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

   1 #ifndef NO_BCACHEFS_FS
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "buckets.h"
   6 #include "clock.h"
   7 #include "error.h"
   8 #include "fs.h"
   9 #include "fs-io.h"
  10 #include "fsck.h"
  11 #include "inode.h"
  12 #include "journal.h"
  13 #include "io.h"
  14 #include "keylist.h"
  15 #include "quota.h"
  16
  17 #include <linux/aio.h>
  18 #include <linux/backing-dev.h>
  19 #include <linux/falloc.h>
  20 #include <linux/migrate.h>
  21 #include <linux/mmu_context.h>
  22 #include <linux/pagevec.h>
  23 #include <linux/task_io_accounting_ops.h>
  24 #include <linux/uio.h>
  25 #include <linux/writeback.h>
  26
  27 #include <trace/events/bcachefs.h>
  28 #include <trace/events/writeback.h>
  29
  30 struct quota_res {
  31         u64                             sectors;
  32 };
  33
  34 struct i_sectors_hook {
  35         struct extent_insert_hook       hook;
  36         struct bch_inode_info           *inode;
  37         struct quota_res                quota_res;
  38         s64                             sectors;
  39         u64                             new_i_size;
  40         unsigned                        flags;
  41         unsigned                        appending:1;
  42 };
  43
  44 struct bchfs_write_op {
  45         struct bch_inode_info           *inode;
  46         s64                             sectors_added;
  47         bool                            is_dio;
  48         bool                            unalloc;
  49         u64                             new_i_size;
  50
  51         /* must be last: */
  52         struct bch_write_op             op;
  53 };
  54
  55 struct bch_writepage_io {
  56         struct closure                  cl;
  57         u64                             new_sectors;
  58
  59         /* must be last: */
  60         struct bchfs_write_op           op;
  61 };
  62
  63 struct dio_write {
  64         struct closure                  cl;
  65         struct kiocb                    *req;
  66         struct task_struct              *task;
  67         unsigned                        loop:1,
  68                                         sync:1,
  69                                         free_iov:1;
  70         struct quota_res                quota_res;
  71
  72         struct iov_iter                 iter;
  73         struct iovec                    inline_vecs[2];
  74
  75         /* must be last: */
  76         struct bchfs_write_op           iop;
  77 };
  78
  79 struct dio_read {
  80         struct closure                  cl;
  81         struct kiocb                    *req;
  82         long                            ret;
  83         struct bch_read_bio             rbio;
  84 };
  85
  86 /* pagecache_block must be held */
  87 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  88                                               loff_t start, loff_t end)
  89 {
  90         int ret;
  91
  92         /*
  93          * XXX: the way this is currently implemented, we can spin if a process
  94          * is continually redirtying a specific page
  95          */
  96         do {
  97                 if (!mapping->nrpages &&
  98                     !mapping->nrexceptional)
  99                         return 0;
 100
 101                 ret = filemap_write_and_wait_range(mapping, start, end);
 102                 if (ret)
 103                         break;
 104
 105                 if (!mapping->nrpages)
 106                         return 0;
 107
 108                 ret = invalidate_inode_pages2_range(mapping,
 109                                 start >> PAGE_SHIFT,
 110                                 end >> PAGE_SHIFT);
 111         } while (ret == -EBUSY);
 112
 113         return ret;
 114 }
 115
 116 /* quotas */
 117
 118 #ifdef CONFIG_BCACHEFS_QUOTA
 119
 120 static void bch2_quota_reservation_put(struct bch_fs *c,
 121                                        struct bch_inode_info *inode,
 122                                        struct quota_res *res)
 123 {
 124         if (!res->sectors)
 125                 return;
 126
 127         mutex_lock(&inode->ei_update_lock);
 128         BUG_ON(res->sectors > inode->ei_quota_reserved);
 129
 130         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 131                         -((s64) res->sectors), BCH_QUOTA_PREALLOC);
 132         inode->ei_quota_reserved -= res->sectors;
 133         mutex_unlock(&inode->ei_update_lock);
 134
 135         res->sectors = 0;
 136 }
 137
 138 static int bch2_quota_reservation_add(struct bch_fs *c,
 139                                       struct bch_inode_info *inode,
 140                                       struct quota_res *res,
 141                                       unsigned sectors,
 142                                       bool check_enospc)
 143 {
 144         int ret;
 145
 146         mutex_lock(&inode->ei_update_lock);
 147         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 148                               check_enospc ? BCH_QUOTA_PREALLOC : BCH_QUOTA_NOCHECK);
 149         if (likely(!ret)) {
 150                 inode->ei_quota_reserved += sectors;
 151                 res->sectors += sectors;
 152         }
 153         mutex_unlock(&inode->ei_update_lock);
 154
 155         return ret;
 156 }
 157
 158 #else
 159
 160 static void bch2_quota_reservation_put(struct bch_fs *c,
 161                                        struct bch_inode_info *inode,
 162                                        struct quota_res *res)
 163 {
 164 }
 165
 166 static int bch2_quota_reservation_add(struct bch_fs *c,
 167                                       struct bch_inode_info *inode,
 168                                       struct quota_res *res,
 169                                       unsigned sectors,
 170                                       bool check_enospc)
 171 {
 172         return 0;
 173 }
 174
 175 #endif
 176
 177 /* i_size updates: */
 178
 179 static int inode_set_size(struct bch_inode_info *inode,
 180                           struct bch_inode_unpacked *bi,
 181                           void *p)
 182 {
 183         loff_t *new_i_size = p;
 184
 185         lockdep_assert_held(&inode->ei_update_lock);
 186
 187         bi->bi_size = *new_i_size;
 188         return 0;
 189 }
 190
 191 static int __must_check bch2_write_inode_size(struct bch_fs *c,
 192                                               struct bch_inode_info *inode,
 193                                               loff_t new_size)
 194 {
 195         return __bch2_write_inode(c, inode, inode_set_size, &new_size);
 196 }
 197
 198 static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 199                              struct quota_res *quota_res, int sectors)
 200 {
 201 #ifdef CONFIG_BCACHEFS_QUOTA
 202         if (quota_res && sectors > 0) {
 203                 BUG_ON(sectors > quota_res->sectors);
 204                 BUG_ON(sectors > inode->ei_quota_reserved);
 205
 206                 quota_res->sectors -= sectors;
 207                 inode->ei_quota_reserved -= sectors;
 208         } else {
 209                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, BCH_QUOTA_WARN);
 210         }
 211 #endif
 212         inode->v.i_blocks += sectors;
 213 }
 214
 215 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 216                            struct quota_res *quota_res, int sectors)
 217 {
 218         mutex_lock(&inode->ei_update_lock);
 219         __i_sectors_acct(c, inode, quota_res, sectors);
 220         mutex_unlock(&inode->ei_update_lock);
 221 }
 222
 223 /* i_sectors accounting: */
 224
 225 static enum btree_insert_ret
 226 i_sectors_hook_fn(struct extent_insert_hook *hook,
 227                   struct bpos committed_pos,
 228                   struct bpos next_pos,
 229                   struct bkey_s_c k,
 230                   const struct bkey_i *insert)
 231 {
 232         struct i_sectors_hook *h = container_of(hook,
 233                                 struct i_sectors_hook, hook);
 234         s64 sectors = next_pos.offset - committed_pos.offset;
 235         int sign = bkey_extent_is_allocation(&insert->k) -
 236                 (k.k && bkey_extent_is_allocation(k.k));
 237
 238         EBUG_ON(!(h->inode->ei_inode.bi_flags & BCH_INODE_I_SECTORS_DIRTY));
 239
 240         h->sectors += sectors * sign;
 241
 242         return BTREE_INSERT_OK;
 243 }
 244
 245 static int i_sectors_dirty_finish_fn(struct bch_inode_info *inode,
 246                                      struct bch_inode_unpacked *bi,
 247                                      void *p)
 248 {
 249         struct i_sectors_hook *h = p;
 250
 251         if (h->new_i_size != U64_MAX &&
 252             (!h->appending ||
 253              h->new_i_size > bi->bi_size))
 254                 bi->bi_size = h->new_i_size;
 255         bi->bi_sectors  += h->sectors;
 256         bi->bi_flags    &= ~h->flags;
 257         return 0;
 258 }
 259
 260 static int i_sectors_dirty_finish(struct bch_fs *c, struct i_sectors_hook *h)
 261 {
 262         int ret;
 263
 264         mutex_lock(&h->inode->ei_update_lock);
 265         if (h->new_i_size != U64_MAX)
 266                 i_size_write(&h->inode->v, h->new_i_size);
 267
 268         __i_sectors_acct(c, h->inode, &h->quota_res, h->sectors);
 269
 270         ret = __bch2_write_inode(c, h->inode, i_sectors_dirty_finish_fn, h);
 271         mutex_unlock(&h->inode->ei_update_lock);
 272
 273         bch2_quota_reservation_put(c, h->inode, &h->quota_res);
 274
 275         h->sectors = 0;
 276
 277         return ret;
 278 }
 279
 280 static int i_sectors_dirty_start_fn(struct bch_inode_info *inode,
 281                                     struct bch_inode_unpacked *bi, void *p)
 282 {
 283         struct i_sectors_hook *h = p;
 284
 285         if (h->flags & BCH_INODE_I_SIZE_DIRTY)
 286                 bi->bi_size = h->new_i_size;
 287
 288         bi->bi_flags |= h->flags;
 289         return 0;
 290 }
 291
 292 static int i_sectors_dirty_start(struct bch_fs *c, struct i_sectors_hook *h)
 293 {
 294         int ret;
 295
 296         mutex_lock(&h->inode->ei_update_lock);
 297         ret = __bch2_write_inode(c, h->inode, i_sectors_dirty_start_fn, h);
 298         mutex_unlock(&h->inode->ei_update_lock);
 299
 300         return ret;
 301 }
 302
 303 static inline struct i_sectors_hook
 304 i_sectors_hook_init(struct bch_inode_info *inode, unsigned flags)
 305 {
 306         return (struct i_sectors_hook) {
 307                 .hook.fn        = i_sectors_hook_fn,
 308                 .inode          = inode,
 309                 .sectors        = 0,
 310                 .new_i_size     = U64_MAX,
 311                 .flags          = flags|BCH_INODE_I_SECTORS_DIRTY,
 312         };
 313 }
 314
 315 /* normal i_size/i_sectors update machinery: */
 316
 317 struct bchfs_extent_trans_hook {
 318         struct bchfs_write_op           *op;
 319         struct extent_insert_hook       hook;
 320
 321         struct bch_inode_unpacked       inode_u;
 322         struct bkey_inode_buf           inode_p;
 323
 324         bool                            need_inode_update;
 325 };
 326
 327 static enum btree_insert_ret
 328 bchfs_extent_update_hook(struct extent_insert_hook *hook,
 329                          struct bpos committed_pos,
 330                          struct bpos next_pos,
 331                          struct bkey_s_c k,
 332                          const struct bkey_i *insert)
 333 {
 334         struct bchfs_extent_trans_hook *h = container_of(hook,
 335                                 struct bchfs_extent_trans_hook, hook);
 336         struct bch_inode_info *inode = h->op->inode;
 337         int sign = bkey_extent_is_allocation(&insert->k) -
 338                 (k.k && bkey_extent_is_allocation(k.k));
 339         s64 sectors = (s64) (next_pos.offset - committed_pos.offset) * sign;
 340         u64 offset = min(next_pos.offset << 9, h->op->new_i_size);
 341         bool do_pack = false;
 342
 343         if (h->op->unalloc &&
 344             !bch2_extent_is_fully_allocated(k))
 345                 return BTREE_INSERT_ENOSPC;
 346
 347         BUG_ON((next_pos.offset << 9) > round_up(offset, PAGE_SIZE));
 348
 349         /* XXX: inode->i_size locking */
 350         if (offset > inode->ei_inode.bi_size) {
 351                 if (!h->need_inode_update) {
