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

   1 #ifndef NO_BCACHEFS_FS
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_update.h"
   6 #include "buckets.h"
   7 #include "clock.h"
   8 #include "error.h"
   9 #include "extents.h"
  10 #include "fs.h"
  11 #include "fs-io.h"
  12 #include "fsck.h"
  13 #include "inode.h"
  14 #include "journal.h"
  15 #include "io.h"
  16 #include "keylist.h"
  17 #include "quota.h"
  18
  19 #include <linux/aio.h>
  20 #include <linux/backing-dev.h>
  21 #include <linux/falloc.h>
  22 #include <linux/migrate.h>
  23 #include <linux/mmu_context.h>
  24 #include <linux/pagevec.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/task_io_accounting_ops.h>
  27 #include <linux/uio.h>
  28 #include <linux/writeback.h>
  29
  30 #include <trace/events/bcachefs.h>
  31 #include <trace/events/writeback.h>
  32
  33 struct quota_res {
  34         u64                             sectors;
  35 };
  36
  37 struct bchfs_write_op {
  38         struct bch_inode_info           *inode;
  39         s64                             sectors_added;
  40         bool                            is_dio;
  41         bool                            unalloc;
  42         u64                             new_i_size;
  43
  44         /* must be last: */
  45         struct bch_write_op             op;
  46 };
  47
  48 struct bch_writepage_io {
  49         struct closure                  cl;
  50         u64                             new_sectors;
  51
  52         /* must be last: */
  53         struct bchfs_write_op           op;
  54 };
  55
  56 struct dio_write {
  57         struct closure                  cl;
  58         struct kiocb                    *req;
  59         struct mm_struct                *mm;
  60         unsigned                        loop:1,
  61                                         sync:1,
  62                                         free_iov:1;
  63         struct quota_res                quota_res;
  64
  65         struct iov_iter                 iter;
  66         struct iovec                    inline_vecs[2];
  67
  68         /* must be last: */
  69         struct bchfs_write_op           iop;
  70 };
  71
  72 struct dio_read {
  73         struct closure                  cl;
  74         struct kiocb                    *req;
  75         long                            ret;
  76         struct bch_read_bio             rbio;
  77 };
  78
  79 /* pagecache_block must be held */
  80 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  81                                               loff_t start, loff_t end)
  82 {
  83         int ret;
  84
  85         /*
  86          * XXX: the way this is currently implemented, we can spin if a process
  87          * is continually redirtying a specific page
  88          */
  89         do {
  90                 if (!mapping->nrpages &&
  91                     !mapping->nrexceptional)
  92                         return 0;
  93
  94                 ret = filemap_write_and_wait_range(mapping, start, end);
  95                 if (ret)
  96                         break;
  97
  98                 if (!mapping->nrpages)
  99                         return 0;
 100
 101                 ret = invalidate_inode_pages2_range(mapping,
 102                                 start >> PAGE_SHIFT,
 103                                 end >> PAGE_SHIFT);
 104         } while (ret == -EBUSY);
 105
 106         return ret;
 107 }
 108
 109 /* quotas */
 110
 111 #ifdef CONFIG_BCACHEFS_QUOTA
 112
 113 static void bch2_quota_reservation_put(struct bch_fs *c,
 114                                        struct bch_inode_info *inode,
 115                                        struct quota_res *res)
 116 {
 117         if (!res->sectors)
 118                 return;
 119
 120         mutex_lock(&inode->ei_quota_lock);
 121         BUG_ON(res->sectors > inode->ei_quota_reserved);
 122
 123         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 124                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 125         inode->ei_quota_reserved -= res->sectors;
 126         mutex_unlock(&inode->ei_quota_lock);
 127
 128         res->sectors = 0;
 129 }
 130
 131 static int bch2_quota_reservation_add(struct bch_fs *c,
 132                                       struct bch_inode_info *inode,
 133                                       struct quota_res *res,
 134                                       unsigned sectors,
 135                                       bool check_enospc)
 136 {
 137         int ret;
 138
 139         mutex_lock(&inode->ei_quota_lock);
 140         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 141                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 142         if (likely(!ret)) {
 143                 inode->ei_quota_reserved += sectors;
 144                 res->sectors += sectors;
 145         }
 146         mutex_unlock(&inode->ei_quota_lock);
 147
 148         return ret;
 149 }
 150
 151 #else
 152
 153 static void bch2_quota_reservation_put(struct bch_fs *c,
 154                                        struct bch_inode_info *inode,
 155                                        struct quota_res *res)
 156 {
 157 }
 158
 159 static int bch2_quota_reservation_add(struct bch_fs *c,
 160                                       struct bch_inode_info *inode,
 161                                       struct quota_res *res,
 162                                       unsigned sectors,
 163                                       bool check_enospc)
 164 {
 165         return 0;
 166 }
 167
 168 #endif
 169
 170 /* i_size updates: */
 171
 172 struct inode_new_size {
 173         loff_t          new_size;
 174         u64             now;
 175         unsigned        fields;
 176 };
 177
 178 static int inode_set_size(struct bch_inode_info *inode,
 179                           struct bch_inode_unpacked *bi,
 180                           void *p)
 181 {
 182         struct inode_new_size *s = p;
 183
 184         bi->bi_size = s->new_size;
 185         if (s->fields & ATTR_ATIME)
 186                 bi->bi_atime = s->now;
 187         if (s->fields & ATTR_MTIME)
 188                 bi->bi_mtime = s->now;
 189         if (s->fields & ATTR_CTIME)
 190                 bi->bi_ctime = s->now;
 191
 192         return 0;
 193 }
 194
 195 static int __must_check bch2_write_inode_size(struct bch_fs *c,
 196                                               struct bch_inode_info *inode,
 197                                               loff_t new_size, unsigned fields)
 198 {
 199         struct inode_new_size s = {
 200                 .new_size       = new_size,
 201                 .now            = bch2_current_time(c),
 202                 .fields         = fields,
 203         };
 204
 205         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 206 }
 207
 208 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 209                            struct quota_res *quota_res, s64 sectors)
 210 {
 211         if (!sectors)
 212                 return;
 213
 214         mutex_lock(&inode->ei_quota_lock);
 215 #ifdef CONFIG_BCACHEFS_QUOTA
 216         if (quota_res && sectors > 0) {
 217                 BUG_ON(sectors > quota_res->sectors);
 218                 BUG_ON(sectors > inode->ei_quota_reserved);
 219
 220                 quota_res->sectors -= sectors;
 221                 inode->ei_quota_reserved -= sectors;
 222         } else {
 223                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 224         }
 225 #endif
 226         inode->v.i_blocks += sectors;
 227         mutex_unlock(&inode->ei_quota_lock);
 228 }
 229
 230 /* normal i_size/i_sectors update machinery: */
 231
 232 static int sum_sector_overwrites(struct btree_trans *trans,
 233                                  struct btree_iter *extent_iter,
 234                                  struct bkey_i *new, bool *allocating,
 235                                  s64 *i_sectors_delta)
 236 {
 237         struct btree_iter *iter = bch2_trans_copy_iter(trans, extent_iter);
 238         struct bkey_s_c old;
 239         s64 delta = 0;
 240
 241         if (IS_ERR(iter))
 242                 return PTR_ERR(iter);
 243
 244         old = bch2_btree_iter_peek_slot(iter);
 245
 246         while (1) {
 247                 /*
 248                  * should not be possible to get an error here, since we're
 249                  * carefully not advancing past @new and thus whatever leaf node
 250                  * @_iter currently points to:
 251                  */
 252                 BUG_ON(btree_iter_err(old));
 253
 254                 if (allocating &&
 255                     !*allocating &&
 256                     bch2_bkey_nr_ptrs_allocated(old) <
 257                     bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(new)))
 258                         *allocating = true;
 259
 260                 delta += (min(new->k.p.offset,
 261                               old.k->p.offset) -
 262                           max(bkey_start_offset(&new->k),
 263                               bkey_start_offset(old.k))) *
 264                         (bkey_extent_is_allocation(&new->k) -
 265                          bkey_extent_is_allocation(old.k));
 266
 267                 if (bkey_cmp(old.k->p, new->k.p) >= 0)
 268                         break;
 269
 270                 old = bch2_btree_iter_next_slot(iter);
 271         }
 272
 273         bch2_trans_iter_free(trans, iter);
 274
 275         *i_sectors_delta = delta;
 276         return 0;
 277 }
 278
 279 static int bch2_extent_update(struct btree_trans *trans,
 280                               struct bch_inode_info *inode,
 281                               struct disk_reservation *disk_res,
 282                               struct quota_res *quota_res,
 283                               struct btree_iter *extent_iter,
 284                               struct bkey_i *k,
 285                               u64 new_i_size,
 286                               bool may_allocate,
 287                               bool direct,
 288                               s64 *total_delta)
 289 {
 290         struct bch_inode_unpacked inode_u;
 291         struct bkey_inode_buf inode_p;
 292         bool allocating = false;
 293         bool extended = false;
 294         bool inode_locked = false;
 295         s64 i_sectors_delta;
 296         int ret;
 297
 298         bch2_trans_begin_updates(trans);
 299
 300         ret = bch2_btree_iter_traverse(extent_iter);
 301         if (ret)
 302                 return ret;
 303
 304         bch2_extent_trim_atomic(k, extent_iter);
 305
 306         ret = sum_sector_overwrites(trans, extent_iter,
 307                                     k, &allocating,
 308                                     &i_sectors_delta);
 309         if (ret)
 310                 return ret;
 311
 312         if (!may_allocate && allocating)
 313                 return -ENOSPC;
 314
 315         bch2_trans_update(trans, BTREE_INSERT_ENTRY(extent_iter, k));
 316
 317         new_i_size = min(k->k.p.offset << 9, new_i_size);
 318
 319         /* XXX: inode->i_size locking */
 320         if (i_sectors_delta ||
 321             new_i_size > inode->ei_inode.bi_size) {
 322                 bch2_btree_iter_unlock(extent_iter);
 323                 mutex_lock(&inode->ei_update_lock);
 324
 325                 if (!bch2_btree_iter_relock(extent_iter)) {
 326                         mutex_unlock(&inode->ei_update_lock);
 327                         return -EINTR;
 328                 }
 329
 330                 inode_locked = true;
 331
 332                 if (!inode->ei_inode_update)
 333                         inode->ei_inode_update =
 334                                 bch2_deferred_update_alloc(trans->c,
 335                                                         BTREE_ID_INODES, 64);
 336
 337                 inode_u = inode->ei_inode;
 338                 inode_u.bi_sectors += i_sectors_delta;
 339
 340                 /* XXX: this is slightly suspect */
 341                 if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
 342                     new_i_size > inode_u.bi_size) {
 343                         inode_u.bi_size = new_i_size;
 344                         extended = true;
 345                 }
 346
 347                 bch2_inode_pack(&inode_p, &inode_u);
 348                 bch2_trans_update(trans,
 349                         BTREE_INSERT_DEFERRED(inode->ei_inode_update,
 350                                               &inode_p.inode.k_i));
 351         }
 352
 353         ret = bch2_trans_commit(trans, disk_res,
 354                                 &inode->ei_journal_seq,
 355                                 BTREE_INSERT_NOFAIL|
 356                                 BTREE_INSERT_ATOMIC|
 357                                 BTREE_INSERT_NOUNLOCK|
 358                                 BTREE_INSERT_USE_RESERVE);
 359         if (ret)
 360                 goto err;
 361
 362         inode->ei_inode.bi_sectors += i_sectors_delta;
 363
 364         EBUG_ON(i_sectors_delta &&
 365                 inode->ei_inode.bi_sectors != inode_u.bi_sectors);
 366
 367         if (extended) {
 368                 inode->ei_inode.bi_size = new_i_size;
 369
 370                 if (direct) {
 371                         spin_lock(&inode->v.i_lock);
 372                         if (new_i_size > inode->v.i_size)
 373                                 i_size_write(&inode->v, new_i_size);
 374                         spin_unlock(&inode->v.i_lock);
 375                 }
 376         }
 377
 378         if (direct)
 379                 i_sectors_acct(trans->c, inode, quota_res, i_sectors_delta);
 380
 381         if (total_delta)
 382                 *total_delta += i_sectors_delta;
 383 err:
 384         if (inode_locked)
 385                 mutex_unlock(&inode->ei_update_lock);
 386
 387         return ret;
 388 }
 389
 390 static int bchfs_write_index_update(struct bch_write_op *wop)
 391 {
 392         struct bch_fs *c = wop->c;
 393         struct bchfs_write_op *op = container_of(wop,
 394                                 struct bchfs_write_op, op);
 395         struct quota_res *quota_res = op->is_dio
 396                 ? &container_of(op, struct dio_write, iop)->quota_res
