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

   1
   2 #include "bcache.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 *bch_writepage_bioset;
  27 struct bio_set *bch_dio_read_bioset;
  28 struct bio_set *bch_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, struct bch_inode *bi,
  63                           void *p)
  64 {
  65         loff_t *new_i_size = p;
  66         unsigned i_flags = le32_to_cpu(bi->i_flags);
  67
  68         lockdep_assert_held(&ei->update_lock);
  69
  70         bi->i_size = cpu_to_le64(*new_i_size);
  71
  72         if (atomic_long_read(&ei->i_size_dirty_count))
  73                 i_flags |= BCH_INODE_I_SIZE_DIRTY;
  74         else
  75                 i_flags &= ~BCH_INODE_I_SIZE_DIRTY;
  76
  77         bi->i_flags = cpu_to_le32(i_flags);
  78
  79         return 0;
  80 }
  81
  82 static int __must_check bch_write_inode_size(struct cache_set *c,
  83                                              struct bch_inode_info *ei,
  84                                              loff_t new_size)
  85 {
  86         return __bch_write_inode(c, ei, inode_set_size, &new_size);
  87 }
  88
  89 static inline void i_size_dirty_put(struct bch_inode_info *ei)
  90 {
  91         atomic_long_dec_bug(&ei->i_size_dirty_count);
  92 }
  93
  94 static inline void i_size_dirty_get(struct bch_inode_info *ei)
  95 {
  96         lockdep_assert_held(&ei->vfs_inode.i_rwsem);
  97
  98         atomic_long_inc(&ei->i_size_dirty_count);
  99 }
 100
 101 /* i_sectors accounting: */
 102
 103 static enum extent_insert_hook_ret
 104 i_sectors_hook_fn(struct extent_insert_hook *hook,
 105                   struct bpos committed_pos,
 106                   struct bpos next_pos,
 107                   struct bkey_s_c k,
 108                   const struct bkey_i *insert)
 109 {
 110         struct i_sectors_hook *h = container_of(hook,
 111                                 struct i_sectors_hook, hook);
 112         s64 sectors = next_pos.offset - committed_pos.offset;
 113         int sign = bkey_extent_is_allocation(&insert->k) -
 114                 (k.k && bkey_extent_is_allocation(k.k));
 115
 116         EBUG_ON(!(h->ei->i_flags & BCH_INODE_I_SECTORS_DIRTY));
 117         EBUG_ON(!atomic_long_read(&h->ei->i_sectors_dirty_count));
 118
 119         h->sectors += sectors * sign;
 120
 121         return BTREE_HOOK_DO_INSERT;
 122 }
 123
 124 static int inode_set_i_sectors_dirty(struct bch_inode_info *ei,
 125                                     struct bch_inode *bi, void *p)
 126 {
 127         BUG_ON(le32_to_cpu(bi->i_flags) & BCH_INODE_I_SECTORS_DIRTY);
 128
 129         bi->i_flags = cpu_to_le32(le32_to_cpu(bi->i_flags)|
 130                                   BCH_INODE_I_SECTORS_DIRTY);
 131         return 0;
 132 }
 133
 134 static int inode_clear_i_sectors_dirty(struct bch_inode_info *ei,
 135                                        struct bch_inode *bi, void *p)
 136 {
 137         BUG_ON(!(le32_to_cpu(bi->i_flags) & BCH_INODE_I_SECTORS_DIRTY));
 138
 139         bi->i_sectors   = cpu_to_le64(atomic64_read(&ei->i_sectors));
 140         bi->i_flags     = cpu_to_le32(le32_to_cpu(bi->i_flags) &
 141                                       ~BCH_INODE_I_SECTORS_DIRTY);
 142         return 0;
 143 }
 144
 145 static void i_sectors_dirty_put(struct bch_inode_info *ei,
 146                                 struct i_sectors_hook *h)
 147 {
 148         struct inode *inode = &ei->vfs_inode;
 149
 150         if (h->sectors) {
 151                 spin_lock(&inode->i_lock);
 152                 inode->i_blocks += h->sectors;
 153                 spin_unlock(&inode->i_lock);
 154
 155                 atomic64_add(h->sectors, &ei->i_sectors);
 156                 EBUG_ON(atomic64_read(&ei->i_sectors) < 0);
 157         }
 158
 159         EBUG_ON(atomic_long_read(&ei->i_sectors_dirty_count) <= 0);
 160
 161         mutex_lock(&ei->update_lock);
 162
 163         if (atomic_long_dec_and_test(&ei->i_sectors_dirty_count)) {
 164                 struct cache_set *c = ei->vfs_inode.i_sb->s_fs_info;
 165                 int ret = __bch_write_inode(c, ei, inode_clear_i_sectors_dirty, NULL);
 166
 167                 ret = ret;
 168         }
 169
 170         mutex_unlock(&ei->update_lock);
 171 }
 172
 173 static int __must_check i_sectors_dirty_get(struct bch_inode_info *ei,
 174                                             struct i_sectors_hook *h)
 175 {
 176         int ret = 0;
 177
 178         h->hook.fn      = i_sectors_hook_fn;
 179         h->sectors      = 0;
 180 #ifdef CONFIG_BCACHE_DEBUG
 181         h->ei           = ei;
 182 #endif
 183
 184         if (atomic_long_inc_not_zero(&ei->i_sectors_dirty_count))
 185                 return 0;
 186
 187         mutex_lock(&ei->update_lock);
 188
 189         if (!(ei->i_flags & BCH_INODE_I_SECTORS_DIRTY)) {
 190                 struct cache_set *c = ei->vfs_inode.i_sb->s_fs_info;
 191
 192                 ret = __bch_write_inode(c, ei, inode_set_i_sectors_dirty, NULL);
 193         }
 194
 195         if (!ret)
 196                 atomic_long_inc(&ei->i_sectors_dirty_count);
 197
 198         mutex_unlock(&ei->update_lock);
 199
 200         return ret;
 201 }
 202
 203 struct bchfs_extent_trans_hook {
 204         struct bchfs_write_op           *op;
 205         struct extent_insert_hook       hook;
 206         struct bkey_i_inode             new_inode;
 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
 226         BUG_ON((next_pos.offset << 9) > round_up(offset, PAGE_SIZE));
 227
 228         /* XXX: ei->i_size locking */
 229         if (offset > ei->i_size) {
 230                 BUG_ON(ei->i_flags & BCH_INODE_I_SIZE_DIRTY);