 352                         h->need_inode_update = true;
 353                         return BTREE_INSERT_NEED_TRAVERSE;
 354                 }
 355
 356                 BUG_ON(h->inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY);
 357
 358                 h->inode_u.bi_size = offset;
 359                 do_pack = true;
 360
 361                 inode->ei_inode.bi_size = offset;
 362
 363                 if (h->op->is_dio)
 364                         i_size_write(&inode->v, offset);
 365         }
 366
 367         if (sectors) {
 368                 if (!h->need_inode_update) {
 369                         h->need_inode_update = true;
 370                         return BTREE_INSERT_NEED_TRAVERSE;
 371                 }
 372
 373                 h->inode_u.bi_sectors += sectors;
 374                 do_pack = true;
 375
 376                 h->op->sectors_added += sectors;
 377         }
 378
 379         if (do_pack)
 380                 bch2_inode_pack(&h->inode_p, &h->inode_u);
 381
 382         return BTREE_INSERT_OK;
 383 }
 384
 385 static int bchfs_write_index_update(struct bch_write_op *wop)
 386 {
 387         struct bchfs_write_op *op = container_of(wop,
 388                                 struct bchfs_write_op, op);
 389         struct keylist *keys = &op->op.insert_keys;
 390         struct btree_iter extent_iter, inode_iter;
 391         struct bchfs_extent_trans_hook hook;
 392         struct bkey_i *k = bch2_keylist_front(keys);
 393         s64 orig_sectors_added = op->sectors_added;
 394         int ret;
 395
 396         BUG_ON(k->k.p.inode != op->inode->v.i_ino);
 397
 398         bch2_btree_iter_init(&extent_iter, wop->c, BTREE_ID_EXTENTS,
 399                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 400                              BTREE_ITER_INTENT);
 401         bch2_btree_iter_init(&inode_iter, wop->c, BTREE_ID_INODES,
 402                              POS(extent_iter.pos.inode, 0),
 403                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 404
 405         hook.op                 = op;
 406         hook.hook.fn            = bchfs_extent_update_hook;
 407         hook.need_inode_update  = false;
 408
 409         do {
 410                 /* XXX: inode->i_size locking */
 411                 k = bch2_keylist_front(keys);
 412                 if (min(k->k.p.offset << 9, op->new_i_size) >
 413                     op->inode->ei_inode.bi_size)
 414                         hook.need_inode_update = true;
 415
 416                 if (hook.need_inode_update) {
 417                         struct bkey_s_c inode;
 418
 419                         if (!btree_iter_linked(&inode_iter))
 420                                 bch2_btree_iter_link(&extent_iter, &inode_iter);
 421
 422                         inode = bch2_btree_iter_peek_slot(&inode_iter);
 423                         if ((ret = btree_iter_err(inode)))
 424                                 goto err;
 425
 426                         if (WARN_ONCE(inode.k->type != BCH_INODE_FS,
 427                                       "inode %llu not found when updating",
 428                                       extent_iter.pos.inode)) {
 429                                 ret = -ENOENT;
 430                                 break;
 431                         }
 432
 433                         if (WARN_ONCE(bkey_bytes(inode.k) >
 434                                       sizeof(hook.inode_p),
 435                                       "inode %llu too big (%zu bytes, buf %zu)",
 436                                       extent_iter.pos.inode,
 437                                       bkey_bytes(inode.k),
 438                                       sizeof(hook.inode_p))) {
 439                                 ret = -ENOENT;
 440                                 break;
 441                         }
 442
 443                         bkey_reassemble(&hook.inode_p.inode.k_i, inode);
 444                         ret = bch2_inode_unpack(bkey_s_c_to_inode(inode),
 445                                                &hook.inode_u);
 446                         if (WARN_ONCE(ret,
 447                                       "error %i unpacking inode %llu",
 448                                       ret, extent_iter.pos.inode)) {
 449                                 ret = -ENOENT;
 450                                 break;
 451                         }
 452
 453                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 454                                         &hook.hook, op_journal_seq(wop),
 455                                         BTREE_INSERT_NOFAIL|
 456                                         BTREE_INSERT_ATOMIC|
 457                                         BTREE_INSERT_USE_RESERVE,
 458                                         BTREE_INSERT_ENTRY(&extent_iter, k),
 459                                         BTREE_INSERT_ENTRY_EXTRA_RES(&inode_iter,
 460                                                         &hook.inode_p.inode.k_i, 2));
 461                 } else {
 462                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 463                                         &hook.hook, op_journal_seq(wop),
 464                                         BTREE_INSERT_NOFAIL|
 465                                         BTREE_INSERT_ATOMIC|
 466                                         BTREE_INSERT_USE_RESERVE,
 467                                         BTREE_INSERT_ENTRY(&extent_iter, k));
 468                 }
 469
 470                 BUG_ON(bkey_cmp(extent_iter.pos, bkey_start_pos(&k->k)));
 471                 BUG_ON(!ret != !k->k.size);
 472 err:
 473                 if (ret == -EINTR)
 474                         continue;
 475                 if (ret)
 476                         break;
 477
 478                 BUG_ON(bkey_cmp(extent_iter.pos, k->k.p) < 0);
 479                 bch2_keylist_pop_front(keys);
 480         } while (!bch2_keylist_empty(keys));
 481
 482         bch2_btree_iter_unlock(&extent_iter);
 483         bch2_btree_iter_unlock(&inode_iter);
 484
 485         if (op->is_dio) {
 486                 struct dio_write *dio = container_of(op, struct dio_write, iop);
 487
 488                 i_sectors_acct(wop->c, op->inode, &dio->quota_res,
 489                                op->sectors_added - orig_sectors_added);
 490         }
 491
 492         return ret;
 493 }
 494
 495 static inline void bch2_fswrite_op_init(struct bchfs_write_op *op,
 496                                         struct bch_fs *c,
 497                                         struct bch_inode_info *inode,
 498                                         struct bch_io_opts opts,
 499                                         bool is_dio)
 500 {
 501         op->inode               = inode;
 502         op->sectors_added       = 0;
 503         op->is_dio              = is_dio;
 504         op->unalloc             = false;
 505         op->new_i_size          = U64_MAX;
 506
 507         bch2_write_op_init(&op->op, c);
 508         op->op.csum_type        = bch2_data_checksum_type(c, opts.data_checksum);
 509         op->op.compression_type = bch2_compression_opt_to_type[opts.compression];
 510         op->op.devs             = c->fastest_devs;
 511         op->op.index_update_fn  = bchfs_write_index_update;
 512         op_journal_seq_set(&op->op, &inode->ei_journal_seq);
 513 }
 514
 515 static inline struct bch_io_opts io_opts(struct bch_fs *c, struct bch_inode_info *inode)
 516 {
 517         struct bch_io_opts opts = bch2_opts_to_inode_opts(c->opts);
 518
 519         bch2_io_opts_apply(&opts, bch2_inode_opts_get(&inode->ei_inode));
 520         return opts;
 521 }
 522
 523 /* page state: */
 524
 525 /* stored in page->private: */
 526
 527 /*
 528  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 529  * almost protected it with the page lock, except that bch2_writepage_io_done has
 530  * to update the sector counts (and from interrupt/bottom half context).
 531  */
 532 struct bch_page_state {
 533 union { struct {
 534         /* existing data: */
 535         unsigned                sectors:PAGE_SECTOR_SHIFT + 1;
 536         unsigned                nr_replicas:4;
 537         unsigned                compressed:1;
 538
 539         /* Owns PAGE_SECTORS sized reservation: */
 540         unsigned                reserved:1;
 541         unsigned                reservation_replicas:4;
 542
 543         /* Owns PAGE_SECTORS sized quota reservation: */
 544         unsigned                quota_reserved:1;
 545
 546         /*
 547          * Number of sectors on disk - for i_blocks
 548          * Uncompressed size, not compressed size:
 549          */
 550         unsigned                dirty_sectors:PAGE_SECTOR_SHIFT + 1;
 551 };
 552         /* for cmpxchg: */
 553         unsigned long           v;
 554 };
 555 };
 556
 557 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 558 ({                                                                      \
 559         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 560         struct bch_page_state _old;                                     \
 561                                                                         \
 562         do {                                                            \
 563                 _old.v = _new.v = _v;                                   \
 564                 _expr;                                                  \
 565                                                                         \
 566                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 567         } while (_old.v != _new.v &&                                    \
 568                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 569                                                                         \
 570         _old;                                                           \
 571 })
 572
 573 static inline struct bch_page_state *page_state(struct page *page)
 574 {
 575         struct bch_page_state *s = (void *) &page->private;
 576
 577         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 578
 579         if (!PagePrivate(page))
 580                 SetPagePrivate(page);
 581
 582         return s;
 583 }
 584
 585 static inline unsigned page_res_sectors(struct bch_page_state s)
 586 {
 587
 588         return s.reserved ? s.reservation_replicas * PAGE_SECTORS : 0;
 589 }
 590
 591 static void __bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 592                                         struct bch_page_state s)
 593 {
 594         struct disk_reservation res = { .sectors = page_res_sectors(s) };
 595         struct quota_res quota_res = { .sectors = s.quota_reserved ? PAGE_SECTORS : 0 };
 596
 597         bch2_quota_reservation_put(c, inode, &quota_res);
 598         bch2_disk_reservation_put(c, &res);
 599 }
 600
 601 static void bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 602                                       struct page *page)
 603 {
 604         struct bch_page_state s;
 605
 606         s = page_state_cmpxchg(page_state(page), s, {
 607                 s.reserved              = 0;
 608                 s.quota_reserved        = 0;
 609         });
 610
 611         __bch2_put_page_reservation(c, inode, s);
 612 }
 613
 614 static int bch2_get_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 615                                      struct page *page, bool check_enospc)
 616 {
 617         struct bch_page_state *s = page_state(page), new, old;
 618         struct disk_reservation disk_res = bch2_disk_reservation_init(c,
 619                                                 READ_ONCE(c->opts.data_replicas));
 620         struct quota_res quota_res = { 0 };
 621         int ret = 0;
 622
 623         /*
 624          * XXX: this could likely be quite a bit simpler, page reservations
 625          * _should_ only be manipulated with page locked:
 626          */
 627
 628         old = page_state_cmpxchg(s, new, {
 629                 if (new.reserved
 630                     ? (new.reservation_replicas < disk_res.nr_replicas)
 631                     : (new.sectors < PAGE_SECTORS ||
 632                        new.nr_replicas < disk_res.nr_replicas ||
 633                        new.compressed)) {
 634                         int sectors = (disk_res.nr_replicas * PAGE_SECTORS -
 635                                        page_res_sectors(new) -
