fsdata)
1517 {
1518         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1519         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1520         struct bch2_page_reservation *res;
1521         pgoff_t index = pos >> PAGE_SHIFT;
1522         unsigned offset = pos & (PAGE_SIZE - 1);
1523         struct page *page;
1524         int ret = -ENOMEM;
1525
1526         res = kmalloc(sizeof(*res), GFP_KERNEL);
1527         if (!res)
1528                 return -ENOMEM;
1529
1530         bch2_page_reservation_init(c, inode, res);
1531         *fsdata = res;
1532
1533         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1534
1535         page = grab_cache_page_write_begin(mapping, index, flags);
1536         if (!page)
1537                 goto err_unlock;
1538
1539         if (PageUptodate(page))
1540                 goto out;
1541
1542         /* If we're writing entire page, don't need to read it in first: */
1543         if (len == PAGE_SIZE)
1544                 goto out;
1545
1546         if (!offset && pos + len >= inode->v.i_size) {
1547                 zero_user_segment(page, len, PAGE_SIZE);
1548                 flush_dcache_page(page);
1549                 goto out;
1550         }
1551
1552         if (index > inode->v.i_size >> PAGE_SHIFT) {
1553                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1554                 flush_dcache_page(page);
1555                 goto out;
1556         }
1557 readpage:
1558         ret = bch2_read_single_page(page, mapping);
1559         if (ret)
1560                 goto err;
1561 out:
1562         if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1563                 ret = bch2_page_state_set(c, inode_inum(inode), &page, 1);
1564                 if (ret)
1565                         goto out;
1566         }
1567
1568         ret = bch2_page_reservation_get(c, inode, page, res,
1569                                         offset, len, true);
1570         if (ret) {
1571                 if (!PageUptodate(page)) {
1572                         /*
1573                          * If the page hasn't been read in, we won't know if we
1574                          * actually need a reservation - we don't actually need
1575                          * to read here, we just need to check if the page is
1576                          * fully backed by uncompressed data:
1577                          */
1578                         goto readpage;
1579                 }
1580
1581                 goto err;
1582         }
1583
1584         *pagep = page;
1585         return 0;
1586 err:
1587         unlock_page(page);
1588         put_page(page);
1589         *pagep = NULL;
1590 err_unlock:
1591         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1592         kfree(res);
1593         *fsdata = NULL;
1594         return ret;
1595 }
1596
1597 int bch2_write_end(struct file *file, struct address_space *mapping,
1598                    loff_t pos, unsigned len, unsigned copied,
1599                    struct page *page, void *fsdata)
1600 {
1601         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1602         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1603         struct bch2_page_reservation *res = fsdata;
1604         unsigned offset = pos & (PAGE_SIZE - 1);
1605
1606         lockdep_assert_held(&inode->v.i_rwsem);
1607
1608         if (unlikely(copied < len && !PageUptodate(page))) {
1609                 /*
1610                  * The page needs to be read in, but that would destroy
1611                  * our partial write - simplest thing is to just force
1612                  * userspace to redo the write:
1613                  */
1614                 zero_user(page, 0, PAGE_SIZE);
1615                 flush_dcache_page(page);
1616                 copied = 0;
1617         }
1618
1619         spin_lock(&inode->v.i_lock);
1620         if (pos + copied > inode->v.i_size)
1621                 i_size_write(&inode->v, pos + copied);
1622         spin_unlock(&inode->v.i_lock);
1623
1624         if (copied) {
1625                 if (!PageUptodate(page))
1626                         SetPageUptodate(page);
1627
1628                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1629
1630                 inode->ei_last_dirtied = (unsigned long) current;
1631         }
1632
1633         unlock_page(page);
1634         put_page(page);
1635         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1636
1637         bch2_page_reservation_put(c, inode, res);
1638         kfree(res);
1639
1640         return copied;
1641 }
1642
1643 #define WRITE_BATCH_PAGES       32
1644
1645 static int __bch2_buffered_write(struct bch_inode_info *inode,
1646                                  struct address_space *mapping,
1647                                  struct iov_iter *iter,
1648                                  loff_t pos, unsigned len)
1649 {
1650         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1651         struct page *pages[WRITE_BATCH_PAGES];
1652         struct bch2_page_reservation res;
1653         unsigned long index = pos >> PAGE_SHIFT;
1654         unsigned offset = pos & (PAGE_SIZE - 1);
1655         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1656         unsigned i, reserved = 0, set_dirty = 0;
1657         unsigned copied = 0, nr_pages_copied = 0;
1658         int ret = 0;
1659
1660         BUG_ON(!len);
1661         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1662
1663         bch2_page_reservation_init(c, inode, &res);
1664
1665         for (i = 0; i < nr_pages; i++) {
1666                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1667                 if (!pages[i]) {
1668                         nr_pages = i;
1669                         if (!i) {
1670                                 ret = -ENOMEM;
1671                                 goto out;
1672                         }
1673                         len = min_t(unsigned, len,
1674                                     nr_pages * PAGE_SIZE - offset);
1675                         break;
1676                 }
1677         }
1678
1679         if (offset && !PageUptodate(pages[0])) {
1680                 ret = bch2_read_single_page(pages[0], mapping);
1681                 if (ret)
1682                         goto out;
1683         }
1684
1685         if ((pos + len) & (PAGE_SIZE - 1) &&
1686             !PageUptodate(pages[nr_pages - 1])) {
1687                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1688                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1689                 } else {
1690                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1691                         if (ret)
1692                                 goto out;
1693                 }
1694         }
1695
1696         while (reserved < len) {
1697                 unsigned i = (offset + reserved) >> PAGE_SHIFT;
1698                 struct page *page = pages[i];
1699                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1700                 unsigned pg_len = min_t(unsigned, len - reserved,
1701                                         PAGE_SIZE - pg_offset);
1702
1703                 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1704                         ret = bch2_page_state_set(c, inode_inum(inode),
1705                                                   pages + i, nr_pages - i);
1706                         if (ret)
1707                                 goto out;
1708                 }
1709
1710                 ret = bch2_page_reservation_get(c, inode, page, &res,
1711                                                 pg_offset, pg_len, true);
1712                 if (ret)
1713                         goto out;
1714
1715                 reserved += pg_len;
1716         }
1717
1718         if (mapping_writably_mapped(mapping))
1719                 for (i = 0; i < nr_pages; i++)
1720                         flush_dcache_page(pages[i]);
1721
1722         while (copied < len) {
1723                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1724                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1725                 unsigned pg_len = min_t(unsigned, len - copied,
1726                                         PAGE_SIZE - pg_offset);
1727                 unsigned pg_copied = copy_page_from_iter_atomic(page,
1728                                                 pg_offset, pg_len,iter);
1729
1730                 if (!pg_copied)
1731                         break;
1732
1733                 if (!PageUptodate(page) &&
1734                     pg_copied != PAGE_SIZE &&
1735                     pos + copied + pg_copied < inode->v.i_size) {
1736                         zero_user(page, 0, PAGE_SIZE);
1737                         break;
1738                 }
1739
1740                 flush_dcache_page(page);
1741                 copied += pg_copied;
1742
1743                 if (pg_copied != pg_len)
1744                         break;
1745         }
1746
1747         if (!copied)
1748                 goto out;
1749
1750         spin_lock(&inode->v.i_lock);
1751         if (pos + copied > inode->v.i_size)
1752                 i_size_write(&inode->v, pos + copied);
1753         spin_unlock(&inode->v.i_lock);
1754
1755         while (set_dirty < copied) {
1756                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1757                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1758                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1759                                         PAGE_SIZE - pg_offset);
1760
1761                 if (!PageUptodate(page))
1762                         SetPageUptodate(page);
1763
1764                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1765                 unlock_page(page);
1766                 put_page(page);
1767
1768                 set_dirty += pg_len;
1769         }
1770
1771         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1772         inode->ei_last_dirtied = (unsigned long) current;
1773 out:
1774         for (i = nr_pages_copied; i < nr_pages; i++) {
1775                 unlock_page(pages[i]);
1776                 put_page(pages[i]);
1777         }
1778
1779         bch2_page_reservation_put(c, inode, &res);
1780
1781         return copied ?: ret;
1782 }
1783
1784 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1785 {
1786         struct file *file = iocb->ki_filp;
1787         struct address_space *mapping = file->f_mapping;
1788         struct bch_inode_info *inode = file_bch_inode(file);
1789         loff_t pos = iocb->ki_pos;
1790         ssize_t written = 0;
1791         int ret = 0;
1792
1793         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1794
1795         do {
1796                 unsigned offset = pos & (PAGE_SIZE - 1);
1797                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1798                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1799 again:
1800                 /*
1801                  * Bring in the user page that we will copy from _first_.