 231
 232                 if (!h->need_inode_update) {
 233                         h->need_inode_update = true;
 234                         return BTREE_HOOK_RESTART_TRANS;
 235                 }
 236
 237                 h->new_inode.v.i_size = cpu_to_le64(offset);
 238                 ei->i_size = offset;
 239
 240                 if (h->op->is_dio)
 241                         i_size_write(inode, offset);
 242         }
 243
 244         if (sectors) {
 245                 if (!h->need_inode_update) {
 246                         h->need_inode_update = true;
 247                         return BTREE_HOOK_RESTART_TRANS;
 248                 }
 249
 250                 le64_add_cpu(&h->new_inode.v.i_sectors, sectors);
 251                 atomic64_add(sectors, &ei->i_sectors);
 252
 253                 h->op->sectors_added += sectors;
 254
 255                 if (h->op->is_dio) {
 256                         spin_lock(&inode->i_lock);
 257                         inode->i_blocks += sectors;
 258                         spin_unlock(&inode->i_lock);
 259                 }
 260         }
 261
 262         return BTREE_HOOK_DO_INSERT;
 263 }
 264
 265 static int bchfs_write_index_update(struct bch_write_op *wop)
 266 {
 267         struct bchfs_write_op *op = container_of(wop,
 268                                 struct bchfs_write_op, op);
 269         struct keylist *keys = &op->op.insert_keys;
 270         struct btree_iter extent_iter, inode_iter;
 271         struct bchfs_extent_trans_hook hook;
 272         struct bkey_i *k = bch_keylist_front(keys);
 273         int ret;
 274
 275         BUG_ON(k->k.p.inode != op->ei->vfs_inode.i_ino);
 276
 277         bch_btree_iter_init_intent(&extent_iter, wop->c, BTREE_ID_EXTENTS,
 278                                    bkey_start_pos(&bch_keylist_front(keys)->k));
 279         bch_btree_iter_init_intent(&inode_iter, wop->c, BTREE_ID_INODES,
 280                                    POS(extent_iter.pos.inode, 0));
 281
 282         hook.op                 = op;
 283         hook.hook.fn            = bchfs_extent_update_hook;
 284         hook.need_inode_update  = false;
 285
 286         do {
 287                 ret = bch_btree_iter_traverse(&extent_iter);
 288                 if (ret)
 289                         goto err;
 290
 291                 /* XXX: ei->i_size locking */
 292                 k = bch_keylist_front(keys);
 293                 if (min(k->k.p.offset << 9, op->new_i_size) > op->ei->i_size)
 294                         hook.need_inode_update = true;
 295
 296                 if (hook.need_inode_update) {
 297                         struct bkey_s_c inode;
 298
 299                         if (!btree_iter_linked(&inode_iter))
 300                                 bch_btree_iter_link(&extent_iter, &inode_iter);
 301
 302                         inode = bch_btree_iter_peek_with_holes(&inode_iter);
 303                         if ((ret = btree_iter_err(inode)))
 304                                 goto err;
 305
 306                         if (WARN_ONCE(inode.k->type != BCH_INODE_FS,
 307                                       "inode %llu not found when updating",
 308                                       extent_iter.pos.inode)) {
 309                                 ret = -ENOENT;
 310                                 break;
 311                         }
 312
 313                         bkey_reassemble(&hook.new_inode.k_i, inode);
 314
 315                         ret = bch_btree_insert_at(wop->c, &wop->res,
 316                                         &hook.hook, op_journal_seq(wop),
 317                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 318                                         BTREE_INSERT_ENTRY(&extent_iter, k),
 319                                         BTREE_INSERT_ENTRY(&inode_iter, &hook.new_inode.k_i));
 320                 } else {
 321                         ret = bch_btree_insert_at(wop->c, &wop->res,
 322                                         &hook.hook, op_journal_seq(wop),
 323                                         BTREE_INSERT_NOFAIL|BTREE_INSERT_ATOMIC,
 324                                         BTREE_INSERT_ENTRY(&extent_iter, k));
 325                 }
 326 err:
 327                 if (ret == -EINTR)
 328                         continue;
 329                 if (ret)
 330                         break;
 331
 332                 bch_keylist_pop_front(keys);
 333         } while (!bch_keylist_empty(keys));
 334
 335         bch_btree_iter_unlock(&extent_iter);
 336         bch_btree_iter_unlock(&inode_iter);
 337
 338         return ret;
 339 }
 340
 341 /* page state: */
 342
 343 /* stored in page->private: */
 344
 345 /*
 346  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 347  * almost protected it with the page lock, except that bch_writepage_io_done has
 348  * to update the sector counts (and from interrupt/bottom half context).
 349  */
 350 struct bch_page_state {
 351 union { struct {
 352         /*
 353          * BCH_PAGE_ALLOCATED: page is _fully_ written on disk, and not
 354          * compressed - which means to write this page we don't have to reserve
 355          * space (the new write will never take up more space on disk than what
 356          * it's overwriting)
 357          *
 358          * BCH_PAGE_UNALLOCATED: page is not fully written on disk, or is
 359          * compressed - before writing we have to reserve space with
 360          * bch_reserve_sectors()
 361          *
 362          * BCH_PAGE_RESERVED: page has space reserved on disk (reservation will
 363          * be consumed when the page is written).