 397                 : NULL;
 398         struct bch_inode_info *inode = op->inode;
 399         struct keylist *keys = &op->op.insert_keys;
 400         struct bkey_i *k = bch2_keylist_front(keys);
 401         struct btree_trans trans;
 402         struct btree_iter *iter;
 403         int ret;
 404
 405         BUG_ON(k->k.p.inode != inode->v.i_ino);
 406
 407         bch2_trans_init(&trans, c);
 408         bch2_trans_preload_iters(&trans);
 409
 410         iter = bch2_trans_get_iter(&trans,
 411                                 BTREE_ID_EXTENTS,
 412                                 bkey_start_pos(&k->k),
 413                                 BTREE_ITER_INTENT);
 414
 415         do {
 416                 BKEY_PADDED(k) tmp;
 417
 418                 bkey_copy(&tmp.k, bch2_keylist_front(keys));
 419
 420                 ret = bch2_extent_update(&trans, inode,
 421                                 &wop->res, quota_res,
 422                                 iter, &tmp.k,
 423                                 op->new_i_size,
 424                                 !op->unalloc,
 425                                 op->is_dio,
 426                                 &op->sectors_added);
 427                 if (ret == -EINTR)
 428                         continue;
 429                 if (ret)
 430                         break;
 431
 432                 if (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) < 0)
 433                         bch2_cut_front(iter->pos, bch2_keylist_front(keys));
 434                 else
 435                         bch2_keylist_pop_front(keys);
 436         } while (!bch2_keylist_empty(keys));
 437
 438         bch2_trans_exit(&trans);
 439
 440         return ret;
 441 }
 442
 443 static inline void bch2_fswrite_op_init(struct bchfs_write_op *op,
 444                                         struct bch_fs *c,
 445                                         struct bch_inode_info *inode,
 446                                         struct bch_io_opts opts,
 447                                         bool is_dio)
 448 {
 449         op->inode               = inode;
 450         op->sectors_added       = 0;
 451         op->is_dio              = is_dio;
 452         op->unalloc             = false;
 453         op->new_i_size          = U64_MAX;
 454
 455         bch2_write_op_init(&op->op, c, opts);
 456         op->op.target           = opts.foreground_target;
 457         op->op.index_update_fn  = bchfs_write_index_update;
 458         op_journal_seq_set(&op->op, &inode->ei_journal_seq);
 459 }
 460
 461 static inline struct bch_io_opts io_opts(struct bch_fs *c, struct bch_inode_info *inode)
 462 {
 463         struct bch_io_opts opts = bch2_opts_to_inode_opts(c->opts);
 464
 465         bch2_io_opts_apply(&opts, bch2_inode_opts_get(&inode->ei_inode));
 466         return opts;
 467 }
 468
 469 /* page state: */
 470
 471 /* stored in page->private: */
 472
 473 /*
 474  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 475  * almost protected it with the page lock, except that bch2_writepage_io_done has
 476  * to update the sector counts (and from interrupt/bottom half context).
 477  */
 478 struct bch_page_state {
 479 union { struct {
 480         /* existing data: */
 481         unsigned                sectors:PAGE_SECTOR_SHIFT + 1;
 482
 483         /* Uncompressed, fully allocated replicas: */
 484         unsigned                nr_replicas:4;
 485
 486         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 487         unsigned                replicas_reserved:4;
 488
 489         /* Owns PAGE_SECTORS sized quota reservation: */
 490         unsigned                quota_reserved:1;
 491
 492         /*
 493          * Number of sectors on disk - for i_blocks
 494          * Uncompressed size, not compressed size:
 495          */
 496         unsigned                dirty_sectors:PAGE_SECTOR_SHIFT + 1;
 497 };
 498         /* for cmpxchg: */
 499         unsigned long           v;
 500 };
 501 };
 502
 503 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 504 ({                                                                      \
 505         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 506         struct bch_page_state _old;                                     \
 507                                                                         \
 508         do {                                                            \
 509                 _old.v = _new.v = _v;                                   \
 510                 _expr;                                                  \
 511                                                                         \
 512                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 513         } while (_old.v != _new.v &&                                    \
 514                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 515                                                                         \
 516         _old;                                                           \
 517 })
 518
 519 static inline struct bch_page_state *page_state(struct page *page)
 520 {
 521         struct bch_page_state *s = (void *) &page->private;
 522
 523         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 524
 525         if (!PagePrivate(page))
 526                 SetPagePrivate(page);
 527
 528         return s;
 529 }
 530
 531 static inline unsigned page_res_sectors(struct bch_page_state s)
 532 {
 533
 534         return s.replicas_reserved * PAGE_SECTORS;
 535 }
 536
 537 static void __bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 538                                         struct bch_page_state s)
 539 {
 540         struct disk_reservation res = { .sectors = page_res_sectors(s) };
 541         struct quota_res quota_res = { .sectors = s.quota_reserved ? PAGE_SECTORS : 0 };
 542
 543         bch2_quota_reservation_put(c, inode, &quota_res);
 544         bch2_disk_reservation_put(c, &res);
 545 }
 546
 547 static void bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 548                                       struct page *page)
 549 {
 550         struct bch_page_state s;
 551
 552         EBUG_ON(!PageLocked(page));
 553
 554         s = page_state_cmpxchg(page_state(page), s, {
 555                 s.replicas_reserved     = 0;
 556                 s.quota_reserved        = 0;
 557         });
 558
 559         __bch2_put_page_reservation(c, inode, s);
 560 }
 561
 562 static int bch2_get_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 563                                      struct page *page, bool check_enospc)
 564 {
 565         struct bch_page_state *s = page_state(page), new;
 566
 567         /* XXX: this should not be open coded */
 568         unsigned nr_replicas = inode->ei_inode.bi_data_replicas
 569                 ? inode->ei_inode.bi_data_replicas - 1
 570                 : c->opts.data_replicas;
 571         struct disk_reservation disk_res;
 572         struct quota_res quota_res = { 0 };
 573         int ret;
 574
 575         EBUG_ON(!PageLocked(page));
 576
 577         if (s->replicas_reserved < nr_replicas) {
 578                 ret = bch2_disk_reservation_get(c, &disk_res, PAGE_SECTORS,
 579                                 nr_replicas - s->replicas_reserved,
 580                                 !check_enospc ? BCH_DISK_RESERVATION_NOFAIL : 0);
 581                 if (unlikely(ret))
 582                         return ret;
 583
 584                 page_state_cmpxchg(s, new, ({
 585                         BUG_ON(new.replicas_reserved +
 586                                disk_res.nr_replicas != nr_replicas);
 587                         new.replicas_reserved += disk_res.nr_replicas;
 588                 }));
 589         }
 590
 591         if (!s->quota_reserved &&
 592             s->sectors + s->dirty_sectors < PAGE_SECTORS) {
 593                 ret = bch2_quota_reservation_add(c, inode, &quota_res,
 594                                                  PAGE_SECTORS,
 595                                                  check_enospc);
 596                 if (unlikely(ret))
 597                         return ret;
 598
 599                 page_state_cmpxchg(s, new, ({
 600                         BUG_ON(new.quota_reserved);
 601                         new.quota_reserved = 1;
 602                 }));
 603         }
 604
 605         return ret;
 606 }
 607
 608 static void bch2_clear_page_bits(struct page *page)
 609 {
 610         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 611         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 612         struct bch_page_state s;
 613
 614         EBUG_ON(!PageLocked(page));
 615
 616         if (!PagePrivate(page))
 617                 return;
 618
 619         s.v = xchg(&page_state(page)->v, 0);
 620         ClearPagePrivate(page);
 621
 622         if (s.dirty_sectors)
 623                 i_sectors_acct(c, inode, NULL, -s.dirty_sectors);
 624
 625         __bch2_put_page_reservation(c, inode, s);
 626 }
 627
 628 int bch2_set_page_dirty(struct page *page)
 629 {
 630         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 631         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 632         struct quota_res quota_res = { 0 };
 633         struct bch_page_state old, new;
 634
 635         old = page_state_cmpxchg(page_state(page), new,
 636                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 637                 new.quota_reserved = 0;
 638         );
 639
 640         quota_res.sectors += old.quota_reserved * PAGE_SECTORS;
 641
 642         if (old.dirty_sectors != new.dirty_sectors)
 643                 i_sectors_acct(c, inode, &quota_res,
 644                                new.dirty_sectors - old.dirty_sectors);
 645         bch2_quota_reservation_put(c, inode, &quota_res);
 646
 647         return __set_page_dirty_nobuffers(page);
 648 }
 649
 650 int bch2_page_mkwrite(struct vm_fault *vmf)
 651 {
 652         struct page *page = vmf->page;
 653         struct file *file = vmf->vma->vm_file;
 654         struct bch_inode_info *inode = file_bch_inode(file);
 655         struct address_space *mapping = inode->v.i_mapping;
 656         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 657         int ret = VM_FAULT_LOCKED;
 658
 659         sb_start_pagefault(inode->v.i_sb);
 660         file_update_time(file);
 661
 662         /*
 663          * Not strictly necessary, but helps avoid dio writes livelocking in
 664          * write_invalidate_inode_pages_range() - can drop this if/when we get
 665          * a write_invalidate_inode_pages_range() that works without dropping
 666          * page lock before invalidating page
 667          */
 668         if (current->pagecache_lock != &mapping->add_lock)
 669                 pagecache_add_get(&mapping->add_lock);
 670
 671         lock_page(page);
 672         if (page->mapping != mapping ||
 673             page_offset(page) > i_size_read(&inode->v)) {
 674                 unlock_page(page);
 675                 ret = VM_FAULT_NOPAGE;
 676                 goto out;
 677         }
 678
 679         if (bch2_get_page_reservation(c, inode, page, true)) {
 680                 unlock_page(page);
 681                 ret = VM_FAULT_SIGBUS;
 682                 goto out;
 683         }
 684
 685         if (!PageDirty(page))
 686                 set_page_dirty(page);
 687         wait_for_stable_page(page);
 688 out:
 689         if (current->pagecache_lock != &mapping->add_lock)
 690                 pagecache_add_put(&mapping->add_lock);
 691         sb_end_pagefault(inode->v.i_sb);
 692         return ret;
 693 }
 694
 695 void bch2_invalidatepage(struct page *page, unsigned int offset,
 696                          unsigned int length)
 697 {
 698         EBUG_ON(!PageLocked(page));
 699         EBUG_ON(PageWriteback(page));
 700
 701         if (offset || length < PAGE_SIZE)
 702                 return;
 703
 704         bch2_clear_page_bits(page);
 705 }
 706
 707 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 708 {
 709         /* XXX: this can't take locks that are held while we allocate memory */
 710         EBUG_ON(!PageLocked(page));
 711         EBUG_ON(PageWriteback(page));
 712
 713         if (PageDirty(page))
 714                 return 0;
 715
 716         bch2_clear_page_bits(page);
 717         return 1;
 718 }
 719
 720 #ifdef CONFIG_MIGRATION
 721 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 722                       struct page *page, enum migrate_mode mode)
 723 {
 724         int ret;
 725
 726         EBUG_ON(!PageLocked(page));
 727         EBUG_ON(!PageLocked(newpage));
 728
 729         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 730         if (ret != MIGRATEPAGE_SUCCESS)
 731                 return ret;
 732
 733         if (PagePrivate(page)) {
 734                 *page_state(newpage) = *page_state(page);
 735                 ClearPagePrivate(page);
 736         }
 737
 738         migrate_page_copy(newpage, page);
 739         return MIGRATEPAGE_SUCCESS;
 740 }
 741 #endif
 742
 743 /* readpages/writepages: */
 744
 745 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 746 {
 747         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 748
 749         return bio->bi_vcnt < bio->bi_max_vecs &&
 750                 bio_end_sector(bio) == offset;
 751 }
 752
 753 static int bio_add_page_contig(struct bio *bio, struct page *page)
 754 {
 755         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 756
 757         EBUG_ON(!bio->bi_max_vecs);
 758
 759         if (!bio->bi_vcnt)
 760                 bio->bi_iter.bi_sector = offset;
 761         else if (!bio_can_add_page_contig(bio, page))
 762                 return -1;
 763
 764         __bio_add_page(bio, page, PAGE_SIZE, 0);
 765         return 0;
 766 }
 767
 768 /* readpage(s): */
 769
 770 static void bch2_readpages_end_io(struct bio *bio)
 771 {
 772         struct bio_vec *bv;
 773         int i;
 774
 775         bio_for_each_segment_all(bv, bio, i) {
 776                 struct page *page = bv->bv_page;
 777
 778                 if (!bio->bi_status) {
 779                         SetPageUptodate(page);
 780                 } else {
 781                         ClearPageUptodate(page);