 636                                        disk_res.sectors);
 637
 638                         if (sectors > 0) {
 639                                 ret = bch2_disk_reservation_add(c, &disk_res, sectors,
 640                                                 !check_enospc
 641                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 642                                 if (unlikely(ret))
 643                                         goto err;
 644                         }
 645
 646                         new.reserved = 1;
 647                         new.reservation_replicas = disk_res.nr_replicas;
 648                 }
 649
 650                 if (!new.quota_reserved &&
 651                     new.sectors + new.dirty_sectors < PAGE_SECTORS) {
 652                         ret = bch2_quota_reservation_add(c, inode, &quota_res,
 653                                                 PAGE_SECTORS - quota_res.sectors,
 654                                                 check_enospc);
 655                         if (unlikely(ret))
 656                                 goto err;
 657
 658                         new.quota_reserved = 1;
 659                 }
 660         });
 661
 662         quota_res.sectors -= (new.quota_reserved - old.quota_reserved) * PAGE_SECTORS;
 663         disk_res.sectors -= page_res_sectors(new) - page_res_sectors(old);
 664 err:
 665         bch2_quota_reservation_put(c, inode, &quota_res);
 666         bch2_disk_reservation_put(c, &disk_res);
 667         return ret;
 668 }
 669
 670 static void bch2_clear_page_bits(struct page *page)
 671 {
 672         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 673         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 674         struct bch_page_state s;
 675
 676         if (!PagePrivate(page))
 677                 return;
 678
 679         s = xchg(page_state(page), (struct bch_page_state) { .v = 0 });
 680         ClearPagePrivate(page);
 681
 682         if (s.dirty_sectors)
 683                 i_sectors_acct(c, inode, NULL, -s.dirty_sectors);
 684
 685         __bch2_put_page_reservation(c, inode, s);
 686 }
 687
 688 int bch2_set_page_dirty(struct page *page)
 689 {
 690         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 691         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 692         struct quota_res quota_res = { 0 };
 693         struct bch_page_state old, new;
 694
 695         old = page_state_cmpxchg(page_state(page), new,
 696                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 697                 new.quota_reserved = 0;
 698         );
 699
 700         quota_res.sectors += old.quota_reserved * PAGE_SECTORS;
 701
 702         if (old.dirty_sectors != new.dirty_sectors)
 703                 i_sectors_acct(c, inode, &quota_res,
 704                                new.dirty_sectors - old.dirty_sectors);
 705         bch2_quota_reservation_put(c, inode, &quota_res);
 706
 707         return __set_page_dirty_nobuffers(page);
 708 }
 709
 710 int bch2_page_mkwrite(struct vm_fault *vmf)
 711 {
 712         struct page *page = vmf->page;
 713         struct file *file = vmf->vma->vm_file;
 714         struct bch_inode_info *inode = file_bch_inode(file);
 715         struct address_space *mapping = inode->v.i_mapping;
 716         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 717         int ret = VM_FAULT_LOCKED;
 718
 719         sb_start_pagefault(inode->v.i_sb);
 720         file_update_time(file);
 721
 722         /*
 723          * Not strictly necessary, but helps avoid dio writes livelocking in
 724          * write_invalidate_inode_pages_range() - can drop this if/when we get
 725          * a write_invalidate_inode_pages_range() that works without dropping
 726          * page lock before invalidating page
 727          */
 728         if (current->pagecache_lock != &mapping->add_lock)
 729                 pagecache_add_get(&mapping->add_lock);
 730
 731         lock_page(page);
 732         if (page->mapping != mapping ||
 733             page_offset(page) > i_size_read(&inode->v)) {
 734                 unlock_page(page);
 735                 ret = VM_FAULT_NOPAGE;
 736                 goto out;
 737         }
 738
 739         if (bch2_get_page_reservation(c, inode, page, true)) {
 740                 unlock_page(page);
 741                 ret = VM_FAULT_SIGBUS;
 742                 goto out;
 743         }
 744
 745         if (!PageDirty(page))
 746                 set_page_dirty(page);
 747         wait_for_stable_page(page);
 748 out:
 749         if (current->pagecache_lock != &mapping->add_lock)
 750                 pagecache_add_put(&mapping->add_lock);
 751         sb_end_pagefault(inode->v.i_sb);
 752         return ret;
 753 }
 754
 755 void bch2_invalidatepage(struct page *page, unsigned int offset,
 756                          unsigned int length)
 757 {
 758         EBUG_ON(!PageLocked(page));
 759         EBUG_ON(PageWriteback(page));
 760
 761         if (offset || length < PAGE_SIZE)
 762                 return;
 763
 764         bch2_clear_page_bits(page);
 765 }
 766
 767 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 768 {
 769         EBUG_ON(!PageLocked(page));
 770         EBUG_ON(PageWriteback(page));
 771
 772         if (PageDirty(page))
 773                 return 0;
 774
 775         bch2_clear_page_bits(page);
 776         return 1;
 777 }
 778
 779 #ifdef CONFIG_MIGRATION
 780 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 781                       struct page *page, enum migrate_mode mode)
 782 {
 783         int ret;
 784
 785         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 786         if (ret != MIGRATEPAGE_SUCCESS)
 787                 return ret;
 788
 789         if (PagePrivate(page)) {
 790                 *page_state(newpage) = *page_state(page);
 791                 ClearPagePrivate(page);
 792         }
 793
 794         migrate_page_copy(newpage, page);
 795         return MIGRATEPAGE_SUCCESS;
 796 }
 797 #endif
 798
 799 /* readpages/writepages: */
 800
 801 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 802 {
 803         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 804
 805         return bio->bi_vcnt < bio->bi_max_vecs &&
 806                 bio_end_sector(bio) == offset;
 807 }
 808
 809 static void __bio_add_page(struct bio *bio, struct page *page)
 810 {
 811         bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
 812                 .bv_page = page,
 813                 .bv_len = PAGE_SIZE,
 814                 .bv_offset = 0,
 815         };
 816
 817         bio->bi_iter.bi_size += PAGE_SIZE;
 818 }
 819
 820 static int bio_add_page_contig(struct bio *bio, struct page *page)
 821 {
 822         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 823
 824         EBUG_ON(!bio->bi_max_vecs);
 825
 826         if (!bio->bi_vcnt)
 827                 bio->bi_iter.bi_sector = offset;
 828         else if (!bio_can_add_page_contig(bio, page))
 829                 return -1;
 830
 831         __bio_add_page(bio, page);
 832         return 0;
 833 }
 834
 835 /* readpage(s): */
 836
 837 static void bch2_readpages_end_io(struct bio *bio)
 838 {
 839         struct bio_vec *bv;
 840         int i;
 841
 842         bio_for_each_segment_all(bv, bio, i) {
 843                 struct page *page = bv->bv_page;
 844
 845                 if (!bio->bi_status) {
 846                         SetPageUptodate(page);
 847                 } else {
 848                         ClearPageUptodate(page);
 849                         SetPageError(page);
 850                 }
 851                 unlock_page(page);
 852         }
 853
 854         bio_put(bio);
 855 }
 856
 857 struct readpages_iter {
 858         struct address_space    *mapping;
 859         struct list_head        pages;
 860         unsigned                nr_pages;
 861 };
 862
 863 static inline void page_state_init_for_read(struct page *page)
 864 {
 865         struct bch_page_state *s = page_state(page);
 866
 867         BUG_ON(s->reserved);
 868         s->sectors      = 0;
 869         s->compressed   = 0;
 870 }
 871
 872 static int readpage_add_page(struct readpages_iter *iter, struct page *page)
 873 {
 874         int ret;
 875
 876         prefetchw(&page->flags);
 877         page_state_init_for_read(page);
 878
 879         ret = add_to_page_cache_lru(page, iter->mapping,
 880                                     page->index, GFP_NOFS);
 881         put_page(page);
 882         return ret;
 883 }
 884
 885 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 886 {
 887         while (iter->nr_pages) {
 888                 struct page *page =
 889                         list_last_entry(&iter->pages, struct page, lru);
 890
 891                 prefetchw(&page->flags);
 892                 list_del(&page->lru);
 893                 iter->nr_pages--;
 894
 895                 if (!readpage_add_page(iter, page))
 896                         return page;
 897         }
 898
 899         return NULL;
 900 }
 901
 902 #define for_each_readpage_page(_iter, _page)                            \
 903         for (;                                                          \
 904              ((_page) = __readpage_next_page(&(_iter)));)               \
 905
 906 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 907 {
 908         struct bvec_iter iter;
 909         struct bio_vec bv;
 910         bool compressed = bch2_extent_is_compressed(k);
 911         unsigned nr_ptrs = bch2_extent_nr_dirty_ptrs(k);
 912
 913         bio_for_each_segment(bv, bio, iter) {
 914                 struct bch_page_state *s = page_state(bv.bv_page);
 915
 916                 /* sectors in @k from the start of this page: */
 917                 unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset);
 918
 919                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 920
 921                 s->nr_replicas = !s->sectors
 922                         ? nr_ptrs
 923                         : min_t(unsigned, s->nr_replicas, nr_ptrs);
 924
 925                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 926                 s->sectors += page_sectors;
 927
 928                 s->compressed |= compressed;
 929         }
 930 }
 931
 932 static void readpage_bio_extend(struct readpages_iter *iter,
 933                                 struct bio *bio, u64 offset,
 934                                 bool get_more)
 935 {
 936         struct page *page;
 937         pgoff_t page_offset;
 938         int ret;
 939
 940         while (bio_end_sector(bio) < offset &&
 941                bio->bi_vcnt < bio->bi_max_vecs) {
 942                 page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 943
 944                 if (iter->nr_pages) {
 945                         page = list_last_entry(&iter->pages, struct page, lru);
 946                         if (page->index != page_offset)
 947                                 break;
 948
 949                         list_del(&page->lru);
 950                         iter->nr_pages--;
 951                 } else if (get_more) {
 952                         rcu_read_lock();
 953                         page = radix_tree_lookup(&iter->mapping->page_tree, page_offset);
 954                         rcu_read_unlock();
 955
 956                         if (page && !radix_tree_exceptional_entry(page))
 957                                 break;
 958
 959                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 960                         if (!page)
 961                                 break;
 962
 963                         page->index = page_offset;
 964                         ClearPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 965                 } else {
 966                         break;
 967                 }
 968
 969                 ret = readpage_add_page(iter, page);
 970                 if (ret)
 971                         break;
 972
 973                 __bio_add_page(bio, page);
 974         }
 975
 976         if (!iter->nr_pages)
 977                 SetPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 978 }
 979
 980 static void bchfs_read(struct bch_fs *c, struct btree_iter *iter,
 981                        struct bch_read_bio *rbio, u64 inum,
 982                        struct readpages_iter *readpages_iter)
 983 {
 984         struct bio *bio = &rbio->bio;
 985         int flags = BCH_READ_RETRY_IF_STALE|