1802                  * Otherwise there's a nasty deadlock on copying from the
1803                  * same page as we're writing to, without it being marked
1804                  * up-to-date.
1805                  *
1806                  * Not only is this an optimisation, but it is also required
1807                  * to check that the address is actually valid, when atomic
1808                  * usercopies are used, below.
1809                  */
1810                 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1811                         bytes = min_t(unsigned long, iov_iter_count(iter),
1812                                       PAGE_SIZE - offset);
1813
1814                         if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1815                                 ret = -EFAULT;
1816                                 break;
1817                         }
1818                 }
1819
1820                 if (unlikely(fatal_signal_pending(current))) {
1821                         ret = -EINTR;
1822                         break;
1823                 }
1824
1825                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1826                 if (unlikely(ret < 0))
1827                         break;
1828
1829                 cond_resched();
1830
1831                 if (unlikely(ret == 0)) {
1832                         /*
1833                          * If we were unable to copy any data at all, we must
1834                          * fall back to a single segment length write.
1835                          *
1836                          * If we didn't fallback here, we could livelock
1837                          * because not all segments in the iov can be copied at
1838                          * once without a pagefault.
1839                          */
1840                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1841                                       iov_iter_single_seg_count(iter));
1842                         goto again;
1843                 }
1844                 pos += ret;
1845                 written += ret;
1846                 ret = 0;
1847
1848                 balance_dirty_pages_ratelimited(mapping);
1849         } while (iov_iter_count(iter));
1850
1851         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1852
1853         return written ? written : ret;
1854 }
1855
1856 /* O_DIRECT reads */
1857
1858 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1859 {
1860         if (check_dirty) {
1861                 bio_check_pages_dirty(bio);
1862         } else {
1863                 bio_release_pages(bio, false);
1864                 bio_put(bio);
1865         }
1866 }
1867
1868 static void bch2_dio_read_complete(struct closure *cl)
1869 {
1870         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1871
1872         dio->req->ki_complete(dio->req, dio->ret);
1873         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1874 }
1875
1876 static void bch2_direct_IO_read_endio(struct bio *bio)
1877 {
1878         struct dio_read *dio = bio->bi_private;
1879
1880         if (bio->bi_status)
1881                 dio->ret = blk_status_to_errno(bio->bi_status);
1882
1883         closure_put(&dio->cl);
1884 }
1885
1886 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1887 {
1888         struct dio_read *dio = bio->bi_private;
1889         bool should_dirty = dio->should_dirty;
1890
1891         bch2_direct_IO_read_endio(bio);
1892         bio_check_or_release(bio, should_dirty);
1893 }
1894
1895 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1896 {
1897         struct file *file = req->ki_filp;
1898         struct bch_inode_info *inode = file_bch_inode(file);
1899         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1900         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1901         struct dio_read *dio;
1902         struct bio *bio;
1903         loff_t offset = req->ki_pos;
1904         bool sync = is_sync_kiocb(req);
1905         size_t shorten;
1906         ssize_t ret;
1907
1908         if ((offset|iter->count) & (block_bytes(c) - 1))
1909                 return -EINVAL;
1910
1911         ret = min_t(loff_t, iter->count,
1912                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1913
1914         if (!ret)
1915                 return ret;
1916
1917         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1918         iter->count -= shorten;
1919
1920         bio = bio_alloc_bioset(NULL,
1921                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1922                                REQ_OP_READ,
1923                                GFP_KERNEL,
1924                                &c->dio_read_bioset);
1925
1926         bio->bi_end_io = bch2_direct_IO_read_endio;
1927
1928         dio = container_of(bio, struct dio_read, rbio.bio);
1929         closure_init(&dio->cl, NULL);
1930
1931         /*
1932          * this is a _really_ horrible hack just to avoid an atomic sub at the
1933          * end:
1934          */
1935         if (!sync) {
1936                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1937                 atomic_set(&dio->cl.remaining,
1938                            CLOSURE_REMAINING_INITIALIZER -
1939                            CLOSURE_RUNNING +
1940                            CLOSURE_DESTRUCTOR);
1941         } else {
1942                 atomic_set(&dio->cl.remaining,
1943                            CLOSURE_REMAINING_INITIALIZER + 1);
1944         }
1945
1946         dio->req        = req;
1947         dio->ret        = ret;
1948         /*
1949          * This is one of the sketchier things I've encountered: we have to skip
1950          * the dirtying of requests that are internal from the kernel (i.e. from
1951          * loopback), because we'll deadlock on page_lock.