 364          */
 365         enum {
 366                 BCH_PAGE_UNALLOCATED    = 0,
 367                 BCH_PAGE_ALLOCATED,
 368         }                       alloc_state:2;
 369
 370         /* Owns PAGE_SECTORS sized reservation: */
 371         unsigned                reserved:1;
 372
 373         /*
 374          * Number of sectors on disk - for i_blocks
 375          * Uncompressed size, not compressed size:
 376          */
 377         u8                      sectors;
 378         u8                      dirty_sectors;
 379 };
 380         /* for cmpxchg: */
 381         unsigned long           v;
 382 };
 383 };
 384
 385 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 386 ({                                                                      \
 387         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 388         struct bch_page_state _old;                                     \
 389                                                                         \
 390         do {                                                            \
 391                 _old.v = _new.v = _v;                                   \
 392                 _expr;                                                  \
 393                                                                         \
 394                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 395         } while (_old.v != _new.v &&                                    \
 396                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 397                                                                         \
 398         _old;                                                           \
 399 })
 400
 401 static inline struct bch_page_state *page_state(struct page *page)
 402 {
 403         struct bch_page_state *s = (void *) &page->private;
 404
 405         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 406
 407         if (!PagePrivate(page))
 408                 SetPagePrivate(page);
 409
 410         return s;
 411 }
 412
 413 static void bch_put_page_reservation(struct cache_set *c, struct page *page)
 414 {
 415         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 416         struct bch_page_state s;
 417
 418         s = page_state_cmpxchg(page_state(page), s, {
 419                 if (!s.reserved)
 420                         return;
 421                 s.reserved = 0;
 422         });
 423
 424         bch_disk_reservation_put(c, &res);
 425 }
 426
 427 static int bch_get_page_reservation(struct cache_set *c, struct page *page,
 428                                     bool check_enospc)
 429 {
 430         struct bch_page_state *s = page_state(page), new;
 431         struct disk_reservation res;
 432         int ret = 0;
 433
 434         BUG_ON(s->alloc_state == BCH_PAGE_ALLOCATED &&
 435                s->sectors != PAGE_SECTORS);
 436
 437         if (s->reserved ||
 438             s->alloc_state == BCH_PAGE_ALLOCATED)
 439                 return 0;
 440
 441         ret = bch_disk_reservation_get(c, &res, PAGE_SECTORS, !check_enospc
 442                                        ? BCH_DISK_RESERVATION_NOFAIL : 0);
 443         if (ret)
 444                 return ret;
 445
 446         page_state_cmpxchg(s, new, {
 447                 if (new.reserved) {
 448                         bch_disk_reservation_put(c, &res);
 449                         return 0;
 450                 }
 451                 new.reserved = 1;
 452         });
 453
 454         return 0;
 455 }
 456
 457 static void bch_clear_page_bits(struct page *page)
 458 {
 459         struct inode *inode = page->mapping->host;
 460         struct cache_set *c = inode->i_sb->s_fs_info;
 461         struct disk_reservation res = { .sectors = PAGE_SECTORS };
 462         struct bch_page_state s;
 463
 464         if (!PagePrivate(page))
 465                 return;
 466
 467         s = xchg(page_state(page), (struct bch_page_state) { .v = 0 });
 468         ClearPagePrivate(page);
 469
 470         if (s.dirty_sectors) {
 471                 spin_lock(&inode->i_lock);
 472                 inode->i_blocks -= s.dirty_sectors;
 473                 spin_unlock(&inode->i_lock);
 474         }
 475
 476         if (s.reserved)
 477                 bch_disk_reservation_put(c, &res);
 478 }
 479
 480 int bch_set_page_dirty(struct page *page)
 481 {
 482         struct bch_page_state old, new;
 483
 484         old = page_state_cmpxchg(page_state(page), new,
 485                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 486         );
 487
 488         if (old.dirty_sectors != new.dirty_sectors) {
 489                 struct inode *inode = page->mapping->host;
 490
 491                 spin_lock(&inode->i_lock);
 492                 inode->i_blocks += new.dirty_sectors - old.dirty_sectors;
 493                 spin_unlock(&inode->i_lock);
 494         }
 495
 496         return __set_page_dirty_nobuffers(page);
 497 }
 498
 499 /* readpages/writepages: */
 500
 501 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 502 {
 503         sector_t offset = (sector_t) page->index << (PAGE_SHIFT - 9);
 504
 505         return bio->bi_vcnt < bio->bi_max_vecs &&
 506                 bio_end_sector(bio) == offset;
 507 }
 508
 509 static int bio_add_page_contig(struct bio *bio, struct page *page)
 510 {
 511         sector_t offset = (sector_t) page->index << (PAGE_SHIFT - 9);
 512
 513         BUG_ON(!bio->bi_max_vecs);
 514
 515         if (!bio->bi_vcnt)
 516                 bio->bi_iter.bi_sector = offset;
 517         else if (!bio_can_add_page_contig(bio, page))
 518                 return -1;
 519
 520         bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
 521                 .bv_page = page,
 522                 .bv_len = PAGE_SIZE,
 523                 .bv_offset = 0,
 524         };
 525
 526         bio->bi_iter.bi_size += PAGE_SIZE;
 527
 528         return 0;
 529 }
 530
 531 static void bch_readpages_end_io(struct bio *bio)
 532 {
 533         struct bio_vec *bv;
 534         int i;
 535
 536         bio_for_each_segment_all(bv, bio, i) {
 537                 struct page *page = bv->bv_page;
 538
 539                 if (!bio->bi_error) {
 540                         SetPageUptodate(page);
 541                 } else {
 542                         ClearPageUptodate(page);
 543                         SetPageError(page);
 544                 }
 545                 unlock_page(page);
 546         }
 547
 548         bio_put(bio);
 549 }
 550
 551 static inline struct page *__readpage_next_page(struct address_space *mapping,
