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

   1
   2 #include "bcachefs.h"
   3 #include "btree_update.h"
   4 #include "buckets.h"
   5 #include "clock.h"
   6 #include "error.h"
   7 #include "fs.h"
   8 #include "fs-gc.h"
   9 #include "fs-io.h"
  10 #include "inode.h"
  11 #include "journal.h"
  12 #include "io.h"
  13 #include "keylist.h"
  14
  15 #include <linux/aio.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/falloc.h>
  18 #include <linux/migrate.h>
  19 #include <linux/mmu_context.h>
  20 #include <linux/pagevec.h>
  21 #include <linux/task_io_accounting_ops.h>
  22 #include <linux/uio.h>
  23 #include <linux/writeback.h>
  24 #include <trace/events/writeback.h>
  25
  26 struct bio_set *bch2_writepage_bioset;
  27 struct bio_set *bch2_dio_read_bioset;
  28 struct bio_set *bch2_dio_write_bioset;
  29
  30 /* pagecache_block must be held */
  31 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  32                                               loff_t start, loff_t end)
  33 {
  34         int ret;
  35
  36         /*
  37          * XXX: the way this is currently implemented, we can spin if a process
  38          * is continually redirtying a specific page
  39          */
  40         do {
  41                 if (!mapping->nrpages &&
  42                     !mapping->nrexceptional)
  43                         return 0;
  44
  45                 ret = filemap_write_and_wait_range(mapping, start, end);
  46                 if (ret)
  47                         break;
  48
  49                 if (!mapping->nrpages)
  50                         return 0;
  51
  52                 ret = invalidate_inode_pages2_range(mapping,
  53                                 start >> PAGE_SHIFT,
  54                                 end >> PAGE_SHIFT);
  55         } while (ret == -EBUSY);
  56
  57         return ret;
  58 }
  59
  60 /* i_size updates: */
  61
  62 static int inode_set_size(struct bch_inode_info *ei,
  63                           struct bch_inode_unpacked *bi,
  64                           void *p)
  65 {
  66         loff_t *new_i_size = p;
  67
  68         lockdep_assert_held(&ei->update_lock);
  69
  70         bi->i_size = *new_i_size;
  71
  72         if (atomic_long_read(&ei->i_size_dirty_count))
  73                 bi->i_flags |= BCH_INODE_I_SIZE_DIRTY;
  74         else
  75                 bi->i_flags &= ~BCH_INODE_I_SIZE_DIRTY;
  76
  77         return 0;
  78 }
  79
  80 static int __must_check bch2_write_inode_size(struct bch_fs *c,
  81                                              struct bch_inode_info *ei,
  82                                              loff_t new_size)
  83 {
  84         return __bch2_write_inode(c, ei, inode_set_size, &new_size);
  85 }
  86
  87 static inline void i_size_dirty_put(struct bch_inode_info *ei)
  88 {
  89         atomic_long_dec_bug(&ei->i_size_dirty_count);
  90 }
  91
  92 static inline void i_size_dirty_get(struct bch_inode_info *ei)
  93 {
  94         lockdep_assert_held(&ei->vfs_inode.i_rwsem);
  95
  96         atomic_long_inc(&ei->i_size_dirty_count);
  97 }
  98
  99 /* i_sectors accounting: */
 100
 101 static enum extent_insert_hook_ret
 102 i_sectors_hook_fn(struct extent_insert_hook *hook,
 103                   struct bpos committed_pos,
 104                   struct bpos next_pos,
 105                   struct bkey_s_c k,
 106                   const struct bkey_i *insert)
 107 {
 108         struct i_sectors_hook *h = container_of(hook,
 109                                 struct i_sectors_hook, hook);
 110         s64 sectors = next_pos.offset - committed_pos.offset;
 111         int sign = bkey_extent_is_allocation(&insert->k) -
 112                 (k.k && bkey_extent_is_allocation(k.k));
 113
 114         EBUG_ON(!(h->ei->i_flags & BCH_INODE_I_SECTORS_DIRTY));
 115         EBUG_ON(!atomic_long_read(&h->ei->i_sectors_dirty_count));
 116
 117         h->sectors += sectors * sign;
 118
 119         return BTREE_HOOK_DO_INSERT;
 120 }
 121
 122 static int inode_set_i_sectors_dirty(struct bch_inode_info *ei,
 123                                     struct bch_inode_unpacked *bi, void *p)
 124 {
 125         BUG_ON(bi->i_flags & BCH_INODE_I_SECTORS_DIRTY);
 126
 127         bi->i_flags |= BCH_INODE_I_SECTORS_DIRTY;
 128         return 0;
 129 }
 130
 131 static int inode_clear_i_sectors_dirty(struct bch_inode_info *ei,
 132                                        struct bch_inode_unpacked *bi,
 133                                        void *p)
 134 {
 135         BUG_ON(!(bi->i_flags & BCH_INODE_I_SECTORS_DIRTY));
 136
 137         bi->i_sectors   = atomic64_read(&ei->i_sectors);
 138         bi->i_flags     &= ~BCH_INODE_I_SECTORS_DIRTY;
 139         return 0;
 140 }
 141
 142 static void i_sectors_dirty_put(struct bch_inode_info *ei,
 143                                 struct i_sectors_hook *h)
 144 {
 145         struct inode *inode = &ei->vfs_inode;
 146
 147         if (h->sectors) {
 148                 spin_lock(&inode->i_lock);
 149                 inode->i_blocks += h->sectors;
 150                 spin_unlock(&inode->i_lock);
 151
 152                 atomic64_add(h->sectors, &ei->i_sectors);
 153                 EBUG_ON(atomic64_read(&ei->i_sectors) < 0);
 154         }
 155
 156         EBUG_ON(atomic_long_read(&ei->i_sectors_dirty_count) <= 0);
 157
 158         mutex_lock(&ei->update_lock);
 159
 160         if (atomic_long_dec_and_test(&ei->i_sectors_dirty_count)) {
 161                 struct bch_fs *c = ei->vfs_inode.i_sb->s_fs_info;
 162                 int ret = __bch2_write_inode(c, ei, inode_clear_i_sectors_dirty, NULL);
 163
 164                 ret = ret;
 165         }
 166
 167         mutex_unlock(&ei->update_lock);
 168 }
 169
 170 static int __must_check i_sectors_dirty_get(struct bch_inode_info *ei,
 171                                             struct i_sectors_hook *h)
 172 {
 173         int ret = 0;
 174
 175         h->hook.fn      = i_sectors_hook_fn;
 176         h->sectors      = 0;
 177 #ifdef CONFIG_BCACHEFS_DEBUG
 178         h->ei           = ei;
 179 #endif
 180
 181         if (atomic_long_inc_not_zero(&ei->i_sectors_dirty_count))
 182                 return 0;
 183
 184         mutex_lock(&ei->update_lock);
 185
 186         if (!(ei->i_flags & BCH_INODE_I_SECTORS_DIRTY)) {
 187                 struct bch_fs *c = ei->vfs_inode.i_sb->s_fs_info;
 188
 189                 ret = __bch2_write_inode(c, ei, inode_set_i_sectors_dirty, NULL);
 190         }
 191
 192         if (!ret)
 193                 atomic_long_inc(&ei->i_sectors_dirty_count);
 194
 195         mutex_unlock(&ei->update_lock);
 196
 197         return ret;
 198 }
 199
 200 struct bchfs_extent_trans_hook {
 201         struct bchfs_write_op           *op;
 202         struct extent_insert_hook       hook;
 203
 204         struct bch_inode_unpacked       inode_u;
 205         struct bkey_inode_buf           inode_p;
 206
 207         bool                            need_inode_update;
 208 };
 209
 210 static enum extent_insert_hook_ret
 211 bchfs_extent_update_hook(struct extent_insert_hook *hook,
 212                          struct bpos committed_pos,
 213                          struct bpos next_pos,
 214                          struct bkey_s_c k,
 215                          const struct bkey_i *insert)
 216 {
 217         struct bchfs_extent_trans_hook *h = container_of(hook,
 218                                 struct bchfs_extent_trans_hook, hook);
 219         struct bch_inode_info *ei = h->op->ei;
 220         struct inode *inode = &ei->vfs_inode;
 221         int sign = bkey_extent_is_allocation(&insert->k) -
 222                 (k.k && bkey_extent_is_allocation(k.k));
 223         s64 sectors = (s64) (next_pos.offset - committed_pos.offset) * sign;
 224         u64 offset = min(next_pos.offset << 9, h->op->new_i_size);
 225         bool do_pack = false;
 226
 227         BUG_ON((next_pos.offset << 9) > round_up(offset, PAGE_SIZE));
 228
 229         /* XXX: ei->i_size locking */
 230         if (offset > ei->i_size) {
 231                 BUG_ON(ei->i_flags & BCH_INODE_I_SIZE_DIRTY);
 232
 233                 if (!h->need_inode_update) {
 234                         h->need_inode_update = true;
 235                         return BTREE_HOOK_RESTART_TRANS;
 236                 }
 237
 238                 h->inode_u.i_size = offset;
 239                 do_pack = true;
 240
 241                 ei->i_size = offset;
 242
 243                 if (h->op->is_dio)
 244                         i_size_write(inode, offset);
 245         }
 246
 247         if (sectors) {
 248                 if (!h->need_inode_update) {
 249                         h->need_inode_update = true;
 250                         return BTREE_HOOK_RESTART_TRANS;
 251                 }
 252
 253                 h->inode_u.i_sectors += sectors;
 254                 do_pack = true;
 255
 256                 atomic64_add(sectors, &ei->i_sectors);
 257
 258                 h->op->sectors_added += sectors;
 259
 260                 if (h->op->is_dio) {
 261                         spin_lock(&inode->i_lock);
 262                         inode->i_blocks += sectors;
 263                         spin_unlock(&inode->i_lock);
 264                 }
 265         }
 266
 267         if (do_pack)
 268                 bch2_inode_pack(&h->inode_p, &h->inode_u);
 269
 270         return BTREE_HOOK_DO_INSERT;
 271 }
 272
 273 static int bchfs_write_index_update(struct bch_write_op *wop)
 274 {
 275         struct bchfs_write_op *op = container_of(wop,
 276                                 struct bchfs_write_op, op);
 277         struct keylist *keys = &op->op.insert_keys;
 278         struct btree_iter extent_iter, inode_iter;
 279         struct bchfs_extent_trans_hook hook;
 280         struct bkey_i *k = bch2_keylist_front(keys);
 281         int ret;
 282
 283         BUG_ON(k->k.p.inode != op->ei->vfs_inode.i_ino);
 284
 285         bch2_btree_iter_init_intent(&extent_iter, wop->c, BTREE_ID_EXTENTS,
 286                                    bkey_start_pos(&bch2_keylist_front(keys)->k));
 287         bch2_btree_iter_init_intent(&inode_iter, wop->c,        BTREE_ID_INODES,
 288                                    POS(extent_iter.pos.inode, 0));
 289
 290         hook.op                 = op;
 291         hook.hook.fn            = bchfs_extent_update_hook;
 292         hook.need_inode_update  = false;
 293
 294         do {
 295                 ret = bch2_btree_iter_traverse(&extent_iter);
 296                 if (ret)
 297                         goto err;
 298
 299                 /* XXX: ei->i_size locking */
 300                 k = bch2_keylist_front(keys);
 301                 if (min(k->k.p.offset << 9, op->new_i_size) > op->ei->i_size)
 302                         hook.need_inode_update = true;
 303
 304                 if (hook.need_inode_update) {
 305                         struct bkey_s_c inode;
 306
 307                         if (!btree_iter_linked(&inode_iter))
 308                                 bch2_btree_iter_link(&extent_iter, &inode_iter);
 309
 310                         inode = bch2_btree_iter_peek_with_holes(&inode_iter);
 311                         if ((ret = btree_iter_err(inode)))
 312                                 goto err;
 313
 314                         if (WARN_ONCE(inode.k->type != BCH_INODE_FS,
 315                                       "inode %llu not found when updating",
 316                                       extent_iter.pos.inode)) {
 317                                 ret = -ENOENT;
 318                                 break;
 319                         }
 320
 321                         if (WARN_ONCE(bkey_bytes(inode.k) >
 322                                       sizeof(hook.inode_p),
 323                                       "inode %llu too big (%zu bytes, buf %zu)",
 324                                       extent_iter.pos.inode,
 325                                       bkey_bytes(inode.k),
 326                                       sizeof(hook.inode_p))) {
 327                                 ret = -ENOENT;
 328                                 break;
 329                         }
 330
 331                         bkey_reassemble(&hook.inode_p.inode.k_i, inode);
 332                         ret = bch2_inode_unpack(bkey_s_c_to_inode(inode),
 333                                                &hook.inode_u);
 334                         if (WARN_ONCE(ret,
 335                                       "error %i unpacking inode %llu",
 336                                       ret, extent_iter.pos.inode)) {
 337                                 ret = -ENOENT;
 338                                 break;
 339                         }
 340
 341                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 342                                         &hook.hook, op_journal_seq(wop),
 343                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 344                                         BTREE_INSERT_ENTRY(&extent_iter, k),
 345                                         BTREE_INSERT_ENTRY_EXTRA_RES(&inode_iter,
 346                                                         &hook.inode_p.inode.k_i, 2));
 347                 } else {
 348                         ret = bch2_btree_insert_at(wop->c, &wop->res,
 349                                         &hook.hook, op_journal_seq(wop),
 350                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 351                                         BTREE_INSERT_ENTRY(&extent_iter, k));
 352                 }
 353 err:
 354                 if (ret == -EINTR)
 355                         continue;
 356                 if (ret)
 357                         break;
 358
 359                 bch2_keylist_pop_front(keys);
 360         } while (!bch2_keylist_empty(keys));
 361
 362         bch2_btree_iter_unlock(&extent_iter);
 363         bch2_btree_iter_unlock(&inode_iter);
 364
 365         return ret;
 366 }
 367
 368 /* page state: */
 369
 370 /* stored in page->private: */
 371
 372 /*
 373  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 374  * almost protected it with the page lock, except that bch2_writepage_io_done has
 375  * to update the sector counts (and from interrupt/bottom half context).
