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 i_sectors_hook {
  31         struct extent_insert_hook       hook;
  32         struct bch_inode_info           *inode;
  33         s64                             sectors;
  34         u64                             new_i_size;
  35         unsigned                        flags;
  36         unsigned                        appending:1;
  37 };
  38
  39 struct bchfs_write_op {
  40         struct bch_inode_info           *inode;
  41         s64                             sectors_added;
  42         bool                            is_dio;
  43         bool                            unalloc;
  44         u64                             new_i_size;
  45
  46         /* must be last: */
  47         struct bch_write_op             op;
  48 };
  49
  50 struct bch_writepage_io {
  51         struct closure                  cl;
  52
  53         /* must be last: */
  54         struct bchfs_write_op           op;
  55 };
  56
  57 struct dio_write {
  58         struct closure                  cl;
  59         struct kiocb                    *req;
  60         struct task_struct              *task;
  61         unsigned                        loop:1,
  62                                         sync:1,
  63                                         free_iov:1;
  64
  65         struct iov_iter                 iter;
  66         struct iovec                    inline_vecs[2];
  67
  68         /* must be last: */
  69         struct bchfs_write_op           iop;
  70 };
  71
  72 struct dio_read {
  73         struct closure                  cl;
  74         struct kiocb                    *req;
  75         long                            ret;
  76         struct bch_read_bio             rbio;
  77 };
  78
  79 /* pagecache_block must be held */
  80 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  81                                               loff_t start, loff_t end)
  82 {
  83         int ret;
  84
  85         /*
  86          * XXX: the way this is currently implemented, we can spin if a process
  87          * is continually redirtying a specific page
  88          */
  89         do {
  90                 if (!mapping->nrpages &&
  91                     !mapping->nrexceptional)
  92                         return 0;
  93
  94                 ret = filemap_write_and_wait_range(mapping, start, end);
  95                 if (ret)
  96                         break;
  97
  98                 if (!mapping->nrpages)
  99                         return 0;
 100
 101                 ret = invalidate_inode_pages2_range(mapping,
 102                                 start >> PAGE_SHIFT,
 103                                 end >> PAGE_SHIFT);
 104         } while (ret == -EBUSY);
 105
 106         return ret;
 107 }
 108
 109 /* i_size updates: */
 110
 111 static int inode_set_size(struct bch_inode_info *inode,
 112                           struct bch_inode_unpacked *bi,
 113                           void *p)
 114 {
 115         loff_t *new_i_size = p;
 116
 117         lockdep_assert_held(&inode->ei_update_lock);
 118
 119         bi->bi_size = *new_i_size;
 120         return 0;
 121 }
 122
 123 static int __must_check bch2_write_inode_size(struct bch_fs *c,
 124                                               struct bch_inode_info *inode,
 125                                               loff_t new_size)
 126 {
 127         return __bch2_write_inode(c, inode, inode_set_size, &new_size);
 128 }
 129
 130 static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, int sectors)
 131 {
 132         inode->v.i_blocks += sectors;
 133         bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, BCH_QUOTA_WARN);
 134 }
 135
 136 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode, int sectors)
 137 {
 138         mutex_lock(&inode->ei_update_lock);
 139         __i_sectors_acct(c, inode, sectors);
 140         mutex_unlock(&inode->ei_update_lock);
 141 }
 142
 143 /* i_sectors accounting: */
 144
 145 static enum btree_insert_ret
 146 i_sectors_hook_fn(struct extent_insert_hook *hook,
 147                   struct bpos committed_pos,
 148                   struct bpos next_pos,
 149                   struct bkey_s_c k,
 150                   const struct bkey_i *insert)
 151 {
 152         struct i_sectors_hook *h = container_of(hook,
 153                                 struct i_sectors_hook, hook);
 154         s64 sectors = next_pos.offset - committed_pos.offset;
 155         int sign = bkey_extent_is_allocation(&insert->k) -
 156                 (k.k && bkey_extent_is_allocation(k.k));
 157
 158         EBUG_ON(!(h->inode->ei_inode.bi_flags & BCH_INODE_I_SECTORS_DIRTY));
 159
 160         h->sectors += sectors * sign;
 161
 162         return BTREE_INSERT_OK;
 163 }
 164
 165 static int i_sectors_dirty_finish_fn(struct bch_inode_info *inode,
 166                                      struct bch_inode_unpacked *bi,
 167                                      void *p)
 168 {
 169         struct i_sectors_hook *h = p;
 170
 171         if (h->new_i_size != U64_MAX &&
 172             (!h->appending ||
 173              h->new_i_size > bi->bi_size))
 174                 bi->bi_size = h->new_i_size;
 175         bi->bi_sectors  += h->sectors;
 176         bi->bi_flags    &= ~h->flags;
 177         return 0;
 178 }
 179
 180 static int i_sectors_dirty_finish(struct bch_fs *c, struct i_sectors_hook *h)
 181 {
 182         int ret;
 183
 184         mutex_lock(&h->inode->ei_update_lock);
 185         if (h->new_i_size != U64_MAX)
 186                 i_size_write(&h->inode->v, h->new_i_size);
 187
 188         __i_sectors_acct(c, h->inode, h->sectors);
 189
 190         ret = __bch2_write_inode(c, h->inode, i_sectors_dirty_finish_fn, h);
 191         mutex_unlock(&h->inode->ei_update_lock);
 192
 193         h->sectors = 0;
 194
 195         return ret;
 196 }
 197
 198 static int i_sectors_dirty_start_fn(struct bch_inode_info *inode,
 199                                     struct bch_inode_unpacked *bi, void *p)
 200 {
 201         struct i_sectors_hook *h = p;
 202
 203         if (h->flags & BCH_INODE_I_SIZE_DIRTY)
 204                 bi->bi_size = h->new_i_size;
 205
 206         bi->bi_flags |= h->flags;
 207         return 0;
 208 }
 209
 210 static int i_sectors_dirty_start(struct bch_fs *c, struct i_sectors_hook *h)
 211 {
 212         int ret;
 213
 214         mutex_lock(&h->inode->ei_update_lock);
 215         ret = __bch2_write_inode(c, h->inode, i_sectors_dirty_start_fn, h);
 216         mutex_unlock(&h->inode->ei_update_lock);
 217
 218         return ret;
 219 }
 220
 221 static inline struct i_sectors_hook
 222 i_sectors_hook_init(struct bch_inode_info *inode, unsigned flags)
 223 {
 224         return (struct i_sectors_hook) {
 225                 .hook.fn        = i_sectors_hook_fn,
 226                 .inode          = inode,
 227                 .sectors        = 0,
 228                 .new_i_size     = U64_MAX,
 229                 .flags          = flags|BCH_INODE_I_SECTORS_DIRTY,
 230         };
 231 }
 232
 233 /* normal i_size/i_sectors update machinery: */
 234
 235 struct bchfs_extent_trans_hook {
 236         struct bchfs_write_op           *op;
 237         struct extent_insert_hook       hook;
 238
 239         struct bch_inode_unpacked       inode_u;
 240         struct bkey_inode_buf           inode_p;
 241
 242         bool                            need_inode_update;
 243 };
 244
 245 static enum btree_insert_ret
 246 bchfs_extent_update_hook(struct extent_insert_hook *hook,
 247                          struct bpos committed_pos,
 248                          struct bpos next_pos,
 249                          struct bkey_s_c k,
 250                          const struct bkey_i *insert)
 251 {
 252         struct bchfs_extent_trans_hook *h = container_of(hook,
 253                                 struct bchfs_extent_trans_hook, hook);
 254         struct bch_inode_info *inode = h->op->inode;
 255         int sign = bkey_extent_is_allocation(&insert->k) -
 256                 (k.k && bkey_extent_is_allocation(k.k));
 257         s64 sectors = (s64) (next_pos.offset - committed_pos.offset) * sign;
 258         u64 offset = min(next_pos.offset << 9, h->op->new_i_size);
 259         bool do_pack = false;
 260
 261         if (h->op->unalloc &&
 262             !bch2_extent_is_fully_allocated(k))
 263                 return BTREE_INSERT_ENOSPC;
 264
 265         BUG_ON((next_pos.offset << 9) > round_up(offset, PAGE_SIZE));
 266
 267         /* XXX: inode->i_size locking */
 268         if (offset > inode->ei_inode.bi_size) {
 269                 if (!h->need_inode_update) {
 270                         h->need_inode_update = true;
 271                         return BTREE_INSERT_NEED_TRAVERSE;
 272                 }
 273
 274                 BUG_ON(h->inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY);
 275
 276                 h->inode_u.bi_size = offset;
 277                 do_pack = true;
 278
 279                 inode->ei_inode.bi_size = offset;
 280
 281                 if (h->op->is_dio)
 282                         i_size_write(&inode->v, offset);
 283         }
 284
 285         if (sectors) {
 286                 if (!h->need_inode_update) {
 287                         h->need_inode_update = true;
 288                         return BTREE_INSERT_NEED_TRAVERSE;
 289                 }
 290
 291                 h->inode_u.bi_sectors += sectors;
 292                 do_pack = true;
 293
 294                 h->op->sectors_added += sectors;
 295         }
 296
 297         if (do_pack)
 298                 bch2_inode_pack(&h->inode_p, &h->inode_u);
 299
 300         return BTREE_INSERT_OK;
 301 }
 302
 303 static int bchfs_write_index_update(struct bch_write_op *wop)
 304 {
 305         struct bchfs_write_op *op = container_of(wop,
 306                                 struct bchfs_write_op, op);
 307         struct keylist *keys = &op->op.insert_keys;
 308         struct btree_iter extent_iter, inode_iter;
 309         struct bchfs_extent_trans_hook hook;
 310         struct bkey_i *k = bch2_keylist_front(keys);
 311         s64 orig_sectors_added = op->sectors_added;
 312         int ret;
 313
 314         BUG_ON(k->k.p.inode != op->inode->v.i_ino);
 315
 316         bch2_btree_iter_init(&extent_iter, wop->c, BTREE_ID_EXTENTS,
 317                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 318                              BTREE_ITER_INTENT);
 319         bch2_btree_iter_init(&inode_iter, wop->c, BTREE_ID_INODES,
 320                              POS(extent_iter.pos.inode, 0),
 321                              BTREE_ITER_INTENT);
 322
 323         hook.op                 = op;
 324         hook.hook.fn            = bchfs_extent_update_hook;
 325         hook.need_inode_update  = false;
 326
 327         do {
 328                 ret = bch2_btree_iter_traverse(&extent_iter);
 329                 if (ret)
 330                         goto err;
 331
 332                 /* XXX: inode->i_size locking */
 333                 k = bch2_keylist_front(keys);
 334                 if (min(k->k.p.offset << 9, op->new_i_size) >
 335                     op->inode->ei_inode.bi_size)
 336                         hook.need_inode_update = true;
 337
 338                 if (hook.need_inode_update) {
 339                         struct bkey_s_c inode;
 340
 341                         if (!btree_iter_linked(&inode_iter))
 342                                 bch2_btree_iter_link(&extent_iter, &inode_iter);
 343
 344                         inode = bch2_btree_iter_peek_with_holes(&inode_iter);
 345                         if ((ret = btree_iter_err(inode)))
 346                                 goto err;
 347
 348                         if (WARN_ONCE(inode.k->type != BCH_INODE_FS,
 349                                       "inode %llu not found when updating",
 350                                       extent_iter.pos.inode)) {
 351                                 ret = -ENOENT;
 352                                 break;
 353                         }
 354
 355                         if (WARN_ONCE(bkey_bytes(inode.k) >
 356                                       sizeof(hook.inode_p),
 357                                       "inode %llu too big (%zu bytes, buf %zu)",
 358                                       extent_iter.pos.inode,
 359                                       bkey_bytes(inode.k),
 360                                       sizeof(hook.inode_p))) {
 361                                 ret = -ENOENT;
 362                                 break;
 363                         }
 364
 365                         bkey_reassemble(&hook.inode_p.inode.k_i, inode);
 366                         ret = bch2_inode_unpack(bkey_s_c_to_inode(inode),
 367                                                &hook.inode_u);
 368                         if (WARN_ONCE(ret,
 369                                       "error %i unpacking inode %llu",
 370                                       ret, extent_iter.pos.inode)) {
 371                                 ret = -ENOENT;
 372                                 break;
 373                         }
 374
 375                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 376                                         &hook.hook, op_journal_seq(wop),
 377                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 378                                         BTREE_INSERT_ENTRY(&extent_iter, k),