 552                                                 struct list_head *pages,
 553                                                 unsigned *nr_pages)
 554 {
 555         struct page *page;
 556         int ret;
 557
 558         while (*nr_pages) {
 559                 page = list_entry(pages->prev, struct page, lru);
 560                 prefetchw(&page->flags);
 561                 list_del(&page->lru);
 562
 563                 ret = add_to_page_cache_lru(page, mapping, page->index, GFP_NOFS);
 564
 565                 /* if add_to_page_cache_lru() succeeded, page is locked: */
 566                 put_page(page);
 567
 568                 if (!ret)
 569                         return page;
 570
 571                 (*nr_pages)--;
 572         }
 573
 574         return NULL;
 575 }
 576
 577 #define for_each_readpage_page(_mapping, _pages, _nr_pages, _page)      \
 578         for (;                                                          \
 579              ((_page) = __readpage_next_page(_mapping, _pages, &(_nr_pages)));\
 580              (_nr_pages)--)
 581
 582 static void bch_mark_pages_unalloc(struct bio *bio)
 583 {
 584         struct bvec_iter iter;
 585         struct bio_vec bv;
 586
 587         bio_for_each_segment(bv, bio, iter)
 588                 page_state(bv.bv_page)->alloc_state = BCH_PAGE_UNALLOCATED;
 589 }
 590
 591 static void bch_add_page_sectors(struct bio *bio, const struct bkey *k)
 592 {
 593         struct bvec_iter iter;
 594         struct bio_vec bv;
 595
 596         bio_for_each_segment(bv, bio, iter) {
 597                 struct bch_page_state *s = page_state(bv.bv_page);
 598
 599                 /* sectors in @k from the start of this page: */
 600                 unsigned k_sectors = k->size - (iter.bi_sector - k->p.offset);
 601
 602                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 603
 604                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 605
 606                 s->sectors += page_sectors;
 607         }
 608 }
 609
 610 static void bchfs_read(struct cache_set *c, struct bch_read_bio *rbio, u64 inode)
 611 {
 612         struct bio *bio = &rbio->bio;
 613         struct btree_iter iter;
 614         struct bkey_s_c k;
 615         struct bio_vec *bv;
 616         unsigned i;
 617         int ret;
 618
 619         bch_increment_clock(c, bio_sectors(bio), READ);
 620
 621         /*
 622          * Initialize page state:
 623          * If a page is partly allocated and partly a hole, we want it to be
 624          * marked BCH_PAGE_UNALLOCATED - so we initially mark all pages
 625          * allocated and then mark them unallocated as we find holes:
 626          *
 627          * Note that the bio hasn't been split yet - it's the only bio that
 628          * points to these pages. As we walk extents and split @bio, that
 629          * necessarily be true, the splits won't necessarily be on page
 630          * boundaries:
 631          */
 632         bio_for_each_segment_all(bv, bio, i) {
 633                 struct bch_page_state *s = page_state(bv->bv_page);
 634
 635                 EBUG_ON(s->reserved);
 636
 637                 s->alloc_state = BCH_PAGE_ALLOCATED;
 638                 s->sectors = 0;
 639         }
 640
 641         for_each_btree_key_with_holes(&iter, c, BTREE_ID_EXTENTS,
 642                                       POS(inode, bio->bi_iter.bi_sector), k) {
 643                 BKEY_PADDED(k) tmp;
 644                 struct extent_pick_ptr pick;
 645                 unsigned bytes, sectors;
 646                 bool is_last;
 647
 648                 bkey_reassemble(&tmp.k, k);
 649                 bch_btree_iter_unlock(&iter);
 650                 k = bkey_i_to_s_c(&tmp.k);
 651
 652                 if (!bkey_extent_is_allocation(k.k) ||
 653                     bkey_extent_is_compressed(c, k))
 654                         bch_mark_pages_unalloc(bio);
 655
 656                 bch_extent_pick_ptr(c, k, &pick);
 657                 if (IS_ERR(pick.ca)) {
 658                         bcache_io_error(c, bio, "no device to read from");
 659                         bio_endio(bio);
 660                         return;
 661                 }
 662
 663                 sectors = min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
 664                         bio->bi_iter.bi_sector;
 665                 bytes = sectors << 9;
 666                 is_last = bytes == bio->bi_iter.bi_size;
 667                 swap(bio->bi_iter.bi_size, bytes);
 668
 669                 if (bkey_extent_is_allocation(k.k))
 670                         bch_add_page_sectors(bio, k.k);
 671
 672                 if (pick.ca) {
 673                         PTR_BUCKET(pick.ca, &pick.ptr)->read_prio =
 674                                 c->prio_clock[READ].hand;
 675
 676                         bch_read_extent(c, rbio, k, &pick,
 677                                         BCH_READ_RETRY_IF_STALE|
 678                                         BCH_READ_PROMOTE|
 679                                         (is_last ? BCH_READ_IS_LAST : 0));
 680                 } else {
 681                         zero_fill_bio_iter(bio, bio->bi_iter);
 682
 683                         if (is_last)
 684                                 bio_endio(bio);
 685                 }
 686
 687                 if (is_last)
 688                         return;
 689
 690                 swap(bio->bi_iter.bi_size, bytes);
 691                 bio_advance(bio, bytes);
 692         }
 693
 694         /*
 695          * If we get here, it better have been because there was an error
 696          * reading a btree node
 697          */
 698         ret = bch_btree_iter_unlock(&iter);
 699         BUG_ON(!ret);
 700         bcache_io_error(c, bio, "btree IO error %i", ret);
 701         bio_endio(bio);
 702 }
 703
 704 int bch_readpages(struct file *file, struct address_space *mapping,
 705                   struct list_head *pages, unsigned nr_pages)
 706 {
 707         struct inode *inode = mapping->host;
 708         struct cache_set *c = inode->i_sb->s_fs_info;
 709         struct bch_read_bio *rbio = NULL;
 710         struct page *page;
 711
 712         pr_debug("reading %u pages", nr_pages);
 713
 714         if (current->pagecache_lock != &mapping->add_lock)
 715                 pagecache_add_get(&mapping->add_lock);
 716
 717         for_each_readpage_page(mapping, pages, nr_pages, page) {
 718 again:
 719                 if (!rbio) {
 720                         rbio = container_of(bio_alloc_bioset(GFP_NOFS,
 721                                                 min_t(unsigned, nr_pages,