 986                 BCH_READ_MAY_PROMOTE;
 987
 988         while (1) {
 989                 struct extent_pick_ptr pick;
 990                 BKEY_PADDED(k) tmp;
 991                 struct bkey_s_c k;
 992                 unsigned bytes;
 993                 bool is_last;
 994
 995                 bch2_btree_iter_set_pos(iter, POS(inum, bio->bi_iter.bi_sector));
 996
 997                 k = bch2_btree_iter_peek_slot(iter);
 998                 BUG_ON(!k.k);
 999
1000                 if (IS_ERR(k.k)) {
1001                         int ret = bch2_btree_iter_unlock(iter);
1002                         BUG_ON(!ret);
1003                         bcache_io_error(c, bio, "btree IO error %i", ret);
1004                         bio_endio(bio);
1005                         return;
1006                 }
1007
1008                 bkey_reassemble(&tmp.k, k);
1009                 bch2_btree_iter_unlock(iter);
1010                 k = bkey_i_to_s_c(&tmp.k);
1011
1012                 bch2_extent_pick_ptr(c, k, NULL, &pick);
1013                 if (IS_ERR(pick.ca)) {
1014                         bcache_io_error(c, bio, "no device to read from");
1015                         bio_endio(bio);
1016                         return;
1017                 }
1018
1019                 if (readpages_iter)
1020                         readpage_bio_extend(readpages_iter,
1021                                             bio, k.k->p.offset,
1022                                             pick.ca &&
1023                                             (pick.crc.csum_type ||
1024                                              pick.crc.compression_type));
1025
1026                 bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
1027                          bio->bi_iter.bi_sector) << 9;
1028                 is_last = bytes == bio->bi_iter.bi_size;
1029                 swap(bio->bi_iter.bi_size, bytes);
1030
1031                 if (bkey_extent_is_allocation(k.k))
1032                         bch2_add_page_sectors(bio, k);
1033
1034                 if (pick.ca) {
1035                         if (!is_last) {
1036                                 bio_inc_remaining(&rbio->bio);
1037                                 flags |= BCH_READ_MUST_CLONE;
1038                                 trace_read_split(&rbio->bio);
1039                         }
1040
1041                         bch2_read_extent(c, rbio, bkey_s_c_to_extent(k),
1042                                          &pick, flags);
1043                 } else {
1044                         zero_fill_bio(bio);
1045
1046                         if (is_last)
1047                                 bio_endio(bio);
1048                 }
1049
1050                 if (is_last)
1051                         return;
1052
1053                 swap(bio->bi_iter.bi_size, bytes);
1054                 bio_advance(bio, bytes);
1055         }
1056 }
1057
1058 int bch2_readpages(struct file *file, struct address_space *mapping,
1059                    struct list_head *pages, unsigned nr_pages)
1060 {
1061         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1062         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1063         struct bch_io_opts opts = io_opts(c, inode);
1064         struct btree_iter iter;
1065         struct page *page;
1066         struct readpages_iter readpages_iter = {
1067                 .mapping = mapping, .nr_pages = nr_pages
1068         };
1069
1070         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1071                              BTREE_ITER_SLOTS);
1072
1073         INIT_LIST_HEAD(&readpages_iter.pages);
1074         list_add(&readpages_iter.pages, pages);
1075         list_del_init(pages);
1076
1077         if (current->pagecache_lock != &mapping->add_lock)
1078                 pagecache_add_get(&mapping->add_lock);
1079
1080         while ((page = readpage_iter_next(&readpages_iter))) {
1081                 unsigned n = max_t(unsigned,
1082                                    min_t(unsigned, readpages_iter.nr_pages + 1,
1083                                          BIO_MAX_PAGES),
1084                                    c->sb.encoded_extent_max >> PAGE_SECTOR_SHIFT);
1085
1086                 struct bch_read_bio *rbio =
1087                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
1088                                   opts);
1089
1090                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1091                 bio_add_page_contig(&rbio->bio, page);
1092                 bchfs_read(c, &iter, rbio, inode->v.i_ino, &readpages_iter);
1093         }
1094
1095         if (current->pagecache_lock != &mapping->add_lock)
1096                 pagecache_add_put(&mapping->add_lock);
1097
1098         return 0;
1099 }
1100
1101 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1102                              u64 inum, struct page *page)
1103 {
1104         struct btree_iter iter;
1105
1106         page_state_init_for_read(page);
1107
1108         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1109         bio_add_page_contig(&rbio->bio, page);
1110
1111         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1112                              BTREE_ITER_SLOTS);
1113         bchfs_read(c, &iter, rbio, inum, NULL);
1114 }
1115
1116 int bch2_readpage(struct file *file, struct page *page)
1117 {
1118         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1119         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1120         struct bch_io_opts opts = io_opts(c, inode);
1121         struct bch_read_bio *rbio;
1122
1123         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
1124         rbio->bio.bi_end_io = bch2_readpages_end_io;
1125
1126         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1127         return 0;
1128 }
1129
1130 static void bch2_read_single_page_end_io(struct bio *bio)
1131 {
1132         complete(bio->bi_private);
1133 }
1134
1135 static int bch2_read_single_page(struct page *page,
1136                                  struct address_space *mapping)
1137 {
1138         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1139         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1140         struct bch_read_bio *rbio;
1141         int ret;
1142         DECLARE_COMPLETION_ONSTACK(done);
1143
1144         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1145                          io_opts(c, inode));
1146         rbio->bio.bi_private = &done;
1147         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1148
1149         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1150         wait_for_completion(&done);
1151
1152         ret = blk_status_to_errno(rbio->bio.bi_status);
1153         bio_put(&rbio->bio);
1154
1155         if (ret < 0)
1156                 return ret;
1157
1158         SetPageUptodate(page);
1159         return 0;
1160 }
1161
1162 /* writepages: */
1163
1164 struct bch_writepage_state {
1165         struct bch_writepage_io *io;
1166         struct bch_io_opts      opts;
1167 };
1168
1169 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1170                                                                   struct bch_inode_info *inode)
1171 {
1172         return (struct bch_writepage_state) { .opts = io_opts(c, inode) };
1173 }
1174
1175 static void bch2_writepage_io_free(struct closure *cl)
1176 {
1177         struct bch_writepage_io *io = container_of(cl,
1178                                         struct bch_writepage_io, cl);
1179
1180         bio_put(&io->op.op.wbio.bio);
1181 }
1182
1183 static void bch2_writepage_io_done(struct closure *cl)
1184 {
1185         struct bch_writepage_io *io = container_of(cl,
1186                                         struct bch_writepage_io, cl);
1187         struct bch_fs *c = io->op.op.c;
1188         struct bio *bio = &io->op.op.wbio.bio;
1189         struct bio_vec *bvec;
1190         unsigned i;
1191
1192         atomic_sub(bio->bi_vcnt, &c->writeback_pages);
1193         wake_up(&c->writeback_wait);
1194
1195         if (io->op.op.error) {
1196                 bio_for_each_segment_all(bvec, bio, i)
1197                         SetPageError(bvec->bv_page);
1198                 set_bit(AS_EIO, &io->op.inode->v.i_mapping->flags);
1199         }
1200
1201         /*
1202          * racing with fallocate can cause us to add fewer sectors than
1203          * expected - but we shouldn't add more sectors than expected:
1204          */
1205         BUG_ON(io->op.sectors_added > (s64) io->new_sectors);
1206
1207         /*
1208          * (error (due to going RO) halfway through a page can screw that up
1209          * slightly)
1210          * XXX wtf?
1211            BUG_ON(io->op.sectors_added - io->new_sectors >= (s64) PAGE_SECTORS);
1212          */
1213
1214         /*
1215          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1216          * before calling end_page_writeback:
1217          */
1218         if (io->op.sectors_added != io->new_sectors)
1219                 i_sectors_acct(c, io->op.inode, NULL,
1220                                io->op.sectors_added - (s64) io->new_sectors);
1221
1222         bio_for_each_segment_all(bvec, bio, i)
1223                 end_page_writeback(bvec->bv_page);
1224
1225         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1226 }
1227
1228 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1229 {
1230         struct bch_writepage_io *io = w->io;
1231         struct bio *bio = &io->op.op.wbio.bio;
1232
1233         w->io = NULL;
1234         atomic_add(bio->bi_vcnt, &io->op.op.c->writeback_pages);
1235
1236         closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl);
1237         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1238 }
1239
1240 /*
1241  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1242  * possible, else allocating a new one:
1243  */
1244 static void bch2_writepage_io_alloc(struct bch_fs *c,
1245                                     struct bch_writepage_state *w,
1246                                     struct bch_inode_info *inode,
1247                                     struct page *page,
1248                                     unsigned nr_replicas)
1249 {
1250         struct bch_write_op *op;
1251         u64 offset = (u64) page->index << PAGE_SECTOR_SHIFT;
1252
1253         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1254                                               BIO_MAX_PAGES,
1255                                               &c->writepage_bioset),
1256                              struct bch_writepage_io, op.op.wbio.bio);
1257
1258         closure_init(&w->io->cl, NULL);
1259         w->io->new_sectors      = 0;
1260         bch2_fswrite_op_init(&w->io->op, c, inode, w->opts, false);
1261         op                      = &w->io->op.op;
1262         op->nr_replicas         = nr_replicas;
1263         op->res.nr_replicas     = nr_replicas;
1264         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1265         op->pos                 = POS(inode->v.i_ino, offset);
1266         op->wbio.bio.bi_iter.bi_sector = offset;
1267 }
1268
1269 static int __bch2_writepage(struct bch_fs *c, struct page *page,
1270                             struct writeback_control *wbc,
1271                             struct bch_writepage_state *w)
1272 {
1273         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1274         struct bch_page_state new, old;
1275         unsigned offset;
1276         loff_t i_size = i_size_read(&inode->v);
1277         pgoff_t end_index = i_size >> PAGE_SHIFT;
1278
1279         EBUG_ON(!PageUptodate(page));
1280
1281         /* Is the page fully inside i_size? */
1282         if (page->index < end_index)
1283                 goto do_io;
1284
1285         /* Is the page fully outside i_size? (truncate in progress) */
1286         offset = i_size & (PAGE_SIZE - 1);
1287         if (page->index > end_index || !offset) {
1288                 unlock_page(page);
1289                 return 0;
1290         }
1291
1292         /*
1293          * The page straddles i_size.  It must be zeroed out on each and every
1294          * writepage invocation because it may be mmapped.  "A file is mapped
1295          * in multiples of the page size.  For a file that is not a multiple of
1296          * the  page size, the remaining memory is zeroed when mapped, and
1297          * writes to that region are not written out to the file."