 379                                         BTREE_INSERT_ENTRY_EXTRA_RES(&inode_iter,
 380                                                         &hook.inode_p.inode.k_i, 2));
 381                 } else {
 382                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 383                                         &hook.hook, op_journal_seq(wop),
 384                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 385                                         BTREE_INSERT_ENTRY(&extent_iter, k));
 386                 }
 387
 388                 BUG_ON(bkey_cmp(extent_iter.pos, bkey_start_pos(&k->k)));
 389                 BUG_ON(!ret != !k->k.size);
 390 err:
 391                 if (ret == -EINTR)
 392                         continue;
 393                 if (ret)
 394                         break;
 395
 396                 BUG_ON(bkey_cmp(extent_iter.pos, k->k.p) < 0);
 397                 bch2_keylist_pop_front(keys);
 398         } while (!bch2_keylist_empty(keys));
 399
 400         bch2_btree_iter_unlock(&extent_iter);
 401         bch2_btree_iter_unlock(&inode_iter);
 402
 403         if (op->is_dio)
 404                 i_sectors_acct(wop->c, op->inode,
 405                                op->sectors_added - orig_sectors_added);
 406
 407         return ret;
 408 }
 409
 410 static inline void bch2_fswrite_op_init(struct bchfs_write_op *op,
 411                                         struct bch_fs *c,
 412                                         struct bch_inode_info *inode,
 413                                         struct bch_io_opts opts,
 414                                         bool is_dio)
 415 {
 416         op->inode               = inode;
 417         op->sectors_added       = 0;
 418         op->is_dio              = is_dio;
 419         op->unalloc             = false;
 420         op->new_i_size          = U64_MAX;
 421
 422         bch2_write_op_init(&op->op, c);
 423         op->op.csum_type        = bch2_data_checksum_type(c, opts.data_checksum);
 424         op->op.compression_type = bch2_compression_opt_to_type(opts.compression);
 425         op->op.devs             = c->fastest_devs;
 426         op->op.index_update_fn  = bchfs_write_index_update;
 427         op_journal_seq_set(&op->op, &inode->ei_journal_seq);
 428 }
 429
 430 static inline struct bch_io_opts io_opts(struct bch_fs *c, struct bch_inode_info *inode)
 431 {
 432         struct bch_io_opts opts = bch2_opts_to_inode_opts(c->opts);
 433
 434         bch2_io_opts_apply(&opts, bch2_inode_opts_get(&inode->ei_inode));
 435         return opts;
 436 }
 437
 438 /* page state: */
 439
 440 /* stored in page->private: */
 441
 442 /*
 443  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 444  * almost protected it with the page lock, except that bch2_writepage_io_done has
 445  * to update the sector counts (and from interrupt/bottom half context).
 446  */
 447 struct bch_page_state {
 448 union { struct {
 449         /*
 450          * page is _fully_ written on disk, and not compressed - which means to
 451          * write this page we don't have to reserve space (the new write will
 452          * never take up more space on disk than what it's overwriting)
 453          */
 454         unsigned allocated:1;
 455
 456         /* Owns PAGE_SECTORS sized reservation: */
 457         unsigned                reserved:1;
 458         unsigned                nr_replicas:4;
 459
 460         /*
 461          * Number of sectors on disk - for i_blocks
 462          * Uncompressed size, not compressed size:
 463          */
 464         u8                      sectors;
 465         u8                      dirty_sectors;
 466 };
 467         /* for cmpxchg: */
 468         unsigned long           v;
 469 };
 470 };
 471
 472 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 473 ({                                                                      \
 474         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 475         struct bch_page_state _old;                                     \
 476                                                                         \
 477         do {                                                            \
 478                 _old.v = _new.v = _v;                                   \
 479                 _expr;                                                  \
 480                                                                         \
 481                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 482         } while (_old.v != _new.v &&                                    \
 483                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 484                                                                         \
 485         _old;                                                           \
 486 })
 487
 488 static inline struct bch_page_state *page_state(struct page *page)
 489 {
 490         struct bch_page_state *s = (void *) &page->private;
 491
 492         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 493
 494         if (!PagePrivate(page))
 495                 SetPagePrivate(page);
 496
 497         return s;
 498 }
 499
 500 static void bch2_put_page_reservation(struct bch_fs *c, struct page *page)
 501 {
 502         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 503         struct bch_page_state s;
 504
 505         s = page_state_cmpxchg(page_state(page), s, {
 506                 if (!s.reserved)
 507                         return;
 508                 s.reserved = 0;
 509         });
 510
 511         bch2_disk_reservation_put(c, &res);
 512 }
 513
 514 static int bch2_get_page_reservation(struct bch_fs *c, struct page *page,
 515                                     bool check_enospc)
 516 {
 517         struct bch_page_state *s = page_state(page), new;
 518         struct disk_reservation res;
 519         int ret = 0;
 520
 521         BUG_ON(s->allocated && s->sectors != PAGE_SECTORS);
 522
 523         if (s->allocated || s->reserved)
 524                 return 0;
 525
 526         ret = bch2_disk_reservation_get(c, &res, PAGE_SECTORS, !check_enospc
 527                                        ? BCH_DISK_RESERVATION_NOFAIL : 0);
 528         if (ret)
 529                 return ret;
 530
 531         page_state_cmpxchg(s, new, {
 532                 if (new.reserved) {
 533                         bch2_disk_reservation_put(c, &res);
 534                         return 0;
 535                 }
 536                 new.reserved    = 1;
 537                 new.nr_replicas = res.nr_replicas;
 538         });
 539
 540         return 0;
 541 }
 542
 543 static void bch2_clear_page_bits(struct page *page)
 544 {
 545         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 546         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 547         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 548         struct bch_page_state s;
 549
 550         if (!PagePrivate(page))
 551                 return;
 552
 553         s = xchg(page_state(page), (struct bch_page_state) { .v = 0 });
 554         ClearPagePrivate(page);
 555
 556         if (s.dirty_sectors)
 557                 i_sectors_acct(c, inode, -s.dirty_sectors);
 558
 559         if (s.reserved)
 560                 bch2_disk_reservation_put(c, &res);
 561 }
 562
 563 int bch2_set_page_dirty(struct page *page)
 564 {
 565         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 566         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 567         struct bch_page_state old, new;
 568
 569         old = page_state_cmpxchg(page_state(page), new,
 570                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 571         );
 572
 573         if (old.dirty_sectors != new.dirty_sectors)
 574                 i_sectors_acct(c, inode, new.dirty_sectors - old.dirty_sectors);
 575
 576         return __set_page_dirty_nobuffers(page);
 577 }
 578
 579 /* readpages/writepages: */
 580
 581 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 582 {
 583         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 584
 585         return bio->bi_vcnt < bio->bi_max_vecs &&
 586                 bio_end_sector(bio) == offset;
 587 }
 588
 589 static void __bio_add_page(struct bio *bio, struct page *page)
 590 {
 591         bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
 592                 .bv_page = page,
 593                 .bv_len = PAGE_SIZE,
 594                 .bv_offset = 0,
 595         };
 596
 597         bio->bi_iter.bi_size += PAGE_SIZE;
 598 }
 599
 600 static int bio_add_page_contig(struct bio *bio, struct page *page)
 601 {
 602         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 603
 604         BUG_ON(!bio->bi_max_vecs);
 605
 606         if (!bio->bi_vcnt)
 607                 bio->bi_iter.bi_sector = offset;
 608         else if (!bio_can_add_page_contig(bio, page))
 609                 return -1;
 610
 611         __bio_add_page(bio, page);
 612         return 0;
 613 }
 614
 615 static void bch2_readpages_end_io(struct bio *bio)
 616 {
 617         struct bio_vec *bv;
 618         int i;
 619
 620         bio_for_each_segment_all(bv, bio, i) {
 621                 struct page *page = bv->bv_page;
 622
 623                 if (!bio->bi_status) {
 624                         SetPageUptodate(page);
 625                 } else {
 626                         ClearPageUptodate(page);
 627                         SetPageError(page);
 628                 }
 629                 unlock_page(page);
 630         }
 631
 632         bio_put(bio);
 633 }
 634
 635 struct readpages_iter {
 636         struct address_space    *mapping;
 637         struct list_head        pages;
 638         unsigned                nr_pages;
 639 };
 640
 641 static int readpage_add_page(struct readpages_iter *iter, struct page *page)
 642 {
 643         struct bch_page_state *s = page_state(page);
 644         int ret;
 645
 646         BUG_ON(s->reserved);
 647         s->allocated = 1;
 648         s->sectors = 0;
 649
 650         prefetchw(&page->flags);
 651         ret = add_to_page_cache_lru(page, iter->mapping,
 652                                     page->index, GFP_NOFS);
 653         put_page(page);
 654         return ret;
 655 }
 656
 657 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 658 {
 659         while (iter->nr_pages) {
 660                 struct page *page =
 661                         list_last_entry(&iter->pages, struct page, lru);
 662
 663                 prefetchw(&page->flags);
 664                 list_del(&page->lru);
 665                 iter->nr_pages--;
 666
 667                 if (!readpage_add_page(iter, page))
 668                         return page;
 669         }
 670
 671         return NULL;
 672 }
 673
 674 #define for_each_readpage_page(_iter, _page)                            \
 675         for (;                                                          \
 676              ((_page) = __readpage_next_page(&(_iter)));)               \
 677
 678 static void bch2_mark_pages_unalloc(struct bio *bio)
 679 {
 680         struct bvec_iter iter;
 681         struct bio_vec bv;
 682
 683         bio_for_each_segment(bv, bio, iter)
 684                 page_state(bv.bv_page)->allocated = 0;
 685 }
 686
 687 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 688 {
 689         struct bvec_iter iter;
 690         struct bio_vec bv;
 691
 692         bio_for_each_segment(bv, bio, iter) {
 693                 struct bch_page_state *s = page_state(bv.bv_page);
 694
 695                 /* sectors in @k from the start of this page: */
 696                 unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset);
 697
 698                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 699
 700                 if (!s->sectors)
 701                         s->nr_replicas = bch2_extent_nr_dirty_ptrs(k);
 702                 else
 703                         s->nr_replicas = min_t(unsigned, s->nr_replicas,
 704                                                bch2_extent_nr_dirty_ptrs(k));
 705
 706                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 707                 s->sectors += page_sectors;
 708         }
 709 }
 710
 711 static void readpage_bio_extend(struct readpages_iter *iter,
 712                                 struct bio *bio, u64 offset,
 713                                 bool get_more)
 714 {
 715         struct page *page;
 716         pgoff_t page_offset;
 717         int ret;
 718
 719         while (bio_end_sector(bio) < offset &&
 720                bio->bi_vcnt < bio->bi_max_vecs) {
 721                 page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 722
 723                 if (iter->nr_pages) {
 724                         page = list_last_entry(&iter->pages, struct page, lru);
 725                         if (page->index != page_offset)
 726                                 break;
 727
 728                         list_del(&page->lru);
 729                         iter->nr_pages--;
 730                 } else if (get_more) {
 731                         rcu_read_lock();
 732                         page = radix_tree_lookup(&iter->mapping->page_tree, page_offset);
 733                         rcu_read_unlock();
 734
 735                         if (page && !radix_tree_exceptional_entry(page))
 736                                 break;
 737