 376  */
 377 struct bch_page_state {
 378 union { struct {
 379         /*
 380          * page is _fully_ written on disk, and not compressed - which means to
 381          * write this page we don't have to reserve space (the new write will
 382          * never take up more space on disk than what it's overwriting)
 383          */
 384         unsigned allocated:1;
 385
 386         /* Owns PAGE_SECTORS sized reservation: */
 387         unsigned                reserved:1;
 388         unsigned                nr_replicas:4;
 389
 390         /*
 391          * Number of sectors on disk - for i_blocks
 392          * Uncompressed size, not compressed size:
 393          */
 394         u8                      sectors;
 395         u8                      dirty_sectors;
 396 };
 397         /* for cmpxchg: */
 398         unsigned long           v;
 399 };
 400 };
 401
 402 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 403 ({                                                                      \
 404         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 405         struct bch_page_state _old;                                     \
 406                                                                         \
 407         do {                                                            \
 408                 _old.v = _new.v = _v;                                   \
 409                 _expr;                                                  \
 410                                                                         \
 411                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 412         } while (_old.v != _new.v &&                                    \
 413                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 414                                                                         \
 415         _old;                                                           \
 416 })
 417
 418 static inline struct bch_page_state *page_state(struct page *page)
 419 {
 420         struct bch_page_state *s = (void *) &page->private;
 421
 422         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 423
 424         if (!PagePrivate(page))
 425                 SetPagePrivate(page);
 426
 427         return s;
 428 }
 429
 430 static void bch2_put_page_reservation(struct bch_fs *c, struct page *page)
 431 {
 432         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 433         struct bch_page_state s;
 434
 435         s = page_state_cmpxchg(page_state(page), s, {
 436                 if (!s.reserved)
 437                         return;
 438                 s.reserved = 0;
 439         });
 440
 441         bch2_disk_reservation_put(c, &res);
 442 }
 443
 444 static int bch2_get_page_reservation(struct bch_fs *c, struct page *page,
 445                                     bool check_enospc)
 446 {
 447         struct bch_page_state *s = page_state(page), new;
 448         struct disk_reservation res;
 449         int ret = 0;
 450
 451         BUG_ON(s->allocated && s->sectors != PAGE_SECTORS);
 452
 453         if (s->allocated || s->reserved)
 454                 return 0;
 455
 456         ret = bch2_disk_reservation_get(c, &res, PAGE_SECTORS, !check_enospc
 457                                        ? BCH_DISK_RESERVATION_NOFAIL : 0);
 458         if (ret)
 459                 return ret;
 460
 461         page_state_cmpxchg(s, new, {
 462                 if (new.reserved) {
 463                         bch2_disk_reservation_put(c, &res);
 464                         return 0;
 465                 }
 466                 new.reserved    = 1;
 467                 new.nr_replicas = res.nr_replicas;
 468         });
 469
 470         return 0;
 471 }
 472
 473 static void bch2_clear_page_bits(struct page *page)
 474 {
 475         struct inode *inode = page->mapping->host;
 476         struct bch_fs *c = inode->i_sb->s_fs_info;
 477         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 478         struct bch_page_state s;
 479
 480         if (!PagePrivate(page))
 481                 return;
 482
 483         s = xchg(page_state(page), (struct bch_page_state) { .v = 0 });
 484         ClearPagePrivate(page);
 485
 486         if (s.dirty_sectors) {
 487                 spin_lock(&inode->i_lock);
 488                 inode->i_blocks -= s.dirty_sectors;
 489                 spin_unlock(&inode->i_lock);
 490         }
 491
 492         if (s.reserved)
 493                 bch2_disk_reservation_put(c, &res);
 494 }
 495
 496 int bch2_set_page_dirty(struct page *page)
 497 {
 498         struct bch_page_state old, new;
 499
 500         old = page_state_cmpxchg(page_state(page), new,
 501                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 502         );
 503
 504         if (old.dirty_sectors != new.dirty_sectors) {
 505                 struct inode *inode = page->mapping->host;
 506
 507                 spin_lock(&inode->i_lock);
 508                 inode->i_blocks += new.dirty_sectors - old.dirty_sectors;
 509                 spin_unlock(&inode->i_lock);
 510         }
 511
 512         return __set_page_dirty_nobuffers(page);
 513 }
 514
 515 /* readpages/writepages: */
 516
 517 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 518 {
 519         sector_t offset = (sector_t) page->index << (PAGE_SHIFT - 9);
 520
 521         return bio->bi_vcnt < bio->bi_max_vecs &&
 522                 bio_end_sector(bio) == offset;
 523 }
 524
 525 static void __bio_add_page(struct bio *bio, struct page *page)
 526 {
 527         bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
 528                 .bv_page = page,
 529                 .bv_len = PAGE_SIZE,
 530                 .bv_offset = 0,
 531         };
 532
 533         bio->bi_iter.bi_size += PAGE_SIZE;
 534 }
 535
 536 static int bio_add_page_contig(struct bio *bio, struct page *page)
 537 {
 538         sector_t offset = (sector_t) page->index << (PAGE_SHIFT - 9);
 539
 540         BUG_ON(!bio->bi_max_vecs);
 541
 542         if (!bio->bi_vcnt)
 543                 bio->bi_iter.bi_sector = offset;
 544         else if (!bio_can_add_page_contig(bio, page))
 545                 return -1;
 546
 547         __bio_add_page(bio, page);
 548         return 0;
 549 }
 550
 551 static void bch2_readpages_end_io(struct bio *bio)
 552 {
 553         struct bio_vec *bv;
 554         int i;
 555
 556         bio_for_each_segment_all(bv, bio, i) {
 557                 struct page *page = bv->bv_page;
 558
 559                 if (!bio->bi_error) {
 560                         SetPageUptodate(page);
 561                 } else {
 562                         ClearPageUptodate(page);
 563                         SetPageError(page);
 564                 }
 565                 unlock_page(page);
 566         }
 567
 568         bio_put(bio);
 569 }
 570
 571 struct readpages_iter {
 572         struct address_space    *mapping;
 573         struct list_head        pages;
 574         unsigned                nr_pages;
 575 };
 576
 577 static int readpage_add_page(struct readpages_iter *iter, struct page *page)
 578 {
 579         struct bch_page_state *s = page_state(page);
 580         int ret;
 581
 582         BUG_ON(s->reserved);
 583         s->allocated = 1;
 584         s->sectors = 0;
 585
 586         prefetchw(&page->flags);
 587         ret = add_to_page_cache_lru(page, iter->mapping,
 588                                     page->index, GFP_NOFS);
 589         put_page(page);
 590         return ret;
 591 }
 592
 593 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 594 {
 595         while (iter->nr_pages) {
 596                 struct page *page =
 597                         list_last_entry(&iter->pages, struct page, lru);
 598
 599                 prefetchw(&page->flags);
 600                 list_del(&page->lru);
 601                 iter->nr_pages--;
 602
 603                 if (!readpage_add_page(iter, page))
 604                         return page;
 605         }
 606
 607         return NULL;
 608 }
 609
 610 #define for_each_readpage_page(_iter, _page)                            \
 611         for (;                                                          \
 612              ((_page) = __readpage_next_page(&(_iter)));)               \
 613
 614 static void bch2_mark_pages_unalloc(struct bio *bio)
 615 {
 616         struct bvec_iter iter;
 617         struct bio_vec bv;
 618
 619         bio_for_each_segment(bv, bio, iter)
 620                 page_state(bv.bv_page)->allocated = 0;
 621 }
 622
 623 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 624 {
 625         struct bvec_iter iter;
 626         struct bio_vec bv;
 627
 628         bio_for_each_segment(bv, bio, iter) {
 629                 struct bch_page_state *s = page_state(bv.bv_page);
 630
 631                 /* sectors in @k from the start of this page: */
 632                 unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset);
 633
 634                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 635
 636                 if (!s->sectors)
 637                         s->nr_replicas = bch2_extent_nr_dirty_ptrs(k);
 638                 else
 639                         s->nr_replicas = min_t(unsigned, s->nr_replicas,
 640                                                bch2_extent_nr_dirty_ptrs(k));
 641
 642                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 643                 s->sectors += page_sectors;
 644         }
 645 }
 646
 647 static void readpage_bio_extend(struct readpages_iter *iter,
 648                                 struct bio *bio, u64 offset,
 649                                 bool get_more)
 650 {
 651         struct page *page;
 652         pgoff_t page_offset;
 653         int ret;
 654
 655         while (bio_end_sector(bio) < offset &&
 656                bio->bi_vcnt < bio->bi_max_vecs) {
 657                 page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 658
 659                 if (iter->nr_pages) {
 660                         page = list_last_entry(&iter->pages, struct page, lru);
 661                         if (page->index != page_offset)
 662                                 break;
 663
 664                         list_del(&page->lru);
 665                         iter->nr_pages--;
 666                 } else if (get_more) {
 667                         rcu_read_lock();
 668                         page = radix_tree_lookup(&iter->mapping->page_tree, page_offset);
 669                         rcu_read_unlock();
 670
 671                         if (page && !radix_tree_exceptional_entry(page))
 672                                 break;
 673
 674                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 675                         if (!page)
 676                                 break;
 677
 678                         page->index = page_offset;
 679                         ClearPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 680                 } else {
 681                         break;
 682                 }
 683
 684                 ret = readpage_add_page(iter, page);
 685                 if (ret)
 686                         break;
 687
 688                 __bio_add_page(bio, page);
 689         }
 690
 691         if (!iter->nr_pages)
 692                 SetPageReadahead(bio->bi_io_vec[bio->bi_vcnt - 1].bv_page);
 693 }
 694
 695 static void bchfs_read(struct bch_fs *c, struct btree_iter *iter,
 696                        struct bch_read_bio *rbio, u64 inode,
 697                        struct readpages_iter *readpages_iter)
 698 {
 699         struct bio *bio = &rbio->bio;
 700         int flags = BCH_READ_RETRY_IF_STALE|
 701                 BCH_READ_PROMOTE|
 702                 BCH_READ_MAY_REUSE_BIO;
 703
 704         while (1) {