 782                         SetPageError(page);
 783                 }
 784                 unlock_page(page);
 785         }
 786
 787         bio_put(bio);
 788 }
 789
 790 static inline void page_state_init_for_read(struct page *page)
 791 {
 792         SetPagePrivate(page);
 793         page->private = 0;
 794 }
 795
 796 struct readpages_iter {
 797         struct address_space    *mapping;
 798         struct page             **pages;
 799         unsigned                nr_pages;
 800         unsigned                nr_added;
 801         unsigned                idx;
 802         pgoff_t                 offset;
 803 };
 804
 805 static int readpages_iter_init(struct readpages_iter *iter,
 806                                struct address_space *mapping,
 807                                struct list_head *pages, unsigned nr_pages)
 808 {
 809         memset(iter, 0, sizeof(*iter));
 810
 811         iter->mapping   = mapping;
 812         iter->offset    = list_last_entry(pages, struct page, lru)->index;
 813
 814         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 815         if (!iter->pages)
 816                 return -ENOMEM;
 817
 818         while (!list_empty(pages)) {
 819                 struct page *page = list_last_entry(pages, struct page, lru);
 820
 821                 prefetchw(&page->flags);
 822                 iter->pages[iter->nr_pages++] = page;
 823                 list_del(&page->lru);
 824         }
 825
 826         return 0;
 827 }
 828
 829 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 830 {
 831         struct page *page;
 832         unsigned i;
 833         int ret;
 834
 835         BUG_ON(iter->idx > iter->nr_added);
 836         BUG_ON(iter->nr_added > iter->nr_pages);
 837
 838         if (iter->idx < iter->nr_added)
 839                 goto out;
 840
 841         while (1) {
 842                 if (iter->idx == iter->nr_pages)
 843                         return NULL;
 844
 845                 ret = add_to_page_cache_lru_vec(iter->mapping,
 846                                 iter->pages     + iter->nr_added,
 847                                 iter->nr_pages  - iter->nr_added,
 848                                 iter->offset    + iter->nr_added,
 849                                 GFP_NOFS);
 850                 if (ret > 0)
 851                         break;
 852
 853                 page = iter->pages[iter->nr_added];
 854                 iter->idx++;
 855                 iter->nr_added++;
 856
 857                 put_page(page);
 858         }
 859
 860         iter->nr_added += ret;
 861
 862         for (i = iter->idx; i < iter->nr_added; i++)
 863                 put_page(iter->pages[i]);
 864 out:
 865         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 866
 867         page_state_init_for_read(iter->pages[iter->idx]);
 868         return iter->pages[iter->idx];
 869 }
 870
 871 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 872 {
 873         struct bvec_iter iter;
 874         struct bio_vec bv;
 875         unsigned nr_ptrs = bch2_bkey_nr_ptrs_allocated(k);
 876
 877         bio_for_each_segment(bv, bio, iter) {
 878                 /* brand new pages, don't need to be locked: */
 879
 880                 struct bch_page_state *s = page_state(bv.bv_page);
 881
 882                 /* sectors in @k from the start of this page: */
 883                 unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset);
 884
 885                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 886
 887                 s->nr_replicas = page_sectors == PAGE_SECTORS
 888                         ? nr_ptrs : 0;
 889
 890                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 891                 s->sectors += page_sectors;
 892         }
 893 }
 894
 895 static void readpage_bio_extend(struct readpages_iter *iter,
 896                                 struct bio *bio, u64 offset,
 897                                 bool get_more)
 898 {
 899         while (bio_end_sector(bio) < offset &&
 900                bio->bi_vcnt < bio->bi_max_vecs) {
 901                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 902                 struct page *page = readpage_iter_next(iter);
 903                 int ret;
 904
 905                 if (page) {
 906                         if (iter->offset + iter->idx != page_offset)
 907                                 break;
 908
 909                         iter->idx++;
 910                 } else {
 911                         if (!get_more)
 912                                 break;
 913
 914                         page = xa_load(&iter->mapping->i_pages, page_offset);
 915                         if (page && !xa_is_value(page))
 916                                 break;
 917
 918                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 919                         if (!page)
 920                                 break;
 921
 922                         page_state_init_for_read(page);
 923
 924                         ret = add_to_page_cache_lru(page, iter->mapping,
 925                                                     page_offset, GFP_NOFS);
 926                         if (ret) {
 927                                 ClearPagePrivate(page);
 928                                 put_page(page);
 929                                 break;
 930                         }
 931
 932                         put_page(page);
 933                 }
 934
 935                 __bio_add_page(bio, page, PAGE_SIZE, 0);
 936         }
 937 }
 938
 939 static void bchfs_read(struct bch_fs *c, struct btree_iter *iter,
 940                        struct bch_read_bio *rbio, u64 inum,
 941                        struct readpages_iter *readpages_iter)
 942 {
 943         struct bio *bio = &rbio->bio;
 944         int flags = BCH_READ_RETRY_IF_STALE|
 945                 BCH_READ_MAY_PROMOTE;
 946
 947         rbio->c = c;
 948         rbio->start_time = local_clock();
 949
 950         while (1) {
 951                 BKEY_PADDED(k) tmp;
 952                 struct bkey_s_c k;
 953                 unsigned bytes;
 954
 955                 bch2_btree_iter_set_pos(iter, POS(inum, bio->bi_iter.bi_sector));
 956
 957                 k = bch2_btree_iter_peek_slot(iter);
 958                 BUG_ON(!k.k);
 959
 960                 if (IS_ERR(k.k)) {
 961                         int ret = bch2_btree_iter_unlock(iter);
 962                         BUG_ON(!ret);
 963                         bcache_io_error(c, bio, "btree IO error %i", ret);
 964                         bio_endio(bio);
 965                         return;
 966                 }
 967
 968                 bkey_reassemble(&tmp.k, k);
 969                 bch2_btree_iter_unlock(iter);
 970                 k = bkey_i_to_s_c(&tmp.k);
 971
 972                 if (readpages_iter) {
 973                         bool want_full_extent = false;
 974
 975                         if (bkey_extent_is_data(k.k)) {
 976                                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 977                                 const union bch_extent_entry *i;
 978                                 struct extent_ptr_decoded p;
 979
 980                                 extent_for_each_ptr_decode(e, p, i)
 981                                         want_full_extent |= ((p.crc.csum_type != 0) |
 982                                                              (p.crc.compression_type != 0));
 983                         }
 984
 985                         readpage_bio_extend(readpages_iter,
 986                                             bio, k.k->p.offset,
 987                                             want_full_extent);
 988                 }
 989
 990                 bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
 991                          bio->bi_iter.bi_sector) << 9;
 992                 swap(bio->bi_iter.bi_size, bytes);
 993
 994                 if (bytes == bio->bi_iter.bi_size)
 995                         flags |= BCH_READ_LAST_FRAGMENT;
 996
 997                 if (bkey_extent_is_allocation(k.k))
 998                         bch2_add_page_sectors(bio, k);
 999
1000                 bch2_read_extent(c, rbio, k, flags);
1001
1002                 if (flags & BCH_READ_LAST_FRAGMENT)
1003                         return;
1004
1005                 swap(bio->bi_iter.bi_size, bytes);
1006                 bio_advance(bio, bytes);
1007         }
1008 }
1009
1010 int bch2_readpages(struct file *file, struct address_space *mapping,
1011                    struct list_head *pages, unsigned nr_pages)
1012 {
1013         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1014         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1015         struct bch_io_opts opts = io_opts(c, inode);
1016         struct btree_iter iter;
1017         struct page *page;
1018         struct readpages_iter readpages_iter;
1019         int ret;
1020
1021         ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
1022         BUG_ON(ret);
1023
1024         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1025                              BTREE_ITER_SLOTS);
1026
1027         if (current->pagecache_lock != &mapping->add_lock)
1028                 pagecache_add_get(&mapping->add_lock);
1029
1030         while ((page = readpage_iter_next(&readpages_iter))) {
1031                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1032                 unsigned n = min_t(unsigned,
1033                                    readpages_iter.nr_pages -
1034                                    readpages_iter.idx,
1035                                    BIO_MAX_PAGES);
1036                 struct bch_read_bio *rbio =
1037                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
1038                                   opts);
1039
1040                 readpages_iter.idx++;
1041
1042                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
1043                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
1044                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1045                 __bio_add_page(&rbio->bio, page, PAGE_SIZE, 0);
1046
1047                 bchfs_read(c, &iter, rbio, inode->v.i_ino, &readpages_iter);
1048         }
1049
1050         if (current->pagecache_lock != &mapping->add_lock)
1051                 pagecache_add_put(&mapping->add_lock);
1052
1053         kfree(readpages_iter.pages);
1054
1055         return 0;
1056 }
1057
1058 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1059                              u64 inum, struct page *page)
1060 {
1061         struct btree_iter iter;
1062
1063         page_state_init_for_read(page);
1064
1065         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1066         bio_add_page_contig(&rbio->bio, page);
1067
1068         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1069                              BTREE_ITER_SLOTS);
1070         bchfs_read(c, &iter, rbio, inum, NULL);
1071 }
1072
1073 int bch2_readpage(struct file *file, struct page *page)
1074 {
1075         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1076         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1077         struct bch_io_opts opts = io_opts(c, inode);
1078         struct bch_read_bio *rbio;
1079
1080         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
1081         rbio->bio.bi_end_io = bch2_readpages_end_io;
1082
1083         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1084         return 0;
1085 }
1086
1087 static void bch2_read_single_page_end_io(struct bio *bio)
1088 {
1089         complete(bio->bi_private);
1090 }
1091
1092 static int bch2_read_single_page(struct page *page,
1093                                  struct address_space *mapping)
1094 {
1095         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1096         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1097         struct bch_read_bio *rbio;
1098         int ret;
1099         DECLARE_COMPLETION_ONSTACK(done);
1100
1101         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1102                          io_opts(c, inode));
1103         rbio->bio.bi_private = &done;
1104         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1105
1106         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1107         wait_for_completion(&done);
1108
1109         ret = blk_status_to_errno(rbio->bio.bi_status);
1110         bio_put(&rbio->bio);
1111
1112         if (ret < 0)
1113                 return ret;
1114
1115         SetPageUptodate(page);
1116         return 0;
1117 }
1118
1119 /* writepages: */
1120
1121 struct bch_writepage_state {
1122         struct bch_writepage_io *io;
1123         struct bch_io_opts      opts;
1124 };
1125
1126 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1127                                                                   struct bch_inode_info *inode)
1128 {
1129         return (struct bch_writepage_state) { .opts = io_opts(c, inode) };
1130 }
1131
1132 static void bch2_writepage_io_free(struct closure *cl)
1133 {
1134         struct bch_writepage_io *io = container_of(cl,
1135                                         struct bch_writepage_io, cl);
1136
1137         bio_put(&io->op.op.wbio.bio);
1138 }
1139
1140 static void bch2_writepage_io_done(struct closure *cl)
1141 {
1142         struct bch_writepage_io *io = container_of(cl,
1143                                         struct bch_writepage_io, cl);
1144         struct bch_fs *c = io->op.op.c;
1145         struct bio *bio = &io->op.op.wbio.bio;
1146         struct bio_vec *bvec;
1147         unsigned i;
1148
1149         if (io->op.op.error) {
1150                 bio_for_each_segment_all(bvec, bio, i)
1151                         SetPageError(bvec->bv_page);
1152                 set_bit(AS_EIO, &io->op.inode->v.i_mapping->flags);
1153         }
1154
1155         /*
1156          * racing with fallocate can cause us to add fewer sectors than
1157          * expected - but we shouldn't add more sectors than expected:
1158          */
1159         BUG_ON(io->op.sectors_added > (s64) io->new_sectors);
1160
1161         /*
1162          * (error (due to going RO) halfway through a page can screw that up
1163          * slightly)
1164          * XXX wtf?