 738                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 739                         if (!page)
 740                                 break;
 741
 742                         page->index = page_offset;
 743                         ClearPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 744                 } else {
 745                         break;
 746                 }
 747
 748                 ret = readpage_add_page(iter, page);
 749                 if (ret)
 750                         break;
 751
 752                 __bio_add_page(bio, page);
 753         }
 754
 755         if (!iter->nr_pages)
 756                 SetPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 757 }
 758
 759 static void bchfs_read(struct bch_fs *c, struct btree_iter *iter,
 760                        struct bch_read_bio *rbio, u64 inum,
 761                        struct readpages_iter *readpages_iter)
 762 {
 763         struct bio *bio = &rbio->bio;
 764         int flags = BCH_READ_RETRY_IF_STALE|
 765                 BCH_READ_MAY_PROMOTE;
 766
 767         while (1) {
 768                 struct extent_pick_ptr pick;
 769                 BKEY_PADDED(k) tmp;
 770                 struct bkey_s_c k;
 771                 unsigned bytes;
 772                 bool is_last;
 773
 774                 bch2_btree_iter_set_pos(iter, POS(inum, bio->bi_iter.bi_sector));
 775
 776                 k = bch2_btree_iter_peek_with_holes(iter);
 777                 BUG_ON(!k.k);
 778
 779                 if (IS_ERR(k.k)) {
 780                         int ret = bch2_btree_iter_unlock(iter);
 781                         BUG_ON(!ret);
 782                         bcache_io_error(c, bio, "btree IO error %i", ret);
 783                         bio_endio(bio);
 784                         return;
 785                 }
 786
 787                 bkey_reassemble(&tmp.k, k);
 788                 bch2_btree_iter_unlock(iter);
 789                 k = bkey_i_to_s_c(&tmp.k);
 790
 791                 bch2_extent_pick_ptr(c, k, NULL, &pick);
 792                 if (IS_ERR(pick.ca)) {
 793                         bcache_io_error(c, bio, "no device to read from");
 794                         bio_endio(bio);
 795                         return;
 796                 }
 797
 798                 if (readpages_iter)
 799                         readpage_bio_extend(readpages_iter,
 800                                             bio, k.k->p.offset,
 801                                             pick.ca &&
 802                                             (pick.crc.csum_type ||
 803                                              pick.crc.compression_type));
 804
 805                 bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
 806                          bio->bi_iter.bi_sector) << 9;
 807                 is_last = bytes == bio->bi_iter.bi_size;
 808                 swap(bio->bi_iter.bi_size, bytes);
 809
 810                 if (bkey_extent_is_allocation(k.k))
 811                         bch2_add_page_sectors(bio, k);
 812
 813                 if (!bch2_extent_is_fully_allocated(k))
 814                         bch2_mark_pages_unalloc(bio);
 815
 816                 if (pick.ca) {
 817                         if (!is_last) {
 818                                 bio_inc_remaining(&rbio->bio);
 819                                 flags |= BCH_READ_MUST_CLONE;
 820                                 trace_read_split(&rbio->bio);
 821                         }
 822
 823                         bch2_read_extent(c, rbio, bkey_s_c_to_extent(k),
 824                                          &pick, flags);
 825                 } else {
 826                         zero_fill_bio(bio);
 827
 828                         if (is_last)
 829                                 bio_endio(bio);
 830                 }
 831
 832                 if (is_last)
 833                         return;
 834
 835                 swap(bio->bi_iter.bi_size, bytes);
 836                 bio_advance(bio, bytes);
 837         }
 838 }
 839
 840 int bch2_readpages(struct file *file, struct address_space *mapping,
 841                    struct list_head *pages, unsigned nr_pages)
 842 {
 843         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 844         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 845         struct bch_io_opts opts = io_opts(c, inode);
 846         struct btree_iter iter;
 847         struct page *page;
 848         struct readpages_iter readpages_iter = {
 849                 .mapping = mapping, .nr_pages = nr_pages
 850         };
 851
 852         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN, 0);
 853
 854         INIT_LIST_HEAD(&readpages_iter.pages);
 855         list_add(&readpages_iter.pages, pages);
 856         list_del_init(pages);
 857
 858         if (current->pagecache_lock != &mapping->add_lock)
 859                 pagecache_add_get(&mapping->add_lock);
 860
 861         while ((page = readpage_iter_next(&readpages_iter))) {
 862                 unsigned n = max_t(unsigned,
 863                                    min_t(unsigned, readpages_iter.nr_pages + 1,
 864                                          BIO_MAX_PAGES),
 865                                    c->sb.encoded_extent_max >> PAGE_SECTOR_SHIFT);
 866
 867                 struct bch_read_bio *rbio =
 868                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 869                                   opts);
 870
 871                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 872                 bio_add_page_contig(&rbio->bio, page);
 873                 bchfs_read(c, &iter, rbio, inode->v.i_ino, &readpages_iter);
 874         }
 875
 876         if (current->pagecache_lock != &mapping->add_lock)
 877                 pagecache_add_put(&mapping->add_lock);
 878
 879         return 0;
 880 }
 881
 882 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 883                              u64 inum, struct page *page)
 884 {
 885         struct btree_iter iter;
 886
 887         /*
 888          * Initialize page state:
 889          * If a page is partly allocated and partly a hole, we want it to be
 890          * marked BCH_PAGE_UNALLOCATED - so we initially mark all pages
 891          * allocated and then mark them unallocated as we find holes:
 892          *
 893          * Note that the bio hasn't been split yet - it's the only bio that
 894          * points to these pages. As we walk extents and split @bio, that
 895          * necessarily be true, the splits won't necessarily be on page
 896          * boundaries:
 897          */
 898         struct bch_page_state *s = page_state(page);
 899
 900         EBUG_ON(s->reserved);
 901         s->allocated = 1;
 902         s->sectors = 0;
 903
 904         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 905         bio_add_page_contig(&rbio->bio, page);
 906
 907         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN, 0);
 908         bchfs_read(c, &iter, rbio, inum, NULL);
 909 }
 910
 911 int bch2_readpage(struct file *file, struct page *page)
 912 {
 913         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 914         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 915         struct bch_io_opts opts = io_opts(c, inode);
 916         struct bch_read_bio *rbio;
 917
 918         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 919         rbio->bio.bi_end_io = bch2_readpages_end_io;
 920
 921         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 922         return 0;
 923 }
 924
 925 struct bch_writepage_state {
 926         struct bch_writepage_io *io;
 927         struct bch_io_opts      opts;
 928 };
 929
 930 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 931                                                                   struct bch_inode_info *inode)
 932 {
 933         return (struct bch_writepage_state) { .opts = io_opts(c, inode) };
 934 }
 935
 936 static void bch2_writepage_io_free(struct closure *cl)
 937 {
 938         struct bch_writepage_io *io = container_of(cl,
 939                                         struct bch_writepage_io, cl);
 940
 941         bio_put(&io->op.op.wbio.bio);
 942 }
 943
 944 static void bch2_writepage_io_done(struct closure *cl)
 945 {
 946         struct bch_writepage_io *io = container_of(cl,
 947                                         struct bch_writepage_io, cl);
 948         struct bch_fs *c = io->op.op.c;
 949         struct bio *bio = &io->op.op.wbio.bio;
 950         struct bio_vec *bvec;
 951         unsigned i;
 952
 953         atomic_sub(bio->bi_vcnt, &c->writeback_pages);
 954         wake_up(&c->writeback_wait);
 955
 956         bio_for_each_segment_all(bvec, bio, i) {
 957                 struct page *page = bvec->bv_page;
 958
 959                 if (io->op.op.error) {
 960                         SetPageError(page);
 961                         if (page->mapping)
 962                                 set_bit(AS_EIO, &page->mapping->flags);
 963                 }
 964
 965                 if (io->op.op.written >= PAGE_SECTORS) {
 966                         struct bch_page_state old, new;
 967
 968                         old = page_state_cmpxchg(page_state(page), new, {
 969                                 new.sectors = PAGE_SECTORS;
 970                                 new.dirty_sectors = 0;
 971                         });
 972
 973                         io->op.sectors_added -= old.dirty_sectors;
 974                         io->op.op.written -= PAGE_SECTORS;
 975                 }
 976         }
 977
 978         /*
 979          * racing with fallocate can cause us to add fewer sectors than
 980          * expected - but we shouldn't add more sectors than expected:
 981          *
 982          * (error (due to going RO) halfway through a page can screw that up
 983          * slightly)
 984          */
 985         BUG_ON(io->op.sectors_added >= (s64) PAGE_SECTORS);
 986
 987         /*
 988          * PageWriteback is effectively our ref on the inode - fixup i_blocks
 989          * before calling end_page_writeback:
 990          */
 991         if (io->op.sectors_added)
 992                 i_sectors_acct(c, io->op.inode, io->op.sectors_added);
 993
 994         bio_for_each_segment_all(bvec, bio, i)
 995                 end_page_writeback(bvec->bv_page);
 996
 997         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
 998 }
 999
1000 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1001 {
1002         struct bch_writepage_io *io = w->io;
1003         struct bio *bio = &io->op.op.wbio.bio;
1004
1005         w->io = NULL;
1006         atomic_add(bio->bi_vcnt, &io->op.op.c->writeback_pages);
1007
1008         closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl);
1009         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1010 }
1011
1012 /*
1013  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1014  * possible, else allocating a new one:
1015  */
1016 static void bch2_writepage_io_alloc(struct bch_fs *c,
1017                                     struct bch_writepage_state *w,
1018                                     struct bch_inode_info *inode,
1019                                     struct page *page,
1020                                     struct bch_page_state s)
1021 {
1022         struct bch_write_op *op;
1023         u64 offset = (u64) page->index << PAGE_SECTOR_SHIFT;
1024
1025         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1026                                               BIO_MAX_PAGES,
1027                                               &c->writepage_bioset),
1028                              struct bch_writepage_io, op.op.wbio.bio);
1029         op = &w->io->op.op;
1030
1031         closure_init(&w->io->cl, NULL);
1032
1033         bch2_fswrite_op_init(&w->io->op, c, inode, w->opts, false);
1034         op->nr_replicas         = s.nr_replicas;
1035         op->res.nr_replicas     = s.nr_replicas;
1036         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1037         op->pos                 = POS(inode->v.i_ino, offset);
1038         op->wbio.bio.bi_iter.bi_sector = offset;
1039 }
1040
1041 static int __bch2_writepage(struct bch_fs *c, struct page *page,
1042                             struct writeback_control *wbc,
1043                             struct bch_writepage_state *w)
1044 {
1045         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1046         struct bch_page_state new, old;
1047         unsigned offset;
1048         loff_t i_size = i_size_read(&inode->v);
1049         pgoff_t end_index = i_size >> PAGE_SHIFT;
1050
1051         EBUG_ON(!PageUptodate(page));
1052
1053         /* Is the page fully inside i_size? */
1054         if (page->index < end_index)
1055                 goto do_io;
1056
1057         /* Is the page fully outside i_size? (truncate in progress) */
1058         offset = i_size & (PAGE_SIZE - 1);
1059         if (page->index > end_index || !offset) {
1060                 unlock_page(page);
1061                 return 0;
1062         }
1063
1064         /*
1065          * The page straddles i_size.  It must be zeroed out on each and every
1066          * writepage invocation because it may be mmapped.  "A file is mapped
1067          * in multiples of the page size.  For a file that is not a multiple of
1068          * the  page size, the remaining memory is zeroed when mapped, and
1069          * writes to that region are not written out to the file."