1298          */
1299         zero_user_segment(page, offset, PAGE_SIZE);
1300 do_io:
1301         /* Before unlocking the page, transfer reservation to w->io: */
1302         old = page_state_cmpxchg(page_state(page), new, {
1303                 EBUG_ON(!new.reserved &&
1304                         (new.sectors != PAGE_SECTORS ||
1305                         new.compressed));
1306
1307                 if (new.reserved)
1308                         new.nr_replicas = new.reservation_replicas;
1309                 new.reserved = 0;
1310
1311                 new.compressed |= w->opts.compression != 0;
1312
1313                 new.sectors += new.dirty_sectors;
1314                 new.dirty_sectors = 0;
1315         });
1316
1317         if (w->io &&
1318             (w->io->op.op.res.nr_replicas != new.nr_replicas ||
1319              !bio_can_add_page_contig(&w->io->op.op.wbio.bio, page)))
1320                 bch2_writepage_do_io(w);
1321
1322         if (!w->io)
1323                 bch2_writepage_io_alloc(c, w, inode, page, new.nr_replicas);
1324
1325         w->io->new_sectors += new.sectors - old.sectors;
1326
1327         BUG_ON(inode != w->io->op.inode);
1328         BUG_ON(bio_add_page_contig(&w->io->op.op.wbio.bio, page));
1329
1330         if (old.reserved)
1331                 w->io->op.op.res.sectors += old.reservation_replicas * PAGE_SECTORS;
1332
1333         /* while page is locked: */
1334         w->io->op.new_i_size = i_size;
1335
1336         if (wbc->sync_mode == WB_SYNC_ALL)
1337                 w->io->op.op.wbio.bio.bi_opf |= REQ_SYNC;
1338
1339         BUG_ON(PageWriteback(page));
1340         set_page_writeback(page);
1341         unlock_page(page);
1342
1343         return 0;
1344 }
1345
1346 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1347 {
1348         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1349         struct bch_writepage_state w =
1350                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1351         struct pagecache_iter iter;
1352         struct page *page;
1353         int ret = 0;
1354         int done = 0;
1355         pgoff_t uninitialized_var(writeback_index);
1356         pgoff_t index;
1357         pgoff_t end;            /* Inclusive */
1358         pgoff_t done_index;
1359         int cycled;
1360         int range_whole = 0;
1361         int tag;
1362
1363         if (wbc->range_cyclic) {
1364                 writeback_index = mapping->writeback_index; /* prev offset */
1365                 index = writeback_index;
1366                 if (index == 0)
1367                         cycled = 1;
1368                 else
1369                         cycled = 0;
1370                 end = -1;
1371         } else {
1372                 index = wbc->range_start >> PAGE_SHIFT;
1373                 end = wbc->range_end >> PAGE_SHIFT;
1374                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1375                         range_whole = 1;
1376                 cycled = 1; /* ignore range_cyclic tests */
1377         }
1378         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1379                 tag = PAGECACHE_TAG_TOWRITE;
1380         else
1381                 tag = PAGECACHE_TAG_DIRTY;
1382 retry:
1383         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1384                 tag_pages_for_writeback(mapping, index, end);
1385
1386         done_index = index;
1387 get_pages:
1388         for_each_pagecache_tag(&iter, mapping, tag, index, end, page) {
1389                 done_index = page->index;
1390
1391                 if (w.io &&
1392                     !bio_can_add_page_contig(&w.io->op.op.wbio.bio, page))
1393                         bch2_writepage_do_io(&w);
1394
1395                 if (!w.io &&
1396                     atomic_read(&c->writeback_pages) >=
1397                     c->writeback_pages_max) {
1398                         /* don't sleep with pages pinned: */
1399                         pagecache_iter_release(&iter);
1400
1401                         __wait_event(c->writeback_wait,
1402                                      atomic_read(&c->writeback_pages) <
1403                                      c->writeback_pages_max);
1404                         goto get_pages;
1405                 }
1406
1407                 lock_page(page);
1408
1409                 /*
1410                  * Page truncated or invalidated. We can freely skip it
1411                  * then, even for data integrity operations: the page
1412                  * has disappeared concurrently, so there could be no
1413                  * real expectation of this data interity operation
1414                  * even if there is now a new, dirty page at the same
1415                  * pagecache address.
1416                  */
1417                 if (unlikely(page->mapping != mapping)) {
1418 continue_unlock:
1419                         unlock_page(page);
1420                         continue;
1421                 }
1422
1423                 if (!PageDirty(page)) {
1424                         /* someone wrote it for us */
1425                         goto continue_unlock;
1426                 }
1427
1428                 if (PageWriteback(page)) {
1429                         if (wbc->sync_mode != WB_SYNC_NONE)
1430                                 wait_on_page_writeback(page);
1431                         else
1432                                 goto continue_unlock;
1433                 }
1434
1435                 BUG_ON(PageWriteback(page));
1436                 if (!clear_page_dirty_for_io(page))
1437                         goto continue_unlock;
1438
1439                 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
1440                 ret = __bch2_writepage(c, page, wbc, &w);
1441                 if (unlikely(ret)) {
1442                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
1443                                 unlock_page(page);
1444                                 ret = 0;
1445                         } else {
1446                                 /*
1447                                  * done_index is set past this page,
1448                                  * so media errors will not choke
1449                                  * background writeout for the entire
1450                                  * file. This has consequences for
1451                                  * range_cyclic semantics (ie. it may
1452                                  * not be suitable for data integrity
1453                                  * writeout).
1454                                  */
1455                                 done_index = page->index + 1;
1456                                 done = 1;
1457                                 break;
1458                         }
1459                 }
1460
1461                 /*
1462                  * We stop writing back only if we are not doing
1463                  * integrity sync. In case of integrity sync we have to
1464                  * keep going until we have written all the pages
1465                  * we tagged for writeback prior to entering this loop.
1466                  */
1467                 if (--wbc->nr_to_write <= 0 &&
1468                     wbc->sync_mode == WB_SYNC_NONE) {
1469                         done = 1;
1470                         break;
1471                 }
1472         }
1473         pagecache_iter_release(&iter);
1474
1475         if (w.io)
1476                 bch2_writepage_do_io(&w);
1477
1478         if (!cycled && !done) {
1479                 /*
1480                  * range_cyclic:
1481                  * We hit the last page and there is more work to be done: wrap
1482                  * back to the start of the file
1483                  */
1484                 cycled = 1;
1485                 index = 0;
1486                 end = writeback_index - 1;
1487                 goto retry;
1488         }
1489         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1490                 mapping->writeback_index = done_index;
1491
1492         return ret;
1493 }
1494
1495 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1496 {
1497         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1498         struct bch_writepage_state w =
1499                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1500         int ret;
1501
1502         ret = __bch2_writepage(c, page, wbc, &w);
1503         if (w.io)
1504                 bch2_writepage_do_io(&w);
1505
1506         return ret;
1507 }
1508
1509 /* buffered writes: */
1510
1511 int bch2_write_begin(struct file *file, struct address_space *mapping,
1512                      loff_t pos, unsigned len, unsigned flags,
1513                      struct page **pagep, void **fsdata)
1514 {
1515         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1516         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1517         pgoff_t index = pos >> PAGE_SHIFT;
1518         unsigned offset = pos & (PAGE_SIZE - 1);
1519         struct page *page;
1520         int ret = -ENOMEM;
1521
1522         BUG_ON(inode_unhashed(&inode->v));
1523
1524         /* Not strictly necessary - same reason as mkwrite(): */
1525         pagecache_add_get(&mapping->add_lock);
1526
1527         page = grab_cache_page_write_begin(mapping, index, flags);
1528         if (!page)
1529                 goto err_unlock;
1530
1531         if (PageUptodate(page))
1532                 goto out;
1533
1534         /* If we're writing entire page, don't need to read it in first: */
1535         if (len == PAGE_SIZE)
1536                 goto out;
1537
1538         if (!offset && pos + len >= inode->v.i_size) {
1539                 zero_user_segment(page, len, PAGE_SIZE);
1540                 flush_dcache_page(page);
1541                 goto out;
1542         }
1543
1544         if (index > inode->v.i_size >> PAGE_SHIFT) {
1545                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1546                 flush_dcache_page(page);
1547                 goto out;
1548         }
1549 readpage:
1550         ret = bch2_read_single_page(page, mapping);
1551         if (ret)
1552                 goto err;
1553 out:
1554         ret = bch2_get_page_reservation(c, inode, page, true);
1555         if (ret) {
1556                 if (!PageUptodate(page)) {
1557                         /*
1558                          * If the page hasn't been read in, we won't know if we
1559                          * actually need a reservation - we don't actually need
1560                          * to read here, we just need to check if the page is
1561                          * fully backed by uncompressed data:
1562                          */
1563                         goto readpage;
1564                 }
1565
1566                 goto err;
1567         }
1568
1569         *pagep = page;
1570         return 0;
1571 err:
1572         unlock_page(page);
1573         put_page(page);
1574         *pagep = NULL;
1575 err_unlock:
1576         pagecache_add_put(&mapping->add_lock);
1577         return ret;
1578 }
1579
1580 int bch2_write_end(struct file *file, struct address_space *mapping,
1581                    loff_t pos, unsigned len, unsigned copied,
1582                    struct page *page, void *fsdata)
1583 {
1584         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1585         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1586
1587         lockdep_assert_held(&inode->v.i_rwsem);
1588
1589         if (unlikely(copied < len && !PageUptodate(page))) {
1590                 /*
1591                  * The page needs to be read in, but that would destroy
1592                  * our partial write - simplest thing is to just force
1593                  * userspace to redo the write:
1594                  */
1595                 zero_user(page, 0, PAGE_SIZE);
1596                 flush_dcache_page(page);
1597                 copied = 0;
1598         }
1599
1600         if (pos + copied > inode->v.i_size)
1601                 i_size_write(&inode->v, pos + copied);
1602
1603         if (copied) {
1604                 if (!PageUptodate(page))
1605                         SetPageUptodate(page);
1606                 if (!PageDirty(page))
1607                         set_page_dirty(page);
1608
1609                 inode->ei_last_dirtied = (unsigned long) current;
1610         } else {
1611                 bch2_put_page_reservation(c, inode, page);
1612         }
1613
1614         unlock_page(page);
1615         put_page(page);
1616         pagecache_add_put(&mapping->add_lock);
1617
1618         return copied;
1619 }
1620
1621 /* O_DIRECT reads */
1622
1623 static void bch2_dio_read_complete(struct closure *cl)
1624 {
1625         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1626
1627         dio->req->ki_complete(dio->req, dio->ret, 0);
1628         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1629 }
1630
1631 static void bch2_direct_IO_read_endio(struct bio *bio)
1632 {
1633         struct dio_read *dio = bio->bi_private;
1634
1635         if (bio->bi_status)
1636                 dio->ret = blk_status_to_errno(bio->bi_status);
1637
1638         closure_put(&dio->cl);