 705                 struct extent_pick_ptr pick;
 706                 BKEY_PADDED(k) tmp;
 707                 struct bkey_s_c k;
 708                 unsigned bytes;
 709                 bool is_last;
 710
 711                 bch2_btree_iter_set_pos(iter, POS(inode, bio->bi_iter.bi_sector));
 712
 713                 k = bch2_btree_iter_peek_with_holes(iter);
 714                 BUG_ON(!k.k);
 715
 716                 if (IS_ERR(k.k)) {
 717                         int ret = bch2_btree_iter_unlock(iter);
 718                         BUG_ON(!ret);
 719                         bcache_io_error(c, bio, "btree IO error %i", ret);
 720                         bio_endio(bio);
 721                         return;
 722                 }
 723
 724                 bkey_reassemble(&tmp.k, k);
 725                 bch2_btree_iter_unlock(iter);
 726                 k = bkey_i_to_s_c(&tmp.k);
 727
 728                 bch2_extent_pick_ptr(c, k, &pick);
 729                 if (IS_ERR(pick.ca)) {
 730                         bcache_io_error(c, bio, "no device to read from");
 731                         bio_endio(bio);
 732                         return;
 733                 }
 734
 735                 if (readpages_iter)
 736                         readpage_bio_extend(readpages_iter,
 737                                             bio, k.k->p.offset,
 738                                             pick.ca &&
 739                                             (pick.crc.csum_type ||
 740                                              pick.crc.compression_type));
 741
 742                 bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
 743                          bio->bi_iter.bi_sector) << 9;
 744                 is_last = bytes == bio->bi_iter.bi_size;
 745                 swap(bio->bi_iter.bi_size, bytes);
 746
 747                 if (bkey_extent_is_allocation(k.k))
 748                         bch2_add_page_sectors(bio, k);
 749
 750                 if (!bkey_extent_is_allocation(k.k) ||
 751                     bkey_extent_is_compressed(k))
 752                         bch2_mark_pages_unalloc(bio);
 753
 754                 if (is_last)
 755                         flags |= BCH_READ_IS_LAST;
 756
 757                 if (pick.ca) {
 758                         PTR_BUCKET(pick.ca, &pick.ptr)->read_prio =
 759                                 c->prio_clock[READ].hand;
 760
 761                         bch2_read_extent(c, rbio, k, &pick, flags);
 762                         flags &= ~BCH_READ_MAY_REUSE_BIO;
 763                 } else {
 764                         zero_fill_bio(bio);
 765
 766                         if (is_last)
 767                                 bio_endio(bio);
 768                 }
 769
 770                 if (is_last)
 771                         return;
 772
 773                 swap(bio->bi_iter.bi_size, bytes);
 774                 bio_advance(bio, bytes);
 775         }
 776 }
 777
 778 int bch2_readpages(struct file *file, struct address_space *mapping,
 779                   struct list_head *pages, unsigned nr_pages)
 780 {
 781         struct inode *inode = mapping->host;
 782         struct bch_fs *c = inode->i_sb->s_fs_info;
 783         struct btree_iter iter;
 784         struct page *page;
 785         struct readpages_iter readpages_iter = {
 786                 .mapping = mapping, .nr_pages = nr_pages
 787         };
 788
 789         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN);
 790
 791         INIT_LIST_HEAD(&readpages_iter.pages);
 792         list_add(&readpages_iter.pages, pages);
 793         list_del_init(pages);
 794
 795         if (current->pagecache_lock != &mapping->add_lock)
 796                 pagecache_add_get(&mapping->add_lock);
 797
 798         while ((page = readpage_iter_next(&readpages_iter))) {
 799                 unsigned n = max(min_t(unsigned, readpages_iter.nr_pages + 1,
 800                                        BIO_MAX_PAGES),
 801                                  BCH_ENCODED_EXTENT_MAX >> PAGE_SECTOR_SHIFT);
 802
 803                 struct bch_read_bio *rbio =
 804                         container_of(bio_alloc_bioset(GFP_NOFS, n,
 805                                                       &c->bio_read),
 806                                      struct bch_read_bio, bio);
 807
 808                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 809                 bio_add_page_contig(&rbio->bio, page);
 810                 bchfs_read(c, &iter, rbio, inode->i_ino, &readpages_iter);
 811         }
 812
 813         if (current->pagecache_lock != &mapping->add_lock)
 814                 pagecache_add_put(&mapping->add_lock);
 815
 816         return 0;
 817 }
 818
 819 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 820                              u64 inode, struct page *page)
 821 {
 822         struct btree_iter iter;
 823
 824         /*
 825          * Initialize page state:
 826          * If a page is partly allocated and partly a hole, we want it to be
 827          * marked BCH_PAGE_UNALLOCATED - so we initially mark all pages
 828          * allocated and then mark them unallocated as we find holes:
 829          *
 830          * Note that the bio hasn't been split yet - it's the only bio that
 831          * points to these pages. As we walk extents and split @bio, that
 832          * necessarily be true, the splits won't necessarily be on page
 833          * boundaries:
 834          */
 835         struct bch_page_state *s = page_state(page);
 836
 837         EBUG_ON(s->reserved);
 838         s->allocated = 1;
 839         s->sectors = 0;
 840
 841         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 842         bio_add_page_contig(&rbio->bio, page);
 843
 844         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN);
 845         bchfs_read(c, &iter, rbio, inode, NULL);
 846 }
 847
 848 int bch2_readpage(struct file *file, struct page *page)
 849 {
 850         struct address_space *mapping = page->mapping;
 851         struct inode *inode = mapping->host;
 852         struct bch_fs *c = inode->i_sb->s_fs_info;
 853         struct bch_read_bio *rbio;
 854
 855         rbio = container_of(bio_alloc_bioset(GFP_NOFS, 1,
 856                                             &c->bio_read),
 857                            struct bch_read_bio, bio);
 858         rbio->bio.bi_end_io = bch2_readpages_end_io;
 859
 860         __bchfs_readpage(c, rbio, inode->i_ino, page);
 861         return 0;
 862 }
 863
 864 struct bch_writepage_state {
 865         struct bch_writepage_io *io;
 866 };
 867
 868 static void bch2_writepage_io_free(struct closure *cl)
 869 {
 870         struct bch_writepage_io *io = container_of(cl,
 871                                         struct bch_writepage_io, cl);
 872         struct bio *bio = &io->bio.bio;
 873
 874         bio_put(bio);
 875 }
 876
 877 static void bch2_writepage_io_done(struct closure *cl)
 878 {
 879         struct bch_writepage_io *io = container_of(cl,
 880                                         struct bch_writepage_io, cl);
 881         struct bch_fs *c = io->op.op.c;
 882         struct bio *bio = &io->bio.bio;
 883         struct bio_vec *bvec;
 884         unsigned i;
 885
 886         atomic_sub(bio->bi_vcnt, &c->writeback_pages);
 887         wake_up(&c->writeback_wait);
 888
 889         bio_for_each_segment_all(bvec, bio, i) {
 890                 struct page *page = bvec->bv_page;
 891
 892                 if (io->op.op.error) {
 893                         SetPageError(page);
 894                         if (page->mapping)
 895                                 set_bit(AS_EIO, &page->mapping->flags);
 896                 }
 897
 898                 if (io->op.op.written >= PAGE_SECTORS) {
 899                         struct bch_page_state old, new;
 900
 901                         old = page_state_cmpxchg(page_state(page), new, {
 902                                 new.sectors = PAGE_SECTORS;
 903                                 new.dirty_sectors = 0;
 904                         });
 905
 906                         io->op.sectors_added -= old.dirty_sectors;
 907                         io->op.op.written -= PAGE_SECTORS;
 908                 }
 909         }
 910
 911         /*
 912          * racing with fallocate can cause us to add fewer sectors than
 913          * expected - but we shouldn't add more sectors than expected:
 914          *
 915          * (error (due to going RO) halfway through a page can screw that up
 916          * slightly)
 917          */
 918         BUG_ON(io->op.sectors_added >= (s64) PAGE_SECTORS);
 919
 920         /*
 921          * PageWriteback is effectively our ref on the inode - fixup i_blocks
 922          * before calling end_page_writeback:
 923          */
 924         if (io->op.sectors_added) {
 925                 struct inode *inode = &io->op.ei->vfs_inode;
 926
 927                 spin_lock(&inode->i_lock);
 928                 inode->i_blocks += io->op.sectors_added;
 929                 spin_unlock(&inode->i_lock);
 930         }
 931
 932         bio_for_each_segment_all(bvec, bio, i)
 933                 end_page_writeback(bvec->bv_page);
 934
 935         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
 936 }
 937
 938 static void bch2_writepage_do_io(struct bch_writepage_state *w)
 939 {
 940         struct bch_writepage_io *io = w->io;
 941
 942         w->io = NULL;
 943         atomic_add(io->bio.bio.bi_vcnt, &io->op.op.c->writeback_pages);
 944
 945         io->op.op.pos.offset = io->bio.bio.bi_iter.bi_sector;
 946
 947         closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl);
 948         continue_at(&io->cl, bch2_writepage_io_done, NULL);
 949 }
 950
 951 /*
 952  * Get a bch_writepage_io and add @page to it - appending to an existing one if
 953  * possible, else allocating a new one:
 954  */
 955 static void bch2_writepage_io_alloc(struct bch_fs *c,
 956                                     struct bch_writepage_state *w,
 957                                     struct bch_inode_info *ei,
 958                                     struct page *page)
 959 {
 960         u64 inum = ei->vfs_inode.i_ino;
 961         unsigned nr_replicas = page_state(page)->nr_replicas;
 962
 963         EBUG_ON(!nr_replicas);
 964         /* XXX: disk_reservation->gen isn't plumbed through */
 965
 966         if (!w->io) {
 967 alloc_io:
 968                 w->io = container_of(bio_alloc_bioset(GFP_NOFS,
 969                                                       BIO_MAX_PAGES,
 970                                                       bch2_writepage_bioset),
 971                                      struct bch_writepage_io, bio.bio);
 972
 973                 closure_init(&w->io->cl, NULL);
 974                 w->io->op.ei            = ei;
 975                 w->io->op.sectors_added = 0;
 976                 w->io->op.is_dio        = false;
 977                 bch2_write_op_init(&w->io->op.op, c, &w->io->bio,
 978                                   (struct disk_reservation) {
 979                                         .nr_replicas = c->opts.data_replicas,
 980                                   },
 981                                   foreground_write_point(c, inum),
 982                                   POS(inum, 0),
 983                                   &ei->journal_seq, 0);
 984                 w->io->op.op.index_update_fn = bchfs_write_index_update;
 985         }
 986
 987         if (w->io->op.op.res.nr_replicas != nr_replicas ||
 988             bio_add_page_contig(&w->io->bio.bio, page)) {
 989                 bch2_writepage_do_io(w);
 990                 goto alloc_io;
 991         }
 992
 993         /*
 994          * We shouldn't ever be handed pages for multiple inodes in a single
 995          * pass - right?