1070          */
1071         zero_user_segment(page, offset, PAGE_SIZE);
1072 do_io:
1073         /* Before unlocking the page, transfer reservation to w->io: */
1074         old = page_state_cmpxchg(page_state(page), new, {
1075                 EBUG_ON(!new.reserved &&
1076                         (new.sectors != PAGE_SECTORS ||
1077                         !new.allocated));
1078
1079                 if (new.allocated && w->opts.compression)
1080                         new.allocated = 0;
1081                 else if (!new.reserved)
1082                         break;
1083                 new.reserved = 0;
1084         });
1085
1086         if (w->io &&
1087             (w->io->op.op.res.nr_replicas != old.nr_replicas ||
1088              !bio_can_add_page_contig(&w->io->op.op.wbio.bio, page)))
1089                 bch2_writepage_do_io(w);
1090
1091         if (!w->io)
1092                 bch2_writepage_io_alloc(c, w, inode, page, old);
1093
1094         BUG_ON(inode != w->io->op.inode);
1095         BUG_ON(bio_add_page_contig(&w->io->op.op.wbio.bio, page));
1096
1097         if (old.reserved)
1098                 w->io->op.op.res.sectors += old.nr_replicas * PAGE_SECTORS;
1099
1100         /* while page is locked: */
1101         w->io->op.new_i_size = i_size;
1102
1103         if (wbc->sync_mode == WB_SYNC_ALL)
1104                 w->io->op.op.wbio.bio.bi_opf |= REQ_SYNC;
1105
1106         BUG_ON(PageWriteback(page));
1107         set_page_writeback(page);
1108         unlock_page(page);
1109
1110         return 0;
1111 }
1112
1113 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1114 {
1115         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1116         struct bch_writepage_state w =
1117                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1118         struct pagecache_iter iter;
1119         struct page *page;
1120         int ret = 0;
1121         int done = 0;
1122         pgoff_t uninitialized_var(writeback_index);
1123         pgoff_t index;
1124         pgoff_t end;            /* Inclusive */
1125         pgoff_t done_index;
1126         int cycled;
1127         int range_whole = 0;
1128         int tag;
1129
1130         if (wbc->range_cyclic) {
1131                 writeback_index = mapping->writeback_index; /* prev offset */
1132                 index = writeback_index;
1133                 if (index == 0)
1134                         cycled = 1;
1135                 else
1136                         cycled = 0;
1137                 end = -1;
1138         } else {
1139                 index = wbc->range_start >> PAGE_SHIFT;
1140                 end = wbc->range_end >> PAGE_SHIFT;
1141                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1142                         range_whole = 1;
1143                 cycled = 1; /* ignore range_cyclic tests */
1144         }
1145         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1146                 tag = PAGECACHE_TAG_TOWRITE;
1147         else
1148                 tag = PAGECACHE_TAG_DIRTY;
1149 retry:
1150         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1151                 tag_pages_for_writeback(mapping, index, end);
1152
1153         done_index = index;
1154 get_pages:
1155         for_each_pagecache_tag(&iter, mapping, tag, index, end, page) {
1156                 done_index = page->index;
1157
1158                 if (w.io &&
1159                     !bio_can_add_page_contig(&w.io->op.op.wbio.bio, page))
1160                         bch2_writepage_do_io(&w);
1161
1162                 if (!w.io &&
1163                     atomic_read(&c->writeback_pages) >=
1164                     c->writeback_pages_max) {
1165                         /* don't sleep with pages pinned: */
1166                         pagecache_iter_release(&iter);
1167
1168                         __wait_event(c->writeback_wait,
1169                                      atomic_read(&c->writeback_pages) <
1170                                      c->writeback_pages_max);
1171                         goto get_pages;
1172                 }
1173
1174                 lock_page(page);
1175
1176                 /*
1177                  * Page truncated or invalidated. We can freely skip it
1178                  * then, even for data integrity operations: the page
1179                  * has disappeared concurrently, so there could be no
1180                  * real expectation of this data interity operation
1181                  * even if there is now a new, dirty page at the same
1182                  * pagecache address.
1183                  */
1184                 if (unlikely(page->mapping != mapping)) {
1185 continue_unlock:
1186                         unlock_page(page);
1187                         continue;
1188                 }
1189
1190                 if (!PageDirty(page)) {
1191                         /* someone wrote it for us */
1192                         goto continue_unlock;
1193                 }
1194
1195                 if (PageWriteback(page)) {
1196                         if (wbc->sync_mode != WB_SYNC_NONE)
1197                                 wait_on_page_writeback(page);
1198                         else
1199                                 goto continue_unlock;
1200                 }
1201
1202                 BUG_ON(PageWriteback(page));
1203                 if (!clear_page_dirty_for_io(page))
1204                         goto continue_unlock;
1205
1206                 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
1207                 ret = __bch2_writepage(c, page, wbc, &w);
1208                 if (unlikely(ret)) {
1209                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
1210                                 unlock_page(page);
1211                                 ret = 0;
1212                         } else {
1213                                 /*
1214                                  * done_index is set past this page,
1215                                  * so media errors will not choke
1216                                  * background writeout for the entire
1217                                  * file. This has consequences for
1218                                  * range_cyclic semantics (ie. it may
1219                                  * not be suitable for data integrity
1220                                  * writeout).
1221                                  */
1222                                 done_index = page->index + 1;
1223                                 done = 1;
1224                                 break;
1225                         }
1226                 }
1227
1228                 /*
1229                  * We stop writing back only if we are not doing
1230                  * integrity sync. In case of integrity sync we have to
1231                  * keep going until we have written all the pages
1232                  * we tagged for writeback prior to entering this loop.