1639 }
1640
1641 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1642 {
1643         bch2_direct_IO_read_endio(bio);
1644         bio_check_pages_dirty(bio);     /* transfers ownership */
1645 }
1646
1647 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1648 {
1649         struct file *file = req->ki_filp;
1650         struct bch_inode_info *inode = file_bch_inode(file);
1651         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1652         struct bch_io_opts opts = io_opts(c, inode);
1653         struct dio_read *dio;
1654         struct bio *bio;
1655         loff_t offset = req->ki_pos;
1656         bool sync = is_sync_kiocb(req);
1657         ssize_t ret;
1658
1659         if ((offset|iter->count) & (block_bytes(c) - 1))
1660                 return -EINVAL;
1661
1662         ret = min_t(loff_t, iter->count,
1663                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1664         iov_iter_truncate(iter, round_up(ret, block_bytes(c)));
1665
1666         if (!ret)
1667                 return ret;
1668
1669         bio = bio_alloc_bioset(GFP_KERNEL,
1670                                iov_iter_npages(iter, BIO_MAX_PAGES),
1671                                &c->dio_read_bioset);
1672
1673         bio->bi_end_io = bch2_direct_IO_read_endio;
1674
1675         dio = container_of(bio, struct dio_read, rbio.bio);
1676         closure_init(&dio->cl, NULL);
1677
1678         /*
1679          * this is a _really_ horrible hack just to avoid an atomic sub at the
1680          * end:
1681          */
1682         if (!sync) {
1683                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1684                 atomic_set(&dio->cl.remaining,
1685                            CLOSURE_REMAINING_INITIALIZER -
1686                            CLOSURE_RUNNING +
1687                            CLOSURE_DESTRUCTOR);
1688         } else {
1689                 atomic_set(&dio->cl.remaining,
1690                            CLOSURE_REMAINING_INITIALIZER + 1);
1691         }
1692
1693         dio->req        = req;
1694         dio->ret        = ret;
1695
1696         goto start;
1697         while (iter->count) {
1698                 bio = bio_alloc_bioset(GFP_KERNEL,
1699                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1700                                        &c->bio_read);
1701                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1702 start:
1703                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1704                 bio->bi_iter.bi_sector  = offset >> 9;
1705                 bio->bi_private         = dio;
1706
1707                 ret = bio_iov_iter_get_pages(bio, iter);
1708                 if (ret < 0) {
1709                         /* XXX: fault inject this path */
1710                         bio->bi_status = BLK_STS_RESOURCE;
1711                         bio_endio(bio);
1712                         break;
1713                 }
1714
1715                 offset += bio->bi_iter.bi_size;
1716                 bio_set_pages_dirty(bio);
1717
1718                 if (iter->count)
1719                         closure_get(&dio->cl);
1720
1721                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1722         }
1723
1724         if (sync) {
1725                 closure_sync(&dio->cl);
1726                 closure_debug_destroy(&dio->cl);
1727                 ret = dio->ret;
1728                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1729                 return ret;
1730         } else {
1731                 return -EIOCBQUEUED;
1732         }
1733 }
1734
1735 /* O_DIRECT writes */
1736
1737 static void bch2_dio_write_loop_async(struct closure *);
1738
1739 static long bch2_dio_write_loop(struct dio_write *dio)
1740 {
1741         struct kiocb *req = dio->req;
1742         struct address_space *mapping = req->ki_filp->f_mapping;
1743         struct bch_inode_info *inode = dio->iop.inode;
1744         struct bio *bio = &dio->iop.op.wbio.bio;
1745         struct bio_vec *bv;
1746         bool sync;
1747         long ret;
1748         int i;
1749
1750         if (dio->loop)
1751                 goto loop;
1752
1753         inode_dio_begin(&inode->v);
1754         __pagecache_block_get(&mapping->add_lock);
1755
1756         /* Write and invalidate pagecache range that we're writing to: */
1757         ret = write_invalidate_inode_pages_range(mapping, req->ki_pos,
1758                                 req->ki_pos + iov_iter_count(&dio->iter) - 1);
1759         if (unlikely(ret))
1760                 goto err;
1761
1762         while (1) {
1763                 BUG_ON(current->pagecache_lock);
1764                 current->pagecache_lock = &mapping->add_lock;
1765                 if (current != dio->task)
1766                         use_mm(dio->task->mm);
1767
1768                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1769
1770                 if (current != dio->task)
1771                         unuse_mm(dio->task->mm);
1772                 current->pagecache_lock = NULL;
1773
1774                 if (unlikely(ret < 0))
1775                         goto err;
1776
1777                 dio->iop.op.pos = POS(inode->v.i_ino,
1778                                 (req->ki_pos >> 9) + dio->iop.op.written);
1779
1780                 task_io_account_write(bio->bi_iter.bi_size);
1781
1782                 closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl);
1783
1784                 if (!dio->sync && !dio->loop && dio->iter.count) {
1785                         struct iovec *iov = dio->inline_vecs;
1786
1787                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1788                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1789                                               GFP_KERNEL);
1790                                 if (unlikely(!iov)) {
1791                                         dio->iop.op.error = -ENOMEM;
1792                                         goto err_wait_io;
1793                                 }
1794
1795                                 dio->free_iov = true;
1796                         }
1797
1798                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1799                         dio->iter.iov = iov;
1800                 }
1801 err_wait_io:
1802                 dio->loop = true;
1803
1804                 if (!dio->sync) {
1805                         continue_at_noreturn(&dio->cl,
1806                                         bch2_dio_write_loop_async, NULL);
1807                         return -EIOCBQUEUED;
1808                 }
1809
1810                 closure_sync(&dio->cl);
1811 loop:
1812                 bio_for_each_segment_all(bv, bio, i)
1813                         put_page(bv->bv_page);
1814                 if (!dio->iter.count || dio->iop.op.error)
1815                         break;
1816                 bio_reset(bio);
1817         }
1818
1819         ret = dio->iop.op.error ?: ((long) dio->iop.op.written << 9);
1820 err:
1821         __pagecache_block_put(&mapping->add_lock);
1822         bch2_disk_reservation_put(dio->iop.op.c, &dio->iop.op.res);
1823         bch2_quota_reservation_put(dio->iop.op.c, inode, &dio->quota_res);
1824
1825         if (dio->free_iov)
1826                 kfree(dio->iter.iov);
1827
1828         closure_debug_destroy(&dio->cl);
1829
1830         sync = dio->sync;
1831         bio_put(bio);
1832
1833         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1834         inode_dio_end(&inode->v);
1835
1836         if (!sync) {
1837                 req->ki_complete(req, ret, 0);
1838                 ret = -EIOCBQUEUED;
1839         }
1840         return ret;
1841 }
1842
1843 static void bch2_dio_write_loop_async(struct closure *cl)
1844 {
1845         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1846
1847         bch2_dio_write_loop(dio);
1848 }
1849
1850 static int bch2_direct_IO_write(struct kiocb *req,
1851                                 struct iov_iter *iter,
1852                                 bool swap)
1853 {
1854         struct file *file = req->ki_filp;
1855         struct bch_inode_info *inode = file_bch_inode(file);
1856         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1857         struct dio_write *dio;
1858         struct bio *bio;
1859         loff_t offset = req->ki_pos;
1860         ssize_t ret;
1861
1862         lockdep_assert_held(&inode->v.i_rwsem);
1863
1864         if (unlikely(!iter->count))
1865                 return 0;
1866
1867         if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
1868                 return -EINVAL;
1869
1870         bio = bio_alloc_bioset(GFP_KERNEL,
1871                                iov_iter_npages(iter, BIO_MAX_PAGES),
1872                                &c->dio_write_bioset);
1873         dio = container_of(bio, struct dio_write, iop.op.wbio.bio);
1874         closure_init(&dio->cl, NULL);
1875         dio->req                = req;
1876         dio->task               = current;
1877         dio->loop               = false;
1878         dio->sync               = is_sync_kiocb(req) ||
1879                 offset + iter->count > inode->v.i_size;
1880         dio->free_iov           = false;
1881         dio->quota_res.sectors  = 0;
1882         dio->iter               = *iter;
1883         bch2_fswrite_op_init(&dio->iop, c, inode, io_opts(c, inode), true);
1884         dio->iop.op.write_point = writepoint_hashed((unsigned long) dio->task);
1885         dio->iop.op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1886
1887         if ((req->ki_flags & IOCB_DSYNC) &&
1888             !c->opts.journal_flush_disabled)
1889                 dio->iop.op.flags |= BCH_WRITE_FLUSH;
1890
1891         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
1892                                          iter->count >> 9, true);
1893         if (unlikely(ret))
1894                 goto err;
1895
1896         ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9,
1897                                         c->opts.data_replicas, 0);
1898         if (unlikely(ret)) {
1899                 if (bch2_check_range_allocated(c, POS(inode->v.i_ino,
1900                                                       offset >> 9),
1901                                                iter->count >> 9))
1902                         goto err;
1903
1904                 dio->iop.unalloc = true;
1905         }
1906
1907         dio->iop.op.nr_replicas = dio->iop.op.res.nr_replicas;
1908
1909         return bch2_dio_write_loop(dio);
1910 err:
1911         bch2_disk_reservation_put(c, &dio->iop.op.res);
1912         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1913         closure_debug_destroy(&dio->cl);
1914         bio_put(bio);
1915         return ret;
1916 }
1917
1918 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1919 {
1920         struct blk_plug plug;
1921         ssize_t ret;
1922
1923         blk_start_plug(&plug);
1924         ret = iov_iter_rw(iter) == WRITE
1925                 ? bch2_direct_IO_write(req, iter, false)
1926                 : bch2_direct_IO_read(req, iter);
1927         blk_finish_plug(&plug);
1928
1929         return ret;
1930 }
1931
1932 static ssize_t
1933 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1934 {
1935         return bch2_direct_IO_write(iocb, iter, true);
1936 }
1937
1938 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1939 {
1940         struct file *file = iocb->ki_filp;
1941         struct bch_inode_info *inode = file_bch_inode(file);
1942         ssize_t ret;
1943
1944         /* We can write back this queue in page reclaim */
1945         current->backing_dev_info = inode_to_bdi(&inode->v);
1946         ret = file_remove_privs(file);
1947         if (ret)
1948                 goto out;
1949
1950         ret = file_update_time(file);
1951         if (ret)
1952                 goto out;
1953
1954         ret = iocb->ki_flags & IOCB_DIRECT
1955                 ? bch2_direct_write(iocb, from)
1956                 : generic_perform_write(file, from, iocb->ki_pos);
1957
1958         if (likely(ret > 0))
1959                 iocb->ki_pos += ret;
1960 out:
1961         current->backing_dev_info = NULL;
1962         return ret;
1963 }
1964
1965 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1966 {
1967         struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp);
1968         bool direct = iocb->ki_flags & IOCB_DIRECT;
1969         ssize_t ret;
1970
1971         inode_lock(&inode->v);
1972         ret = generic_write_checks(iocb, from);
1973         if (ret > 0)
1974                 ret = __bch2_write_iter(iocb, from);
1975         inode_unlock(&inode->v);
1976
1977         if (ret > 0 && !direct)
1978                 ret = generic_write_sync(iocb, ret);
1979
1980         return ret;
1981 }
1982
1983 /* fsync: */
1984
1985 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1986 {
1987         struct bch_inode_info *inode = file_bch_inode(file);
1988         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1989         int ret;
1990
1991         ret = filemap_write_and_wait_range(inode->v.i_mapping, start, end);
1992         if (ret)
1993                 return ret;