 722                                                       BIO_MAX_PAGES),
 723                                                 &c->bio_read),
 724                                            struct bch_read_bio, bio);
 725
 726                         rbio->bio.bi_end_io = bch_readpages_end_io;
 727                 }
 728
 729                 if (bio_add_page_contig(&rbio->bio, page)) {
 730                         bchfs_read(c, rbio, inode->i_ino);
 731                         rbio = NULL;
 732                         goto again;
 733                 }
 734         }
 735
 736         if (rbio)
 737                 bchfs_read(c, rbio, inode->i_ino);
 738
 739         if (current->pagecache_lock != &mapping->add_lock)
 740                 pagecache_add_put(&mapping->add_lock);
 741
 742         pr_debug("success");
 743         return 0;
 744 }
 745
 746 int bch_readpage(struct file *file, struct page *page)
 747 {
 748         struct address_space *mapping = page->mapping;
 749         struct inode *inode = mapping->host;
 750         struct cache_set *c = inode->i_sb->s_fs_info;
 751         struct bch_read_bio *rbio;
 752
 753         rbio = container_of(bio_alloc_bioset(GFP_NOFS, 1,
 754                                             &c->bio_read),
 755                            struct bch_read_bio, bio);
 756         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 757         rbio->bio.bi_end_io = bch_readpages_end_io;
 758
 759         bio_add_page_contig(&rbio->bio, page);
 760         bchfs_read(c, rbio, inode->i_ino);
 761
 762         return 0;
 763 }
 764
 765 struct bch_writepage_state {
 766         struct bch_writepage_io *io;
 767 };
 768
 769 static void bch_writepage_io_free(struct closure *cl)
 770 {
 771         struct bch_writepage_io *io = container_of(cl,
 772                                         struct bch_writepage_io, cl);
 773         struct bio *bio = &io->bio.bio;
 774
 775         bio_put(bio);
 776 }
 777
 778 static void bch_writepage_io_done(struct closure *cl)
 779 {
 780         struct bch_writepage_io *io = container_of(cl,
 781                                         struct bch_writepage_io, cl);
 782         struct cache_set *c = io->op.op.c;
 783         struct bio *bio = &io->bio.bio;
 784         struct bio_vec *bvec;
 785         unsigned i;
 786
 787         atomic_sub(bio->bi_vcnt, &c->writeback_pages);
 788         wake_up(&c->writeback_wait);
 789
 790         bio_for_each_segment_all(bvec, bio, i) {
 791                 struct page *page = bvec->bv_page;
 792
 793                 if (io->op.op.error) {
 794                         SetPageError(page);
 795                         if (page->mapping)
 796                                 set_bit(AS_EIO, &page->mapping->flags);
 797                 }
 798
 799                 if (io->op.op.written >= PAGE_SECTORS) {
 800                         struct bch_page_state old, new;
 801
 802                         old = page_state_cmpxchg(page_state(page), new, {
 803                                 new.sectors = PAGE_SECTORS;
 804                                 new.dirty_sectors = 0;
 805                         });
 806
 807                         io->op.sectors_added -= old.dirty_sectors;
 808                         io->op.op.written -= PAGE_SECTORS;
 809                 }
 810         }
 811
 812         /*
 813          * racing with fallocate can cause us to add fewer sectors than
 814          * expected - but we shouldn't add more sectors than expected:
 815          *
 816          * (error (due to going RO) halfway through a page can screw that up
 817          * slightly)
 818          */
 819         BUG_ON(io->op.sectors_added >= (s64) PAGE_SECTORS);
 820
 821         /*
 822          * PageWriteback is effectively our ref on the inode - fixup i_blocks
 823          * before calling end_page_writeback:
 824          */
 825         if (io->op.sectors_added) {
 826                 struct inode *inode = &io->op.ei->vfs_inode;
 827
 828                 spin_lock(&inode->i_lock);
 829                 inode->i_blocks += io->op.sectors_added;
 830                 spin_unlock(&inode->i_lock);
 831         }
 832
 833         bio_for_each_segment_all(bvec, bio, i)
 834                 end_page_writeback(bvec->bv_page);
 835
 836         closure_return_with_destructor(&io->cl, bch_writepage_io_free);
 837 }
 838
 839 static void bch_writepage_do_io(struct bch_writepage_state *w)
 840 {
 841         struct bch_writepage_io *io = w->io;
 842
 843         w->io = NULL;
 844         atomic_add(io->bio.bio.bi_vcnt, &io->op.op.c->writeback_pages);
 845
 846         io->op.op.pos.offset = io->bio.bio.bi_iter.bi_sector;
 847
 848         closure_call(&io->op.op.cl, bch_write, NULL, &io->cl);
 849         continue_at(&io->cl, bch_writepage_io_done, NULL);
 850 }
 851
 852 /*
 853  * Get a bch_writepage_io and add @page to it - appending to an existing one if
 854  * possible, else allocating a new one:
 855  */
 856 static void bch_writepage_io_alloc(struct cache_set *c,
 857                                    struct bch_writepage_state *w,
 858                                    struct bch_inode_info *ei,
 859                                    struct page *page)
 860 {
 861         u64 inum = ei->vfs_inode.i_ino;
 862
 863         if (!w->io) {
 864 alloc_io:
 865                 w->io = container_of(bio_alloc_bioset(GFP_NOFS,
 866                                                       BIO_MAX_PAGES,
 867                                                       bch_writepage_bioset),
 868                                      struct bch_writepage_io, bio.bio);
 869
 870                 closure_init(&w->io->cl, NULL);
 871                 w->io->op.ei            = ei;
 872                 w->io->op.sectors_added = 0;
 873                 w->io->op.is_dio        = false;
 874                 bch_write_op_init(&w->io->op.op, c, &w->io->bio,
 875                                   (struct disk_reservation) {
 876                                         .nr_replicas = c->opts.data_replicas,
 877                                   },
 878                                   foreground_write_point(c, inum),
 879                                   POS(inum, 0),
 880                                   &ei->journal_seq, 0);
 881                 w->io->op.op.index_update_fn = bchfs_write_index_update;
 882         }
 883
 884         if (bio_add_page_contig(&w->io->bio.bio, page)) {
 885                 bch_writepage_do_io(w);
 886                 goto alloc_io;
 887         }
 888
 889         /*
 890          * We shouldn't ever be handed pages for multiple inodes in a single
 891          * pass - right?