1165            BUG_ON(io->op.sectors_added - io->new_sectors >= (s64) PAGE_SECTORS);
1166          */
1167
1168         /*
1169          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1170          * before calling end_page_writeback:
1171          */
1172         if (io->op.sectors_added != io->new_sectors)
1173                 i_sectors_acct(c, io->op.inode, NULL,
1174                                io->op.sectors_added - (s64) io->new_sectors);
1175
1176         bio_for_each_segment_all(bvec, bio, i)
1177                 end_page_writeback(bvec->bv_page);
1178
1179         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1180 }
1181
1182 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1183 {
1184         struct bch_writepage_io *io = w->io;
1185
1186         w->io = NULL;
1187         closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl);
1188         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1189 }
1190
1191 /*
1192  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1193  * possible, else allocating a new one:
1194  */
1195 static void bch2_writepage_io_alloc(struct bch_fs *c,
1196                                     struct bch_writepage_state *w,
1197                                     struct bch_inode_info *inode,
1198                                     struct page *page,
1199                                     unsigned nr_replicas)
1200 {
1201         struct bch_write_op *op;
1202         u64 offset = (u64) page->index << PAGE_SECTOR_SHIFT;
1203
1204         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1205                                               BIO_MAX_PAGES,
1206                                               &c->writepage_bioset),
1207                              struct bch_writepage_io, op.op.wbio.bio);
1208
1209         closure_init(&w->io->cl, NULL);
1210         w->io->new_sectors      = 0;
1211         bch2_fswrite_op_init(&w->io->op, c, inode, w->opts, false);
1212         op                      = &w->io->op.op;
1213         op->nr_replicas         = nr_replicas;
1214         op->res.nr_replicas     = nr_replicas;
1215         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1216         op->pos                 = POS(inode->v.i_ino, offset);
1217         op->wbio.bio.bi_iter.bi_sector = offset;
1218 }
1219
1220 static int __bch2_writepage(struct page *page,
1221                             struct writeback_control *wbc,
1222                             void *data)
1223 {
1224         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1225         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1226         struct bch_writepage_state *w = data;
1227         struct bch_page_state new, old;
1228         unsigned offset, nr_replicas_this_write;
1229         loff_t i_size = i_size_read(&inode->v);
1230         pgoff_t end_index = i_size >> PAGE_SHIFT;
1231
1232         EBUG_ON(!PageUptodate(page));
1233
1234         /* Is the page fully inside i_size? */
1235         if (page->index < end_index)
1236                 goto do_io;
1237
1238         /* Is the page fully outside i_size? (truncate in progress) */
1239         offset = i_size & (PAGE_SIZE - 1);
1240         if (page->index > end_index || !offset) {
1241                 unlock_page(page);
1242                 return 0;
1243         }
1244
1245         /*
1246          * The page straddles i_size.  It must be zeroed out on each and every
1247          * writepage invocation because it may be mmapped.  "A file is mapped
1248          * in multiples of the page size.  For a file that is not a multiple of
1249          * the  page size, the remaining memory is zeroed when mapped, and
1250          * writes to that region are not written out to the file."
1251          */
1252         zero_user_segment(page, offset, PAGE_SIZE);
1253 do_io:
1254         EBUG_ON(!PageLocked(page));
1255
1256         /* Before unlocking the page, transfer reservation to w->io: */
1257         old = page_state_cmpxchg(page_state(page), new, {
1258                 /*
1259                  * If we didn't get a reservation, we can only write out the
1260                  * number of (fully allocated) replicas that currently exist,
1261                  * and only if the entire page has been written:
1262                  */
1263                 nr_replicas_this_write =
1264                         max_t(unsigned,
1265                               new.replicas_reserved,
1266                               (new.sectors == PAGE_SECTORS
1267                                ? new.nr_replicas : 0));
1268
1269                 BUG_ON(!nr_replicas_this_write);
1270
1271                 new.nr_replicas = w->opts.compression
1272                         ? 0
1273                         : nr_replicas_this_write;
1274
1275                 new.replicas_reserved = 0;
1276
1277                 new.sectors += new.dirty_sectors;
1278                 BUG_ON(new.sectors != PAGE_SECTORS);
1279                 new.dirty_sectors = 0;
1280         });
1281
1282         BUG_ON(PageWriteback(page));
1283         set_page_writeback(page);
1284         unlock_page(page);
1285
1286         if (w->io &&
1287             (w->io->op.op.res.nr_replicas != nr_replicas_this_write ||
1288              !bio_can_add_page_contig(&w->io->op.op.wbio.bio, page)))
1289                 bch2_writepage_do_io(w);
1290
1291         if (!w->io)
1292                 bch2_writepage_io_alloc(c, w, inode, page,
1293                                         nr_replicas_this_write);
1294
1295         w->io->new_sectors += new.sectors - old.sectors;
1296
1297         BUG_ON(inode != w->io->op.inode);
1298         BUG_ON(bio_add_page_contig(&w->io->op.op.wbio.bio, page));
1299
1300         w->io->op.op.res.sectors += old.replicas_reserved * PAGE_SECTORS;
1301         w->io->op.new_i_size = i_size;
1302
1303         if (wbc->sync_mode == WB_SYNC_ALL)
1304                 w->io->op.op.wbio.bio.bi_opf |= REQ_SYNC;
1305
1306         return 0;
1307 }
1308
1309 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1310 {
1311         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1312         struct bch_writepage_state w =
1313                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1314         struct blk_plug plug;
1315         int ret;
1316
1317         blk_start_plug(&plug);
1318         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1319         if (w.io)
1320                 bch2_writepage_do_io(&w);
1321         blk_finish_plug(&plug);
1322         return ret;
1323 }
1324
1325 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1326 {
1327         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1328         struct bch_writepage_state w =
1329                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1330         int ret;
1331
1332         ret = __bch2_writepage(page, wbc, &w);
1333         if (w.io)
1334                 bch2_writepage_do_io(&w);
1335
1336         return ret;
1337 }
1338
1339 /* buffered writes: */
1340
1341 int bch2_write_begin(struct file *file, struct address_space *mapping,
1342                      loff_t pos, unsigned len, unsigned flags,
1343                      struct page **pagep, void **fsdata)
1344 {
1345         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1346         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1347         pgoff_t index = pos >> PAGE_SHIFT;
1348         unsigned offset = pos & (PAGE_SIZE - 1);
1349         struct page *page;
1350         int ret = -ENOMEM;
1351
1352         BUG_ON(inode_unhashed(&inode->v));
1353
1354         /* Not strictly necessary - same reason as mkwrite(): */
1355         pagecache_add_get(&mapping->add_lock);
1356
1357         page = grab_cache_page_write_begin(mapping, index, flags);
1358         if (!page)
1359                 goto err_unlock;
1360
1361         if (PageUptodate(page))
1362                 goto out;
1363
1364         /* If we're writing entire page, don't need to read it in first: */
1365         if (len == PAGE_SIZE)
1366                 goto out;
1367
1368         if (!offset && pos + len >= inode->v.i_size) {
1369                 zero_user_segment(page, len, PAGE_SIZE);
1370                 flush_dcache_page(page);
1371                 goto out;
1372         }
1373
1374         if (index > inode->v.i_size >> PAGE_SHIFT) {
1375                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1376                 flush_dcache_page(page);
1377                 goto out;
1378         }
1379 readpage:
1380         ret = bch2_read_single_page(page, mapping);
1381         if (ret)
1382                 goto err;
1383 out:
1384         ret = bch2_get_page_reservation(c, inode, page, true);
1385         if (ret) {
1386                 if (!PageUptodate(page)) {
1387                         /*
1388                          * If the page hasn't been read in, we won't know if we
1389                          * actually need a reservation - we don't actually need
1390                          * to read here, we just need to check if the page is
1391                          * fully backed by uncompressed data:
1392                          */
1393                         goto readpage;
1394                 }
1395
1396                 goto err;
1397         }
1398
1399         *pagep = page;
1400         return 0;
1401 err:
1402         unlock_page(page);
1403         put_page(page);
1404         *pagep = NULL;
1405 err_unlock:
1406         pagecache_add_put(&mapping->add_lock);
1407         return ret;
1408 }
1409
1410 int bch2_write_end(struct file *file, struct address_space *mapping,
1411                    loff_t pos, unsigned len, unsigned copied,
1412                    struct page *page, void *fsdata)
1413 {
1414         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1415         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1416
1417         lockdep_assert_held(&inode->v.i_rwsem);
1418
1419         if (unlikely(copied < len && !PageUptodate(page))) {
1420                 /*
1421                  * The page needs to be read in, but that would destroy
1422                  * our partial write - simplest thing is to just force
1423                  * userspace to redo the write:
1424                  */
1425                 zero_user(page, 0, PAGE_SIZE);
1426                 flush_dcache_page(page);
1427                 copied = 0;
1428         }
1429
1430         spin_lock(&inode->v.i_lock);
1431         if (pos + copied > inode->v.i_size)
1432                 i_size_write(&inode->v, pos + copied);
1433         spin_unlock(&inode->v.i_lock);
1434
1435         if (copied) {
1436                 if (!PageUptodate(page))
1437                         SetPageUptodate(page);
1438                 if (!PageDirty(page))
1439                         set_page_dirty(page);
1440
1441                 inode->ei_last_dirtied = (unsigned long) current;
1442         } else {
1443                 bch2_put_page_reservation(c, inode, page);
1444         }
1445
1446         unlock_page(page);
1447         put_page(page);
1448         pagecache_add_put(&mapping->add_lock);
1449
1450         return copied;
1451 }
1452
1453 #define WRITE_BATCH_PAGES       32
1454
1455 static int __bch2_buffered_write(struct bch_inode_info *inode,
1456                                  struct address_space *mapping,
1457                                  struct iov_iter *iter,
1458                                  loff_t pos, unsigned len)
1459 {
1460         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1461         struct page *pages[WRITE_BATCH_PAGES];
1462         unsigned long index = pos >> PAGE_SHIFT;
1463         unsigned offset = pos & (PAGE_SIZE - 1);
1464         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1465         unsigned i, copied = 0, nr_pages_copied = 0;
1466         int ret = 0;
1467
1468         BUG_ON(!len);
1469         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1470
1471         for (i = 0; i < nr_pages; i++) {
1472                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1473                 if (!pages[i]) {
1474                         nr_pages = i;
1475                         ret = -ENOMEM;