1994
1995         if (c->opts.journal_flush_disabled)
1996                 return 0;
1997
1998         return bch2_journal_flush_seq(&c->journal, inode->ei_journal_seq);
1999 }
2000
2001 /* truncate: */
2002
2003 static int __bch2_truncate_page(struct bch_inode_info *inode,
2004                                 pgoff_t index, loff_t start, loff_t end)
2005 {
2006         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2007         struct address_space *mapping = inode->v.i_mapping;
2008         unsigned start_offset = start & (PAGE_SIZE - 1);
2009         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2010         struct page *page;
2011         int ret = 0;
2012
2013         /* Page boundary? Nothing to do */
2014         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2015               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2016                 return 0;
2017
2018         /* Above i_size? */
2019         if (index << PAGE_SHIFT >= inode->v.i_size)
2020                 return 0;
2021
2022         page = find_lock_page(mapping, index);
2023         if (!page) {
2024                 struct btree_iter iter;
2025                 struct bkey_s_c k = bkey_s_c_null;
2026
2027                 /*
2028                  * XXX: we're doing two index lookups when we end up reading the
2029                  * page
2030                  */
2031                 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2032                                    POS(inode->v.i_ino,
2033                                        index << PAGE_SECTOR_SHIFT), 0, k) {
2034                         if (bkey_cmp(bkey_start_pos(k.k),
2035                                      POS(inode->v.i_ino,
2036                                          (index + 1) << PAGE_SECTOR_SHIFT)) >= 0)
2037                                 break;
2038
2039                         if (k.k->type != KEY_TYPE_DISCARD &&
2040                             k.k->type != BCH_RESERVATION) {
2041                                 bch2_btree_iter_unlock(&iter);
2042                                 goto create;
2043                         }
2044                 }
2045                 bch2_btree_iter_unlock(&iter);
2046                 return 0;
2047 create:
2048                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2049                 if (unlikely(!page)) {
2050                         ret = -ENOMEM;
2051                         goto out;
2052                 }
2053         }
2054
2055         if (!PageUptodate(page)) {
2056                 ret = bch2_read_single_page(page, mapping);
2057                 if (ret)
2058                         goto unlock;
2059         }
2060
2061         /*
2062          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2063          *
2064          * XXX: because we aren't currently tracking whether the page has actual
2065          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2066          */
2067         ret = bch2_get_page_reservation(c, inode, page, false);
2068         BUG_ON(ret);
2069
2070         if (index == start >> PAGE_SHIFT &&
2071             index == end >> PAGE_SHIFT)
2072                 zero_user_segment(page, start_offset, end_offset);
2073         else if (index == start >> PAGE_SHIFT)
2074                 zero_user_segment(page, start_offset, PAGE_SIZE);
2075         else if (index == end >> PAGE_SHIFT)
2076                 zero_user_segment(page, 0, end_offset);
2077
2078         if (!PageDirty(page))
2079                 set_page_dirty(page);
2080 unlock:
2081         unlock_page(page);
2082         put_page(page);
2083 out:
2084         return ret;
2085 }
2086
2087 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2088 {
2089         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2090                                     from, from + PAGE_SIZE);
2091 }
2092
2093 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2094 {
2095         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2096         struct address_space *mapping = inode->v.i_mapping;
2097         bool shrink = iattr->ia_size <= inode->v.i_size;
2098         struct i_sectors_hook i_sectors_hook =
2099                 i_sectors_hook_init(inode, BCH_INODE_I_SIZE_DIRTY);
2100         int ret = 0;
2101
2102         inode_dio_wait(&inode->v);
2103         pagecache_block_get(&mapping->add_lock);
2104
2105         truncate_setsize(&inode->v, iattr->ia_size);
2106
2107         /* sync appends.. */
2108         /* XXX what protects inode->i_size? */
2109         if (iattr->ia_size > inode->ei_inode.bi_size)
2110                 ret = filemap_write_and_wait_range(mapping,
2111                                                    inode->ei_inode.bi_size, S64_MAX);
2112         if (ret)
2113                 goto err_put_pagecache;
2114
2115         i_sectors_hook.new_i_size = iattr->ia_size;
2116
2117         ret = i_sectors_dirty_start(c, &i_sectors_hook);
2118         if (unlikely(ret))
2119                 goto err;
2120
2121         /*
2122          * There might be persistent reservations (from fallocate())
2123          * above i_size, which bch2_inode_truncate() will discard - we're
2124          * only supposed to discard them if we're doing a real truncate
2125          * here (new i_size < current i_size):
2126          */
2127         if (shrink) {
2128                 ret = bch2_truncate_page(inode, iattr->ia_size);
2129                 if (unlikely(ret))
2130                         goto err;
2131
2132                 ret = bch2_inode_truncate(c, inode->v.i_ino,
2133                                           round_up(iattr->ia_size, PAGE_SIZE) >> 9,
2134                                           &i_sectors_hook.hook,
2135                                           &inode->ei_journal_seq);
2136                 if (unlikely(ret))
2137                         goto err;
2138         }
2139
2140         setattr_copy(&inode->v, iattr);
2141         inode->v.i_mtime = inode->v.i_ctime = current_time(&inode->v);
2142 err:
2143         /*
2144          * On error - in particular, bch2_truncate_page() error - don't clear
2145          * I_SIZE_DIRTY, as we've left data above i_size!:
2146          */
2147         if (ret)
2148                 i_sectors_hook.flags &= ~BCH_INODE_I_SIZE_DIRTY;
2149
2150         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2151 err_put_pagecache:
2152         pagecache_block_put(&mapping->add_lock);
2153         return ret;
2154 }
2155
2156 /* fallocate: */
2157
2158 static long bch2_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2159 {
2160         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2161         struct address_space *mapping = inode->v.i_mapping;
2162         u64 ino = inode->v.i_ino;
2163         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
2164         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
2165         int ret = 0;
2166
2167         inode_lock(&inode->v);
2168         inode_dio_wait(&inode->v);
2169         pagecache_block_get(&mapping->add_lock);
2170
2171         ret = __bch2_truncate_page(inode,
2172                                    offset >> PAGE_SHIFT,
2173                                    offset, offset + len);
2174         if (unlikely(ret))
2175                 goto err;
2176
2177         if (offset >> PAGE_SHIFT !=
2178             (offset + len) >> PAGE_SHIFT) {
2179                 ret = __bch2_truncate_page(inode,
2180                                            (offset + len) >> PAGE_SHIFT,
2181                                            offset, offset + len);
2182                 if (unlikely(ret))
2183                         goto err;
2184         }
2185
2186         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2187
2188         if (discard_start < discard_end) {
2189                 /*
2190                  * We need to pass in a disk reservation here because we might
2191                  * be splitting a compressed extent into two. This isn't a
2192                  * problem with truncate because truncate will never split an
2193                  * extent, only truncate it...
2194                  */
2195                 struct disk_reservation disk_res =
2196                         bch2_disk_reservation_init(c, 0);
2197                 struct i_sectors_hook i_sectors_hook =
2198                         i_sectors_hook_init(inode, 0);
2199                 int ret;
2200
2201                 ret = i_sectors_dirty_start(c, &i_sectors_hook);
2202                 if (unlikely(ret))
2203                         goto err;
2204
2205                 ret = bch2_btree_delete_range(c,
2206                                 BTREE_ID_EXTENTS,
2207                                 POS(ino, discard_start),
2208                                 POS(ino, discard_end),
2209                                 ZERO_VERSION,
2210                                 &disk_res,
2211                                 &i_sectors_hook.hook,
2212                                 &inode->ei_journal_seq);
2213
2214                 ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2215         }
2216 err:
2217         pagecache_block_put(&mapping->add_lock);
2218         inode_unlock(&inode->v);
2219
2220         return ret;
2221 }
2222
2223 static long bch2_fcollapse(struct bch_inode_info *inode,
2224                            loff_t offset, loff_t len)
2225 {
2226         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2227         struct address_space *mapping = inode->v.i_mapping;
2228         struct btree_iter src;
2229         struct btree_iter dst;
2230         BKEY_PADDED(k) copy;
2231         struct bkey_s_c k;
2232         struct i_sectors_hook i_sectors_hook = i_sectors_hook_init(inode, 0);
2233         loff_t new_size;
2234         int ret;
2235
2236         if ((offset | len) & (PAGE_SIZE - 1))
2237                 return -EINVAL;
2238
2239         bch2_btree_iter_init(&dst, c, BTREE_ID_EXTENTS,
2240                              POS(inode->v.i_ino, offset >> 9),
2241                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2242         /* position will be set from dst iter's position: */
2243         bch2_btree_iter_init(&src, c, BTREE_ID_EXTENTS, POS_MIN,
2244                              BTREE_ITER_SLOTS);
2245         bch2_btree_iter_link(&src, &dst);
2246
2247         /*
2248          * We need i_mutex to keep the page cache consistent with the extents
2249          * btree, and the btree consistent with i_size - we don't need outside
2250          * locking for the extents btree itself, because we're using linked
2251          * iterators
2252          */
2253         inode_lock(&inode->v);
2254         inode_dio_wait(&inode->v);
2255         pagecache_block_get(&mapping->add_lock);
2256
2257         ret = -EINVAL;
2258         if (offset + len >= inode->v.i_size)
2259                 goto err;
2260
2261         if (inode->v.i_size < len)
2262                 goto err;
2263
2264         new_size = inode->v.i_size - len;
2265
2266         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2267         if (ret)
2268                 goto err;
2269
2270         ret = i_sectors_dirty_start(c, &i_sectors_hook);
2271         if (ret)
2272                 goto err;
2273
2274         while (bkey_cmp(dst.pos,
2275                         POS(inode->v.i_ino,
2276                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2277                 struct disk_reservation disk_res;
2278
2279                 bch2_btree_iter_set_pos(&src,
2280                         POS(dst.pos.inode, dst.pos.offset + (len >> 9)));
2281
2282                 k = bch2_btree_iter_peek_slot(&src);
2283                 if ((ret = btree_iter_err(k)))
2284                         goto btree_iter_err;
2285
2286                 bkey_reassemble(&copy.k, k);
2287
2288                 if (bkey_deleted(&copy.k.k))
2289                         copy.k.k.type = KEY_TYPE_DISCARD;
2290
2291                 bch2_cut_front(src.pos, &copy.k);
2292                 copy.k.k.p.offset -= len >> 9;
2293
2294                 BUG_ON(bkey_cmp(dst.pos, bkey_start_pos(&copy.k.k)));
2295
2296                 ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size,
2297                                 bch2_extent_nr_dirty_ptrs(bkey_i_to_s_c(&copy.k)),
2298                                 BCH_DISK_RESERVATION_NOFAIL);
2299                 BUG_ON(ret);
2300
2301                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2302                                            &inode->ei_journal_seq,
2303                                            BTREE_INSERT_ATOMIC|
2304                                            BTREE_INSERT_NOFAIL,
2305                                            BTREE_INSERT_ENTRY(&dst, &copy.k));
2306                 bch2_disk_reservation_put(c, &disk_res);
2307 btree_iter_err:
2308                 if (ret == -EINTR)
2309                         ret = 0;
2310                 if (ret)
2311                         goto err_put_sectors_dirty;
2312                 /*
2313                  * XXX: if we error here we've left data with multiple
2314                  * pointers... which isn't a _super_ serious problem...