1952          */
1953         dio->should_dirty = iter_is_iovec(iter);
1954
1955         goto start;
1956         while (iter->count) {
1957                 bio = bio_alloc_bioset(NULL,
1958                                        bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1959                                        REQ_OP_READ,
1960                                        GFP_KERNEL,
1961                                        &c->bio_read);
1962                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1963 start:
1964                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1965                 bio->bi_iter.bi_sector  = offset >> 9;
1966                 bio->bi_private         = dio;
1967
1968                 ret = bio_iov_iter_get_pages(bio, iter);
1969                 if (ret < 0) {
1970                         /* XXX: fault inject this path */
1971                         bio->bi_status = BLK_STS_RESOURCE;
1972                         bio_endio(bio);
1973                         break;
1974                 }
1975
1976                 offset += bio->bi_iter.bi_size;
1977
1978                 if (dio->should_dirty)
1979                         bio_set_pages_dirty(bio);
1980
1981                 if (iter->count)
1982                         closure_get(&dio->cl);
1983
1984                 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
1985         }
1986
1987         iter->count += shorten;
1988
1989         if (sync) {
1990                 closure_sync(&dio->cl);
1991                 closure_debug_destroy(&dio->cl);
1992                 ret = dio->ret;
1993                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1994                 return ret;
1995         } else {
1996                 return -EIOCBQUEUED;
1997         }
1998 }
1999
2000 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2001 {
2002         struct file *file = iocb->ki_filp;
2003         struct bch_inode_info *inode = file_bch_inode(file);
2004         struct address_space *mapping = file->f_mapping;
2005         size_t count = iov_iter_count(iter);
2006         ssize_t ret;
2007
2008         if (!count)
2009                 return 0; /* skip atime */
2010
2011         if (iocb->ki_flags & IOCB_DIRECT) {
2012                 struct blk_plug plug;
2013
2014                 ret = filemap_write_and_wait_range(mapping,
2015                                         iocb->ki_pos,
2016                                         iocb->ki_pos + count - 1);
2017                 if (ret < 0)
2018                         return ret;
2019
2020                 file_accessed(file);
2021
2022                 blk_start_plug(&plug);
2023                 ret = bch2_direct_IO_read(iocb, iter);
2024                 blk_finish_plug(&plug);
2025
2026                 if (ret >= 0)
2027                         iocb->ki_pos += ret;
2028         } else {
2029                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
2030                 ret = generic_file_read_iter(iocb, iter);
2031                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
2032         }
2033
2034         return ret;
2035 }
2036
2037 /* O_DIRECT writes */
2038
2039 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2040                                        u64 offset, u64 size,
2041                                        unsigned nr_replicas, bool compressed)
2042 {
2043         struct btree_trans trans;
2044         struct btree_iter iter;
2045         struct bkey_s_c k;
2046         u64 end = offset + size;
2047         u32 snapshot;
2048         bool ret = true;
2049         int err;
2050
2051         bch2_trans_init(&trans, c, 0, 0);
2052 retry:
2053         bch2_trans_begin(&trans);
2054
2055         err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2056         if (err)
2057                 goto err;
2058
2059         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2060                            SPOS(inum.inum, offset, snapshot),
2061                            BTREE_ITER_SLOTS, k, err) {
2062                 if (bkey_cmp(bkey_start_pos(k.k), POS(inum.inum, end)) >= 0)
2063                         break;
2064
2065                 if (k.k->p.snapshot != snapshot ||
2066                     nr_replicas > bch2_bkey_replicas(c, k) ||
2067                     (!compressed && bch2_bkey_sectors_compressed(k))) {
2068                         ret = false;
2069                         break;
2070                 }
2071         }
2072
2073         offset = iter.pos.offset;
2074         bch2_trans_iter_exit(&trans, &iter);
2075 err:
2076         if (err == -EINTR)
2077                 goto retry;
2078         bch2_trans_exit(&trans);
2079
2080         return err ? false : ret;
2081 }
2082
2083 static void bch2_dio_write_loop_async(struct bch_write_op *);
2084
2085 static long bch2_dio_write_loop(struct dio_write *dio)
2086 {
2087         bool kthread = (current->flags & PF_KTHREAD) != 0;
2088         struct kiocb *req = dio->req;
2089         struct address_space *mapping = req->ki_filp->f_mapping;
2090         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
2091         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2092         struct bio *bio = &dio->op.wbio.bio;
2093         struct bvec_iter_all iter;
2094         struct bio_vec *bv;
2095         unsigned unaligned, iter_count;
2096         bool sync = dio->sync, dropped_locks;
2097         long ret;
2098
2099         if (dio->loop)
2100                 goto loop;
2101
2102         down(&c->io_in_flight);
2103
2104         while (1) {
2105                 iter_count = dio->iter.count;
2106
2107                 if (kthread && dio->mm)
2108                         kthread_use_mm(dio->mm);
2109                 BUG_ON(current->faults_disabled_mapping);
2110                 current->faults_disabled_mapping = mapping;
2111
2112                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2113
2114                 dropped_locks = fdm_dropped_locks();
2115
2116                 current->faults_disabled_mapping = NULL;
2117                 if (kthread && dio->mm)
2118                         kthread_unuse_mm(dio->mm);
2119
2120                 /*
2121                  * If the fault handler returned an error but also signalled
2122                  * that it dropped & retook ei_pagecache_lock, we just need to
2123                  * re-shoot down the page cache and retry:
2124                  */
2125                 if (dropped_locks && ret)
2126                         ret = 0;
2127
2128                 if (unlikely(ret < 0))
2129                         goto err;
2130
2131                 if (unlikely(dropped_locks)) {
2132                         ret = write_invalidate_inode_pages_range(mapping,
2133                                         req->ki_pos,
2134                                         req->ki_pos + iter_count - 1);
2135                         if (unlikely(ret))
2136                                 goto err;
2137
2138                         if (!bio->bi_iter.bi_size)
2139                                 continue;
2140                 }
2141
2142                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2143                 bio->bi_iter.bi_size -= unaligned;
2144                 iov_iter_revert(&dio->iter, unaligned);
2145
2146                 if (!bio->bi_iter.bi_size) {
2147                         /*
2148                          * bio_iov_iter_get_pages was only able to get <
2149                          * blocksize worth of pages:
2150                          */
2151                         ret = -EFAULT;
2152                         goto err;
2153                 }
2154
2155                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
2156                 dio->op.end_io          = bch2_dio_write_loop_async;
2157                 dio->op.target          = dio->op.opts.foreground_target;
2158                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
2159                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
2160                 dio->op.subvol          = inode->ei_subvol;
2161                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2162
2163                 if ((req->ki_flags & IOCB_DSYNC) &&
2164                     !c->opts.journal_flush_disabled)
2165                         dio->op.flags |= BCH_WRITE_FLUSH;
2166                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2167
2168                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2169                                                 dio->op.opts.data_replicas, 0);
2170                 if (unlikely(ret) &&
2171                     !bch2_check_range_allocated(c, inode_inum(inode),
2172                                 dio->op.pos.offset, bio_sectors(bio),
2173                                 dio->op.opts.data_replicas,
2174                                 dio->op.opts.compression != 0))
2175                         goto err;
2176
2177                 task_io_account_write(bio->bi_iter.bi_size);
2178
2179                 if (!dio->sync && !dio->loop && dio->iter.count) {
2180                         struct iovec *iov = dio->inline_vecs;
2181
2182                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2183                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
2184                                               GFP_KERNEL);
2185                                 if (unlikely(!iov)) {
2186                                         dio->sync = sync = true;
2187                                         goto do_io;
2188                                 }
2189
2190                                 dio->free_iov = true;
2191                         }
2192
2193                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2194                         dio->iter.iov = iov;
2195                 }
2196 do_io:
2197                 dio->loop = true;
2198                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2199