1476                         goto out;
1477                 }
1478         }
1479
1480         if (offset && !PageUptodate(pages[0])) {
1481                 ret = bch2_read_single_page(pages[0], mapping);
1482                 if (ret)
1483                         goto out;
1484         }
1485
1486         if ((pos + len) & (PAGE_SIZE - 1) &&
1487             !PageUptodate(pages[nr_pages - 1])) {
1488                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1489                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1490                 } else {
1491                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1492                         if (ret)
1493                                 goto out;
1494                 }
1495         }
1496
1497         for (i = 0; i < nr_pages; i++) {
1498                 ret = bch2_get_page_reservation(c, inode, pages[i], true);
1499
1500                 if (ret && !PageUptodate(pages[i])) {
1501                         ret = bch2_read_single_page(pages[i], mapping);
1502                         if (ret)
1503                                 goto out;
1504
1505                         ret = bch2_get_page_reservation(c, inode, pages[i], true);
1506                 }
1507
1508                 if (ret)
1509                         goto out;
1510         }
1511
1512         if (mapping_writably_mapped(mapping))
1513                 for (i = 0; i < nr_pages; i++)
1514                         flush_dcache_page(pages[i]);
1515
1516         while (copied < len) {
1517                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1518                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1519                 unsigned pg_bytes = min_t(unsigned, len - copied,
1520                                           PAGE_SIZE - pg_offset);
1521                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1522                                                 iter, pg_offset, pg_bytes);
1523
1524                 if (!pg_copied)
1525                         break;
1526
1527                 flush_dcache_page(page);
1528                 iov_iter_advance(iter, pg_copied);
1529                 copied += pg_copied;
1530         }
1531
1532         if (!copied)
1533                 goto out;
1534
1535         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1536         inode->ei_last_dirtied = (unsigned long) current;
1537
1538         spin_lock(&inode->v.i_lock);
1539         if (pos + copied > inode->v.i_size)
1540                 i_size_write(&inode->v, pos + copied);
1541         spin_unlock(&inode->v.i_lock);
1542
1543         if (copied < len &&
1544             ((offset + copied) & (PAGE_SIZE - 1))) {
1545                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1546
1547                 if (!PageUptodate(page)) {
1548                         zero_user(page, 0, PAGE_SIZE);
1549                         copied -= (offset + copied) & (PAGE_SIZE - 1);
1550                 }
1551         }
1552 out:
1553         for (i = 0; i < nr_pages_copied; i++) {
1554                 if (!PageUptodate(pages[i]))
1555                         SetPageUptodate(pages[i]);
1556                 if (!PageDirty(pages[i]))
1557                         set_page_dirty(pages[i]);
1558                 unlock_page(pages[i]);
1559                 put_page(pages[i]);
1560         }
1561
1562         for (i = nr_pages_copied; i < nr_pages; i++) {
1563                 if (!PageDirty(pages[i]))
1564                         bch2_put_page_reservation(c, inode, pages[i]);
1565                 unlock_page(pages[i]);
1566                 put_page(pages[i]);
1567         }
1568
1569         return copied ?: ret;
1570 }
1571
1572 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1573 {
1574         struct file *file = iocb->ki_filp;
1575         struct address_space *mapping = file->f_mapping;
1576         struct bch_inode_info *inode = file_bch_inode(file);
1577         loff_t pos = iocb->ki_pos;
1578         ssize_t written = 0;
1579         int ret = 0;
1580
1581         pagecache_add_get(&mapping->add_lock);
1582
1583         do {
1584                 unsigned offset = pos & (PAGE_SIZE - 1);
1585                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1586                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1587 again:
1588                 /*
1589                  * Bring in the user page that we will copy from _first_.
1590                  * Otherwise there's a nasty deadlock on copying from the
1591                  * same page as we're writing to, without it being marked
1592                  * up-to-date.
1593                  *
1594                  * Not only is this an optimisation, but it is also required
1595                  * to check that the address is actually valid, when atomic
1596                  * usercopies are used, below.
1597                  */
1598                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1599                         bytes = min_t(unsigned long, iov_iter_count(iter),
1600                                       PAGE_SIZE - offset);
1601
1602                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1603                                 ret = -EFAULT;
1604                                 break;
1605                         }
1606                 }
1607
1608                 if (unlikely(fatal_signal_pending(current))) {
1609                         ret = -EINTR;
1610                         break;
1611                 }
1612
1613                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1614                 if (unlikely(ret < 0))
1615                         break;
1616
1617                 cond_resched();
1618
1619                 if (unlikely(ret == 0)) {
1620                         /*
1621                          * If we were unable to copy any data at all, we must
1622                          * fall back to a single segment length write.
1623                          *
1624                          * If we didn't fallback here, we could livelock
1625                          * because not all segments in the iov can be copied at
1626                          * once without a pagefault.
1627                          */
1628                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1629                                       iov_iter_single_seg_count(iter));
1630                         goto again;
1631                 }
1632                 pos += ret;
1633                 written += ret;
1634
1635                 balance_dirty_pages_ratelimited(mapping);
1636         } while (iov_iter_count(iter));
1637
1638         pagecache_add_put(&mapping->add_lock);
1639
1640         return written ? written : ret;
1641 }
1642
1643 /* O_DIRECT reads */
1644
1645 static void bch2_dio_read_complete(struct closure *cl)
1646 {
1647         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1648
1649         dio->req->ki_complete(dio->req, dio->ret, 0);
1650         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1651 }
1652
1653 static void bch2_direct_IO_read_endio(struct bio *bio)
1654 {
1655         struct dio_read *dio = bio->bi_private;
1656
1657         if (bio->bi_status)
1658                 dio->ret = blk_status_to_errno(bio->bi_status);
1659
1660         closure_put(&dio->cl);
1661 }
1662
1663 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1664 {
1665         bch2_direct_IO_read_endio(bio);
1666         bio_check_pages_dirty(bio);     /* transfers ownership */
1667 }
1668
1669 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1670 {
1671         struct file *file = req->ki_filp;
1672         struct bch_inode_info *inode = file_bch_inode(file);
1673         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1674         struct bch_io_opts opts = io_opts(c, inode);
1675         struct dio_read *dio;
1676         struct bio *bio;
1677         loff_t offset = req->ki_pos;
1678         bool sync = is_sync_kiocb(req);
1679         size_t shorten;
1680         ssize_t ret;
1681
1682         if ((offset|iter->count) & (block_bytes(c) - 1))
1683                 return -EINVAL;
1684
1685         ret = min_t(loff_t, iter->count,
1686                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1687
1688         if (!ret)
1689                 return ret;
1690
1691         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1692         iter->count -= shorten;
1693
1694         bio = bio_alloc_bioset(GFP_KERNEL,
1695                                iov_iter_npages(iter, BIO_MAX_PAGES),
1696                                &c->dio_read_bioset);
1697
1698         bio->bi_end_io = bch2_direct_IO_read_endio;
1699
1700         dio = container_of(bio, struct dio_read, rbio.bio);
1701         closure_init(&dio->cl, NULL);
1702
1703         /*
1704          * this is a _really_ horrible hack just to avoid an atomic sub at the
1705          * end:
1706          */
1707         if (!sync) {
1708                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1709                 atomic_set(&dio->cl.remaining,
1710                            CLOSURE_REMAINING_INITIALIZER -
1711                            CLOSURE_RUNNING +
1712                            CLOSURE_DESTRUCTOR);
1713         } else {
1714                 atomic_set(&dio->cl.remaining,
1715                            CLOSURE_REMAINING_INITIALIZER + 1);
1716         }
1717
1718         dio->req        = req;
1719         dio->ret        = ret;
1720
1721         goto start;
1722         while (iter->count) {
1723                 bio = bio_alloc_bioset(GFP_KERNEL,
1724                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1725                                        &c->bio_read);
1726                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1727 start:
1728                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1729                 bio->bi_iter.bi_sector  = offset >> 9;
1730                 bio->bi_private         = dio;
1731
1732                 ret = bio_iov_iter_get_pages(bio, iter);
1733                 if (ret < 0) {
1734                         /* XXX: fault inject this path */
1735                         bio->bi_status = BLK_STS_RESOURCE;
1736                         bio_endio(bio);
1737                         break;
1738                 }
1739
1740                 offset += bio->bi_iter.bi_size;
1741                 bio_set_pages_dirty(bio);
1742
1743                 if (iter->count)
1744                         closure_get(&dio->cl);
1745
1746                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1747         }
1748
1749         iter->count += shorten;
1750
1751         if (sync) {
1752                 closure_sync(&dio->cl);
1753                 closure_debug_destroy(&dio->cl);
1754                 ret = dio->ret;
1755                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1756                 return ret;
1757         } else {
1758                 return -EIOCBQUEUED;
1759         }
1760 }
1761
1762 /* O_DIRECT writes */
1763
1764 static void bch2_dio_write_loop_async(struct closure *);
1765
1766 static long bch2_dio_write_loop(struct dio_write *dio)
1767 {
1768         bool kthread = (current->flags & PF_KTHREAD) != 0;
1769         struct kiocb *req = dio->req;
1770         struct address_space *mapping = req->ki_filp->f_mapping;
1771         struct bch_inode_info *inode = dio->iop.inode;
1772         struct bio *bio = &dio->iop.op.wbio.bio;
1773         struct bio_vec *bv;
1774         loff_t offset;
1775         bool sync;
1776         long ret;
1777         int i;
1778
1779         if (dio->loop)
1780                 goto loop;
1781
1782         inode_dio_begin(&inode->v);
1783         __pagecache_block_get(&mapping->add_lock);
1784
1785         /* Write and invalidate pagecache range that we're writing to: */
1786         offset = req->ki_pos + (dio->iop.op.written << 9);
1787         ret = write_invalidate_inode_pages_range(mapping,
1788                                         offset,
1789                                         offset + iov_iter_count(&dio->iter) - 1);
1790         if (unlikely(ret))
1791                 goto err;