 892          */
 893         BUG_ON(ei != w->io->op.ei);
 894 }
 895
 896 static int __bch_writepage(struct cache_set *c, struct page *page,
 897                            struct writeback_control *wbc,
 898                            struct bch_writepage_state *w)
 899 {
 900         struct inode *inode = page->mapping->host;
 901         struct bch_inode_info *ei = to_bch_ei(inode);
 902         struct bch_page_state new, old;
 903         unsigned offset;
 904         loff_t i_size = i_size_read(inode);
 905         pgoff_t end_index = i_size >> PAGE_SHIFT;
 906
 907         EBUG_ON(!PageUptodate(page));
 908
 909         /* Is the page fully inside i_size? */
 910         if (page->index < end_index)
 911                 goto do_io;
 912
 913         /* Is the page fully outside i_size? (truncate in progress) */
 914         offset = i_size & (PAGE_SIZE - 1);
 915         if (page->index > end_index || !offset) {
 916                 unlock_page(page);
 917                 return 0;
 918         }
 919
 920         /*
 921          * The page straddles i_size.  It must be zeroed out on each and every
 922          * writepage invocation because it may be mmapped.  "A file is mapped
 923          * in multiples of the page size.  For a file that is not a multiple of
 924          * the  page size, the remaining memory is zeroed when mapped, and
 925          * writes to that region are not written out to the file."
 926          */
 927         zero_user_segment(page, offset, PAGE_SIZE);
 928 do_io:
 929         bch_writepage_io_alloc(c, w, ei, page);
 930
 931         /* while page is locked: */
 932         w->io->op.new_i_size = i_size;
 933
 934         if (wbc->sync_mode == WB_SYNC_ALL)
 935                 w->io->bio.bio.bi_opf |= WRITE_SYNC;
 936
 937         /* Before unlocking the page, transfer reservation to w->io: */
 938         old = page_state_cmpxchg(page_state(page), new, {
 939                 BUG_ON(!new.reserved &&
 940                        (new.sectors != PAGE_SECTORS ||
 941                         new.alloc_state != BCH_PAGE_ALLOCATED));
 942
 943                 if (new.alloc_state == BCH_PAGE_ALLOCATED &&
 944                     w->io->op.op.compression_type != BCH_COMPRESSION_NONE)
 945                         new.alloc_state = BCH_PAGE_UNALLOCATED;
 946                 else if (!new.reserved)
 947                         goto out;
 948                 new.reserved = 0;
 949         });
 950
 951         w->io->op.op.res.sectors += PAGE_SECTORS * (old.reserved - new.reserved);
 952 out:
 953         BUG_ON(PageWriteback(page));
 954         set_page_writeback(page);
 955         unlock_page(page);
 956
 957         return 0;
 958 }
 959
 960 int bch_writepages(struct address_space *mapping, struct writeback_control *wbc)
 961 {
 962         struct cache_set *c = mapping->host->i_sb->s_fs_info;
 963         struct bch_writepage_state w = { NULL };
 964         struct pagecache_iter iter;
 965         struct page *page;
 966         int ret = 0;
 967         int done = 0;
 968         pgoff_t uninitialized_var(writeback_index);
 969         pgoff_t index;
 970         pgoff_t end;            /* Inclusive */
 971         pgoff_t done_index;
 972         int cycled;
 973         int range_whole = 0;
 974         int tag;
 975
 976         if (wbc->range_cyclic) {
 977                 writeback_index = mapping->writeback_index; /* prev offset */
 978                 index = writeback_index;
 979                 if (index == 0)
 980                         cycled = 1;
 981                 else
 982                         cycled = 0;
 983                 end = -1;
 984         } else {
 985                 index = wbc->range_start >> PAGE_SHIFT;
 986                 end = wbc->range_end >> PAGE_SHIFT;
 987                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 988                         range_whole = 1;
 989                 cycled = 1; /* ignore range_cyclic tests */
 990         }
 991         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 992                 tag = PAGECACHE_TAG_TOWRITE;
 993         else
 994                 tag = PAGECACHE_TAG_DIRTY;
 995 retry:
 996         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 997                 tag_pages_for_writeback(mapping, index, end);
 998
 999         done_index = index;
1000 get_pages:
1001         for_each_pagecache_tag(&iter, mapping, tag, index, end, page) {
1002                 done_index = page->index;
1003
1004                 if (w.io &&
1005                     !bio_can_add_page_contig(&w.io->bio.bio, page))
1006                         bch_writepage_do_io(&w);
1007
1008                 if (!w.io &&
1009                     atomic_read(&c->writeback_pages) >=
1010                     c->writeback_pages_max) {
1011                         /* don't sleep with pages pinned: */
1012                         pagecache_iter_release(&iter);
1013
1014                         __wait_event(c->writeback_wait,
1015                                      atomic_read(&c->writeback_pages) <
1016                                      c->writeback_pages_max);
1017                         goto get_pages;
1018                 }
1019
1020                 lock_page(page);
1021
1022                 /*
1023                  * Page truncated or invalidated. We can freely skip it
1024                  * then, even for data integrity operations: the page
1025                  * has disappeared concurrently, so there could be no
1026                  * real expectation of this data interity operation
1027                  * even if there is now a new, dirty page at the same
1028                  * pagecache address.