2200                 if (sync)
2201                         wait_for_completion(&dio->done);
2202                 else
2203                         return -EIOCBQUEUED;
2204 loop:
2205                 i_sectors_acct(c, inode, &dio->quota_res,
2206                                dio->op.i_sectors_delta);
2207                 req->ki_pos += (u64) dio->op.written << 9;
2208                 dio->written += dio->op.written;
2209
2210                 spin_lock(&inode->v.i_lock);
2211                 if (req->ki_pos > inode->v.i_size)
2212                         i_size_write(&inode->v, req->ki_pos);
2213                 spin_unlock(&inode->v.i_lock);
2214
2215                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2216                         bio_for_each_segment_all(bv, bio, iter)
2217                                 put_page(bv->bv_page);
2218                 bio->bi_vcnt = 0;
2219
2220                 if (dio->op.error) {
2221                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
2222                         break;
2223                 }
2224
2225                 if (!dio->iter.count)
2226                         break;
2227
2228                 bio_reset(bio, NULL, REQ_OP_WRITE);
2229                 reinit_completion(&dio->done);
2230         }
2231
2232         ret = dio->op.error ?: ((long) dio->written << 9);
2233 err:
2234         up(&c->io_in_flight);
2235         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2236         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2237
2238         if (dio->free_iov)
2239                 kfree(dio->iter.iov);
2240
2241         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2242                 bio_for_each_segment_all(bv, bio, iter)
2243                         put_page(bv->bv_page);
2244         bio_put(bio);
2245
2246         /* inode->i_dio_count is our ref on inode and thus bch_fs */
2247         inode_dio_end(&inode->v);
2248
2249         if (!sync) {
2250                 req->ki_complete(req, ret);
2251                 ret = -EIOCBQUEUED;
2252         }
2253         return ret;
2254 }
2255
2256 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2257 {
2258         struct dio_write *dio = container_of(op, struct dio_write, op);
2259
2260         if (dio->sync)
2261                 complete(&dio->done);
2262         else
2263                 bch2_dio_write_loop(dio);
2264 }
2265
2266 static noinline
2267 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2268 {
2269         struct file *file = req->ki_filp;
2270         struct address_space *mapping = file->f_mapping;
2271         struct bch_inode_info *inode = file_bch_inode(file);
2272         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2273         struct dio_write *dio;
2274         struct bio *bio;
2275         bool locked = true, extending;
2276         ssize_t ret;
2277
2278         prefetch(&c->opts);
2279         prefetch((void *) &c->opts + 64);
2280         prefetch(&inode->ei_inode);
2281         prefetch((void *) &inode->ei_inode + 64);
2282
2283         inode_lock(&inode->v);
2284
2285         ret = generic_write_checks(req, iter);
2286         if (unlikely(ret <= 0))
2287                 goto err;
2288
2289         ret = file_remove_privs(file);
2290         if (unlikely(ret))
2291                 goto err;
2292
2293         ret = file_update_time(file);
2294         if (unlikely(ret))
2295                 goto err;
2296
2297         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2298                 goto err;
2299
2300         inode_dio_begin(&inode->v);
2301         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2302
2303         extending = req->ki_pos + iter->count > inode->v.i_size;
2304         if (!extending) {
2305                 inode_unlock(&inode->v);
2306                 locked = false;
2307         }
2308
2309         bio = bio_alloc_bioset(NULL,
2310                                bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2311                                REQ_OP_WRITE,
2312                                GFP_KERNEL,
2313                                &c->dio_write_bioset);
2314         dio = container_of(bio, struct dio_write, op.wbio.bio);
2315         init_completion(&dio->done);
2316         dio->req                = req;
2317         dio->mm                 = current->mm;
2318         dio->loop               = false;
2319         dio->sync               = is_sync_kiocb(req) || extending;
2320         dio->free_iov           = false;
2321         dio->quota_res.sectors  = 0;
2322         dio->written            = 0;
2323         dio->iter               = *iter;
2324
2325         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2326                                          iter->count >> 9, true);
2327         if (unlikely(ret))
2328                 goto err_put_bio;
2329
2330         ret = write_invalidate_inode_pages_range(mapping,
2331                                         req->ki_pos,
2332                                         req->ki_pos + iter->count - 1);
2333         if (unlikely(ret))
2334                 goto err_put_bio;
2335
2336         ret = bch2_dio_write_loop(dio);
2337 err:
2338         if (locked)
2339                 inode_unlock(&inode->v);
2340         return ret;
2341 err_put_bio:
2342         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2343         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2344         bio_put(bio);
2345         inode_dio_end(&inode->v);
2346         goto err;
2347 }
2348
2349 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2350 {
2351         struct file *file = iocb->ki_filp;
2352         struct bch_inode_info *inode = file_bch_inode(file);
2353         ssize_t ret;
2354
2355         if (iocb->ki_flags & IOCB_DIRECT)
2356                 return bch2_direct_write(iocb, from);
2357
2358         /* We can write back this queue in page reclaim */
2359         current->backing_dev_info = inode_to_bdi(&inode->v);
2360         inode_lock(&inode->v);
2361
2362         ret = generic_write_checks(iocb, from);
2363         if (ret <= 0)
2364                 goto unlock;
2365
2366         ret = file_remove_privs(file);
2367         if (ret)
2368                 goto unlock;
2369
2370         ret = file_update_time(file);
2371         if (ret)
2372                 goto unlock;
2373
2374         ret = bch2_buffered_write(iocb, from);
2375         if (likely(ret > 0))
2376                 iocb->ki_pos += ret;
2377 unlock:
2378         inode_unlock(&inode->v);
2379         current->backing_dev_info = NULL;
2380
2381         if (ret > 0)
2382                 ret = generic_write_sync(iocb, ret);
2383
2384         return ret;
2385 }
2386
2387 /* fsync: */
2388
2389 /*
2390  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2391  * insert trigger: look up the btree inode instead
2392  */
2393 static int bch2_flush_inode(struct bch_fs *c, subvol_inum inum)
2394 {
2395         struct bch_inode_unpacked inode;
2396         int ret;
2397
2398         if (c->opts.journal_flush_disabled)
2399                 return 0;
2400
2401         ret = bch2_inode_find_by_inum(c, inum, &inode);
2402         if (ret)
2403                 return ret;
2404
2405         return bch2_journal_flush_seq(&c->journal, inode.bi_journal_seq);
2406 }
2407
2408 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2409 {
2410         struct bch_inode_info *inode = file_bch_inode(file);
2411         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2412         int ret, ret2, ret3;
2413
2414         ret = file_write_and_wait_range(file, start, end);
2415         ret2 = sync_inode_metadata(&inode->v, 1);
2416         ret3 = bch2_flush_inode(c, inode_inum(inode));
2417
2418         return ret ?: ret2 ?: ret3;
2419 }
2420
2421 /* truncate: */
2422
2423 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2424                                  struct bpos start,
2425                                  struct bpos end)
2426 {
2427         struct btree_trans trans;
2428         struct btree_iter iter;
2429         struct bkey_s_c k;
2430         int ret = 0;
2431
2432         bch2_trans_init(&trans, c, 0, 0);
2433 retry:
2434         bch2_trans_begin(&trans);
2435
2436         ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2437         if (ret)
2438                 goto err;
2439
2440         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2441                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2442                         break;
2443
2444                 if (bkey_extent_is_data(k.k)) {
2445                         ret = 1;
2446                         break;
2447                 }
2448         }
2449         start = iter.pos;
2450         bch2_trans_iter_exit(&trans, &iter);
2451 err:
2452         if (ret == -EINTR)
2453                 goto retry;
2454
2455         bch2_trans_exit(&trans);
2456         return ret;
2457 }
2458
2459 static int __bch2_truncate_page(struct bch_inode_info *inode,
2460                                 pgoff_t index, loff_t start, loff_t end)
2461 {
2462         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2463         struct address_space *mapping = inode->v.i_mapping;
2464         struct bch_page_state *s;
2465         unsigned start_offset = start & (PAGE_SIZE - 1);
2466         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2467         unsigned i;
2468         struct page *page;
2469         s64 i_sectors_delta = 0;
2470         int ret = 0;
2471
2472         /* Page boundary? Nothing to do */
2473         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2474               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2475                 return 0;
2476
2477         /* Above i_size? */
2478         if (index << PAGE_SHIFT >= inode->v.i_size)
2479                 return 0;
2480
2481         page = find_lock_page(mapping, index);
2482         if (!page) {
2483                 /*