1792
1793         while (1) {
1794                 offset = req->ki_pos + (dio->iop.op.written << 9);
1795
1796                 BUG_ON(current->pagecache_lock);
1797                 current->pagecache_lock = &mapping->add_lock;
1798                 if (kthread)
1799                         use_mm(dio->mm);
1800
1801                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1802
1803                 if (kthread)
1804                         unuse_mm(dio->mm);
1805                 current->pagecache_lock = NULL;
1806
1807                 if (unlikely(ret < 0))
1808                         goto err;
1809
1810                 /* gup might have faulted pages back in: */
1811                 ret = write_invalidate_inode_pages_range(mapping,
1812                                 offset,
1813                                 offset + bio->bi_iter.bi_size - 1);
1814                 if (unlikely(ret))
1815                         goto err;
1816
1817                 dio->iop.op.pos = POS(inode->v.i_ino, offset >> 9);
1818
1819                 task_io_account_write(bio->bi_iter.bi_size);
1820
1821                 closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl);
1822
1823                 if (!dio->sync && !dio->loop && dio->iter.count) {
1824                         struct iovec *iov = dio->inline_vecs;
1825
1826                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1827                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1828                                               GFP_KERNEL);
1829                                 if (unlikely(!iov)) {
1830                                         dio->iop.op.error = -ENOMEM;
1831                                         goto err_wait_io;
1832                                 }
1833
1834                                 dio->free_iov = true;
1835                         }
1836
1837                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1838                         dio->iter.iov = iov;
1839                 }
1840 err_wait_io:
1841                 dio->loop = true;
1842
1843                 if (!dio->sync) {
1844                         continue_at(&dio->cl, bch2_dio_write_loop_async, NULL);
1845                         return -EIOCBQUEUED;
1846                 }
1847
1848                 closure_sync(&dio->cl);
1849 loop:
1850                 bio_for_each_segment_all(bv, bio, i)
1851                         put_page(bv->bv_page);
1852                 if (!dio->iter.count || dio->iop.op.error)
1853                         break;
1854                 bio_reset(bio);
1855         }
1856
1857         ret = dio->iop.op.error ?: ((long) dio->iop.op.written << 9);
1858 err:
1859         __pagecache_block_put(&mapping->add_lock);
1860         bch2_disk_reservation_put(dio->iop.op.c, &dio->iop.op.res);
1861         bch2_quota_reservation_put(dio->iop.op.c, inode, &dio->quota_res);
1862
1863         if (dio->free_iov)
1864                 kfree(dio->iter.iov);
1865
1866         closure_debug_destroy(&dio->cl);
1867
1868         sync = dio->sync;
1869         bio_put(bio);
1870
1871         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1872         inode_dio_end(&inode->v);
1873
1874         if (!sync) {
1875                 req->ki_complete(req, ret, 0);
1876                 ret = -EIOCBQUEUED;
1877         }
1878         return ret;
1879 }
1880
1881 static void bch2_dio_write_loop_async(struct closure *cl)
1882 {
1883         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1884
1885         bch2_dio_write_loop(dio);
1886 }
1887
1888 static int bch2_direct_IO_write(struct kiocb *req,
1889                                 struct iov_iter *iter,
1890                                 bool swap)
1891 {
1892         struct file *file = req->ki_filp;
1893         struct bch_inode_info *inode = file_bch_inode(file);
1894         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1895         struct dio_write *dio;
1896         struct bio *bio;
1897         ssize_t ret;
1898
1899         lockdep_assert_held(&inode->v.i_rwsem);
1900
1901         if (unlikely(!iter->count))
1902                 return 0;
1903
1904         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1905                 return -EINVAL;
1906
1907         bio = bio_alloc_bioset(GFP_KERNEL,
1908                                iov_iter_npages(iter, BIO_MAX_PAGES),
1909                                &c->dio_write_bioset);
1910         dio = container_of(bio, struct dio_write, iop.op.wbio.bio);
1911         closure_init(&dio->cl, NULL);
1912         dio->req                = req;
1913         dio->mm                 = current->mm;
1914         dio->loop               = false;
1915         dio->sync               = is_sync_kiocb(req) ||
1916                 req->ki_pos + iter->count > inode->v.i_size;
1917         dio->free_iov           = false;
1918         dio->quota_res.sectors  = 0;
1919         dio->iter               = *iter;
1920         bch2_fswrite_op_init(&dio->iop, c, inode, io_opts(c, inode), true);
1921         dio->iop.op.write_point = writepoint_hashed((unsigned long) current);
1922         dio->iop.op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1923
1924         if ((req->ki_flags & IOCB_DSYNC) &&
1925             !c->opts.journal_flush_disabled)
1926                 dio->iop.op.flags |= BCH_WRITE_FLUSH;
1927
1928         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
1929                                          iter->count >> 9, true);
1930         if (unlikely(ret))
1931                 goto err;
1932
1933         dio->iop.op.nr_replicas = dio->iop.op.opts.data_replicas;
1934
1935         ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9,
1936                                         dio->iop.op.opts.data_replicas, 0);
1937         if (unlikely(ret)) {
1938                 if (!bch2_check_range_allocated(c, POS(inode->v.i_ino,
1939                                                        req->ki_pos >> 9),
1940                                                 iter->count >> 9,
1941                                                 dio->iop.op.opts.data_replicas))
1942                         goto err;
1943
1944                 dio->iop.unalloc = true;
1945         }
1946
1947         return bch2_dio_write_loop(dio);
1948 err:
1949         bch2_disk_reservation_put(c, &dio->iop.op.res);
1950         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1951         closure_debug_destroy(&dio->cl);
1952         bio_put(bio);
1953         return ret;
1954 }
1955
1956 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1957 {
1958         struct blk_plug plug;
1959         ssize_t ret;
1960
1961         blk_start_plug(&plug);
1962         ret = iov_iter_rw(iter) == WRITE
1963                 ? bch2_direct_IO_write(req, iter, false)
1964                 : bch2_direct_IO_read(req, iter);
1965         blk_finish_plug(&plug);
1966
1967         return ret;
1968 }
1969
1970 static ssize_t
1971 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1972 {
1973         return bch2_direct_IO_write(iocb, iter, true);
1974 }
1975
1976 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1977 {
1978         struct file *file = iocb->ki_filp;
1979         struct bch_inode_info *inode = file_bch_inode(file);
1980         ssize_t ret;
1981
1982         /* We can write back this queue in page reclaim */
1983         current->backing_dev_info = inode_to_bdi(&inode->v);
1984         ret = file_remove_privs(file);
1985         if (ret)
1986                 goto out;
1987
1988         ret = file_update_time(file);
1989         if (ret)
1990                 goto out;
1991
1992         ret = iocb->ki_flags & IOCB_DIRECT
1993                 ? bch2_direct_write(iocb, from)
1994                 : bch2_buffered_write(iocb, from);
1995
1996         if (likely(ret > 0))
1997                 iocb->ki_pos += ret;
1998 out:
1999         current->backing_dev_info = NULL;
2000         return ret;
2001 }
2002
2003 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2004 {
2005         struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp);
2006         bool direct = iocb->ki_flags & IOCB_DIRECT;
2007         ssize_t ret;
2008
2009         inode_lock(&inode->v);
2010         ret = generic_write_checks(iocb, from);
2011         if (ret > 0)
2012                 ret = __bch2_write_iter(iocb, from);
2013         inode_unlock(&inode->v);
2014
2015         if (ret > 0 && !direct)
2016                 ret = generic_write_sync(iocb, ret);
2017
2018         return ret;
2019 }
2020
2021 /* fsync: */
2022
2023 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2024 {
2025         struct bch_inode_info *inode = file_bch_inode(file);
2026         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2027         int ret, ret2;
2028
2029         ret = file_write_and_wait_range(file, start, end);
2030         if (ret)
2031                 return ret;
2032
2033         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2034                 goto out;
2035
2036         ret = sync_inode_metadata(&inode->v, 1);
2037         if (ret)
2038                 return ret;
2039 out:
2040         if (c->opts.journal_flush_disabled)
2041                 return 0;
2042
2043         ret = bch2_journal_flush_seq(&c->journal, inode->ei_journal_seq);
2044         ret2 = file_check_and_advance_wb_err(file);
2045
2046         return ret ?: ret2;
2047 }
2048
2049 /* truncate: */
2050
2051 static int __bch2_fpunch(struct bch_fs *c, struct bch_inode_info *inode,
2052                          u64 start_offset, u64 end_offset, u64 *journal_seq)
2053 {
2054         struct bpos start       = POS(inode->v.i_ino, start_offset);
2055         struct bpos end         = POS(inode->v.i_ino, end_offset);
2056         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
2057         struct btree_trans trans;
2058         struct btree_iter *iter;
2059         struct bkey_s_c k;
2060         int ret = 0;
2061
2062         bch2_trans_init(&trans, c);
2063         bch2_trans_preload_iters(&trans);
2064
2065         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, start,
2066                                    BTREE_ITER_INTENT);
2067
2068         while ((k = bch2_btree_iter_peek(iter)).k &&
2069                !(ret = btree_iter_err(k)) &&
2070                bkey_cmp(iter->pos, end) < 0) {
2071                 struct disk_reservation disk_res =
2072                         bch2_disk_reservation_init(c, 0);
2073                 struct bkey_i delete;
2074
2075                 bkey_init(&delete.k);
2076                 delete.k.p = iter->pos;
2077
2078                 /* create the biggest key we can */
2079                 bch2_key_resize(&delete.k, max_sectors);
2080                 bch2_cut_back(end, &delete.k);
2081
2082                 ret = bch2_extent_update(&trans, inode,
2083                                 &disk_res, NULL, iter, &delete,
2084                                 0, true, true, NULL);
2085                 bch2_disk_reservation_put(c, &disk_res);
2086
2087                 if (ret == -EINTR)
2088                         ret = 0;
2089                 if (ret)
2090                         break;
2091
2092                 bch2_btree_iter_cond_resched(iter);
2093         }
2094
2095         bch2_trans_exit(&trans);
2096
2097         return ret;
2098 }
2099
2100 static inline int range_has_data(struct bch_fs *c,
2101                                   struct bpos start,
2102                                   struct bpos end)
2103 {
2104
2105         struct btree_iter iter;
2106         struct bkey_s_c k;
2107         int ret = 0;
2108
2109         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2110                            start, 0, k) {
2111                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2112                         break;
2113
2114                 if (bkey_extent_is_data(k.k)) {
2115                         ret = 1;
2116                         break;
2117                 }
2118         }
2119
2120         return bch2_btree_iter_unlock(&iter) ?: ret;
2121 }
2122
2123 static int __bch2_truncate_page(struct bch_inode_info *inode,