2315                  */
2316
2317                 bch2_btree_iter_cond_resched(&src);
2318         }
2319
2320         bch2_btree_iter_unlock(&src);
2321         bch2_btree_iter_unlock(&dst);
2322
2323         ret = bch2_inode_truncate(c, inode->v.i_ino,
2324                                  round_up(new_size, PAGE_SIZE) >> 9,
2325                                  &i_sectors_hook.hook,
2326                                  &inode->ei_journal_seq);
2327         if (ret)
2328                 goto err_put_sectors_dirty;
2329
2330         i_size_write(&inode->v, new_size);
2331         i_sectors_hook.new_i_size = new_size;
2332 err_put_sectors_dirty:
2333         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2334 err:
2335         pagecache_block_put(&mapping->add_lock);
2336         inode_unlock(&inode->v);
2337
2338         bch2_btree_iter_unlock(&src);
2339         bch2_btree_iter_unlock(&dst);
2340         return ret;
2341 }
2342
2343 static long bch2_fallocate(struct bch_inode_info *inode, int mode,
2344                            loff_t offset, loff_t len)
2345 {
2346         struct address_space *mapping = inode->v.i_mapping;
2347         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2348         struct i_sectors_hook i_sectors_hook = i_sectors_hook_init(inode, 0);
2349         struct btree_iter iter;
2350         struct bpos end_pos;
2351         loff_t block_start, block_end;
2352         loff_t end = offset + len;
2353         unsigned sectors;
2354         unsigned replicas = READ_ONCE(c->opts.data_replicas);
2355         int ret;
2356
2357         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
2358                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2359
2360         inode_lock(&inode->v);
2361         inode_dio_wait(&inode->v);
2362         pagecache_block_get(&mapping->add_lock);
2363
2364         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2365                 ret = inode_newsize_ok(&inode->v, end);
2366                 if (ret)
2367                         goto err;
2368         }
2369
2370         if (mode & FALLOC_FL_ZERO_RANGE) {
2371                 ret = __bch2_truncate_page(inode,
2372                                            offset >> PAGE_SHIFT,
2373                                            offset, end);
2374
2375                 if (!ret &&
2376                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2377                         ret = __bch2_truncate_page(inode,
2378                                                    end >> PAGE_SHIFT,
2379                                                    offset, end);
2380
2381                 if (unlikely(ret))
2382                         goto err;
2383
2384                 truncate_pagecache_range(&inode->v, offset, end - 1);
2385
2386                 block_start     = round_up(offset, PAGE_SIZE);
2387                 block_end       = round_down(end, PAGE_SIZE);
2388         } else {
2389                 block_start     = round_down(offset, PAGE_SIZE);
2390                 block_end       = round_up(end, PAGE_SIZE);
2391         }
2392
2393         bch2_btree_iter_set_pos(&iter, POS(inode->v.i_ino, block_start >> 9));
2394         end_pos = POS(inode->v.i_ino, block_end >> 9);
2395
2396         ret = i_sectors_dirty_start(c, &i_sectors_hook);
2397         if (unlikely(ret))
2398                 goto err;
2399
2400         while (bkey_cmp(iter.pos, end_pos) < 0) {
2401                 struct disk_reservation disk_res = { 0 };
2402                 struct bkey_i_reservation reservation;
2403                 struct bkey_s_c k;
2404
2405                 k = bch2_btree_iter_peek_slot(&iter);
2406                 if ((ret = btree_iter_err(k)))
2407                         goto btree_iter_err;
2408
2409                 /* already reserved */
2410                 if (k.k->type == BCH_RESERVATION &&
2411                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2412                         bch2_btree_iter_next_slot(&iter);
2413                         continue;
2414                 }
2415
2416                 if (bkey_extent_is_data(k.k)) {
2417                         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
2418                                 bch2_btree_iter_next_slot(&iter);
2419                                 continue;
2420                         }
2421                 }
2422
2423                 bkey_reservation_init(&reservation.k_i);
2424                 reservation.k.type      = BCH_RESERVATION;
2425                 reservation.k.p         = k.k->p;
2426                 reservation.k.size      = k.k->size;
2427
2428                 bch2_cut_front(iter.pos, &reservation.k_i);
2429                 bch2_cut_back(end_pos, &reservation.k);
2430
2431                 sectors = reservation.k.size;
2432                 reservation.v.nr_replicas = bch2_extent_nr_dirty_ptrs(k);
2433
2434                 if (!bkey_extent_is_allocation(k.k)) {
2435                         ret = bch2_quota_reservation_add(c, inode,
2436                                         &i_sectors_hook.quota_res,
2437                                         sectors, true);
2438                         if (unlikely(ret))
2439                                 goto err_put_sectors_dirty;
2440                 }
2441
2442                 if (reservation.v.nr_replicas < replicas ||
2443                     bch2_extent_is_compressed(k)) {
2444                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2445                                                         replicas, 0);
2446                         if (unlikely(ret))
2447                                 goto err_put_sectors_dirty;
2448
2449                         reservation.v.nr_replicas = disk_res.nr_replicas;
2450                 }
2451
2452                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2453                                           &inode->ei_journal_seq,
2454                                           BTREE_INSERT_ATOMIC|
2455                                           BTREE_INSERT_NOFAIL,
2456                                           BTREE_INSERT_ENTRY(&iter, &reservation.k_i));
2457                 bch2_disk_reservation_put(c, &disk_res);
2458 btree_iter_err:
2459                 if (ret < 0 && ret != -EINTR)
2460                         goto err_put_sectors_dirty;
2461
2462         }
2463         bch2_btree_iter_unlock(&iter);
2464
2465         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2466
2467         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2468             end > inode->v.i_size) {
2469                 i_size_write(&inode->v, end);
2470
2471                 mutex_lock(&inode->ei_update_lock);
2472                 ret = bch2_write_inode_size(c, inode, inode->v.i_size);
2473                 mutex_unlock(&inode->ei_update_lock);
2474         }
2475
2476         /* blech */
2477         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2478             (mode & FALLOC_FL_ZERO_RANGE) &&
2479             inode->ei_inode.bi_size != inode->v.i_size) {
2480                 /* sync appends.. */
2481                 ret = filemap_write_and_wait_range(mapping,
2482                                         inode->ei_inode.bi_size, S64_MAX);
2483                 if (ret)
2484                         goto err;
2485
2486                 if (inode->ei_inode.bi_size != inode->v.i_size) {
2487                         mutex_lock(&inode->ei_update_lock);
2488                         ret = bch2_write_inode_size(c, inode, inode->v.i_size);
2489                         mutex_unlock(&inode->ei_update_lock);
2490                 }
2491         }
2492
2493         pagecache_block_put(&mapping->add_lock);
2494         inode_unlock(&inode->v);
2495
2496         return 0;
2497 err_put_sectors_dirty:
2498         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2499 err:
2500         bch2_btree_iter_unlock(&iter);
2501         pagecache_block_put(&mapping->add_lock);
2502         inode_unlock(&inode->v);
2503         return ret;
2504 }
2505
2506 long bch2_fallocate_dispatch(struct file *file, int mode,
2507                              loff_t offset, loff_t len)
2508 {
2509         struct bch_inode_info *inode = file_bch_inode(file);
2510
2511         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2512                 return bch2_fallocate(inode, mode, offset, len);
2513
2514         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2515                 return bch2_fpunch(inode, offset, len);
2516
2517         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2518                 return bch2_fcollapse(inode, offset, len);
2519
2520         return -EOPNOTSUPP;
2521 }
2522
2523 /* fseek: */
2524
2525 static bool page_is_data(struct page *page)
2526 {
2527         /* XXX: should only have to check PageDirty */
2528         return PagePrivate(page) &&
2529                 (page_state(page)->sectors ||
2530                  page_state(page)->dirty_sectors);
2531 }
2532
2533 static loff_t bch2_next_pagecache_data(struct inode *vinode,
2534                                        loff_t start_offset,
2535                                        loff_t end_offset)
2536 {
2537         struct address_space *mapping = vinode->i_mapping;
2538         struct page *page;
2539         pgoff_t index;
2540
2541         for (index = start_offset >> PAGE_SHIFT;
2542              index < end_offset >> PAGE_SHIFT;
2543              index++) {
2544                 if (find_get_pages(mapping, &index, 1, &page)) {
2545                         lock_page(page);
2546
2547                         if (page_is_data(page))
2548                                 end_offset =
2549                                         min(end_offset,
2550                                         max(start_offset,
2551                                             ((loff_t) index) << PAGE_SHIFT));
2552                         unlock_page(page);
2553                         put_page(page);
2554                 } else {
2555                         break;
2556                 }
2557         }
2558
2559         return end_offset;
2560 }
2561
2562 static loff_t bch2_seek_data(struct file *file, u64 offset)
2563 {
2564         struct bch_inode_info *inode = file_bch_inode(file);
2565         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2566         struct btree_iter iter;
2567         struct bkey_s_c k;
2568         u64 isize, next_data = MAX_LFS_FILESIZE;
2569         int ret;
2570
2571         isize = i_size_read(&inode->v);
2572         if (offset >= isize)
2573                 return -ENXIO;
2574
2575         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2576                            POS(inode->v.i_ino, offset >> 9), 0, k) {
2577                 if (k.k->p.inode != inode->v.i_ino) {
2578                         break;
2579                 } else if (bkey_extent_is_data(k.k)) {
2580                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2581                         break;
2582                 } else if (k.k->p.offset >> 9 > isize)
2583                         break;
2584         }
2585
2586         ret = bch2_btree_iter_unlock(&iter);
2587         if (ret)
2588                 return ret;
2589
2590         if (next_data > offset)
2591                 next_data = bch2_next_pagecache_data(&inode->v,
2592                                                      offset, next_data);
2593
2594         if (next_data > isize)
2595                 return -ENXIO;
2596
2597         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2598 }
2599
2600 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2601 {
2602         struct page *page;
2603         bool ret;
2604
2605         page = find_lock_entry(mapping, index);
2606         if (!page || radix_tree_exception(page))
2607                 return false;
2608
2609         ret = page_is_data(page);
2610         unlock_page(page);
2611
2612         return ret;
2613 }
2614
2615 static loff_t bch2_next_pagecache_hole(struct inode *vinode,
2616                                        loff_t start_offset,
2617                                        loff_t end_offset)
2618 {
2619         struct address_space *mapping = vinode->i_mapping;
2620         pgoff_t index;
2621
2622         for (index = start_offset >> PAGE_SHIFT;
2623              index < end_offset >> PAGE_SHIFT;
2624              index++)
2625                 if (!page_slot_is_data(mapping, index))
2626                         end_offset = max(start_offset,
2627                                          ((loff_t) index) << PAGE_SHIFT);
2628
2629         return end_offset;
2630 }
2631
2632 static loff_t bch2_seek_hole(struct file *file, u64 offset)
2633 {
2634         struct bch_inode_info *inode = file_bch_inode(file);
2635         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2636         struct btree_iter iter;
2637         struct bkey_s_c k;
2638         u64 isize, next_hole = MAX_LFS_FILESIZE;
2639         int ret;
2640
2641         isize = i_size_read(&inode->v);
2642         if (offset >= isize)
2643                 return -ENXIO;
2644
2645         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2646                            POS(inode->v.i_ino, offset >> 9),
2647                            BTREE_ITER_SLOTS, k) {
2648                 if (k.k->p.inode != inode->v.i_ino) {
2649                         next_hole = bch2_next_pagecache_hole(&inode->v,
2650                                         offset, MAX_LFS_FILESIZE);
2651                         break;
2652                 } else if (!bkey_extent_is_data(k.k)) {
2653                         next_hole = bch2_next_pagecache_hole(&inode->v,
2654                                         max(offset, bkey_start_offset(k.k) << 9),
2655                                         k.k->p.offset << 9);
2656
2657                         if (next_hole < k.k->p.offset << 9)
2658                                 break;
2659                 } else {
2660                         offset = max(offset, bkey_start_offset(k.k) << 9);
2661                 }
2662         }
2663
2664         ret = bch2_btree_iter_unlock(&iter);
2665         if (ret)
2666                 return ret;
2667
2668         if (next_hole > isize)
2669                 next_hole = isize;
2670
2671         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2672 }
2673
2674 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
2675 {
2676         switch (whence) {
2677         case SEEK_SET:
2678         case SEEK_CUR:
2679         case SEEK_END:
2680                 return generic_file_llseek(file, offset, whence);
2681         case SEEK_DATA:
2682                 return bch2_seek_data(file, offset);
2683         case SEEK_HOLE:
2684                 return bch2_seek_hole(file, offset);
2685         }
2686
2687         return -EINVAL;
2688 }
2689
2690 void bch2_fs_fsio_exit(struct bch_fs *c)
2691 {
2692         bioset_exit(&c->dio_write_bioset);
2693         bioset_exit(&c->dio_read_bioset);
2694         bioset_exit(&c->writepage_bioset);
2695 }
2696
2697 int bch2_fs_fsio_init(struct bch_fs *c)
2698 {
2699         int ret = 0;
2700
2701         pr_verbose_init(c->opts, "");
2702
2703         if (bioset_init(&c->writepage_bioset,
2704                         4, offsetof(struct bch_writepage_io, op.op.wbio.bio),
2705                         BIOSET_NEED_BVECS) ||
2706             bioset_init(&c->dio_read_bioset,
2707                         4, offsetof(struct dio_read, rbio.bio),
2708                         BIOSET_NEED_BVECS) ||
2709             bioset_init(&c->dio_write_bioset,
2710                         4, offsetof(struct dio_write, iop.op.wbio.bio),
2711                         BIOSET_NEED_BVECS))
2712                 ret = -ENOMEM;
2713
2714         pr_verbose_init(c->opts, "ret %i", ret);
2715         return ret;
2716 }
2717
2718 #endif /* NO_BCACHEFS_FS */