2484                  * XXX: we're doing two index lookups when we end up reading the
2485                  * page
2486                  */
2487                 ret = range_has_data(c, inode->ei_subvol,
2488                                 POS(inode->v.i_ino, index << PAGE_SECTORS_SHIFT),
2489                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTORS_SHIFT));
2490                 if (ret <= 0)
2491                         return ret;
2492
2493                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2494                 if (unlikely(!page)) {
2495                         ret = -ENOMEM;
2496                         goto out;
2497                 }
2498         }
2499
2500         s = bch2_page_state_create(page, 0);
2501         if (!s) {
2502                 ret = -ENOMEM;
2503                 goto unlock;
2504         }
2505
2506         if (!PageUptodate(page)) {
2507                 ret = bch2_read_single_page(page, mapping);
2508                 if (ret)
2509                         goto unlock;
2510         }
2511
2512         if (index != start >> PAGE_SHIFT)
2513                 start_offset = 0;
2514         if (index != end >> PAGE_SHIFT)
2515                 end_offset = PAGE_SIZE;
2516
2517         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2518              i < round_down(end_offset, block_bytes(c)) >> 9;
2519              i++) {
2520                 s->s[i].nr_replicas     = 0;
2521                 if (s->s[i].state == SECTOR_DIRTY)
2522                         i_sectors_delta--;
2523                 s->s[i].state           = SECTOR_UNALLOCATED;
2524         }
2525
2526         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2527
2528         /*
2529          * Caller needs to know whether this page will be written out by
2530          * writeback - doing an i_size update if necessary - or whether it will
2531          * be responsible for the i_size update:
2532          */
2533         ret = s->s[(min_t(u64, inode->v.i_size - (index << PAGE_SHIFT),
2534                           PAGE_SIZE) - 1) >> 9].state >= SECTOR_DIRTY;
2535
2536         zero_user_segment(page, start_offset, end_offset);
2537
2538         /*
2539          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2540          *
2541          * XXX: because we aren't currently tracking whether the page has actual
2542          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2543          */
2544         BUG_ON(bch2_get_page_disk_reservation(c, inode, page, false));
2545
2546         /*
2547          * This removes any writeable userspace mappings; we need to force
2548          * .page_mkwrite to be called again before any mmapped writes, to
2549          * redirty the full page:
2550          */
2551         page_mkclean(page);
2552         __set_page_dirty_nobuffers(page);
2553 unlock:
2554         unlock_page(page);
2555         put_page(page);
2556 out:
2557         return ret;
2558 }
2559
2560 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2561 {
2562         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2563                                     from, round_up(from, PAGE_SIZE));
2564 }
2565
2566 static int bch2_truncate_pages(struct bch_inode_info *inode,
2567                                loff_t start, loff_t end)
2568 {
2569         int ret = __bch2_truncate_page(inode, start >> PAGE_SHIFT,
2570                                        start, end);
2571
2572         if (ret >= 0 &&
2573             start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2574                 ret = __bch2_truncate_page(inode,
2575                                            end >> PAGE_SHIFT,
2576                                            start, end);
2577         return ret;
2578 }
2579
2580 static int bch2_extend(struct user_namespace *mnt_userns,
2581                        struct bch_inode_info *inode,
2582                        struct bch_inode_unpacked *inode_u,
2583                        struct iattr *iattr)
2584 {
2585         struct address_space *mapping = inode->v.i_mapping;
2586         int ret;
2587
2588         /*
2589          * sync appends:
2590          *
2591          * this has to be done _before_ extending i_size:
2592          */
2593         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2594         if (ret)
2595                 return ret;
2596
2597         truncate_setsize(&inode->v, iattr->ia_size);
2598
2599         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2600 }
2601
2602 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2603                                    struct bch_inode_unpacked *bi,
2604                                    void *p)
2605 {
2606         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2607         return 0;
2608 }
2609
2610 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2611                                   struct bch_inode_unpacked *bi, void *p)
2612 {
2613         u64 *new_i_size = p;
2614
2615         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2616         bi->bi_size = *new_i_size;
2617         return 0;
2618 }
2619
2620 int bch2_truncate(struct user_namespace *mnt_userns,
2621                   struct bch_inode_info *inode, struct iattr *iattr)
2622 {
2623         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2624         struct address_space *mapping = inode->v.i_mapping;
2625         struct bch_inode_unpacked inode_u;
2626         u64 new_i_size = iattr->ia_size;
2627         s64 i_sectors_delta = 0;
2628         int ret = 0;
2629
2630         /*
2631          * If the truncate call with change the size of the file, the
2632          * cmtimes should be updated. If the size will not change, we
2633          * do not need to update the cmtimes.
2634          */
2635         if (iattr->ia_size != inode->v.i_size) {
2636                 if (!(iattr->ia_valid & ATTR_MTIME))
2637                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2638                 if (!(iattr->ia_valid & ATTR_CTIME))
2639                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2640                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2641         }
2642
2643         inode_dio_wait(&inode->v);
2644         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2645
2646         ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
2647         if (ret)
2648                 goto err;
2649
2650         /*
2651          * check this before next assertion; on filesystem error our normal
2652          * invariants are a bit broken (truncate has to truncate the page cache
2653          * before the inode).
2654          */
2655         ret = bch2_journal_error(&c->journal);
2656         if (ret)
2657                 goto err;
2658
2659         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2660                 inode->v.i_size < inode_u.bi_size);
2661
2662         if (iattr->ia_size > inode->v.i_size) {
2663                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2664                 goto err;
2665         }
2666
2667         iattr->ia_valid &= ~ATTR_SIZE;
2668
2669         ret = bch2_truncate_page(inode, iattr->ia_size);
2670         if (unlikely(ret < 0))
2671                 goto err;
2672
2673         /*
2674          * When extending, we're going to write the new i_size to disk
2675          * immediately so we need to flush anything above the current on disk
2676          * i_size first:
2677          *
2678          * Also, when extending we need to flush the page that i_size currently
2679          * straddles - if it's mapped to userspace, we need to ensure that
2680          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2681          * again to allocate the part of the page that was extended.
2682          */
2683         if (iattr->ia_size > inode_u.bi_size)
2684                 ret = filemap_write_and_wait_range(mapping,
2685                                 inode_u.bi_size,
2686                                 iattr->ia_size - 1);
2687         else if (iattr->ia_size & (PAGE_SIZE - 1))
2688                 ret = filemap_write_and_wait_range(mapping,
2689                                 round_down(iattr->ia_size, PAGE_SIZE),
2690                                 iattr->ia_size - 1);
2691         if (ret)
2692                 goto err;
2693
2694         mutex_lock(&inode->ei_update_lock);
2695         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2696                                &new_i_size, 0);
2697         mutex_unlock(&inode->ei_update_lock);
2698
2699         if (unlikely(ret))
2700                 goto err;
2701
2702         truncate_setsize(&inode->v, iattr->ia_size);
2703
2704         ret = bch2_fpunch(c, inode_inum(inode),
2705                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2706                         U64_MAX, &i_sectors_delta);
2707         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2708
2709         bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
2710                                 !bch2_journal_error(&c->journal), c,
2711                                 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
2712                                 inode->v.i_ino, (u64) inode->v.i_blocks,
2713                                 inode->ei_inode.bi_sectors);
2714         if (unlikely(ret))
2715                 goto err;
2716
2717         mutex_lock(&inode->ei_update_lock);
2718         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2719         mutex_unlock(&inode->ei_update_lock);
2720
2721         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2722 err:
2723         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2724         return ret;
2725 }
2726
2727 /* fallocate: */
2728
2729 static int inode_update_times_fn(struct bch_inode_info *inode,
2730                                  struct bch_inode_unpacked *bi, void *p)
2731 {
2732         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2733