2124                                 pgoff_t index, loff_t start, loff_t end)
2125 {
2126         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2127         struct address_space *mapping = inode->v.i_mapping;
2128         unsigned start_offset = start & (PAGE_SIZE - 1);
2129         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2130         struct page *page;
2131         int ret = 0;
2132
2133         /* Page boundary? Nothing to do */
2134         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2135               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2136                 return 0;
2137
2138         /* Above i_size? */
2139         if (index << PAGE_SHIFT >= inode->v.i_size)
2140                 return 0;
2141
2142         page = find_lock_page(mapping, index);
2143         if (!page) {
2144                 /*
2145                  * XXX: we're doing two index lookups when we end up reading the
2146                  * page
2147                  */
2148                 ret = range_has_data(c,
2149                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2150                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2151                 if (ret <= 0)
2152                         return ret;
2153
2154                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2155                 if (unlikely(!page)) {
2156                         ret = -ENOMEM;
2157                         goto out;
2158                 }
2159         }
2160
2161         if (!PageUptodate(page)) {
2162                 ret = bch2_read_single_page(page, mapping);
2163                 if (ret)
2164                         goto unlock;
2165         }
2166
2167         /*
2168          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2169          *
2170          * XXX: because we aren't currently tracking whether the page has actual
2171          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2172          */
2173         ret = bch2_get_page_reservation(c, inode, page, false);
2174         BUG_ON(ret);
2175
2176         if (index == start >> PAGE_SHIFT &&
2177             index == end >> PAGE_SHIFT)
2178                 zero_user_segment(page, start_offset, end_offset);
2179         else if (index == start >> PAGE_SHIFT)
2180                 zero_user_segment(page, start_offset, PAGE_SIZE);
2181         else if (index == end >> PAGE_SHIFT)
2182                 zero_user_segment(page, 0, end_offset);
2183
2184         if (!PageDirty(page))
2185                 set_page_dirty(page);
2186 unlock:
2187         unlock_page(page);
2188         put_page(page);
2189 out:
2190         return ret;
2191 }
2192
2193 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2194 {
2195         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2196                                     from, from + PAGE_SIZE);
2197 }
2198
2199 static int bch2_extend(struct bch_inode_info *inode, struct iattr *iattr)
2200 {
2201         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2202         struct address_space *mapping = inode->v.i_mapping;
2203         int ret;
2204
2205         ret = filemap_write_and_wait_range(mapping,
2206                         inode->ei_inode.bi_size, S64_MAX);
2207         if (ret)
2208                 return ret;
2209
2210         truncate_setsize(&inode->v, iattr->ia_size);
2211         setattr_copy(&inode->v, iattr);
2212
2213         mutex_lock(&inode->ei_update_lock);
2214         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2215                                     ATTR_MTIME|ATTR_CTIME);
2216         mutex_unlock(&inode->ei_update_lock);
2217
2218         return ret;
2219 }
2220
2221 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2222                                    struct bch_inode_unpacked *bi,
2223                                    void *p)
2224 {
2225         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2226
2227         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2228         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2229         return 0;
2230 }
2231
2232 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2233                                   struct bch_inode_unpacked *bi, void *p)
2234 {
2235         u64 *new_i_size = p;
2236
2237         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2238         bi->bi_size = *new_i_size;
2239         return 0;
2240 }
2241
2242 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2243 {
2244         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2245         struct address_space *mapping = inode->v.i_mapping;
2246         u64 new_i_size = iattr->ia_size;
2247         bool shrink;
2248         int ret = 0;
2249
2250         inode_dio_wait(&inode->v);
2251         pagecache_block_get(&mapping->add_lock);
2252
2253         BUG_ON(inode->v.i_size < inode->ei_inode.bi_size);
2254
2255         shrink = iattr->ia_size <= inode->v.i_size;
2256
2257         if (!shrink) {
2258                 ret = bch2_extend(inode, iattr);
2259                 goto err;
2260         }
2261
2262         ret = bch2_truncate_page(inode, iattr->ia_size);
2263         if (unlikely(ret))
2264                 goto err;
2265
2266         if (iattr->ia_size > inode->ei_inode.bi_size)
2267                 ret = filemap_write_and_wait_range(mapping,
2268                                 inode->ei_inode.bi_size,
2269                                 iattr->ia_size - 1);
2270         else if (iattr->ia_size & (PAGE_SIZE - 1))
2271                 ret = filemap_write_and_wait_range(mapping,
2272                                 round_down(iattr->ia_size, PAGE_SIZE),
2273                                 iattr->ia_size - 1);
2274         if (ret)
2275                 goto err;
2276
2277         mutex_lock(&inode->ei_update_lock);
2278         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2279                                &new_i_size, 0);
2280         mutex_unlock(&inode->ei_update_lock);
2281
2282         if (unlikely(ret))
2283                 goto err;
2284
2285         truncate_setsize(&inode->v, iattr->ia_size);
2286
2287         /*
2288          * XXX: need a comment explaining why PAGE_SIZE and not block_bytes()
2289          * here:
2290          */
2291         ret = __bch2_fpunch(c, inode,
2292                         round_up(iattr->ia_size, PAGE_SIZE) >> 9,
2293                         U64_MAX, &inode->ei_journal_seq);
2294         if (unlikely(ret))
2295                 goto err;
2296
2297         setattr_copy(&inode->v, iattr);
2298
2299         mutex_lock(&inode->ei_update_lock);
2300         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2301                                ATTR_MTIME|ATTR_CTIME);
2302         mutex_unlock(&inode->ei_update_lock);
2303 err:
2304         pagecache_block_put(&mapping->add_lock);
2305         return ret;
2306 }
2307
2308 /* fallocate: */
2309
2310 static long bch2_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2311 {
2312         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2313         struct address_space *mapping = inode->v.i_mapping;
2314         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
2315         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
2316         int ret = 0;
2317
2318         inode_lock(&inode->v);
2319         inode_dio_wait(&inode->v);
2320         pagecache_block_get(&mapping->add_lock);
2321
2322         ret = __bch2_truncate_page(inode,
2323                                    offset >> PAGE_SHIFT,
2324                                    offset, offset + len);
2325         if (unlikely(ret))
2326                 goto err;
2327
2328         if (offset >> PAGE_SHIFT !=
2329             (offset + len) >> PAGE_SHIFT) {
2330                 ret = __bch2_truncate_page(inode,
2331                                            (offset + len) >> PAGE_SHIFT,
2332                                            offset, offset + len);
2333                 if (unlikely(ret))
2334                         goto err;
2335         }
2336
2337         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2338
2339         if (discard_start < discard_end)
2340                 ret = __bch2_fpunch(c, inode, discard_start, discard_end,
2341                                     &inode->ei_journal_seq);
2342 err:
2343         pagecache_block_put(&mapping->add_lock);
2344         inode_unlock(&inode->v);
2345
2346         return ret;
2347 }
2348
2349 static long bch2_fcollapse(struct bch_inode_info *inode,
2350                            loff_t offset, loff_t len)
2351 {
2352         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2353         struct address_space *mapping = inode->v.i_mapping;
2354         struct btree_trans trans;
2355         struct btree_iter *src, *dst;
2356         BKEY_PADDED(k) copy;
2357         struct bkey_s_c k;
2358         loff_t new_size;
2359         int ret;
2360
2361         if ((offset | len) & (block_bytes(c) - 1))
2362                 return -EINVAL;
2363
2364         bch2_trans_init(&trans, c);
2365         bch2_trans_preload_iters(&trans);
2366
2367         /*
2368          * We need i_mutex to keep the page cache consistent with the extents
2369          * btree, and the btree consistent with i_size - we don't need outside
2370          * locking for the extents btree itself, because we're using linked
2371          * iterators
2372          */
2373         inode_lock(&inode->v);
2374         inode_dio_wait(&inode->v);
2375         pagecache_block_get(&mapping->add_lock);
2376
2377         ret = -EINVAL;
2378         if (offset + len >= inode->v.i_size)
2379                 goto err;
2380
2381         if (inode->v.i_size < len)
2382                 goto err;
2383
2384         new_size = inode->v.i_size - len;
2385
2386         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2387         if (ret)
2388                 goto err;
2389
2390         dst = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2391                         POS(inode->v.i_ino, offset >> 9),
2392                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2393         BUG_ON(IS_ERR_OR_NULL(dst));
2394
2395         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2396                         POS_MIN, BTREE_ITER_SLOTS);
2397         BUG_ON(IS_ERR_OR_NULL(src));
2398
2399         while (bkey_cmp(dst->pos,
2400                         POS(inode->v.i_ino,
2401                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2402                 struct disk_reservation disk_res;
2403
2404                 ret = bch2_btree_iter_traverse(dst);
2405                 if (ret)
2406                         goto btree_iter_err;
2407
2408                 bch2_btree_iter_set_pos(src,
2409                         POS(dst->pos.inode, dst->pos.offset + (len >> 9)));
2410
2411                 k = bch2_btree_iter_peek_slot(src);
2412                 if ((ret = btree_iter_err(k)))
2413                         goto btree_iter_err;
2414
2415                 bkey_reassemble(&copy.k, k);
2416
2417                 bch2_cut_front(src->pos, &copy.k);
2418                 copy.k.k.p.offset -= len >> 9;
2419
2420                 bch2_extent_trim_atomic(&copy.k, dst);
2421
2422                 BUG_ON(bkey_cmp(dst->pos, bkey_start_pos(&copy.k.k)));
2423
2424                 ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size,
2425                                 bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&copy.k)),
2426                                 BCH_DISK_RESERVATION_NOFAIL);
2427                 BUG_ON(ret);
2428
2429                 ret = bch2_extent_update(&trans, inode,
2430                                 &disk_res, NULL,
2431                                 dst, &copy.k,
2432                                 0, true, true, NULL);
2433                 bch2_disk_reservation_put(c, &disk_res);
2434 btree_iter_err:
2435                 if (ret == -EINTR)
2436                         ret = 0;
2437                 if (ret)
2438                         goto err;
2439                 /*
2440                  * XXX: if we error here we've left data with multiple
2441                  * pointers... which isn't a _super_ serious problem...