 996          */
 997         BUG_ON(ei != w->io->op.ei);
 998 }
 999
1000 static int __bch2_writepage(struct bch_fs *c, struct page *page,
1001                             struct writeback_control *wbc,
1002                             struct bch_writepage_state *w)
1003 {
1004         struct inode *inode = page->mapping->host;
1005         struct bch_inode_info *ei = to_bch_ei(inode);
1006         struct bch_page_state new, old;
1007         unsigned offset;
1008         loff_t i_size = i_size_read(inode);
1009         pgoff_t end_index = i_size >> PAGE_SHIFT;
1010
1011         EBUG_ON(!PageUptodate(page));
1012
1013         /* Is the page fully inside i_size? */
1014         if (page->index < end_index)
1015                 goto do_io;
1016
1017         /* Is the page fully outside i_size? (truncate in progress) */
1018         offset = i_size & (PAGE_SIZE - 1);
1019         if (page->index > end_index || !offset) {
1020                 unlock_page(page);
1021                 return 0;
1022         }
1023
1024         /*
1025          * The page straddles i_size.  It must be zeroed out on each and every
1026          * writepage invocation because it may be mmapped.  "A file is mapped
1027          * in multiples of the page size.  For a file that is not a multiple of
1028          * the  page size, the remaining memory is zeroed when mapped, and
1029          * writes to that region are not written out to the file."
1030          */
1031         zero_user_segment(page, offset, PAGE_SIZE);
1032 do_io:
1033         bch2_writepage_io_alloc(c, w, ei, page);
1034
1035         /* while page is locked: */
1036         w->io->op.new_i_size = i_size;
1037
1038         if (wbc->sync_mode == WB_SYNC_ALL)
1039                 w->io->bio.bio.bi_opf |= WRITE_SYNC;
1040
1041         /* Before unlocking the page, transfer reservation to w->io: */
1042         old = page_state_cmpxchg(page_state(page), new, {
1043                 EBUG_ON(!new.reserved &&
1044                         (new.sectors != PAGE_SECTORS ||
1045                         !new.allocated));
1046
1047                 if (new.allocated &&
1048                     w->io->op.op.compression_type != BCH_COMPRESSION_NONE)
1049                         new.allocated = 0;
1050                 else if (!new.reserved)
1051                         goto out;
1052                 new.reserved = 0;
1053         });
1054
1055         w->io->op.op.res.sectors += PAGE_SECTORS *
1056                 (old.reserved - new.reserved) *
1057                 old.nr_replicas;
1058 out:
1059         BUG_ON(PageWriteback(page));
1060         set_page_writeback(page);
1061         unlock_page(page);
1062
1063         return 0;
1064 }
1065
1066 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1067 {
1068         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1069         struct bch_writepage_state w = { NULL };
1070         struct pagecache_iter iter;
1071         struct page *page;
1072         int ret = 0;
1073         int done = 0;
1074         pgoff_t uninitialized_var(writeback_index);
1075         pgoff_t index;
1076         pgoff_t end;            /* Inclusive */
1077         pgoff_t done_index;
1078         int cycled;
1079         int range_whole = 0;
1080         int tag;
1081
1082         if (wbc->range_cyclic) {
1083                 writeback_index = mapping->writeback_index; /* prev offset */
1084                 index = writeback_index;
1085                 if (index == 0)
1086                         cycled = 1;
1087                 else
1088                         cycled = 0;
1089                 end = -1;
1090         } else {
1091                 index = wbc->range_start >> PAGE_SHIFT;
1092                 end = wbc->range_end >> PAGE_SHIFT;
1093                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1094                         range_whole = 1;
1095                 cycled = 1; /* ignore range_cyclic tests */
1096         }
1097         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1098                 tag = PAGECACHE_TAG_TOWRITE;
1099         else
1100                 tag = PAGECACHE_TAG_DIRTY;
1101 retry:
1102         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1103                 tag_pages_for_writeback(mapping, index, end);
1104
1105         done_index = index;
1106 get_pages:
1107         for_each_pagecache_tag(&iter, mapping, tag, index, end, page) {
1108                 done_index = page->index;
1109
1110                 if (w.io &&
1111                     !bio_can_add_page_contig(&w.io->bio.bio, page))
1112                         bch2_writepage_do_io(&w);
1113
1114                 if (!w.io &&
1115                     atomic_read(&c->writeback_pages) >=
1116                     c->writeback_pages_max) {
1117                         /* don't sleep with pages pinned: */
1118                         pagecache_iter_release(&iter);
1119
1120                         __wait_event(c->writeback_wait,
1121                                      atomic_read(&c->writeback_pages) <
1122                                      c->writeback_pages_max);
1123                         goto get_pages;
1124                 }
1125
1126                 lock_page(page);
1127
1128                 /*
1129                  * Page truncated or invalidated. We can freely skip it
1130                  * then, even for data integrity operations: the page
1131                  * has disappeared concurrently, so there could be no
1132                  * real expectation of this data interity operation
1133                  * even if there is now a new, dirty page at the same
1134                  * pagecache address.
1135                  */
1136                 if (unlikely(page->mapping != mapping)) {
1137 continue_unlock:
1138                         unlock_page(page);
1139                         continue;
1140                 }
1141
1142                 if (!PageDirty(page)) {
1143                         /* someone wrote it for us */
1144                         goto continue_unlock;
1145                 }
1146
1147                 if (PageWriteback(page)) {
1148                         if (wbc->sync_mode != WB_SYNC_NONE)
1149                                 wait_on_page_writeback(page);
1150                         else
1151                                 goto continue_unlock;
1152                 }
1153
1154                 BUG_ON(PageWriteback(page));
1155                 if (!clear_page_dirty_for_io(page))
1156                         goto continue_unlock;
1157
1158                 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
1159                 ret = __bch2_writepage(c, page, wbc, &w);
1160                 if (unlikely(ret)) {
1161                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
1162                                 unlock_page(page);
1163                                 ret = 0;
1164                         } else {
1165                                 /*
1166                                  * done_index is set past this page,
1167                                  * so media errors will not choke
1168                                  * background writeout for the entire
1169                                  * file. This has consequences for
1170                                  * range_cyclic semantics (ie. it may
1171                                  * not be suitable for data integrity
1172                                  * writeout).
1173                                  */
1174                                 done_index = page->index + 1;
1175                                 done = 1;
1176                                 break;
1177                         }
1178                 }
1179
1180                 /*
1181                  * We stop writing back only if we are not doing
1182                  * integrity sync. In case of integrity sync we have to
1183                  * keep going until we have written all the pages
1184                  * we tagged for writeback prior to entering this loop.