2734         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2735         return 0;
2736 }
2737
2738 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2739 {
2740         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2741         u64 end         = offset + len;
2742         u64 block_start = round_up(offset, block_bytes(c));
2743         u64 block_end   = round_down(end, block_bytes(c));
2744         bool truncated_last_page;
2745         int ret = 0;
2746
2747         ret = bch2_truncate_pages(inode, offset, end);
2748         if (unlikely(ret < 0))
2749                 goto err;
2750
2751         truncated_last_page = ret;
2752
2753         truncate_pagecache_range(&inode->v, offset, end - 1);
2754
2755         if (block_start < block_end ) {
2756                 s64 i_sectors_delta = 0;
2757
2758                 ret = bch2_fpunch(c, inode_inum(inode),
2759                                   block_start >> 9, block_end >> 9,
2760                                   &i_sectors_delta);
2761                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2762         }
2763
2764         mutex_lock(&inode->ei_update_lock);
2765         if (end >= inode->v.i_size && !truncated_last_page) {
2766                 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2767                                             ATTR_MTIME|ATTR_CTIME);
2768         } else {
2769                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2770                                        ATTR_MTIME|ATTR_CTIME);
2771         }
2772         mutex_unlock(&inode->ei_update_lock);
2773 err:
2774         return ret;
2775 }
2776
2777 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2778                                    loff_t offset, loff_t len,
2779                                    bool insert)
2780 {
2781         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2782         struct address_space *mapping = inode->v.i_mapping;
2783         struct bkey_buf copy;
2784         struct btree_trans trans;
2785         struct btree_iter src, dst, del;
2786         loff_t shift, new_size;
2787         u64 src_start;
2788         int ret = 0;
2789
2790         if ((offset | len) & (block_bytes(c) - 1))
2791                 return -EINVAL;
2792
2793         if (insert) {
2794                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2795                         return -EFBIG;
2796
2797                 if (offset >= inode->v.i_size)
2798                         return -EINVAL;
2799
2800                 src_start       = U64_MAX;
2801                 shift           = len;
2802         } else {
2803                 if (offset + len >= inode->v.i_size)
2804                         return -EINVAL;
2805
2806                 src_start       = offset + len;
2807                 shift           = -len;
2808         }
2809
2810         new_size = inode->v.i_size + shift;
2811
2812         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2813         if (ret)
2814                 return ret;
2815
2816         if (insert) {
2817                 i_size_write(&inode->v, new_size);
2818                 mutex_lock(&inode->ei_update_lock);
2819                 ret = bch2_write_inode_size(c, inode, new_size,
2820                                             ATTR_MTIME|ATTR_CTIME);
2821                 mutex_unlock(&inode->ei_update_lock);
2822         } else {
2823                 s64 i_sectors_delta = 0;
2824
2825                 ret = bch2_fpunch(c, inode_inum(inode),
2826                                   offset >> 9, (offset + len) >> 9,
2827                                   &i_sectors_delta);
2828                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2829
2830                 if (ret)
2831                         return ret;
2832         }
2833
2834         bch2_bkey_buf_init(&copy);
2835         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2836         bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
2837                         POS(inode->v.i_ino, src_start >> 9),
2838                         BTREE_ITER_INTENT);
2839         bch2_trans_copy_iter(&dst, &src);
2840         bch2_trans_copy_iter(&del, &src);
2841
2842         while (ret == 0 || ret == -EINTR) {
2843                 struct disk_reservation disk_res =
2844                         bch2_disk_reservation_init(c, 0);
2845                 struct bkey_i delete;
2846                 struct bkey_s_c k;
2847                 struct bpos next_pos;
2848                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2849                 struct bpos atomic_end;
2850                 unsigned trigger_flags = 0;
2851                 u32 snapshot;
2852
2853                 bch2_trans_begin(&trans);
2854
2855                 ret = bch2_subvolume_get_snapshot(&trans,
2856                                         inode->ei_subvol, &snapshot);
2857                 if (ret)
2858                         continue;
2859
2860                 bch2_btree_iter_set_snapshot(&src, snapshot);
2861                 bch2_btree_iter_set_snapshot(&dst, snapshot);
2862                 bch2_btree_iter_set_snapshot(&del, snapshot);
2863
2864                 bch2_trans_begin(&trans);
2865
2866                 k = insert
2867                         ? bch2_btree_iter_peek_prev(&src)
2868                         : bch2_btree_iter_peek(&src);
2869                 if ((ret = bkey_err(k)))
2870                         continue;
2871
2872                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2873                         break;
2874
2875                 if (insert &&
2876                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2877                         break;
2878 reassemble:
2879                 bch2_bkey_buf_reassemble(&copy, c, k);
2880
2881                 if (insert &&
2882                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2883                         bch2_cut_front(move_pos, copy.k);
2884
2885                 copy.k->k.p.offset += shift >> 9;
2886                 bch2_btree_iter_set_pos(&dst, bkey_start_pos(&copy.k->k));
2887
2888                 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
2889                 if (ret)
2890                         continue;
2891
2892                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2893                         if (insert) {
2894                                 move_pos = atomic_end;
2895                                 move_pos.offset -= shift >> 9;
2896                                 goto reassemble;
2897                         } else {
2898                                 bch2_cut_back(atomic_end, copy.k);
2899                         }
2900                 }
2901
2902                 bkey_init(&delete.k);
2903                 delete.k.p = copy.k->k.p;
2904                 delete.k.size = copy.k->k.size;
2905                 delete.k.p.offset -= shift >> 9;
2906                 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
2907
2908                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2909
2910                 if (copy.k->k.size != k.k->size) {
2911                         /* We might end up splitting compressed extents: */
2912                         unsigned nr_ptrs =
2913                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2914
2915                         ret = bch2_disk_reservation_get(c, &disk_res,
2916                                         copy.k->k.size, nr_ptrs,
2917                                         BCH_DISK_RESERVATION_NOFAIL);
2918                         BUG_ON(ret);
2919                 }
2920
2921                 ret =   bch2_btree_iter_traverse(&del) ?:
2922                         bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
2923                         bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
2924                         bch2_trans_commit(&trans, &disk_res, NULL,
2925                                           BTREE_INSERT_NOFAIL);
2926                 bch2_disk_reservation_put(c, &disk_res);
2927
2928                 if (!ret)
2929                         bch2_btree_iter_set_pos(&src, next_pos);
2930         }
2931         bch2_trans_iter_exit(&trans, &del);
2932         bch2_trans_iter_exit(&trans, &dst);
2933         bch2_trans_iter_exit(&trans, &src);
2934         bch2_trans_exit(&trans);
2935         bch2_bkey_buf_exit(&copy, c);
2936
2937         if (ret)
2938                 return ret;
2939
2940         mutex_lock(&inode->ei_update_lock);
2941         if (!insert) {
2942                 i_size_write(&inode->v, new_size);
2943                 ret = bch2_write_inode_size(c, inode, new_size,
2944                                             ATTR_MTIME|ATTR_CTIME);
2945         } else {
2946                 /* We need an inode update to update bi_journal_seq for fsync: */
2947                 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2948                                        ATTR_MTIME|ATTR_CTIME);
2949         }
2950         mutex_unlock(&inode->ei_update_lock);
2951         return ret;
2952 }
2953
2954 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2955                              u64 start_sector, u64 end_sector)
2956 {
2957         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2958         struct btree_trans trans;
2959         struct btree_iter iter;
2960         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2961         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2962         int ret = 0;
2963
2964         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2965
2966         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2967                         POS(inode->v.i_ino, start_sector),
2968                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2969
2970         while (!ret && bkey_cmp(iter.pos, end_pos) < 0) {
2971                 s64 i_sectors_delta = 0;
2972                 struct disk_reservation disk_res = { 0 };
2973                 struct quota_res quota_res = { 0 };