1233                  */
1234                 if (--wbc->nr_to_write <= 0 &&
1235                     wbc->sync_mode == WB_SYNC_NONE) {
1236                         done = 1;
1237                         break;
1238                 }
1239         }
1240         pagecache_iter_release(&iter);
1241
1242         if (w.io)
1243                 bch2_writepage_do_io(&w);
1244
1245         if (!cycled && !done) {
1246                 /*
1247                  * range_cyclic:
1248                  * We hit the last page and there is more work to be done: wrap
1249                  * back to the start of the file
1250                  */
1251                 cycled = 1;
1252                 index = 0;
1253                 end = writeback_index - 1;
1254                 goto retry;
1255         }
1256         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1257                 mapping->writeback_index = done_index;
1258
1259         return ret;
1260 }
1261
1262 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1263 {
1264         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1265         struct bch_writepage_state w =
1266                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1267         int ret;
1268
1269         ret = __bch2_writepage(c, page, wbc, &w);
1270         if (w.io)
1271                 bch2_writepage_do_io(&w);
1272
1273         return ret;
1274 }
1275
1276 static void bch2_read_single_page_end_io(struct bio *bio)
1277 {
1278         complete(bio->bi_private);
1279 }
1280
1281 static int bch2_read_single_page(struct page *page,
1282                                  struct address_space *mapping)
1283 {
1284         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1285         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1286         struct bch_read_bio *rbio;
1287         int ret;
1288         DECLARE_COMPLETION_ONSTACK(done);
1289
1290         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1291                          io_opts(c, inode));
1292         rbio->bio.bi_private = &done;
1293         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1294
1295         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1296         wait_for_completion(&done);
1297
1298         ret = blk_status_to_errno(rbio->bio.bi_status);
1299         bio_put(&rbio->bio);
1300
1301         if (ret < 0)
1302                 return ret;
1303
1304         SetPageUptodate(page);
1305         return 0;
1306 }
1307
1308 int bch2_write_begin(struct file *file, struct address_space *mapping,
1309                      loff_t pos, unsigned len, unsigned flags,
1310                      struct page **pagep, void **fsdata)
1311 {
1312         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1313         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1314         pgoff_t index = pos >> PAGE_SHIFT;
1315         unsigned offset = pos & (PAGE_SIZE - 1);
1316         struct page *page;
1317         int ret = -ENOMEM;
1318
1319         BUG_ON(inode_unhashed(&inode->v));
1320
1321         /* Not strictly necessary - same reason as mkwrite(): */
1322         pagecache_add_get(&mapping->add_lock);
1323
1324         page = grab_cache_page_write_begin(mapping, index, flags);
1325         if (!page)
1326                 goto err_unlock;
1327
1328         if (PageUptodate(page))
1329                 goto out;
1330
1331         /* If we're writing entire page, don't need to read it in first: */
1332         if (len == PAGE_SIZE)
1333                 goto out;
1334
1335         if (!offset && pos + len >= inode->v.i_size) {
1336                 zero_user_segment(page, len, PAGE_SIZE);
1337                 flush_dcache_page(page);
1338                 goto out;
1339         }
1340
1341         if (index > inode->v.i_size >> PAGE_SHIFT) {
1342                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1343                 flush_dcache_page(page);
1344                 goto out;
1345         }
1346 readpage:
1347         ret = bch2_read_single_page(page, mapping);
1348         if (ret)
1349                 goto err;
1350 out:
1351         ret = bch2_get_page_reservation(c, page, true);
1352         if (ret) {
1353                 if (!PageUptodate(page)) {
1354                         /*
1355                          * If the page hasn't been read in, we won't know if we
1356                          * actually need a reservation - we don't actually need
1357                          * to read here, we just need to check if the page is
1358                          * fully backed by uncompressed data:
1359                          */
1360                         goto readpage;
1361                 }
1362
1363                 goto err;
1364         }
1365
1366         *pagep = page;
1367         return 0;
1368 err:
1369         unlock_page(page);
1370         put_page(page);
1371         *pagep = NULL;
1372 err_unlock:
1373         pagecache_add_put(&mapping->add_lock);
1374         return ret;
1375 }
1376
1377 int bch2_write_end(struct file *filp, struct address_space *mapping,
1378                    loff_t pos, unsigned len, unsigned copied,
1379                    struct page *page, void *fsdata)
1380 {
1381         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1382         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1383
1384         lockdep_assert_held(&inode->v.i_rwsem);
1385
1386         if (unlikely(copied < len && !PageUptodate(page))) {
1387                 /*
1388                  * The page needs to be read in, but that would destroy
1389                  * our partial write - simplest thing is to just force
1390                  * userspace to redo the write:
1391                  */
1392                 zero_user(page, 0, PAGE_SIZE);
1393                 flush_dcache_page(page);
1394                 copied = 0;
1395         }
1396
1397         if (pos + copied > inode->v.i_size)
1398                 i_size_write(&inode->v, pos + copied);
1399
1400         if (copied) {
1401                 if (!PageUptodate(page))
1402                         SetPageUptodate(page);
1403                 if (!PageDirty(page))
1404                         set_page_dirty(page);
1405
1406                 inode->ei_last_dirtied = (unsigned long) current;
1407         } else {
1408                 bch2_put_page_reservation(c, page);
1409         }
1410
1411         unlock_page(page);
1412         put_page(page);
1413         pagecache_add_put(&mapping->add_lock);
1414
1415         return copied;
1416 }
1417
1418 /* O_DIRECT */
1419
1420 static void bch2_dio_read_complete(struct closure *cl)
1421 {
1422         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1423
1424         dio->req->ki_complete(dio->req, dio->ret, 0);
1425         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1426 }
1427
1428 static void bch2_direct_IO_read_endio(struct bio *bio)
1429 {
1430         struct dio_read *dio = bio->bi_private;
1431
1432         if (bio->bi_status)
1433                 dio->ret = blk_status_to_errno(bio->bi_status);
1434
1435         closure_put(&dio->cl);
1436 }
1437
1438 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1439 {
1440         bch2_direct_IO_read_endio(bio);
1441         bio_check_pages_dirty(bio);     /* transfers ownership */
1442 }
1443
1444 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1445 {
1446         struct file *file = req->ki_filp;
1447         struct bch_inode_info *inode = file_bch_inode(file);
1448         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1449         struct bch_io_opts opts = io_opts(c, inode);
1450         struct dio_read *dio;
1451         struct bio *bio;
1452         loff_t offset = req->ki_pos;
1453         bool sync = is_sync_kiocb(req);
1454         ssize_t ret;
1455
1456         if ((offset|iter->count) & (block_bytes(c) - 1))
1457                 return -EINVAL;
1458
1459         ret = min_t(loff_t, iter->count,
1460                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1461         iov_iter_truncate(iter, round_up(ret, block_bytes(c)));
1462
1463         if (!ret)
1464                 return ret;
1465
1466         bio = bio_alloc_bioset(GFP_KERNEL,
1467                                iov_iter_npages(iter, BIO_MAX_PAGES),
1468                                &c->dio_read_bioset);
1469
1470         bio->bi_end_io = bch2_direct_IO_read_endio;
1471
1472         dio = container_of(bio, struct dio_read, rbio.bio);
1473         closure_init(&dio->cl, NULL);
1474
1475         /*
1476          * this is a _really_ horrible hack just to avoid an atomic sub at the
1477          * end:
1478          */
1479         if (!sync) {
1480                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1481                 atomic_set(&dio->cl.remaining,
1482                            CLOSURE_REMAINING_INITIALIZER -
1483                            CLOSURE_RUNNING +
1484                            CLOSURE_DESTRUCTOR);
1485         } else {
1486                 atomic_set(&dio->cl.remaining,
1487                            CLOSURE_REMAINING_INITIALIZER + 1);
1488         }
1489
1490         dio->req        = req;
1491         dio->ret        = ret;
1492
1493         goto start;
1494         while (iter->count) {
1495                 bio = bio_alloc_bioset(GFP_KERNEL,
1496                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1497                                        &c->bio_read);
1498                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1499 start:
1500                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1501                 bio->bi_iter.bi_sector  = offset >> 9;
1502                 bio->bi_private         = dio;
1503
1504                 ret = bio_iov_iter_get_pages(bio, iter);
1505                 if (ret < 0) {
1506                         /* XXX: fault inject this path */
1507                         bio->bi_status = BLK_STS_RESOURCE;
1508                         bio_endio(bio);
1509                         break;
1510                 }
1511
1512                 offset += bio->bi_iter.bi_size;
1513                 bio_set_pages_dirty(bio);
1514
1515                 if (iter->count)
1516                         closure_get(&dio->cl);
1517
1518                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1519         }
1520
1521         if (sync) {
1522                 closure_sync(&dio->cl);
1523                 closure_debug_destroy(&dio->cl);
1524                 ret = dio->ret;
1525                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1526                 return ret;
1527         } else {
1528                 return -EIOCBQUEUED;
1529         }
1530 }
1531
1532 static void bch2_dio_write_loop_async(struct closure *);
1533
1534 static long bch2_dio_write_loop(struct dio_write *dio)
1535 {
1536         struct kiocb *req = dio->req;
1537         struct file *file = req->ki_filp;
1538         struct address_space *mapping = file->f_mapping;
1539         struct bch_inode_info *inode = file_bch_inode(file);
1540         struct bio *bio = &dio->iop.op.wbio.bio;
1541         struct bio_vec *bv;
1542         bool sync;
1543         long ret;
1544         int i;
1545
1546         if (dio->loop)
1547                 goto loop;
1548
1549         inode_dio_begin(&inode->v);
1550         __pagecache_block_get(&mapping->add_lock);
1551
1552         /* Write and invalidate pagecache range that we're writing to: */
1553         ret = write_invalidate_inode_pages_range(mapping, req->ki_pos,
1554                                 req->ki_pos + iov_iter_count(&dio->iter) - 1);
1555         if (unlikely(ret))
1556                 goto err;
1557
1558         while (1) {
1559                 BUG_ON(current->pagecache_lock);
1560                 current->pagecache_lock = &mapping->add_lock;
1561                 if (current != dio->task)
1562                         use_mm(dio->task->mm);
1563
1564                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1565
1566                 if (current != dio->task)
1567                         unuse_mm(dio->task->mm);
1568                 current->pagecache_lock = NULL;
1569
1570                 if (unlikely(ret < 0))
1571                         goto err;
1572
1573                 dio->iop.op.pos = POS(inode->v.i_ino,
1574                                 (req->ki_pos >> 9) + dio->iop.op.written);
1575
1576                 task_io_account_write(bio->bi_iter.bi_size);
1577
1578                 closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl);
1579
1580                 if (!dio->sync && !dio->loop && dio->iter.count) {
1581                         struct iovec *iov = dio->inline_vecs;
1582
1583                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1584                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1585                                               GFP_KERNEL);