1185                  */
1186                 if (--wbc->nr_to_write <= 0 &&
1187                     wbc->sync_mode == WB_SYNC_NONE) {
1188                         done = 1;
1189                         break;
1190                 }
1191         }
1192         pagecache_iter_release(&iter);
1193
1194         if (w.io)
1195                 bch2_writepage_do_io(&w);
1196
1197         if (!cycled && !done) {
1198                 /*
1199                  * range_cyclic:
1200                  * We hit the last page and there is more work to be done: wrap
1201                  * back to the start of the file
1202                  */
1203                 cycled = 1;
1204                 index = 0;
1205                 end = writeback_index - 1;
1206                 goto retry;
1207         }
1208         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1209                 mapping->writeback_index = done_index;
1210
1211         return ret;
1212 }
1213
1214 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1215 {
1216         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1217         struct bch_writepage_state w = { NULL };
1218         int ret;
1219
1220         ret = __bch2_writepage(c, page, wbc, &w);
1221         if (w.io)
1222                 bch2_writepage_do_io(&w);
1223
1224         return ret;
1225 }
1226
1227 static void bch2_read_single_page_end_io(struct bio *bio)
1228 {
1229         complete(bio->bi_private);
1230 }
1231
1232 static int bch2_read_single_page(struct page *page,
1233                                  struct address_space *mapping)
1234 {
1235         struct inode *inode = mapping->host;
1236         struct bch_fs *c = inode->i_sb->s_fs_info;
1237         struct bch_read_bio *rbio;
1238         int ret;
1239         DECLARE_COMPLETION_ONSTACK(done);
1240
1241         rbio = container_of(bio_alloc_bioset(GFP_NOFS, 1,
1242                                              &c->bio_read),
1243                             struct bch_read_bio, bio);
1244         rbio->bio.bi_private = &done;
1245         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1246
1247         __bchfs_readpage(c, rbio, inode->i_ino, page);
1248         wait_for_completion(&done);
1249
1250         ret = rbio->bio.bi_error;
1251         bio_put(&rbio->bio);
1252
1253         if (ret < 0)
1254                 return ret;
1255
1256         SetPageUptodate(page);
1257         return 0;
1258 }
1259
1260 int bch2_write_begin(struct file *file, struct address_space *mapping,
1261                      loff_t pos, unsigned len, unsigned flags,
1262                      struct page **pagep, void **fsdata)
1263 {
1264         struct inode *inode = mapping->host;
1265         struct bch_fs *c = inode->i_sb->s_fs_info;
1266         pgoff_t index = pos >> PAGE_SHIFT;
1267         unsigned offset = pos & (PAGE_SIZE - 1);
1268         struct page *page;
1269         int ret = -ENOMEM;
1270
1271         BUG_ON(inode_unhashed(mapping->host));
1272
1273         /* Not strictly necessary - same reason as mkwrite(): */
1274         pagecache_add_get(&mapping->add_lock);
1275
1276         page = grab_cache_page_write_begin(mapping, index, flags);
1277         if (!page)
1278                 goto err_unlock;
1279
1280         if (PageUptodate(page))
1281                 goto out;
1282
1283         /* If we're writing entire page, don't need to read it in first: */
1284         if (len == PAGE_SIZE)
1285                 goto out;
1286
1287         if (!offset && pos + len >= inode->i_size) {
1288                 zero_user_segment(page, len, PAGE_SIZE);
1289                 flush_dcache_page(page);
1290                 goto out;
1291         }
1292
1293         if (index > inode->i_size >> PAGE_SHIFT) {
1294                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1295                 flush_dcache_page(page);
1296                 goto out;
1297         }
1298 readpage:
1299         ret = bch2_read_single_page(page, mapping);
1300         if (ret)
1301                 goto err;
1302 out:
1303         ret = bch2_get_page_reservation(c, page, true);
1304         if (ret) {
1305                 if (!PageUptodate(page)) {
1306                         /*
1307                          * If the page hasn't been read in, we won't know if we
1308                          * actually need a reservation - we don't actually need
1309                          * to read here, we just need to check if the page is
1310                          * fully backed by uncompressed data:
1311                          */
1312                         goto readpage;
1313                 }
1314
1315                 goto err;
1316         }
1317
1318         *pagep = page;
1319         return 0;
1320 err:
1321         unlock_page(page);
1322         put_page(page);
1323         *pagep = NULL;
1324 err_unlock:
1325         pagecache_add_put(&mapping->add_lock);
1326         return ret;
1327 }
1328
1329 int bch2_write_end(struct file *filp, struct address_space *mapping,
1330                    loff_t pos, unsigned len, unsigned copied,
1331                    struct page *page, void *fsdata)
1332 {
1333         struct inode *inode = page->mapping->host;
1334         struct bch_fs *c = inode->i_sb->s_fs_info;
1335
1336         lockdep_assert_held(&inode->i_rwsem);
1337
1338         if (unlikely(copied < len && !PageUptodate(page))) {
1339                 /*
1340                  * The page needs to be read in, but that would destroy
1341                  * our partial write - simplest thing is to just force
1342                  * userspace to redo the write:
1343                  */
1344                 zero_user(page, 0, PAGE_SIZE);
1345                 flush_dcache_page(page);
1346                 copied = 0;
1347         }
1348
1349         if (pos + copied > inode->i_size)
1350                 i_size_write(inode, pos + copied);
1351
1352         if (copied) {
1353                 if (!PageUptodate(page))
1354                         SetPageUptodate(page);
1355                 if (!PageDirty(page))
1356                         set_page_dirty(page);
1357         } else {
1358                 bch2_put_page_reservation(c, page);
1359         }
1360
1361         unlock_page(page);
1362         put_page(page);
1363         pagecache_add_put(&mapping->add_lock);
1364
1365         return copied;
1366 }
1367
1368 /* O_DIRECT */
1369
1370 static void bch2_dio_read_complete(struct closure *cl)
1371 {
1372         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1373
1374         dio->req->ki_complete(dio->req, dio->ret, 0);
1375         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1376 }
1377
1378 static void bch2_direct_IO_read_endio(struct bio *bio)
1379 {
1380         struct dio_read *dio = bio->bi_private;
1381
1382         if (bio->bi_error)
1383                 dio->ret = bio->bi_error;
1384
1385         closure_put(&dio->cl);
1386 }
1387
1388 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1389 {
1390         bch2_direct_IO_read_endio(bio);
1391         bio_check_pages_dirty(bio);     /* transfers ownership */
1392 }
1393
1394 static int bch2_direct_IO_read(struct bch_fs *c, struct kiocb *req,
1395                                struct file *file, struct inode *inode,
1396                                struct iov_iter *iter, loff_t offset)
1397 {
1398         struct dio_read *dio;
1399         struct bio *bio;
1400         bool sync = is_sync_kiocb(req);
1401         ssize_t ret;
1402
1403         if ((offset|iter->count) & (block_bytes(c) - 1))
1404                 return -EINVAL;
1405
1406         ret = min_t(loff_t, iter->count,
1407                     max_t(loff_t, 0, i_size_read(inode) - offset));
1408         iov_iter_truncate(iter, round_up(ret, block_bytes(c)));
1409
1410         if (!ret)
1411                 return ret;
1412
1413         bio = bio_alloc_bioset(GFP_KERNEL,
1414                                iov_iter_npages(iter, BIO_MAX_PAGES),
1415                                bch2_dio_read_bioset);
1416
1417         bio->bi_end_io = bch2_direct_IO_read_endio;
1418
1419         dio = container_of(bio, struct dio_read, rbio.bio);
1420         closure_init(&dio->cl, NULL);
1421
1422         /*
1423          * this is a _really_ horrible hack just to avoid an atomic sub at the
1424          * end:
1425          */
1426         if (!sync) {
1427                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1428                 atomic_set(&dio->cl.remaining,
1429                            CLOSURE_REMAINING_INITIALIZER -
1430                            CLOSURE_RUNNING +
1431                            CLOSURE_DESTRUCTOR);
1432         } else {
1433                 atomic_set(&dio->cl.remaining,
1434                            CLOSURE_REMAINING_INITIALIZER + 1);
1435         }
1436
1437         dio->req        = req;
1438         dio->ret        = ret;
1439
1440         goto start;
1441         while (iter->count) {
1442                 bio = bio_alloc_bioset(GFP_KERNEL,
1443                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1444                                        &c->bio_read);
1445                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1446 start:
1447                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1448                 bio->bi_iter.bi_sector  = offset >> 9;
1449                 bio->bi_private         = dio;
1450
1451                 ret = bio_get_user_pages(bio, iter, 1);
1452                 if (ret < 0) {
1453                         /* XXX: fault inject this path */
1454                         bio->bi_error = ret;
1455                         bio_endio(bio);
1456                         break;
1457                 }
1458
1459                 offset += bio->bi_iter.bi_size;
1460                 bio_set_pages_dirty(bio);
1461
1462                 if (iter->count)
1463                         closure_get(&dio->cl);
1464
1465                 bch2_read(c, container_of(bio,
1466                                 struct bch_read_bio, bio),
1467                          inode->i_ino);
1468         }
1469
1470         if (sync) {
1471                 closure_sync(&dio->cl);
1472                 closure_debug_destroy(&dio->cl);
1473                 ret = dio->ret;
1474                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1475                 return ret;
1476         } else {
1477                 return -EIOCBQUEUED;
1478         }
1479 }
1480
1481 static long __bch2_dio_write_complete(struct dio_write *dio)
1482 {
1483         struct file *file = dio->req->ki_filp;
1484         struct address_space *mapping = file->f_mapping;
1485         struct inode *inode = file->f_inode;
1486         long ret = dio->error ?: dio->written;
1487
1488         bch2_disk_reservation_put(dio->c, &dio->res);
1489
1490         __pagecache_block_put(&mapping->add_lock);
1491         inode_dio_end(inode);
1492
1493         if (dio->iovec && dio->iovec != dio->inline_vecs)
1494                 kfree(dio->iovec);
1495
1496         bio_put(&dio->bio.bio);
1497         return ret;
1498 }
1499
1500 static void bch2_dio_write_complete(struct closure *cl)
1501 {
1502         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1503         struct kiocb *req = dio->req;
1504
1505         req->ki_complete(req, __bch2_dio_write_complete(dio), 0);
1506 }
1507
1508 static void bch2_dio_write_done(struct dio_write *dio)
1509 {
1510         struct bio_vec *bv;
1511         int i;
1512
1513         dio->written += dio->iop.op.written << 9;
1514
1515         if (dio->iop.op.error)
1516                 dio->error = dio->iop.op.error;
1517
1518         bio_for_each_segment_all(bv, &dio->bio.bio, i)
1519                 put_page(bv->bv_page);
1520
1521         if (dio->iter.count)
1522                 bio_reset(&dio->bio.bio);
1523 }
1524
1525 static void bch2_do_direct_IO_write(struct dio_write *dio)
1526 {
1527         struct file *file = dio->req->ki_filp;
1528         struct inode *inode = file->f_inode;
1529         struct bch_inode_info *ei = to_bch_ei(inode);
1530         struct bio *bio = &dio->bio.bio;
1531         unsigned flags = 0;
1532         int ret;
1533
1534         if ((dio->req->ki_flags & IOCB_DSYNC) &&
1535             !dio->c->opts.journal_flush_disabled)
1536                 flags |= BCH_WRITE_FLUSH;
1537
1538         bio->bi_iter.bi_sector = (dio->offset + dio->written) >> 9;
1539
1540         ret = bio_get_user_pages(bio, &dio->iter, 0);
1541         if (ret < 0) {
1542                 /*
1543                  * these didn't get initialized, but bch2_dio_write_done() will
1544                  * look at them:
1545                  */
1546                 dio->iop.op.error = 0;
1547                 dio->iop.op.written = 0;
1548                 dio->error = ret;
1549                 return;
1550         }
1551
1552         dio->iop.ei             = ei;
1553         dio->iop.sectors_added  = 0;
1554         dio->iop.is_dio         = true;
1555         dio->iop.new_i_size     = U64_MAX;
1556         bch2_write_op_init(&dio->iop.op, dio->c, &dio->bio,
1557                           dio->res,
1558                           foreground_write_point(dio->c, inode->i_ino),
1559                           POS(inode->i_ino, bio->bi_iter.bi_sector),
1560                           &ei->journal_seq, flags);
1561         dio->iop.op.index_update_fn = bchfs_write_index_update;
1562
1563         dio->res.sectors -= bio_sectors(bio);
1564         dio->iop.op.res.sectors = bio_sectors(bio);
1565
1566         task_io_account_write(bio->bi_iter.bi_size);
1567
1568         closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl);
1569 }
1570
1571 static void bch2_dio_write_loop_async(struct closure *cl)
1572 {
1573         struct dio_write *dio =
1574                 container_of(cl, struct dio_write, cl);
1575         struct address_space *mapping = dio->req->ki_filp->f_mapping;
1576
1577         bch2_dio_write_done(dio);
1578
1579         if (dio->iter.count && !dio->error) {
1580                 use_mm(dio->mm);
1581                 pagecache_block_get(&mapping->add_lock);
1582
1583                 bch2_do_direct_IO_write(dio);
1584
1585                 pagecache_block_put(&mapping->add_lock);
1586                 unuse_mm(dio->mm);
1587
1588                 continue_at(&dio->cl, bch2_dio_write_loop_async, NULL);
1589         } else {
1590 #if 0
1591                 closure_return_with_destructor(cl, bch2_dio_write_complete);
1592 #else
1593                 closure_debug_destroy(cl);
1594                 bch2_dio_write_complete(cl);
1595 #endif
1596         }
1597 }
1598
1599 static int bch2_direct_IO_write(struct bch_fs *c, struct kiocb *req,
1600                                 struct file *file, struct inode *inode,
1601                                 struct iov_iter *iter, loff_t offset)
1602 {
1603         struct address_space *mapping = file->f_mapping;
1604         struct dio_write *dio;
1605         struct bio *bio;
1606         ssize_t ret;
1607         bool sync = is_sync_kiocb(req);
1608
1609         lockdep_assert_held(&inode->i_rwsem);
1610
1611         if (unlikely(!iter->count))
1612                 return 0;
1613
1614         if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
1615                 return -EINVAL;
1616
1617         bio = bio_alloc_bioset(GFP_KERNEL,
1618                                iov_iter_npages(iter, BIO_MAX_PAGES),
1619                                bch2_dio_write_bioset);
1620         dio = container_of(bio, struct dio_write, bio.bio);
1621         dio->req        = req;
1622         dio->c          = c;
1623         dio->written    = 0;
1624         dio->error      = 0;
1625         dio->offset     = offset;
1626         dio->iovec      = NULL;
1627         dio->iter       = *iter;
1628         dio->mm         = current->mm;
1629         closure_init(&dio->cl, NULL);
1630
1631         if (offset + iter->count > inode->i_size)
1632                 sync = true;
1633
1634         /*
1635          * XXX: we shouldn't return -ENOSPC if we're overwriting existing data -
1636          * if getting a reservation fails we should check if we are doing an
1637          * overwrite.