2974                 struct bkey_i_reservation reservation;
2975                 struct bkey_s_c k;
2976                 unsigned sectors;
2977                 u32 snapshot;
2978
2979                 bch2_trans_begin(&trans);
2980
2981                 ret = bch2_subvolume_get_snapshot(&trans,
2982                                         inode->ei_subvol, &snapshot);
2983                 if (ret)
2984                         goto bkey_err;
2985
2986                 bch2_btree_iter_set_snapshot(&iter, snapshot);
2987
2988                 k = bch2_btree_iter_peek_slot(&iter);
2989                 if ((ret = bkey_err(k)))
2990                         goto bkey_err;
2991
2992                 /* already reserved */
2993                 if (k.k->type == KEY_TYPE_reservation &&
2994                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2995                         bch2_btree_iter_advance(&iter);
2996                         continue;
2997                 }
2998
2999                 if (bkey_extent_is_data(k.k) &&
3000                     !(mode & FALLOC_FL_ZERO_RANGE)) {
3001                         bch2_btree_iter_advance(&iter);
3002                         continue;
3003                 }
3004
3005                 bkey_reservation_init(&reservation.k_i);
3006                 reservation.k.type      = KEY_TYPE_reservation;
3007                 reservation.k.p         = k.k->p;
3008                 reservation.k.size      = k.k->size;
3009
3010                 bch2_cut_front(iter.pos,        &reservation.k_i);
3011                 bch2_cut_back(end_pos,          &reservation.k_i);
3012
3013                 sectors = reservation.k.size;
3014                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
3015
3016                 if (!bkey_extent_is_allocation(k.k)) {
3017                         ret = bch2_quota_reservation_add(c, inode,
3018                                         &quota_res,
3019                                         sectors, true);
3020                         if (unlikely(ret))
3021                                 goto bkey_err;
3022                 }
3023
3024                 if (reservation.v.nr_replicas < replicas ||
3025                     bch2_bkey_sectors_compressed(k)) {
3026                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
3027                                                         replicas, 0);
3028                         if (unlikely(ret))
3029                                 goto bkey_err;
3030
3031                         reservation.v.nr_replicas = disk_res.nr_replicas;
3032                 }
3033
3034                 ret = bch2_extent_update(&trans, inode_inum(inode), &iter,
3035                                          &reservation.k_i,
3036                                 &disk_res, NULL,
3037                                 0, &i_sectors_delta, true);
3038                 if (ret)
3039                         goto bkey_err;
3040                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3041 bkey_err:
3042                 bch2_quota_reservation_put(c, inode, &quota_res);
3043                 bch2_disk_reservation_put(c, &disk_res);
3044                 if (ret == -EINTR)
3045                         ret = 0;
3046         }
3047
3048         bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3049         mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3050
3051         if (ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)) {
3052                 struct quota_res quota_res = { 0 };
3053                 s64 i_sectors_delta = 0;
3054
3055                 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3056                                end_sector, &i_sectors_delta);
3057                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
3058                 bch2_quota_reservation_put(c, inode, &quota_res);
3059         }
3060
3061         bch2_trans_iter_exit(&trans, &iter);
3062         bch2_trans_exit(&trans);
3063         return ret;
3064 }
3065
3066 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3067                             loff_t offset, loff_t len)
3068 {
3069         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3070         u64 end         = offset + len;
3071         u64 block_start = round_down(offset,    block_bytes(c));
3072         u64 block_end   = round_up(end,         block_bytes(c));
3073         bool truncated_last_page = false;
3074         int ret, ret2 = 0;
3075
3076         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3077                 ret = inode_newsize_ok(&inode->v, end);
3078                 if (ret)
3079                         return ret;
3080         }
3081
3082         if (mode & FALLOC_FL_ZERO_RANGE) {
3083                 ret = bch2_truncate_pages(inode, offset, end);
3084                 if (unlikely(ret < 0))
3085                         return ret;
3086
3087                 truncated_last_page = ret;
3088
3089                 truncate_pagecache_range(&inode->v, offset, end - 1);
3090
3091                 block_start     = round_up(offset,      block_bytes(c));
3092                 block_end       = round_down(end,       block_bytes(c));
3093         }
3094
3095         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3096
3097         /*
3098          * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3099          * so that the VFS cache i_size is consistent with the btree i_size:
3100          */
3101         if (ret &&
3102             !(ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)))
3103                 return ret;
3104
3105         if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3106                 end = inode->v.i_size;
3107
3108         if (end >= inode->v.i_size &&
3109             (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3110              !(mode & FALLOC_FL_KEEP_SIZE))) {
3111                 spin_lock(&inode->v.i_lock);
3112                 i_size_write(&inode->v, end);
3113                 spin_unlock(&inode->v.i_lock);
3114
3115                 mutex_lock(&inode->ei_update_lock);
3116                 ret2 = bch2_write_inode_size(c, inode, end, 0);
3117                 mutex_unlock(&inode->ei_update_lock);
3118         }
3119
3120         return ret ?: ret2;
3121 }
3122
3123 long bch2_fallocate_dispatch(struct file *file, int mode,
3124                              loff_t offset, loff_t len)
3125 {
3126         struct bch_inode_info *inode = file_bch_inode(file);
3127         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3128         long ret;
3129
3130         if (!percpu_ref_tryget_live(&c->writes))
3131                 return -EROFS;
3132
3133         inode_lock(&inode->v);
3134         inode_dio_wait(&inode->v);
3135         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
3136
3137         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3138                 ret = bchfs_fallocate(inode, mode, offset, len);
3139         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3140                 ret = bchfs_fpunch(inode, offset, len);
3141         else if (mode == FALLOC_FL_INSERT_RANGE)
3142                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3143         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3144                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3145         else
3146                 ret = -EOPNOTSUPP;
3147
3148
3149         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
3150         inode_unlock(&inode->v);
3151         percpu_ref_put(&c->writes);
3152
3153         return ret;
3154 }
3155
3156 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3157                              struct file *file_dst, loff_t pos_dst,
3158                              loff_t len, unsigned remap_flags)
3159 {
3160         struct bch_inode_info *src = file_bch_inode(file_src);
3161         struct bch_inode_info *dst = file_bch_inode(file_dst);
3162         struct bch_fs *c = src->v.i_sb->s_fs_info;
3163         s64 i_sectors_delta = 0;
3164         u64 aligned_len;
3165         loff_t ret = 0;
3166
3167         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3168                 return -EINVAL;
3169
3170         if (remap_flags & REMAP_FILE_DEDUP)
3171                 return -EOPNOTSUPP;
3172
3173         if ((pos_src & (block_bytes(c) - 1)) ||
3174             (pos_dst & (block_bytes(c) - 1)))
3175                 return -EINVAL;
3176
3177         if (src == dst &&
3178             abs(pos_src - pos_dst) < len)
3179                 return -EINVAL;
3180
3181         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3182
3183         file_update_time(file_dst);
3184
3185         inode_dio_wait(&src->v);
3186         inode_dio_wait(&dst->v);
3187
3188         ret = generic_remap_file_range_prep(file_src, pos_src,
3189                                             file_dst, pos_dst,
3190                                             &len, remap_flags);
3191         if (ret < 0 || len == 0)
3192                 goto err;
3193
3194         aligned_len = round_up((u64) len, block_bytes(c));
3195
3196         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3197                                 pos_dst, pos_dst + len - 1);
3198         if (ret)
3199                 goto err;
3200
3201         mark_pagecache_unallocated(src, pos_src >> 9,
3202                                    (pos_src + aligned_len) >> 9);
3203
3204         ret = bch2_remap_range(c,
3205                                inode_inum(dst), pos_dst >> 9,
3206                                inode_inum(src), pos_src >> 9,
3207                                aligned_len >> 9,
3208                                pos_dst + len, &i_sectors_delta);
3209         if (ret < 0)
3210                 goto err;
3211
3212         /*
3213          * due to alignment, we might have remapped slightly more than requsted
3214          */
3215         ret = min((u64) ret << 9, (u64) len);
3216