1586                                 if (unlikely(!iov)) {
1587                                         dio->iop.op.error = -ENOMEM;
1588                                         goto err_wait_io;
1589                                 }
1590
1591                                 dio->free_iov = true;
1592                         }
1593
1594                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1595                         dio->iter.iov = iov;
1596                 }
1597 err_wait_io:
1598                 dio->loop = true;
1599
1600                 if (!dio->sync) {
1601                         continue_at_noreturn(&dio->cl,
1602                                         bch2_dio_write_loop_async, NULL);
1603                         return -EIOCBQUEUED;
1604                 }
1605
1606                 closure_sync(&dio->cl);
1607 loop:
1608                 bio_for_each_segment_all(bv, bio, i)
1609                         put_page(bv->bv_page);
1610                 if (!dio->iter.count || dio->iop.op.error)
1611                         break;
1612                 bio_reset(bio);
1613         }
1614
1615         ret = dio->iop.op.error ?: ((long) dio->iop.op.written << 9);
1616 err:
1617         __pagecache_block_put(&mapping->add_lock);
1618         inode_dio_end(&inode->v);
1619         bch2_disk_reservation_put(dio->iop.op.c, &dio->iop.op.res);
1620
1621         if (dio->free_iov)
1622                 kfree(dio->iter.iov);
1623
1624         closure_debug_destroy(&dio->cl);
1625
1626         sync = dio->sync;
1627         bio_put(bio);
1628
1629         if (!sync) {
1630                 req->ki_complete(req, ret, 0);
1631                 ret = -EIOCBQUEUED;
1632         }
1633         return ret;
1634 }
1635
1636 static void bch2_dio_write_loop_async(struct closure *cl)
1637 {
1638         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1639
1640         bch2_dio_write_loop(dio);
1641 }
1642
1643 static int bch2_direct_IO_write(struct kiocb *req,
1644                                 struct iov_iter *iter,
1645                                 bool swap)
1646 {
1647         struct file *file = req->ki_filp;
1648         struct bch_inode_info *inode = file_bch_inode(file);
1649         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1650         struct dio_write *dio;
1651         struct bio *bio;
1652         loff_t offset = req->ki_pos;
1653         ssize_t ret;
1654
1655         lockdep_assert_held(&inode->v.i_rwsem);
1656
1657         if (unlikely(!iter->count))
1658                 return 0;
1659
1660         if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
1661                 return -EINVAL;
1662
1663         bio = bio_alloc_bioset(GFP_KERNEL,
1664                                iov_iter_npages(iter, BIO_MAX_PAGES),
1665                                &c->dio_write_bioset);
1666         dio = container_of(bio, struct dio_write, iop.op.wbio.bio);
1667         closure_init(&dio->cl, NULL);
1668         dio->req                = req;
1669         dio->task               = current;
1670         dio->loop               = false;
1671         dio->sync               = is_sync_kiocb(req) ||
1672                 offset + iter->count > inode->v.i_size;
1673         dio->free_iov           = false;
1674         dio->iter               = *iter;
1675         bch2_fswrite_op_init(&dio->iop, c, inode, io_opts(c, inode), true);
1676         dio->iop.op.write_point = writepoint_hashed((unsigned long) dio->task);
1677         dio->iop.op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1678
1679         if ((req->ki_flags & IOCB_DSYNC) &&
1680             !c->opts.journal_flush_disabled)
1681                 dio->iop.op.flags |= BCH_WRITE_FLUSH;
1682
1683         ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9, 0);
1684         if (unlikely(ret)) {
1685                 if (bch2_check_range_allocated(c, POS(inode->v.i_ino,
1686                                                       offset >> 9),
1687                                                iter->count >> 9))
1688                         goto err;
1689
1690                 dio->iop.unalloc = true;
1691         }
1692
1693         dio->iop.op.nr_replicas = dio->iop.op.res.nr_replicas;
1694
1695         return bch2_dio_write_loop(dio);
1696 err:
1697         bch2_disk_reservation_put(c, &dio->iop.op.res);
1698         closure_debug_destroy(&dio->cl);
1699         bio_put(bio);
1700         return ret;
1701 }
1702
1703 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1704 {
1705         struct blk_plug plug;
1706         ssize_t ret;
1707
1708         blk_start_plug(&plug);
1709         ret = iov_iter_rw(iter) == WRITE
1710                 ? bch2_direct_IO_write(req, iter, false)
1711                 : bch2_direct_IO_read(req, iter);
1712         blk_finish_plug(&plug);
1713
1714         return ret;
1715 }
1716
1717 static ssize_t
1718 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1719 {
1720         return bch2_direct_IO_write(iocb, iter, true);
1721 }
1722
1723 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1724 {
1725         struct file *file = iocb->ki_filp;
1726         struct bch_inode_info *inode = file_bch_inode(file);
1727         ssize_t ret;
1728
1729         /* We can write back this queue in page reclaim */
1730         current->backing_dev_info = inode_to_bdi(&inode->v);
1731         ret = file_remove_privs(file);
1732         if (ret)
1733                 goto out;
1734
1735         ret = file_update_time(file);
1736         if (ret)
1737                 goto out;
1738
1739         ret = iocb->ki_flags & IOCB_DIRECT
1740                 ? bch2_direct_write(iocb, from)
1741                 : generic_perform_write(file, from, iocb->ki_pos);
1742
1743         if (likely(ret > 0))
1744                 iocb->ki_pos += ret;
1745 out:
1746         current->backing_dev_info = NULL;
1747         return ret;
1748 }
1749
1750 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1751 {
1752         struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp);
1753         bool direct = iocb->ki_flags & IOCB_DIRECT;
1754         ssize_t ret;
1755
1756         inode_lock(&inode->v);
1757         ret = generic_write_checks(iocb, from);
1758         if (ret > 0)
1759                 ret = __bch2_write_iter(iocb, from);
1760         inode_unlock(&inode->v);
1761
1762         if (ret > 0 && !direct)
1763                 ret = generic_write_sync(iocb, ret);
1764
1765         return ret;
1766 }
1767
1768 int bch2_page_mkwrite(struct vm_fault *vmf)
1769 {
1770         struct page *page = vmf->page;
1771         struct file *file = vmf->vma->vm_file;
1772         struct bch_inode_info *inode = file_bch_inode(file);
1773         struct address_space *mapping = inode->v.i_mapping;
1774         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1775         int ret = VM_FAULT_LOCKED;
1776
1777         sb_start_pagefault(inode->v.i_sb);
1778         file_update_time(file);
1779
1780         /*
1781          * Not strictly necessary, but helps avoid dio writes livelocking in
1782          * write_invalidate_inode_pages_range() - can drop this if/when we get
1783          * a write_invalidate_inode_pages_range() that works without dropping
1784          * page lock before invalidating page
1785          */
1786         if (current->pagecache_lock != &mapping->add_lock)
1787                 pagecache_add_get(&mapping->add_lock);
1788
1789         lock_page(page);
1790         if (page->mapping != mapping ||
1791             page_offset(page) > i_size_read(&inode->v)) {
1792                 unlock_page(page);
1793                 ret = VM_FAULT_NOPAGE;
1794                 goto out;
1795         }
1796
1797         if (bch2_get_page_reservation(c, page, true)) {
1798                 unlock_page(page);
1799                 ret = VM_FAULT_SIGBUS;
1800                 goto out;
1801         }
1802
1803         if (!PageDirty(page))
1804                 set_page_dirty(page);
1805         wait_for_stable_page(page);
1806 out:
1807         if (current->pagecache_lock != &mapping->add_lock)
1808                 pagecache_add_put(&mapping->add_lock);
1809         sb_end_pagefault(inode->v.i_sb);
1810         return ret;
1811 }
1812
1813 void bch2_invalidatepage(struct page *page, unsigned int offset,
1814                          unsigned int length)
1815 {
1816         EBUG_ON(!PageLocked(page));
1817         EBUG_ON(PageWriteback(page));
1818
1819         if (offset || length < PAGE_SIZE)
1820                 return;
1821
1822         bch2_clear_page_bits(page);
1823 }
1824
1825 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
1826 {
1827         EBUG_ON(!PageLocked(page));
1828         EBUG_ON(PageWriteback(page));
1829
1830         if (PageDirty(page))
1831                 return 0;
1832
1833         bch2_clear_page_bits(page);
1834         return 1;
1835 }
1836
1837 #ifdef CONFIG_MIGRATION
1838 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
1839                       struct page *page, enum migrate_mode mode)
1840 {
1841         int ret;
1842
1843         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
1844         if (ret != MIGRATEPAGE_SUCCESS)
1845                 return ret;
1846
1847         if (PagePrivate(page)) {
1848                 *page_state(newpage) = *page_state(page);
1849                 ClearPagePrivate(page);
1850         }
1851
1852         migrate_page_copy(newpage, page);
1853         return MIGRATEPAGE_SUCCESS;
1854 }
1855 #endif
1856
1857 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1858 {
1859         struct bch_inode_info *inode = file_bch_inode(file);
1860         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1861         int ret;
1862
1863         ret = filemap_write_and_wait_range(inode->v.i_mapping, start, end);
1864         if (ret)
1865                 return ret;
1866
1867         if (c->opts.journal_flush_disabled)
1868                 return 0;
1869
1870         return bch2_journal_flush_seq(&c->journal, inode->ei_journal_seq);
1871 }
1872
1873 static int __bch2_truncate_page(struct bch_inode_info *inode,
1874                                 pgoff_t index, loff_t start, loff_t end)
1875 {
1876         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1877         struct address_space *mapping = inode->v.i_mapping;
1878         unsigned start_offset = start & (PAGE_SIZE - 1);
1879         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
1880         struct page *page;
1881         int ret = 0;
1882
1883         /* Page boundary? Nothing to do */
1884         if (!((index == start >> PAGE_SHIFT && start_offset) ||
1885               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
1886                 return 0;
1887
1888         /* Above i_size? */
1889         if (index << PAGE_SHIFT >= inode->v.i_size)
1890                 return 0;
1891
1892         page = find_lock_page(mapping, index);
1893         if (!page) {
1894                 struct btree_iter iter;
1895                 struct bkey_s_c k = bkey_s_c_null;
1896
1897                 /*
1898                  * XXX: we're doing two index lookups when we end up reading the
1899                  * page
1900                  */
1901                 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
1902                                    POS(inode->v.i_ino,
1903                                        index << PAGE_SECTOR_SHIFT), 0, k) {
1904                         if (bkey_cmp(bkey_start_pos(k.k),
1905                                      POS(inode->v.i_ino,
1906                                          (index + 1) << PAGE_SECTOR_SHIFT)) >= 0)
1907                                 break;
1908
1909                         if (k.k->type != KEY_TYPE_DISCARD &&
1910                             k.k->type != BCH_RESERVATION) {
1911                                 bch2_btree_iter_unlock(&iter);
1912                                 goto create;
1913                         }
1914                 }
1915                 bch2_btree_iter_unlock(&iter);
1916                 return 0;
1917 create:
1918                 page = find_or_create_page(mapping, index, GFP_KERNEL);
1919                 if (unlikely(!page)) {
1920                         ret = -ENOMEM;
1921                         goto out;
1922                 }
1923         }
1924
1925         if (!PageUptodate(page)) {
1926                 ret = bch2_read_single_page(page, mapping);
1927                 if (ret)
1928                         goto unlock;
1929         }
1930
1931         /*
1932          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
1933          *
1934          * XXX: because we aren't currently tracking whether the page has actual
1935          * data in it (vs. just 0s, or only partially written) this wrong. ick.