1638          *
1639          * Have to then guard against racing with truncate (deleting data that
1640          * we would have been overwriting)
1641          */
1642         ret = bch2_disk_reservation_get(c, &dio->res, iter->count >> 9, 0);
1643         if (unlikely(ret)) {
1644                 closure_debug_destroy(&dio->cl);
1645                 bio_put(bio);
1646                 return ret;
1647         }
1648
1649         inode_dio_begin(inode);
1650         __pagecache_block_get(&mapping->add_lock);
1651
1652         if (sync) {
1653                 do {
1654                         bch2_do_direct_IO_write(dio);
1655
1656                         closure_sync(&dio->cl);
1657                         bch2_dio_write_done(dio);
1658                 } while (dio->iter.count && !dio->error);
1659
1660                 closure_debug_destroy(&dio->cl);
1661                 return __bch2_dio_write_complete(dio);
1662         } else {
1663                 bch2_do_direct_IO_write(dio);
1664
1665                 if (dio->iter.count && !dio->error) {
1666                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1667                                 dio->iovec = kmalloc(dio->iter.nr_segs *
1668                                                      sizeof(struct iovec),
1669                                                      GFP_KERNEL);
1670                                 if (!dio->iovec)
1671                                         dio->error = -ENOMEM;
1672                         } else {
1673                                 dio->iovec = dio->inline_vecs;
1674                         }
1675
1676                         memcpy(dio->iovec,
1677                                dio->iter.iov,
1678                                dio->iter.nr_segs * sizeof(struct iovec));
1679                         dio->iter.iov = dio->iovec;
1680                 }
1681
1682                 continue_at_noreturn(&dio->cl, bch2_dio_write_loop_async, NULL);
1683                 return -EIOCBQUEUED;
1684         }
1685 }
1686
1687 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1688 {
1689         struct file *file = req->ki_filp;
1690         struct inode *inode = file->f_inode;
1691         struct bch_fs *c = inode->i_sb->s_fs_info;
1692         struct blk_plug plug;
1693         ssize_t ret;
1694
1695         blk_start_plug(&plug);
1696         ret = ((iov_iter_rw(iter) == WRITE)
1697                 ? bch2_direct_IO_write
1698                 : bch2_direct_IO_read)(c, req, file, inode, iter, req->ki_pos);
1699         blk_finish_plug(&plug);
1700
1701         return ret;
1702 }
1703
1704 static ssize_t
1705 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1706 {
1707         struct file *file = iocb->ki_filp;
1708         struct inode *inode = file->f_inode;
1709         struct bch_fs *c = inode->i_sb->s_fs_info;
1710         struct address_space *mapping = file->f_mapping;
1711         loff_t pos = iocb->ki_pos;
1712         ssize_t ret;
1713
1714         pagecache_block_get(&mapping->add_lock);
1715
1716         /* Write and invalidate pagecache range that we're writing to: */
1717         ret = write_invalidate_inode_pages_range(file->f_mapping, pos,
1718                                         pos + iov_iter_count(iter) - 1);
1719         if (unlikely(ret))
1720                 goto err;
1721
1722         ret = bch2_direct_IO_write(c, iocb, file, inode, iter, pos);
1723 err:
1724         pagecache_block_put(&mapping->add_lock);
1725
1726         return ret;
1727 }
1728
1729 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1730 {
1731         struct file *file = iocb->ki_filp;
1732         struct address_space *mapping = file->f_mapping;
1733         struct inode *inode = mapping->host;
1734         ssize_t ret;
1735
1736         /* We can write back this queue in page reclaim */
1737         current->backing_dev_info = inode_to_bdi(inode);
1738         ret = file_remove_privs(file);
1739         if (ret)
1740                 goto out;
1741
1742         ret = file_update_time(file);
1743         if (ret)
1744                 goto out;
1745
1746         ret = iocb->ki_flags & IOCB_DIRECT
1747                 ? bch2_direct_write(iocb, from)
1748                 : generic_perform_write(file, from, iocb->ki_pos);
1749
1750         if (likely(ret > 0))
1751                 iocb->ki_pos += ret;
1752 out:
1753         current->backing_dev_info = NULL;
1754         return ret;
1755 }
1756
1757 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1758 {
1759         struct file *file = iocb->ki_filp;
1760         struct inode *inode = file->f_mapping->host;
1761         bool direct = iocb->ki_flags & IOCB_DIRECT;
1762         ssize_t ret;
1763
1764         inode_lock(inode);
1765         ret = generic_write_checks(iocb, from);
1766         if (ret > 0)
1767                 ret = __bch2_write_iter(iocb, from);
1768         inode_unlock(inode);
1769
1770         if (ret > 0 && !direct)
1771                 ret = generic_write_sync(iocb, ret);
1772
1773         return ret;
1774 }
1775
1776 int bch2_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1777 {
1778         struct page *page = vmf->page;
1779         struct inode *inode = file_inode(vma->vm_file);
1780         struct address_space *mapping = inode->i_mapping;
1781         struct bch_fs *c = inode->i_sb->s_fs_info;
1782         int ret = VM_FAULT_LOCKED;
1783
1784         sb_start_pagefault(inode->i_sb);
1785         file_update_time(vma->vm_file);
1786
1787         /*
1788          * Not strictly necessary, but helps avoid dio writes livelocking in
1789          * write_invalidate_inode_pages_range() - can drop this if/when we get
1790          * a write_invalidate_inode_pages_range() that works without dropping
1791          * page lock before invalidating page
1792          */
1793         if (current->pagecache_lock != &mapping->add_lock)
1794                 pagecache_add_get(&mapping->add_lock);
1795
1796         lock_page(page);
1797         if (page->mapping != mapping ||
1798             page_offset(page) > i_size_read(inode)) {
1799                 unlock_page(page);
1800                 ret = VM_FAULT_NOPAGE;
1801                 goto out;
1802         }
1803
1804         if (bch2_get_page_reservation(c, page, true)) {
1805                 unlock_page(page);
1806                 ret = VM_FAULT_SIGBUS;
1807                 goto out;
1808         }
1809
1810         if (!PageDirty(page))
1811                 set_page_dirty(page);
1812         wait_for_stable_page(page);
1813 out:
1814         if (current->pagecache_lock != &mapping->add_lock)
1815                 pagecache_add_put(&mapping->add_lock);
1816         sb_end_pagefault(inode->i_sb);
1817         return ret;
1818 }
1819
1820 void bch2_invalidatepage(struct page *page, unsigned int offset,
1821                          unsigned int length)
1822 {
1823         EBUG_ON(!PageLocked(page));
1824         EBUG_ON(PageWriteback(page));
1825
1826         if (offset || length < PAGE_SIZE)
1827                 return;
1828
1829         bch2_clear_page_bits(page);
1830 }
1831
1832 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
1833 {
1834         EBUG_ON(!PageLocked(page));
1835         EBUG_ON(PageWriteback(page));
1836
1837         if (PageDirty(page))
1838                 return 0;
1839
1840         bch2_clear_page_bits(page);
1841         return 1;
1842 }
1843
1844 #ifdef CONFIG_MIGRATION
1845 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
1846                       struct page *page, enum migrate_mode mode)
1847 {
1848         int ret;
1849
1850         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
1851         if (ret != MIGRATEPAGE_SUCCESS)
1852                 return ret;
1853
1854         if (PagePrivate(page)) {
1855                 *page_state(newpage) = *page_state(page);
1856                 ClearPagePrivate(page);
1857         }
1858
1859         migrate_page_copy(newpage, page);
1860         return MIGRATEPAGE_SUCCESS;
1861 }
1862 #endif
1863
1864 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1865 {
1866         struct inode *inode = file->f_mapping->host;
1867         struct bch_inode_info *ei = to_bch_ei(inode);
1868         struct bch_fs *c = inode->i_sb->s_fs_info;
1869         int ret;
1870
1871         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1872         if (ret)
1873                 return ret;
1874
1875         if (c->opts.journal_flush_disabled)
1876                 return 0;
1877
1878         return bch2_journal_flush_seq(&c->journal, ei->journal_seq);
1879 }
1880
1881 static int __bch2_truncate_page(struct address_space *mapping,
1882                                 pgoff_t index, loff_t start, loff_t end)
1883 {
1884         struct inode *inode = mapping->host;
1885         struct bch_fs *c = inode->i_sb->s_fs_info;
1886         unsigned start_offset = start & (PAGE_SIZE - 1);
1887         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
1888         struct page *page;
1889         int ret = 0;
1890
1891         /* Page boundary? Nothing to do */
1892         if (!((index == start >> PAGE_SHIFT && start_offset) ||
1893               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
1894                 return 0;
1895
1896         /* Above i_size? */
1897         if (index << PAGE_SHIFT >= inode->i_size)
1898                 return 0;
1899
1900         page = find_lock_page(mapping, index);
1901         if (!page) {
1902                 struct btree_iter iter;
1903                 struct bkey_s_c k = bkey_s_c_null;
1904
1905                 /*
1906                  * XXX: we're doing two index lookups when we end up reading the
1907                  * page
1908                  */
1909                 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
1910                                    POS(inode->i_ino,
1911                                        index << (PAGE_SHIFT - 9)), k) {
1912                         if (bkey_cmp(bkey_start_pos(k.k),
1913                                      POS(inode->i_ino,
1914                                          (index + 1) << (PAGE_SHIFT - 9))) >= 0)
1915                                 break;
1916
1917                         if (k.k->type != KEY_TYPE_DISCARD &&
1918                             k.k->type != BCH_RESERVATION) {
1919                                 bch2_btree_iter_unlock(&iter);
1920                                 goto create;
1921                         }
1922                 }
1923                 bch2_btree_iter_unlock(&iter);
1924                 return 0;
1925 create:
1926                 page = find_or_create_page(mapping, index, GFP_KERNEL);
1927                 if (unlikely(!page)) {
1928                         ret = -ENOMEM;
1929                         goto out;
1930                 }
1931         }
1932
1933         if (!PageUptodate(page)) {
1934                 ret = bch2_read_single_page(page, mapping);
1935                 if (ret)
1936                         goto unlock;
1937         }
1938
1939         /*
1940          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
1941          *
1942          * XXX: because we aren't currently tracking whether the page has actual
1943          * data in it (vs. just 0s, or only partially written) this wrong. ick.