3217         /* XXX get a quota reservation */
3218         i_sectors_acct(c, dst, NULL, i_sectors_delta);
3219
3220         spin_lock(&dst->v.i_lock);
3221         if (pos_dst + ret > dst->v.i_size)
3222                 i_size_write(&dst->v, pos_dst + ret);
3223         spin_unlock(&dst->v.i_lock);
3224
3225         if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3226             IS_SYNC(file_inode(file_dst)))
3227                 ret = bch2_flush_inode(c, inode_inum(dst));
3228 err:
3229         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3230
3231         return ret;
3232 }
3233
3234 /* fseek: */
3235
3236 static int page_data_offset(struct page *page, unsigned offset)
3237 {
3238         struct bch_page_state *s = bch2_page_state(page);
3239         unsigned i;
3240
3241         if (s)
3242                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3243                         if (s->s[i].state >= SECTOR_DIRTY)
3244                                 return i << 9;
3245
3246         return -1;
3247 }
3248
3249 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3250                                        loff_t start_offset,
3251                                        loff_t end_offset)
3252 {
3253         struct address_space *mapping = vinode->i_mapping;
3254         struct page *page;
3255         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
3256         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
3257         pgoff_t index           = start_index;
3258         loff_t ret;
3259         int offset;
3260
3261         while (index <= end_index) {
3262                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
3263                         lock_page(page);
3264
3265                         offset = page_data_offset(page,
3266                                         page->index == start_index
3267                                         ? start_offset & (PAGE_SIZE - 1)
3268                                         : 0);
3269                         if (offset >= 0) {
3270                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
3271                                             offset,
3272                                             start_offset, end_offset);
3273                                 unlock_page(page);
3274                                 put_page(page);
3275                                 return ret;
3276                         }
3277
3278                         unlock_page(page);
3279                         put_page(page);
3280                 } else {
3281                         break;
3282                 }
3283         }
3284
3285         return end_offset;
3286 }
3287
3288 static loff_t bch2_seek_data(struct file *file, u64 offset)
3289 {
3290         struct bch_inode_info *inode = file_bch_inode(file);
3291         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3292         struct btree_trans trans;
3293         struct btree_iter iter;
3294         struct bkey_s_c k;
3295         subvol_inum inum = inode_inum(inode);
3296         u64 isize, next_data = MAX_LFS_FILESIZE;
3297         u32 snapshot;
3298         int ret;
3299
3300         isize = i_size_read(&inode->v);
3301         if (offset >= isize)
3302                 return -ENXIO;
3303
3304         bch2_trans_init(&trans, c, 0, 0);
3305 retry:
3306         bch2_trans_begin(&trans);
3307
3308         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3309         if (ret)
3310                 goto err;
3311
3312         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3313                            SPOS(inode->v.i_ino, offset >> 9, snapshot), 0, k, ret) {
3314                 if (k.k->p.inode != inode->v.i_ino) {
3315                         break;
3316                 } else if (bkey_extent_is_data(k.k)) {
3317                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3318                         break;
3319                 } else if (k.k->p.offset >> 9 > isize)
3320                         break;
3321         }
3322         bch2_trans_iter_exit(&trans, &iter);
3323 err:
3324         if (ret == -EINTR)
3325                 goto retry;
3326
3327         bch2_trans_exit(&trans);
3328         if (ret)
3329                 return ret;
3330
3331         if (next_data > offset)
3332                 next_data = bch2_seek_pagecache_data(&inode->v,
3333                                                      offset, next_data);
3334
3335         if (next_data >= isize)
3336                 return -ENXIO;
3337
3338         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3339 }
3340
3341 static int __page_hole_offset(struct page *page, unsigned offset)
3342 {
3343         struct bch_page_state *s = bch2_page_state(page);
3344         unsigned i;
3345
3346         if (!s)
3347                 return 0;
3348
3349         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3350                 if (s->s[i].state < SECTOR_DIRTY)
3351                         return i << 9;
3352
3353         return -1;
3354 }
3355
3356 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3357 {
3358         pgoff_t index = offset >> PAGE_SHIFT;
3359         struct page *page;
3360         int pg_offset;
3361         loff_t ret = -1;
3362
3363         page = find_lock_page(mapping, index);
3364         if (!page)
3365                 return offset;
3366
3367         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3368         if (pg_offset >= 0)
3369                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3370
3371         unlock_page(page);
3372
3373         return ret;
3374 }
3375
3376 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3377                                        loff_t start_offset,
3378                                        loff_t end_offset)
3379 {
3380         struct address_space *mapping = vinode->i_mapping;
3381         loff_t offset = start_offset, hole;
3382
3383         while (offset < end_offset) {
3384                 hole = page_hole_offset(mapping, offset);
3385                 if (hole >= 0 && hole <= end_offset)
3386                         return max(start_offset, hole);
3387
3388                 offset += PAGE_SIZE;
3389                 offset &= PAGE_MASK;
3390         }
3391
3392         return end_offset;
3393 }
3394
3395 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3396 {
3397         struct bch_inode_info *inode = file_bch_inode(file);
3398         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3399         struct btree_trans trans;
3400         struct btree_iter iter;
3401         struct bkey_s_c k;
3402         subvol_inum inum = inode_inum(inode);
3403         u64 isize, next_hole = MAX_LFS_FILESIZE;
3404         u32 snapshot;
3405         int ret;
3406
3407         isize = i_size_read(&inode->v);
3408         if (offset >= isize)
3409                 return -ENXIO;
3410
3411         bch2_trans_init(&trans, c, 0, 0);
3412 retry:
3413         bch2_trans_begin(&trans);
3414
3415         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3416         if (ret)
3417                 goto err;
3418
3419         for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3420                            SPOS(inode->v.i_ino, offset >> 9, snapshot),
3421                            BTREE_ITER_SLOTS, k, ret) {
3422                 if (k.k->p.inode != inode->v.i_ino) {
3423                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3424                                         offset, MAX_LFS_FILESIZE);
3425                         break;
3426                 } else if (!bkey_extent_is_data(k.k)) {
3427                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3428                                         max(offset, bkey_start_offset(k.k) << 9),
3429                                         k.k->p.offset << 9);
3430
3431                         if (next_hole < k.k->p.offset << 9)
3432                                 break;
3433                 } else {
3434                         offset = max(offset, bkey_start_offset(k.k) << 9);
3435                 }
3436         }
3437         bch2_trans_iter_exit(&trans, &iter);
3438 err:
3439         if (ret == -EINTR)
3440                 goto retry;
3441
3442         bch2_trans_exit(&trans);
3443         if (ret)
3444                 return ret;
3445
3446         if (next_hole > isize)
3447                 next_hole = isize;
3448
3449         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3450 }
3451
3452 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3453 {
3454         switch (whence) {
3455         case SEEK_SET:
3456         case SEEK_CUR:
3457         case SEEK_END:
3458                 return generic_file_llseek(file, offset, whence);
3459         case SEEK_DATA:
3460                 return bch2_seek_data(file, offset);
3461         case SEEK_HOLE:
3462                 return bch2_seek_hole(file, offset);
3463         }
3464
3465         return -EINVAL;
3466 }
3467
3468 void bch2_fs_fsio_exit(struct bch_fs *c)
3469 {
3470         bioset_exit(&c->dio_write_bioset);
3471         bioset_exit(&c->dio_read_bioset);
3472         bioset_exit(&c->writepage_bioset);
3473 }
3474
3475 int bch2_fs_fsio_init(struct bch_fs *c)
3476 {
3477         int ret = 0;
3478
3479         pr_verbose_init(c->opts, "");
3480
3481         if (bioset_init(&c->writepage_bioset,
3482                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3483                         BIOSET_NEED_BVECS) ||
3484             bioset_init(&c->dio_read_bioset,
3485                         4, offsetof(struct dio_read, rbio.bio),
3486                         BIOSET_NEED_BVECS) ||
3487             bioset_init(&c->dio_write_bioset,
3488                         4, offsetof(struct dio_write, op.wbio.bio),
3489                         BIOSET_NEED_BVECS))
3490                 ret = -ENOMEM;
3491
3492         pr_verbose_init(c->opts, "ret %i", ret);
3493         return ret;
3494 }
3495
3496 #endif /* NO_BCACHEFS_FS */