1936          */
1937         ret = bch2_get_page_reservation(c, page, false);
1938         BUG_ON(ret);
1939
1940         if (index == start >> PAGE_SHIFT &&
1941             index == end >> PAGE_SHIFT)
1942                 zero_user_segment(page, start_offset, end_offset);
1943         else if (index == start >> PAGE_SHIFT)
1944                 zero_user_segment(page, start_offset, PAGE_SIZE);
1945         else if (index == end >> PAGE_SHIFT)
1946                 zero_user_segment(page, 0, end_offset);
1947
1948         if (!PageDirty(page))
1949                 set_page_dirty(page);
1950 unlock:
1951         unlock_page(page);
1952         put_page(page);
1953 out:
1954         return ret;
1955 }
1956
1957 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
1958 {
1959         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
1960                                     from, from + PAGE_SIZE);
1961 }
1962
1963 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
1964 {
1965         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1966         struct address_space *mapping = inode->v.i_mapping;
1967         bool shrink = iattr->ia_size <= inode->v.i_size;
1968         struct i_sectors_hook i_sectors_hook =
1969                 i_sectors_hook_init(inode, BCH_INODE_I_SIZE_DIRTY);
1970         int ret = 0;
1971
1972         inode_dio_wait(&inode->v);
1973         pagecache_block_get(&mapping->add_lock);
1974
1975         truncate_setsize(&inode->v, iattr->ia_size);
1976
1977         /* sync appends.. */
1978         /* XXX what protects inode->i_size? */
1979         if (iattr->ia_size > inode->ei_inode.bi_size)
1980                 ret = filemap_write_and_wait_range(mapping,
1981                                                    inode->ei_inode.bi_size, S64_MAX);
1982         if (ret)
1983                 goto err_put_pagecache;
1984
1985         i_sectors_hook.new_i_size = iattr->ia_size;
1986
1987         ret = i_sectors_dirty_start(c, &i_sectors_hook);
1988         if (unlikely(ret))
1989                 goto err;
1990
1991         /*
1992          * There might be persistent reservations (from fallocate())
1993          * above i_size, which bch2_inode_truncate() will discard - we're
1994          * only supposed to discard them if we're doing a real truncate
1995          * here (new i_size < current i_size):
1996          */
1997         if (shrink) {
1998                 ret = bch2_truncate_page(inode, iattr->ia_size);
1999                 if (unlikely(ret))
2000                         goto err;
2001
2002                 ret = bch2_inode_truncate(c, inode->v.i_ino,
2003                                           round_up(iattr->ia_size, PAGE_SIZE) >> 9,
2004                                           &i_sectors_hook.hook,
2005                                           &inode->ei_journal_seq);
2006                 if (unlikely(ret))
2007                         goto err;
2008         }
2009
2010         setattr_copy(&inode->v, iattr);
2011         inode->v.i_mtime = inode->v.i_ctime = current_time(&inode->v);
2012 err:
2013         /*
2014          * On error - in particular, bch2_truncate_page() error - don't clear
2015          * I_SIZE_DIRTY, as we've left data above i_size!:
2016          */
2017         if (ret)
2018                 i_sectors_hook.flags &= ~BCH_INODE_I_SIZE_DIRTY;
2019
2020         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2021 err_put_pagecache:
2022         pagecache_block_put(&mapping->add_lock);
2023         return ret;
2024 }
2025
2026 static long bch2_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2027 {
2028         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2029         struct address_space *mapping = inode->v.i_mapping;
2030         u64 ino = inode->v.i_ino;
2031         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
2032         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
2033         int ret = 0;
2034
2035         inode_lock(&inode->v);
2036         inode_dio_wait(&inode->v);
2037         pagecache_block_get(&mapping->add_lock);
2038
2039         ret = __bch2_truncate_page(inode,
2040                                    offset >> PAGE_SHIFT,
2041                                    offset, offset + len);
2042         if (unlikely(ret))
2043                 goto err;
2044
2045         if (offset >> PAGE_SHIFT !=
2046             (offset + len) >> PAGE_SHIFT) {
2047                 ret = __bch2_truncate_page(inode,
2048                                            (offset + len) >> PAGE_SHIFT,
2049                                            offset, offset + len);
2050                 if (unlikely(ret))
2051                         goto err;
2052         }
2053
2054         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2055
2056         if (discard_start < discard_end) {
2057                 struct disk_reservation disk_res;
2058                 struct i_sectors_hook i_sectors_hook =
2059                         i_sectors_hook_init(inode, 0);
2060                 int ret;
2061
2062                 ret = i_sectors_dirty_start(c, &i_sectors_hook);
2063                 if (unlikely(ret))
2064                         goto err;
2065
2066                 /*
2067                  * We need to pass in a disk reservation here because we might
2068                  * be splitting a compressed extent into two. This isn't a
2069                  * problem with truncate because truncate will never split an
2070                  * extent, only truncate it...
2071                  */
2072                 ret = bch2_disk_reservation_get(c, &disk_res, 0, 0);
2073                 BUG_ON(ret);
2074
2075                 ret = bch2_btree_delete_range(c,
2076                                 BTREE_ID_EXTENTS,
2077                                 POS(ino, discard_start),
2078                                 POS(ino, discard_end),
2079                                 ZERO_VERSION,
2080                                 &disk_res,
2081                                 &i_sectors_hook.hook,
2082                                 &inode->ei_journal_seq);
2083                 bch2_disk_reservation_put(c, &disk_res);
2084
2085                 ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2086         }
2087 err:
2088         pagecache_block_put(&mapping->add_lock);
2089         inode_unlock(&inode->v);
2090
2091         return ret;
2092 }
2093
2094 static long bch2_fcollapse(struct bch_inode_info *inode,
2095                            loff_t offset, loff_t len)
2096 {
2097         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2098         struct address_space *mapping = inode->v.i_mapping;
2099         struct btree_iter src;
2100         struct btree_iter dst;
2101         BKEY_PADDED(k) copy;
2102         struct bkey_s_c k;
2103         struct i_sectors_hook i_sectors_hook = i_sectors_hook_init(inode, 0);
2104         loff_t new_size;
2105         int ret;
2106
2107         if ((offset | len) & (PAGE_SIZE - 1))
2108                 return -EINVAL;
2109
2110         bch2_btree_iter_init(&dst, c, BTREE_ID_EXTENTS,
2111                              POS(inode->v.i_ino, offset >> 9),
2112                              BTREE_ITER_INTENT);
2113         /* position will be set from dst iter's position: */
2114         bch2_btree_iter_init(&src, c, BTREE_ID_EXTENTS, POS_MIN, 0);
2115         bch2_btree_iter_link(&src, &dst);
2116
2117         /*
2118          * We need i_mutex to keep the page cache consistent with the extents
2119          * btree, and the btree consistent with i_size - we don't need outside
2120          * locking for the extents btree itself, because we're using linked
2121          * iterators
2122          */
2123         inode_lock(&inode->v);
2124         inode_dio_wait(&inode->v);
2125         pagecache_block_get(&mapping->add_lock);
2126
2127         ret = -EINVAL;
2128         if (offset + len >= inode->v.i_size)
2129                 goto err;
2130
2131         if (inode->v.i_size < len)
2132                 goto err;
2133
2134         new_size = inode->v.i_size - len;
2135
2136         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2137         if (ret)
2138                 goto err;
2139
2140         ret = i_sectors_dirty_start(c, &i_sectors_hook);
2141         if (ret)
2142                 goto err;
2143
2144         while (bkey_cmp(dst.pos,
2145                         POS(inode->v.i_ino,
2146                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2147                 struct disk_reservation disk_res;
2148
2149                 bch2_btree_iter_set_pos(&src,
2150                         POS(dst.pos.inode, dst.pos.offset + (len >> 9)));
2151
2152                 ret = bch2_btree_iter_traverse(&dst);
2153                 if (ret)
2154                         goto btree_iter_err;
2155
2156                 k = bch2_btree_iter_peek_with_holes(&src);
2157                 if ((ret = btree_iter_err(k)))
2158                         goto btree_iter_err;
2159
2160                 bkey_reassemble(&copy.k, k);
2161
2162                 if (bkey_deleted(&copy.k.k))
2163                         copy.k.k.type = KEY_TYPE_DISCARD;
2164
2165                 bch2_cut_front(src.pos, &copy.k);
2166                 copy.k.k.p.offset -= len >> 9;
2167
2168                 BUG_ON(bkey_cmp(dst.pos, bkey_start_pos(&copy.k.k)));
2169
2170                 ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size,
2171                                                BCH_DISK_RESERVATION_NOFAIL);
2172                 BUG_ON(ret);
2173
2174                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2175                                            &inode->ei_journal_seq,
2176                                            BTREE_INSERT_ATOMIC|
2177                                            BTREE_INSERT_NOFAIL,
2178                                            BTREE_INSERT_ENTRY(&dst, &copy.k));
2179                 bch2_disk_reservation_put(c, &disk_res);
2180 btree_iter_err:
2181                 if (ret == -EINTR)
2182                         ret = 0;
2183                 if (ret)
2184                         goto err_put_sectors_dirty;
2185                 /*
2186                  * XXX: if we error here we've left data with multiple
2187                  * pointers... which isn't a _super_ serious problem...