1944          */
1945         ret = bch2_get_page_reservation(c, page, false);
1946         BUG_ON(ret);
1947
1948         if (index == start >> PAGE_SHIFT &&
1949             index == end >> PAGE_SHIFT)
1950                 zero_user_segment(page, start_offset, end_offset);
1951         else if (index == start >> PAGE_SHIFT)
1952                 zero_user_segment(page, start_offset, PAGE_SIZE);
1953         else if (index == end >> PAGE_SHIFT)
1954                 zero_user_segment(page, 0, end_offset);
1955
1956         if (!PageDirty(page))
1957                 set_page_dirty(page);
1958 unlock:
1959         unlock_page(page);
1960         put_page(page);
1961 out:
1962         return ret;
1963 }
1964
1965 static int bch2_truncate_page(struct address_space *mapping, loff_t from)
1966 {
1967         return __bch2_truncate_page(mapping, from >> PAGE_SHIFT,
1968                                    from, from + PAGE_SIZE);
1969 }
1970
1971 int bch2_truncate(struct inode *inode, struct iattr *iattr)
1972 {
1973         struct address_space *mapping = inode->i_mapping;
1974         struct bch_inode_info *ei = to_bch_ei(inode);
1975         struct bch_fs *c = inode->i_sb->s_fs_info;
1976         bool shrink = iattr->ia_size <= inode->i_size;
1977         int ret = 0;
1978
1979         inode_dio_wait(inode);
1980         pagecache_block_get(&mapping->add_lock);
1981
1982         truncate_setsize(inode, iattr->ia_size);
1983
1984         /* sync appends.. */
1985         /* XXX what protects ei->i_size? */
1986         if (iattr->ia_size > ei->i_size)
1987                 ret = filemap_write_and_wait_range(mapping, ei->i_size, S64_MAX);
1988         if (ret)
1989                 goto err_put_pagecache;
1990
1991         mutex_lock(&ei->update_lock);
1992         i_size_dirty_get(ei);
1993         ret = bch2_write_inode_size(c, ei, inode->i_size);
1994         mutex_unlock(&ei->update_lock);
1995
1996         if (unlikely(ret))
1997                 goto err;
1998
1999         /*
2000          * There might be persistent reservations (from fallocate())
2001          * above i_size, which bch2_inode_truncate() will discard - we're
2002          * only supposed to discard them if we're doing a real truncate
2003          * here (new i_size < current i_size):
2004          */
2005         if (shrink) {
2006                 struct i_sectors_hook i_sectors_hook;
2007                 int ret;
2008
2009                 ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2010                 if (unlikely(ret))
2011                         goto err;
2012
2013                 ret = bch2_truncate_page(inode->i_mapping, iattr->ia_size);
2014                 if (unlikely(ret)) {
2015                         i_sectors_dirty_put(ei, &i_sectors_hook);
2016                         goto err;
2017                 }
2018
2019                 ret = bch2_inode_truncate(c, inode->i_ino,
2020                                          round_up(iattr->ia_size, PAGE_SIZE) >> 9,
2021                                          &i_sectors_hook.hook,
2022                                          &ei->journal_seq);
2023
2024                 i_sectors_dirty_put(ei, &i_sectors_hook);
2025
2026                 if (unlikely(ret))
2027                         goto err;
2028         }
2029
2030         mutex_lock(&ei->update_lock);
2031         setattr_copy(inode, iattr);
2032         inode->i_mtime = inode->i_ctime = current_fs_time(inode->i_sb);
2033
2034         /* clear I_SIZE_DIRTY: */
2035         i_size_dirty_put(ei);
2036         ret = bch2_write_inode_size(c, ei, inode->i_size);
2037         mutex_unlock(&ei->update_lock);
2038
2039         pagecache_block_put(&mapping->add_lock);
2040
2041         return 0;
2042 err:
2043         i_size_dirty_put(ei);
2044 err_put_pagecache:
2045         pagecache_block_put(&mapping->add_lock);
2046         return ret;
2047 }
2048
2049 static long bch2_fpunch(struct inode *inode, loff_t offset, loff_t len)
2050 {
2051         struct address_space *mapping = inode->i_mapping;
2052         struct bch_inode_info *ei = to_bch_ei(inode);
2053         struct bch_fs *c = inode->i_sb->s_fs_info;
2054         u64 ino = inode->i_ino;
2055         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
2056         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
2057         int ret = 0;
2058
2059         inode_lock(inode);
2060         inode_dio_wait(inode);
2061         pagecache_block_get(&mapping->add_lock);
2062
2063         ret = __bch2_truncate_page(inode->i_mapping,
2064                                   offset >> PAGE_SHIFT,
2065                                   offset, offset + len);
2066         if (unlikely(ret))
2067                 goto out;
2068
2069         if (offset >> PAGE_SHIFT !=
2070             (offset + len) >> PAGE_SHIFT) {
2071                 ret = __bch2_truncate_page(inode->i_mapping,
2072                                           (offset + len) >> PAGE_SHIFT,
2073                                           offset, offset + len);
2074                 if (unlikely(ret))
2075                         goto out;
2076         }
2077
2078         truncate_pagecache_range(inode, offset, offset + len - 1);
2079
2080         if (discard_start < discard_end) {
2081                 struct disk_reservation disk_res;
2082                 struct i_sectors_hook i_sectors_hook;
2083                 int ret;
2084
2085                 BUG_ON(bch2_disk_reservation_get(c, &disk_res, 0, 0));
2086
2087                 ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2088                 if (unlikely(ret))
2089                         goto out;
2090
2091                 ret = bch2_discard(c,
2092                                   POS(ino, discard_start),
2093                                   POS(ino, discard_end),
2094                                   ZERO_VERSION,
2095                                   &disk_res,
2096                                   &i_sectors_hook.hook,
2097                                   &ei->journal_seq);
2098
2099                 i_sectors_dirty_put(ei, &i_sectors_hook);
2100                 bch2_disk_reservation_put(c, &disk_res);
2101         }
2102 out:
2103         pagecache_block_put(&mapping->add_lock);
2104         inode_unlock(inode);
2105
2106         return ret;
2107 }
2108
2109 static long bch2_fcollapse(struct inode *inode, loff_t offset, loff_t len)
2110 {
2111         struct address_space *mapping = inode->i_mapping;
2112         struct bch_inode_info *ei = to_bch_ei(inode);
2113         struct bch_fs *c = inode->i_sb->s_fs_info;
2114         struct btree_iter src;
2115         struct btree_iter dst;
2116         BKEY_PADDED(k) copy;
2117         struct bkey_s_c k;
2118         struct i_sectors_hook i_sectors_hook;
2119         loff_t new_size;
2120         int ret;
2121
2122         if ((offset | len) & (PAGE_SIZE - 1))
2123                 return -EINVAL;
2124
2125         bch2_btree_iter_init_intent(&dst, c, BTREE_ID_EXTENTS,
2126                                    POS(inode->i_ino, offset >> 9));
2127         /* position will be set from dst iter's position: */
2128         bch2_btree_iter_init(&src, c, BTREE_ID_EXTENTS, POS_MIN);
2129         bch2_btree_iter_link(&src, &dst);
2130
2131         /*
2132          * We need i_mutex to keep the page cache consistent with the extents
2133          * btree, and the btree consistent with i_size - we don't need outside
2134          * locking for the extents btree itself, because we're using linked
2135          * iterators
2136          */
2137         inode_lock(inode);
2138         inode_dio_wait(inode);
2139         pagecache_block_get(&mapping->add_lock);
2140
2141         ret = -EINVAL;
2142         if (offset + len >= inode->i_size)
2143                 goto err;
2144
2145         if (inode->i_size < len)
2146                 goto err;
2147
2148         new_size = inode->i_size - len;
2149
2150         ret = write_invalidate_inode_pages_range(inode->i_mapping,
2151                                                  offset, LLONG_MAX);
2152         if (ret)
2153                 goto err;
2154
2155         ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2156         if (ret)
2157                 goto err;
2158
2159         while (bkey_cmp(dst.pos,
2160                         POS(inode->i_ino,
2161                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2162                 struct disk_reservation disk_res;
2163
2164                 bch2_btree_iter_set_pos(&src,
2165                         POS(dst.pos.inode, dst.pos.offset + (len >> 9)));
2166
2167                 ret = bch2_btree_iter_traverse(&dst);
2168                 if (ret)
2169                         goto btree_iter_err;
2170
2171                 k = bch2_btree_iter_peek_with_holes(&src);
2172                 if ((ret = btree_iter_err(k)))
2173                         goto btree_iter_err;
2174
2175                 bkey_reassemble(&copy.k, k);
2176
2177                 if (bkey_deleted(&copy.k.k))
2178                         copy.k.k.type = KEY_TYPE_DISCARD;
2179
2180                 bch2_cut_front(src.pos, &copy.k);
2181                 copy.k.k.p.offset -= len >> 9;
2182
2183                 BUG_ON(bkey_cmp(dst.pos, bkey_start_pos(&copy.k.k)));
2184
2185                 ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size,
2186                                                BCH_DISK_RESERVATION_NOFAIL);
2187                 BUG_ON(ret);
2188
2189                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2190                                           &ei->journal_seq,
2191                                           BTREE_INSERT_ATOMIC|
2192                                           BTREE_INSERT_NOFAIL,
2193                                           BTREE_INSERT_ENTRY(&dst, &copy.k));
2194                 bch2_disk_reservation_put(c, &disk_res);
2195 btree_iter_err:
2196                 if (ret < 0 && ret != -EINTR)
2197                         goto err_unwind;
2198
2199                 bch2_btree_iter_cond_resched(&src);
2200         }
2201
2202         bch2_btree_iter_unlock(&src);
2203         bch2_btree_iter_unlock(&dst);
2204
2205         ret = bch2_inode_truncate(c, inode->i_ino,
2206                                  round_up(new_size, PAGE_SIZE) >> 9,
2207                                  &i_sectors_hook.hook,
2208                                  &ei->journal_seq);
2209         if (ret)
2210                 goto err_unwind;
2211
2212         i_sectors_dirty_put(ei, &i_sectors_hook);
2213
2214         mutex_lock(&ei->update_lock);
2215         i_size_write(inode, new_size);
2216         ret = bch2_write_inode_size(c, ei, inode->i_size);
2217         mutex_unlock(&ei->update_lock);
2218
2219         pagecache_block_put(&mapping->add_lock);
2220         inode_unlock(inode);
2221
2222         return ret;
2223 err_unwind:
2224         /*
2225          * XXX: we've left data with multiple pointers... which isn't a _super_
2226          * serious problem...