2442                  */
2443
2444                 bch2_btree_iter_cond_resched(src);
2445         }
2446         bch2_trans_unlock(&trans);
2447
2448         ret = __bch2_fpunch(c, inode,
2449                         round_up(new_size, block_bytes(c)) >> 9,
2450                         U64_MAX, &inode->ei_journal_seq);
2451         if (ret)
2452                 goto err;
2453
2454         i_size_write(&inode->v, new_size);
2455         mutex_lock(&inode->ei_update_lock);
2456         ret = bch2_write_inode_size(c, inode, new_size,
2457                                     ATTR_MTIME|ATTR_CTIME);
2458         mutex_unlock(&inode->ei_update_lock);
2459 err:
2460         bch2_trans_exit(&trans);
2461         pagecache_block_put(&mapping->add_lock);
2462         inode_unlock(&inode->v);
2463         return ret;
2464 }
2465
2466 static long bch2_fallocate(struct bch_inode_info *inode, int mode,
2467                            loff_t offset, loff_t len)
2468 {
2469         struct address_space *mapping = inode->v.i_mapping;
2470         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2471         struct btree_trans trans;
2472         struct btree_iter *iter;
2473         struct bpos end_pos;
2474         loff_t block_start, block_end;
2475         loff_t end = offset + len;
2476         unsigned sectors;
2477         unsigned replicas = io_opts(c, inode).data_replicas;
2478         int ret;
2479
2480         bch2_trans_init(&trans, c);
2481         bch2_trans_preload_iters(&trans);
2482
2483         inode_lock(&inode->v);
2484         inode_dio_wait(&inode->v);
2485         pagecache_block_get(&mapping->add_lock);
2486
2487         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2488                 ret = inode_newsize_ok(&inode->v, end);
2489                 if (ret)
2490                         goto err;
2491         }
2492
2493         if (mode & FALLOC_FL_ZERO_RANGE) {
2494                 ret = __bch2_truncate_page(inode,
2495                                            offset >> PAGE_SHIFT,
2496                                            offset, end);
2497
2498                 if (!ret &&
2499                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2500                         ret = __bch2_truncate_page(inode,
2501                                                    end >> PAGE_SHIFT,
2502                                                    offset, end);
2503
2504                 if (unlikely(ret))
2505                         goto err;
2506
2507                 truncate_pagecache_range(&inode->v, offset, end - 1);
2508
2509                 block_start     = round_up(offset, PAGE_SIZE);
2510                 block_end       = round_down(end, PAGE_SIZE);
2511         } else {
2512                 block_start     = round_down(offset, PAGE_SIZE);
2513                 block_end       = round_up(end, PAGE_SIZE);
2514         }
2515
2516         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2517                         POS(inode->v.i_ino, block_start >> 9),
2518                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2519         end_pos = POS(inode->v.i_ino, block_end >> 9);
2520
2521         while (bkey_cmp(iter->pos, end_pos) < 0) {
2522                 struct disk_reservation disk_res = { 0 };
2523                 struct quota_res quota_res = { 0 };
2524                 struct bkey_i_reservation reservation;
2525                 struct bkey_s_c k;
2526
2527                 k = bch2_btree_iter_peek_slot(iter);
2528                 if ((ret = btree_iter_err(k)))
2529                         goto btree_iter_err;
2530
2531                 /* already reserved */
2532                 if (k.k->type == KEY_TYPE_reservation &&
2533                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2534                         bch2_btree_iter_next_slot(iter);
2535                         continue;
2536                 }
2537
2538                 if (bkey_extent_is_data(k.k) &&
2539                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2540                         bch2_btree_iter_next_slot(iter);
2541                         continue;
2542                 }
2543
2544                 bkey_reservation_init(&reservation.k_i);
2545                 reservation.k.type      = KEY_TYPE_reservation;
2546                 reservation.k.p         = k.k->p;
2547                 reservation.k.size      = k.k->size;
2548
2549                 bch2_cut_front(iter->pos, &reservation.k_i);
2550                 bch2_cut_back(end_pos, &reservation.k);
2551
2552                 sectors = reservation.k.size;
2553                 reservation.v.nr_replicas = bch2_bkey_nr_dirty_ptrs(k);
2554
2555                 if (!bkey_extent_is_allocation(k.k)) {
2556                         ret = bch2_quota_reservation_add(c, inode,
2557                                         &quota_res,
2558                                         sectors, true);
2559                         if (unlikely(ret))
2560                                 goto btree_iter_err;
2561                 }
2562
2563                 if (reservation.v.nr_replicas < replicas ||
2564                     bch2_extent_is_compressed(k)) {
2565                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2566                                                         replicas, 0);
2567                         if (unlikely(ret))
2568                                 goto btree_iter_err;
2569
2570                         reservation.v.nr_replicas = disk_res.nr_replicas;
2571                 }
2572
2573                 ret = bch2_extent_update(&trans, inode,
2574                                 &disk_res, &quota_res,
2575                                 iter, &reservation.k_i,
2576                                 0, true, true, NULL);
2577 btree_iter_err:
2578                 bch2_quota_reservation_put(c, inode, &quota_res);
2579                 bch2_disk_reservation_put(c, &disk_res);
2580                 if (ret == -EINTR)
2581                         ret = 0;
2582                 if (ret)
2583                         goto err;
2584         }
2585         bch2_trans_unlock(&trans);
2586
2587         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2588             end > inode->v.i_size) {
2589                 i_size_write(&inode->v, end);
2590
2591                 mutex_lock(&inode->ei_update_lock);
2592                 ret = bch2_write_inode_size(c, inode, inode->v.i_size, 0);
2593                 mutex_unlock(&inode->ei_update_lock);
2594         }
2595
2596         /* blech */
2597         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2598             (mode & FALLOC_FL_ZERO_RANGE) &&
2599             inode->ei_inode.bi_size != inode->v.i_size) {
2600                 /* sync appends.. */
2601                 ret = filemap_write_and_wait_range(mapping,
2602                                         inode->ei_inode.bi_size, S64_MAX);
2603                 if (ret)
2604                         goto err;
2605
2606                 if (inode->ei_inode.bi_size != inode->v.i_size) {
2607                         mutex_lock(&inode->ei_update_lock);
2608                         ret = bch2_write_inode_size(c, inode,
2609                                                     inode->v.i_size, 0);
2610                         mutex_unlock(&inode->ei_update_lock);
2611                 }
2612         }
2613 err:
2614         bch2_trans_exit(&trans);
2615         pagecache_block_put(&mapping->add_lock);
2616         inode_unlock(&inode->v);
2617         return ret;
2618 }
2619
2620 long bch2_fallocate_dispatch(struct file *file, int mode,
2621                              loff_t offset, loff_t len)
2622 {
2623         struct bch_inode_info *inode = file_bch_inode(file);
2624
2625         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2626                 return bch2_fallocate(inode, mode, offset, len);
2627
2628         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2629                 return bch2_fpunch(inode, offset, len);
2630
2631         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2632                 return bch2_fcollapse(inode, offset, len);
2633
2634         return -EOPNOTSUPP;
2635 }
2636
2637 /* fseek: */
2638
2639 static bool page_is_data(struct page *page)
2640 {
2641         EBUG_ON(!PageLocked(page));
2642
2643         /* XXX: should only have to check PageDirty */
2644         return PagePrivate(page) &&
2645                 (page_state(page)->sectors ||
2646                  page_state(page)->dirty_sectors);
2647 }
2648
2649 static loff_t bch2_next_pagecache_data(struct inode *vinode,
2650                                        loff_t start_offset,
2651                                        loff_t end_offset)
2652 {
2653         struct address_space *mapping = vinode->i_mapping;
2654         struct page *page;
2655         pgoff_t index;
2656
2657         for (index = start_offset >> PAGE_SHIFT;
2658              index < end_offset >> PAGE_SHIFT;
2659              index++) {
2660                 if (find_get_pages(mapping, &index, 1, &page)) {
2661                         lock_page(page);
2662
2663                         if (page_is_data(page))
2664                                 end_offset =
2665                                         min(end_offset,
2666                                         max(start_offset,
2667                                             ((loff_t) index) << PAGE_SHIFT));
2668                         unlock_page(page);
2669                         put_page(page);
2670                 } else {
2671                         break;
2672                 }
2673         }
2674
2675         return end_offset;
2676 }
2677
2678 static loff_t bch2_seek_data(struct file *file, u64 offset)
2679 {
2680         struct bch_inode_info *inode = file_bch_inode(file);
2681         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2682         struct btree_iter iter;
2683         struct bkey_s_c k;
2684         u64 isize, next_data = MAX_LFS_FILESIZE;
2685         int ret;
2686
2687         isize = i_size_read(&inode->v);
2688         if (offset >= isize)
2689                 return -ENXIO;
2690
2691         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2692                            POS(inode->v.i_ino, offset >> 9), 0, k) {
2693                 if (k.k->p.inode != inode->v.i_ino) {
2694                         break;
2695                 } else if (bkey_extent_is_data(k.k)) {
2696                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2697                         break;
2698                 } else if (k.k->p.offset >> 9 > isize)
2699                         break;
2700         }
2701
2702         ret = bch2_btree_iter_unlock(&iter);
2703         if (ret)
2704                 return ret;
2705
2706         if (next_data > offset)
2707                 next_data = bch2_next_pagecache_data(&inode->v,
2708                                                      offset, next_data);
2709
2710         if (next_data > isize)
2711                 return -ENXIO;
2712
2713         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2714 }
2715
2716 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2717 {
2718         struct page *page;
2719         bool ret;
2720
2721         page = find_lock_entry(mapping, index);
2722         if (!page || xa_is_value(page))
2723                 return false;
2724
2725         ret = page_is_data(page);
2726         unlock_page(page);
2727
2728         return ret;
2729 }
2730
2731 static loff_t bch2_next_pagecache_hole(struct inode *vinode,
2732                                        loff_t start_offset,
2733                                        loff_t end_offset)
2734 {
2735         struct address_space *mapping = vinode->i_mapping;
2736         pgoff_t index;
2737
2738         for (index = start_offset >> PAGE_SHIFT;
2739              index < end_offset >> PAGE_SHIFT;
2740              index++)
2741                 if (!page_slot_is_data(mapping, index))
2742                         end_offset = max(start_offset,
2743                                          ((loff_t) index) << PAGE_SHIFT);
2744
2745         return end_offset;
2746 }
2747
2748 static loff_t bch2_seek_hole(struct file *file, u64 offset)
2749 {
2750         struct bch_inode_info *inode = file_bch_inode(file);
2751         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2752         struct btree_iter iter;
2753         struct bkey_s_c k;
2754         u64 isize, next_hole = MAX_LFS_FILESIZE;
2755         int ret;
2756
2757         isize = i_size_read(&inode->v);
2758         if (offset >= isize)
2759                 return -ENXIO;
2760
2761         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2762                            POS(inode->v.i_ino, offset >> 9),
2763                            BTREE_ITER_SLOTS, k) {
2764                 if (k.k->p.inode != inode->v.i_ino) {
2765                         next_hole = bch2_next_pagecache_hole(&inode->v,
2766                                         offset, MAX_LFS_FILESIZE);
2767                         break;
2768                 } else if (!bkey_extent_is_data(k.k)) {
2769                         next_hole = bch2_next_pagecache_hole(&inode->v,
2770                                         max(offset, bkey_start_offset(k.k) << 9),
2771                                         k.k->p.offset << 9);
2772
2773                         if (next_hole < k.k->p.offset << 9)
2774                                 break;
2775                 } else {
2776                         offset = max(offset, bkey_start_offset(k.k) << 9);
2777                 }
2778         }
2779
2780         ret = bch2_btree_iter_unlock(&iter);
2781         if (ret)
2782                 return ret;
2783
2784         if (next_hole > isize)
2785                 next_hole = isize;
2786
2787         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2788 }
2789
2790 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
2791 {
2792         switch (whence) {
2793         case SEEK_SET:
2794         case SEEK_CUR:
2795         case SEEK_END:
2796                 return generic_file_llseek(file, offset, whence);
2797         case SEEK_DATA:
2798                 return bch2_seek_data(file, offset);
2799         case SEEK_HOLE:
2800                 return bch2_seek_hole(file, offset);
2801         }
2802
2803         return -EINVAL;
2804 }
2805
2806 void bch2_fs_fsio_exit(struct bch_fs *c)
2807 {
2808         bioset_exit(&c->dio_write_bioset);
2809         bioset_exit(&c->dio_read_bioset);
2810         bioset_exit(&c->writepage_bioset);
2811 }
2812
2813 int bch2_fs_fsio_init(struct bch_fs *c)
2814 {
2815         int ret = 0;
2816
2817         pr_verbose_init(c->opts, "");
2818
2819         if (bioset_init(&c->writepage_bioset,
2820                         4, offsetof(struct bch_writepage_io, op.op.wbio.bio),
2821                         BIOSET_NEED_BVECS) ||
2822             bioset_init(&c->dio_read_bioset,
2823                         4, offsetof(struct dio_read, rbio.bio),
2824                         BIOSET_NEED_BVECS) ||
2825             bioset_init(&c->dio_write_bioset,
2826                         4, offsetof(struct dio_write, iop.op.wbio.bio),
2827                         BIOSET_NEED_BVECS))
2828                 ret = -ENOMEM;
2829
2830         pr_verbose_init(c->opts, "ret %i", ret);
2831         return ret;
2832 }
2833
2834 #endif /* NO_BCACHEFS_FS */