2188                  */
2189
2190                 bch2_btree_iter_cond_resched(&src);
2191         }
2192
2193         bch2_btree_iter_unlock(&src);
2194         bch2_btree_iter_unlock(&dst);
2195
2196         ret = bch2_inode_truncate(c, inode->v.i_ino,
2197                                  round_up(new_size, PAGE_SIZE) >> 9,
2198                                  &i_sectors_hook.hook,
2199                                  &inode->ei_journal_seq);
2200         if (ret)
2201                 goto err_put_sectors_dirty;
2202
2203         i_size_write(&inode->v, new_size);
2204         i_sectors_hook.new_i_size = new_size;
2205 err_put_sectors_dirty:
2206         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2207 err:
2208         pagecache_block_put(&mapping->add_lock);
2209         inode_unlock(&inode->v);
2210
2211         bch2_btree_iter_unlock(&src);
2212         bch2_btree_iter_unlock(&dst);
2213         return ret;
2214 }
2215
2216 static long bch2_fallocate(struct bch_inode_info *inode, int mode,
2217                            loff_t offset, loff_t len)
2218 {
2219         struct address_space *mapping = inode->v.i_mapping;
2220         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2221         struct i_sectors_hook i_sectors_hook = i_sectors_hook_init(inode, 0);
2222         struct btree_iter iter;
2223         struct bpos end_pos;
2224         loff_t block_start, block_end;
2225         loff_t end = offset + len;
2226         unsigned sectors;
2227         unsigned replicas = READ_ONCE(c->opts.data_replicas);
2228         int ret;
2229
2230         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
2231                              BTREE_ITER_INTENT);
2232
2233         inode_lock(&inode->v);
2234         inode_dio_wait(&inode->v);
2235         pagecache_block_get(&mapping->add_lock);
2236
2237         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2238                 ret = inode_newsize_ok(&inode->v, end);
2239                 if (ret)
2240                         goto err;
2241         }
2242
2243         if (mode & FALLOC_FL_ZERO_RANGE) {
2244                 ret = __bch2_truncate_page(inode,
2245                                            offset >> PAGE_SHIFT,
2246                                            offset, end);
2247
2248                 if (!ret &&
2249                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2250                         ret = __bch2_truncate_page(inode,
2251                                                    end >> PAGE_SHIFT,
2252                                                    offset, end);
2253
2254                 if (unlikely(ret))
2255                         goto err;
2256
2257                 truncate_pagecache_range(&inode->v, offset, end - 1);
2258
2259                 block_start     = round_up(offset, PAGE_SIZE);
2260                 block_end       = round_down(end, PAGE_SIZE);
2261         } else {
2262                 block_start     = round_down(offset, PAGE_SIZE);
2263                 block_end       = round_up(end, PAGE_SIZE);
2264         }
2265
2266         bch2_btree_iter_set_pos(&iter, POS(inode->v.i_ino, block_start >> 9));
2267         end_pos = POS(inode->v.i_ino, block_end >> 9);
2268
2269         ret = i_sectors_dirty_start(c, &i_sectors_hook);
2270         if (unlikely(ret))
2271                 goto err;
2272
2273         while (bkey_cmp(iter.pos, end_pos) < 0) {
2274                 struct disk_reservation disk_res = { 0 };
2275                 struct bkey_i_reservation reservation;
2276                 struct bkey_s_c k;
2277
2278                 k = bch2_btree_iter_peek_with_holes(&iter);
2279                 if ((ret = btree_iter_err(k)))
2280                         goto btree_iter_err;
2281
2282                 /* already reserved */
2283                 if (k.k->type == BCH_RESERVATION &&
2284                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2285                         bch2_btree_iter_advance_pos(&iter);
2286                         continue;
2287                 }
2288
2289                 if (bkey_extent_is_data(k.k)) {
2290                         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
2291                                 bch2_btree_iter_advance_pos(&iter);
2292                                 continue;
2293                         }
2294                 }
2295
2296                 bkey_reservation_init(&reservation.k_i);
2297                 reservation.k.type      = BCH_RESERVATION;
2298                 reservation.k.p         = k.k->p;
2299                 reservation.k.size      = k.k->size;
2300
2301                 bch2_cut_front(iter.pos, &reservation.k_i);
2302                 bch2_cut_back(end_pos, &reservation.k);
2303
2304                 sectors = reservation.k.size;
2305                 reservation.v.nr_replicas = bch2_extent_nr_dirty_ptrs(k);
2306
2307                 if (reservation.v.nr_replicas < replicas ||
2308                     bch2_extent_is_compressed(k)) {
2309                         ret = bch2_disk_reservation_get(c, &disk_res,
2310                                                        sectors, 0);
2311                         if (ret)
2312                                 goto err_put_sectors_dirty;
2313
2314                         reservation.v.nr_replicas = disk_res.nr_replicas;
2315                 }
2316
2317                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2318                                           &inode->ei_journal_seq,
2319                                           BTREE_INSERT_ATOMIC|
2320                                           BTREE_INSERT_NOFAIL,
2321                                           BTREE_INSERT_ENTRY(&iter, &reservation.k_i));
2322                 bch2_disk_reservation_put(c, &disk_res);
2323 btree_iter_err:
2324                 if (ret < 0 && ret != -EINTR)
2325                         goto err_put_sectors_dirty;
2326
2327         }
2328         bch2_btree_iter_unlock(&iter);
2329
2330         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2331
2332         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2333             end > inode->v.i_size) {
2334                 i_size_write(&inode->v, end);
2335
2336                 mutex_lock(&inode->ei_update_lock);
2337                 ret = bch2_write_inode_size(c, inode, inode->v.i_size);
2338                 mutex_unlock(&inode->ei_update_lock);
2339         }
2340
2341         /* blech */
2342         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2343             (mode & FALLOC_FL_ZERO_RANGE) &&
2344             inode->ei_inode.bi_size != inode->v.i_size) {
2345                 /* sync appends.. */
2346                 ret = filemap_write_and_wait_range(mapping,
2347                                         inode->ei_inode.bi_size, S64_MAX);
2348                 if (ret)
2349                         goto err;
2350
2351                 if (inode->ei_inode.bi_size != inode->v.i_size) {
2352                         mutex_lock(&inode->ei_update_lock);
2353                         ret = bch2_write_inode_size(c, inode, inode->v.i_size);
2354                         mutex_unlock(&inode->ei_update_lock);
2355                 }
2356         }
2357
2358         pagecache_block_put(&mapping->add_lock);
2359         inode_unlock(&inode->v);
2360
2361         return 0;
2362 err_put_sectors_dirty:
2363         ret = i_sectors_dirty_finish(c, &i_sectors_hook) ?: ret;
2364 err:
2365         bch2_btree_iter_unlock(&iter);
2366         pagecache_block_put(&mapping->add_lock);
2367         inode_unlock(&inode->v);
2368         return ret;
2369 }
2370
2371 long bch2_fallocate_dispatch(struct file *file, int mode,
2372                              loff_t offset, loff_t len)
2373 {
2374         struct bch_inode_info *inode = file_bch_inode(file);
2375
2376         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2377                 return bch2_fallocate(inode, mode, offset, len);
2378
2379         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2380                 return bch2_fpunch(inode, offset, len);
2381
2382         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2383                 return bch2_fcollapse(inode, offset, len);
2384
2385         return -EOPNOTSUPP;
2386 }
2387
2388 static bool page_is_data(struct page *page)
2389 {
2390         /* XXX: should only have to check PageDirty */
2391         return PagePrivate(page) &&
2392                 (page_state(page)->sectors ||
2393                  page_state(page)->dirty_sectors);
2394 }
2395
2396 static loff_t bch2_next_pagecache_data(struct inode *vinode,
2397                                        loff_t start_offset,
2398                                        loff_t end_offset)
2399 {
2400         struct address_space *mapping = vinode->i_mapping;
2401         struct page *page;
2402         pgoff_t index;
2403
2404         for (index = start_offset >> PAGE_SHIFT;
2405              index < end_offset >> PAGE_SHIFT;
2406              index++) {
2407                 if (find_get_pages(mapping, index, 1, &page)) {
2408                         lock_page(page);
2409                         index = page->index;
2410
2411                         if (page_is_data(page))
2412                                 end_offset =
2413                                         min(end_offset,
2414                                         max(start_offset,
2415                                             ((loff_t) index) << PAGE_SHIFT));
2416                         unlock_page(page);
2417                         put_page(page);
2418                 } else {
2419                         break;
2420                 }
2421         }
2422
2423         return end_offset;
2424 }
2425
2426 static loff_t bch2_seek_data(struct file *file, u64 offset)
2427 {
2428         struct bch_inode_info *inode = file_bch_inode(file);
2429         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2430         struct btree_iter iter;
2431         struct bkey_s_c k;
2432         u64 isize, next_data = MAX_LFS_FILESIZE;
2433         int ret;
2434
2435         isize = i_size_read(&inode->v);
2436         if (offset >= isize)
2437                 return -ENXIO;
2438
2439         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2440                            POS(inode->v.i_ino, offset >> 9), 0, k) {
2441                 if (k.k->p.inode != inode->v.i_ino) {
2442                         break;
2443                 } else if (bkey_extent_is_data(k.k)) {
2444                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2445                         break;
2446                 } else if (k.k->p.offset >> 9 > isize)
2447                         break;
2448         }
2449
2450         ret = bch2_btree_iter_unlock(&iter);
2451         if (ret)
2452                 return ret;
2453
2454         if (next_data > offset)
2455                 next_data = bch2_next_pagecache_data(&inode->v,
2456                                                      offset, next_data);
2457
2458         if (next_data > isize)
2459                 return -ENXIO;
2460
2461         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2462 }
2463
2464 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2465 {
2466         struct page *page;
2467         bool ret;
2468
2469         page = find_lock_entry(mapping, index);
2470         if (!page || radix_tree_exception(page))
2471                 return false;
2472
2473         ret = page_is_data(page);
2474         unlock_page(page);
2475
2476         return ret;
2477 }
2478
2479 static loff_t bch2_next_pagecache_hole(struct inode *vinode,
2480                                        loff_t start_offset,
2481                                        loff_t end_offset)
2482 {
2483         struct address_space *mapping = vinode->i_mapping;
2484         pgoff_t index;
2485
2486         for (index = start_offset >> PAGE_SHIFT;
2487              index < end_offset >> PAGE_SHIFT;
2488              index++)
2489                 if (!page_slot_is_data(mapping, index))
2490                         end_offset = max(start_offset,
2491                                          ((loff_t) index) << PAGE_SHIFT);
2492
2493         return end_offset;
2494 }
2495
2496 static loff_t bch2_seek_hole(struct file *file, u64 offset)
2497 {
2498         struct bch_inode_info *inode = file_bch_inode(file);
2499         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2500         struct btree_iter iter;
2501         struct bkey_s_c k;
2502         u64 isize, next_hole = MAX_LFS_FILESIZE;
2503         int ret;
2504
2505         isize = i_size_read(&inode->v);
2506         if (offset >= isize)
2507                 return -ENXIO;
2508
2509         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2510                            POS(inode->v.i_ino, offset >> 9),
2511                            BTREE_ITER_WITH_HOLES, k) {
2512                 if (k.k->p.inode != inode->v.i_ino) {
2513                         next_hole = bch2_next_pagecache_hole(&inode->v,
2514                                         offset, MAX_LFS_FILESIZE);
2515                         break;
2516                 } else if (!bkey_extent_is_data(k.k)) {
2517                         next_hole = bch2_next_pagecache_hole(&inode->v,
2518                                         max(offset, bkey_start_offset(k.k) << 9),
2519                                         k.k->p.offset << 9);
2520
2521                         if (next_hole < k.k->p.offset << 9)
2522                                 break;
2523                 } else {
2524                         offset = max(offset, bkey_start_offset(k.k) << 9);
2525                 }
2526         }
2527
2528         ret = bch2_btree_iter_unlock(&iter);
2529         if (ret)
2530                 return ret;
2531
2532         if (next_hole > isize)
2533                 next_hole = isize;
2534
2535         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2536 }
2537
2538 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
2539 {
2540         switch (whence) {
2541         case SEEK_SET:
2542         case SEEK_CUR:
2543         case SEEK_END:
2544                 return generic_file_llseek(file, offset, whence);
2545         case SEEK_DATA:
2546                 return bch2_seek_data(file, offset);
2547         case SEEK_HOLE:
2548                 return bch2_seek_hole(file, offset);
2549         }
2550
2551         return -EINVAL;
2552 }
2553
2554 void bch2_fs_fsio_exit(struct bch_fs *c)
2555 {
2556         bioset_exit(&c->dio_write_bioset);
2557         bioset_exit(&c->dio_read_bioset);
2558         bioset_exit(&c->writepage_bioset);
2559 }
2560
2561 int bch2_fs_fsio_init(struct bch_fs *c)
2562 {
2563         if (bioset_init(&c->writepage_bioset,
2564                         4, offsetof(struct bch_writepage_io, op.op.wbio.bio),
2565                         BIOSET_NEED_BVECS) ||
2566             bioset_init(&c->dio_read_bioset,
2567                         4, offsetof(struct dio_read, rbio.bio),
2568                         BIOSET_NEED_BVECS) ||
2569             bioset_init(&c->dio_write_bioset,
2570                         4, offsetof(struct dio_write, iop.op.wbio.bio),
2571                         BIOSET_NEED_BVECS))
2572                 return -ENOMEM;
2573
2574         return 0;
2575 }
2576
2577 #endif /* NO_BCACHEFS_FS */