2227          */
2228         i_sectors_dirty_put(ei, &i_sectors_hook);
2229 err:
2230         bch2_btree_iter_unlock(&src);
2231         bch2_btree_iter_unlock(&dst);
2232         pagecache_block_put(&mapping->add_lock);
2233         inode_unlock(inode);
2234         return ret;
2235 }
2236
2237 static long bch2_fallocate(struct inode *inode, int mode,
2238                            loff_t offset, loff_t len)
2239 {
2240         struct address_space *mapping = inode->i_mapping;
2241         struct bch_inode_info *ei = to_bch_ei(inode);
2242         struct bch_fs *c = inode->i_sb->s_fs_info;
2243         struct i_sectors_hook i_sectors_hook;
2244         struct btree_iter iter;
2245         struct bpos end;
2246         loff_t block_start, block_end;
2247         loff_t new_size = offset + len;
2248         unsigned sectors;
2249         unsigned replicas = READ_ONCE(c->opts.data_replicas);
2250         int ret;
2251
2252         bch2_btree_iter_init_intent(&iter, c, BTREE_ID_EXTENTS, POS_MIN);
2253
2254         inode_lock(inode);
2255         inode_dio_wait(inode);
2256         pagecache_block_get(&mapping->add_lock);
2257
2258         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2259             new_size > inode->i_size) {
2260                 ret = inode_newsize_ok(inode, new_size);
2261                 if (ret)
2262                         goto err;
2263         }
2264
2265         if (mode & FALLOC_FL_ZERO_RANGE) {
2266                 ret = __bch2_truncate_page(inode->i_mapping,
2267                                           offset >> PAGE_SHIFT,
2268                                           offset, offset + len);
2269
2270                 if (!ret &&
2271                     offset >> PAGE_SHIFT !=
2272                     (offset + len) >> PAGE_SHIFT)
2273                         ret = __bch2_truncate_page(inode->i_mapping,
2274                                                   (offset + len) >> PAGE_SHIFT,
2275                                                   offset, offset + len);
2276
2277                 if (unlikely(ret))
2278                         goto err;
2279
2280                 truncate_pagecache_range(inode, offset, offset + len - 1);
2281
2282                 block_start     = round_up(offset, PAGE_SIZE);
2283                 block_end       = round_down(offset + len, PAGE_SIZE);
2284         } else {
2285                 block_start     = round_down(offset, PAGE_SIZE);
2286                 block_end       = round_up(offset + len, PAGE_SIZE);
2287         }
2288
2289         bch2_btree_iter_set_pos(&iter, POS(inode->i_ino, block_start >> 9));
2290         end = POS(inode->i_ino, block_end >> 9);
2291
2292         ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2293         if (unlikely(ret))
2294                 goto err;
2295
2296         while (bkey_cmp(iter.pos, end) < 0) {
2297                 struct disk_reservation disk_res = { 0 };
2298                 struct bkey_i_reservation reservation;
2299                 struct bkey_s_c k;
2300
2301                 k = bch2_btree_iter_peek_with_holes(&iter);
2302                 if ((ret = btree_iter_err(k)))
2303                         goto btree_iter_err;
2304
2305                 /* already reserved */
2306                 if (k.k->type == BCH_RESERVATION &&
2307                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2308                         bch2_btree_iter_advance_pos(&iter);
2309                         continue;
2310                 }
2311
2312                 if (bkey_extent_is_data(k.k)) {
2313                         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
2314                                 bch2_btree_iter_advance_pos(&iter);
2315                                 continue;
2316                         }
2317                 }
2318
2319                 bkey_reservation_init(&reservation.k_i);
2320                 reservation.k.type      = BCH_RESERVATION;
2321                 reservation.k.p         = k.k->p;
2322                 reservation.k.size      = k.k->size;
2323
2324                 bch2_cut_front(iter.pos, &reservation.k_i);
2325                 bch2_cut_back(end, &reservation.k);
2326
2327                 sectors = reservation.k.size;
2328                 reservation.v.nr_replicas = bch2_extent_nr_dirty_ptrs(k);
2329
2330                 if (reservation.v.nr_replicas < replicas ||
2331                     bkey_extent_is_compressed(k)) {
2332                         ret = bch2_disk_reservation_get(c, &disk_res,
2333                                                        sectors, 0);
2334                         if (ret)
2335                                 goto err_put_sectors_dirty;
2336
2337                         reservation.v.nr_replicas = disk_res.nr_replicas;
2338                 }
2339
2340                 ret = bch2_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2341                                           &ei->journal_seq,
2342                                           BTREE_INSERT_ATOMIC|
2343                                           BTREE_INSERT_NOFAIL,
2344                                           BTREE_INSERT_ENTRY(&iter, &reservation.k_i));
2345                 bch2_disk_reservation_put(c, &disk_res);
2346 btree_iter_err:
2347                 if (ret < 0 && ret != -EINTR)
2348                         goto err_put_sectors_dirty;
2349
2350         }
2351         bch2_btree_iter_unlock(&iter);
2352
2353         i_sectors_dirty_put(ei, &i_sectors_hook);
2354
2355         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2356             new_size > inode->i_size) {
2357                 i_size_write(inode, new_size);
2358
2359                 mutex_lock(&ei->update_lock);
2360                 ret = bch2_write_inode_size(c, ei, inode->i_size);
2361                 mutex_unlock(&ei->update_lock);
2362         }
2363
2364         /* blech */
2365         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2366             (mode & FALLOC_FL_ZERO_RANGE) &&
2367             ei->i_size != inode->i_size) {
2368                 /* sync appends.. */
2369                 ret = filemap_write_and_wait_range(mapping, ei->i_size, S64_MAX);
2370                 if (ret)
2371                         goto err;
2372
2373                 if (ei->i_size != inode->i_size) {
2374                         mutex_lock(&ei->update_lock);
2375                         ret = bch2_write_inode_size(c, ei, inode->i_size);
2376                         mutex_unlock(&ei->update_lock);
2377                 }
2378         }
2379
2380         pagecache_block_put(&mapping->add_lock);
2381         inode_unlock(inode);
2382
2383         return 0;
2384 err_put_sectors_dirty:
2385         i_sectors_dirty_put(ei, &i_sectors_hook);
2386 err:
2387         bch2_btree_iter_unlock(&iter);
2388         pagecache_block_put(&mapping->add_lock);
2389         inode_unlock(inode);
2390         return ret;
2391 }
2392
2393 long bch2_fallocate_dispatch(struct file *file, int mode,
2394                              loff_t offset, loff_t len)
2395 {
2396         struct inode *inode = file_inode(file);
2397
2398         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2399                 return bch2_fallocate(inode, mode, offset, len);
2400
2401         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2402                 return bch2_fpunch(inode, offset, len);
2403
2404         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2405                 return bch2_fcollapse(inode, offset, len);
2406
2407         return -EOPNOTSUPP;
2408 }
2409
2410 static bool page_is_data(struct page *page)
2411 {
2412         /* XXX: should only have to check PageDirty */
2413         return PagePrivate(page) &&
2414                 (page_state(page)->sectors ||
2415                  page_state(page)->dirty_sectors);
2416 }
2417
2418 static loff_t bch2_next_pagecache_data(struct inode *inode,
2419                                       loff_t start_offset,
2420                                       loff_t end_offset)
2421 {
2422         struct address_space *mapping = inode->i_mapping;
2423         struct page *page;
2424         pgoff_t index;
2425
2426         for (index = start_offset >> PAGE_SHIFT;
2427              index < end_offset >> PAGE_SHIFT;
2428              index++) {
2429                 if (find_get_pages(mapping, index, 1, &page)) {
2430                         lock_page(page);
2431                         index = page->index;
2432
2433                         if (page_is_data(page))
2434                                 end_offset =
2435                                         min(end_offset,
2436                                         max(start_offset,
2437                                             ((loff_t) index) << PAGE_SHIFT));
2438                         unlock_page(page);
2439                         put_page(page);
2440                 } else {
2441                         break;
2442                 }
2443         }
2444
2445         return end_offset;
2446 }
2447
2448 static loff_t bch2_seek_data(struct file *file, u64 offset)
2449 {
2450         struct inode *inode = file->f_mapping->host;
2451         struct bch_fs *c = inode->i_sb->s_fs_info;
2452         struct btree_iter iter;
2453         struct bkey_s_c k;
2454         u64 isize, next_data = MAX_LFS_FILESIZE;
2455         int ret;
2456
2457         isize = i_size_read(inode);
2458         if (offset >= isize)
2459                 return -ENXIO;
2460
2461         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2462                            POS(inode->i_ino, offset >> 9), k) {
2463                 if (k.k->p.inode != inode->i_ino) {
2464                         break;
2465                 } else if (bkey_extent_is_data(k.k)) {
2466                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2467                         break;
2468                 } else if (k.k->p.offset >> 9 > isize)
2469                         break;
2470         }
2471
2472         ret = bch2_btree_iter_unlock(&iter);
2473         if (ret)
2474                 return ret;
2475
2476         if (next_data > offset)
2477                 next_data = bch2_next_pagecache_data(inode, offset, next_data);
2478
2479         if (next_data > isize)
2480                 return -ENXIO;
2481
2482         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2483 }
2484
2485 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2486 {
2487         struct page *page;
2488         bool ret;
2489
2490         page = find_lock_entry(mapping, index);
2491         if (!page || radix_tree_exception(page))
2492                 return false;
2493
2494         ret = page_is_data(page);
2495         unlock_page(page);
2496
2497         return ret;
2498 }
2499
2500 static loff_t bch2_next_pagecache_hole(struct inode *inode,
2501                                       loff_t start_offset,
2502                                       loff_t end_offset)
2503 {
2504         struct address_space *mapping = inode->i_mapping;
2505         pgoff_t index;
2506
2507         for (index = start_offset >> PAGE_SHIFT;
2508              index < end_offset >> PAGE_SHIFT;
2509              index++)
2510                 if (!page_slot_is_data(mapping, index))
2511                         end_offset = max(start_offset,
2512                                          ((loff_t) index) << PAGE_SHIFT);
2513
2514         return end_offset;
2515 }
2516
2517 static loff_t bch2_seek_hole(struct file *file, u64 offset)
2518 {
2519         struct inode *inode = file->f_mapping->host;
2520         struct bch_fs *c = inode->i_sb->s_fs_info;
2521         struct btree_iter iter;
2522         struct bkey_s_c k;
2523         u64 isize, next_hole = MAX_LFS_FILESIZE;
2524         int ret;
2525
2526         isize = i_size_read(inode);
2527         if (offset >= isize)
2528                 return -ENXIO;
2529
2530         for_each_btree_key_with_holes(&iter, c, BTREE_ID_EXTENTS,
2531                                       POS(inode->i_ino, offset >> 9), k) {
2532                 if (k.k->p.inode != inode->i_ino) {
2533                         next_hole = bch2_next_pagecache_hole(inode,
2534                                         offset, MAX_LFS_FILESIZE);
2535                         break;
2536                 } else if (!bkey_extent_is_data(k.k)) {
2537                         next_hole = bch2_next_pagecache_hole(inode,
2538                                         max(offset, bkey_start_offset(k.k) << 9),
2539                                         k.k->p.offset << 9);
2540
2541                         if (next_hole < k.k->p.offset << 9)
2542                                 break;
2543                 } else {
2544                         offset = max(offset, bkey_start_offset(k.k) << 9);
2545                 }
2546         }
2547
2548         ret = bch2_btree_iter_unlock(&iter);
2549         if (ret)
2550                 return ret;
2551
2552         if (next_hole > isize)
2553                 next_hole = isize;
2554
2555         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2556 }
2557
2558 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
2559 {
2560         switch (whence) {
2561         case SEEK_SET:
2562         case SEEK_CUR:
2563         case SEEK_END:
2564                 return generic_file_llseek(file, offset, whence);
2565         case SEEK_DATA:
2566                 return bch2_seek_data(file, offset);
2567         case SEEK_HOLE:
2568                 return bch2_seek_hole(file, offset);
2569         }
2570
2571         return -EINVAL;
2572 }