1029                  */
1030                 if (unlikely(page->mapping != mapping)) {
1031 continue_unlock:
1032                         unlock_page(page);
1033                         continue;
1034                 }
1035
1036                 if (!PageDirty(page)) {
1037                         /* someone wrote it for us */
1038                         goto continue_unlock;
1039                 }
1040
1041                 if (PageWriteback(page)) {
1042                         if (wbc->sync_mode != WB_SYNC_NONE)
1043                                 wait_on_page_writeback(page);
1044                         else
1045                                 goto continue_unlock;
1046                 }
1047
1048                 BUG_ON(PageWriteback(page));
1049                 if (!clear_page_dirty_for_io(page))
1050                         goto continue_unlock;
1051
1052                 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
1053                 ret = __bch_writepage(c, page, wbc, &w);
1054                 if (unlikely(ret)) {
1055                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
1056                                 unlock_page(page);
1057                                 ret = 0;
1058                         } else {
1059                                 /*
1060                                  * done_index is set past this page,
1061                                  * so media errors will not choke
1062                                  * background writeout for the entire
1063                                  * file. This has consequences for
1064                                  * range_cyclic semantics (ie. it may
1065                                  * not be suitable for data integrity
1066                                  * writeout).
1067                                  */
1068                                 done_index = page->index + 1;
1069                                 done = 1;
1070                                 break;
1071                         }
1072                 }
1073
1074                 /*
1075                  * We stop writing back only if we are not doing
1076                  * integrity sync. In case of integrity sync we have to
1077                  * keep going until we have written all the pages
1078                  * we tagged for writeback prior to entering this loop.
1079                  */
1080                 if (--wbc->nr_to_write <= 0 &&
1081                     wbc->sync_mode == WB_SYNC_NONE) {
1082                         done = 1;
1083                         break;
1084                 }
1085         }
1086         pagecache_iter_release(&iter);
1087
1088         if (w.io)
1089                 bch_writepage_do_io(&w);
1090
1091         if (!cycled && !done) {
1092                 /*
1093                  * range_cyclic:
1094                  * We hit the last page and there is more work to be done: wrap
1095                  * back to the start of the file
1096                  */
1097                 cycled = 1;
1098                 index = 0;
1099                 end = writeback_index - 1;
1100                 goto retry;
1101         }
1102         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1103                 mapping->writeback_index = done_index;
1104
1105         return ret;
1106 }
1107
1108 int bch_writepage(struct page *page, struct writeback_control *wbc)
1109 {
1110         struct cache_set *c = page->mapping->host->i_sb->s_fs_info;
1111         struct bch_writepage_state w = { NULL };
1112         int ret;
1113
1114         ret = __bch_writepage(c, page, wbc, &w);
1115         if (w.io)
1116                 bch_writepage_do_io(&w);
1117
1118         return ret;
1119 }
1120
1121 static void bch_read_single_page_end_io(struct bio *bio)
1122 {
1123         complete(bio->bi_private);
1124 }
1125
1126 static int bch_read_single_page(struct page *page,
1127                                 struct address_space *mapping)
1128 {
1129         struct inode *inode = mapping->host;
1130         struct cache_set *c = inode->i_sb->s_fs_info;
1131         struct bch_read_bio *rbio;
1132         int ret;
1133         DECLARE_COMPLETION_ONSTACK(done);
1134
1135         rbio = container_of(bio_alloc_bioset(GFP_NOFS, 1,
1136                                              &c->bio_read),
1137                             struct bch_read_bio, bio);
1138         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1139         rbio->bio.bi_private = &done;
1140         rbio->bio.bi_end_io = bch_read_single_page_end_io;
1141         bio_add_page_contig(&rbio->bio, page);
1142
1143         bchfs_read(c, rbio, inode->i_ino);
1144         wait_for_completion(&done);
1145
1146         ret = rbio->bio.bi_error;
1147         bio_put(&rbio->bio);
1148
1149         if (ret < 0)
1150                 return ret;
1151
1152         SetPageUptodate(page);
1153         return 0;
1154 }
1155
1156 int bch_write_begin(struct file *file, struct address_space *mapping,
1157                     loff_t pos, unsigned len, unsigned flags,
1158                     struct page **pagep, void **fsdata)
1159 {
1160         struct inode *inode = mapping->host;
1161         struct cache_set *c = inode->i_sb->s_fs_info;
1162         pgoff_t index = pos >> PAGE_SHIFT;
1163         unsigned offset = pos & (PAGE_SIZE - 1);
1164         struct page *page;
1165         int ret = -ENOMEM;
1166
1167         BUG_ON(inode_unhashed(mapping->host));
1168
1169         /* Not strictly necessary - same reason as mkwrite(): */
1170         pagecache_add_get(&mapping->add_lock);
1171
1172         page = grab_cache_page_write_begin(mapping, index, flags);
1173         if (!page)
1174                 goto err_unlock;
1175
1176         if (PageUptodate(page))
1177                 goto out;
1178
1179         /* If we're writing entire page, don't need to read it in first: */
1180         if (len == PAGE_SIZE)
1181                 goto out;
1182
1183         if (!offset && pos + len >= inode->i_size) {
1184                 zero_user_segment(page, len, PAGE_SIZE);
1185                 flush_dcache_page(page);
1186                 goto out;
1187         }
1188
1189         if (index > inode->i_size >> PAGE_SHIFT) {
1190                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1191                 flush_dcache_page(page);
1192                 goto out;
1193         }
1194 readpage:
1195         ret = bch_read_single_page(page, mapping);
1196         if (ret)
1197                 goto err;
1198 out:
1199         ret = bch_get_page_reservation(c, page, true);
1200         if (ret) {
1201                 if (!PageUptodate(page)) {
1202                         /*
1203                          * If the page hasn't been read in, we won't know if we
1204                          * actually need a reservation - we don't actually need
1205                          * to read here, we just need to check if the page is
1206                          * fully backed by uncompressed data:
1207                          */
1208                         goto readpage;
1209                 }
1210
1211                 goto err;
1212         }
1213
1214         *pagep = page;
1215         return 0;
1216 err:
1217         unlock_page(page);
1218         put_page(page);
1219         *pagep = NULL;
1220 err_unlock:
1221         pagecache_add_put(&mapping->add_lock);
1222         return ret;
1223 }
1224
1225 int bch_write_end(struct file *filp, struct address_space *mapping,
1226                   loff_t pos, unsigned len, unsigned copied,
1227                   struct page *page, void *fsdata)
1228 {
1229         struct inode *inode = page->mapping->host;
1230         struct cache_set *c = inode->i_sb->s_fs_info;
1231
1232         lockdep_assert_held(&inode->i_rwsem);
1233
1234         if (unlikely(copied < len && !PageUptodate(page))) {
1235                 /*
1236                  * The page needs to be read in, but that would destroy
1237                  * our partial write - simplest thing is to just force
1238                  * userspace to redo the write:
1239                  */
1240                 zero_user(page, 0, PAGE_SIZE);
1241                 flush_dcache_page(page);
1242                 copied = 0;
1243         }
1244
1245         if (pos + copied > inode->i_size)
1246                 i_size_write(inode, pos + copied);
1247
1248         if (copied) {
1249                 if (!PageUptodate(page))
1250                         SetPageUptodate(page);
1251                 if (!PageDirty(page))
1252                         set_page_dirty(page);
1253         } else {
1254                 bch_put_page_reservation(c, page);
1255         }
1256
1257         unlock_page(page);
1258         put_page(page);
1259         pagecache_add_put(&mapping->add_lock);
1260
1261         return copied;
1262 }
1263
1264 /* O_DIRECT */
1265
1266 static void bch_dio_read_complete(struct closure *cl)
1267 {
1268         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1269
1270         dio->req->ki_complete(dio->req, dio->ret, 0);
1271         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1272 }
1273
1274 static void bch_direct_IO_read_endio(struct bio *bio)
1275 {
1276         struct dio_read *dio = bio->bi_private;
1277
1278         if (bio->bi_error)
1279                 dio->ret = bio->bi_error;
1280
1281         closure_put(&dio->cl);
1282 }
1283
1284 static void bch_direct_IO_read_split_endio(struct bio *bio)
1285 {
1286         bch_direct_IO_read_endio(bio);
1287         bio_check_pages_dirty(bio);     /* transfers ownership */
1288 }
1289
1290 static int bch_direct_IO_read(struct cache_set *c, struct kiocb *req,
1291                               struct file *file, struct inode *inode,
1292                               struct iov_iter *iter, loff_t offset)
1293 {
1294         struct dio_read *dio;
1295         struct bio *bio;
1296         bool sync = is_sync_kiocb(req);
1297         ssize_t ret;
1298
1299         if ((offset|iter->count) & (block_bytes(c) - 1))
1300                 return -EINVAL;
1301
1302         ret = min_t(loff_t, iter->count,
1303                     max_t(loff_t, 0, i_size_read(inode) - offset));
1304         iov_iter_truncate(iter, round_up(ret, block_bytes(c)));
1305
1306         if (!ret)
1307                 return ret;
1308
1309         bio = bio_alloc_bioset(GFP_KERNEL,
1310                                iov_iter_npages(iter, BIO_MAX_PAGES),
1311                                bch_dio_read_bioset);
1312
1313         bio->bi_end_io = bch_direct_IO_read_endio;
1314
1315         dio = container_of(bio, struct dio_read, rbio.bio);
1316         closure_init(&dio->cl, NULL);
1317
1318         /*
1319          * this is a _really_ horrible hack just to avoid an atomic sub at the
1320          * end:
1321          */
1322         if (!sync) {
1323                 set_closure_fn(&dio->cl, bch_dio_read_complete, NULL);
1324                 atomic_set(&dio->cl.remaining,
1325                            CLOSURE_REMAINING_INITIALIZER -
1326                            CLOSURE_RUNNING +
1327                            CLOSURE_DESTRUCTOR);
1328         } else {
1329                 atomic_set(&dio->cl.remaining,
1330                            CLOSURE_REMAINING_INITIALIZER + 1);
1331         }
1332
1333         dio->req        = req;
1334         dio->ret        = ret;
1335
1336         goto start;
1337         while (iter->count) {
1338                 bio = bio_alloc_bioset(GFP_KERNEL,
1339                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1340                                        &c->bio_read);
1341                 bio->bi_end_io          = bch_direct_IO_read_split_endio;
1342 start:
1343                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1344                 bio->bi_iter.bi_sector  = offset >> 9;
1345                 bio->bi_private         = dio;
1346
1347                 ret = bio_get_user_pages(bio, iter, 1);
1348                 if (ret < 0) {
1349                         /* XXX: fault inject this path */
1350                         bio->bi_error = ret;
1351                         bio_endio(bio);
1352                         break;
1353                 }
1354
1355                 offset += bio->bi_iter.bi_size;
1356                 bio_set_pages_dirty(bio);
1357
1358                 if (iter->count)
1359                         closure_get(&dio->cl);
1360
1361                 bch_read(c, container_of(bio,
1362                                 struct bch_read_bio, bio),
1363                          inode->i_ino);
1364         }
1365
1366         if (sync) {
1367                 closure_sync(&dio->cl);
1368                 closure_debug_destroy(&dio->cl);
1369                 ret = dio->ret;
1370                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1371                 return ret;
1372         } else {
1373                 return -EIOCBQUEUED;
1374         }
1375 }
1376
1377 static long __bch_dio_write_complete(struct dio_write *dio)
1378 {
1379         struct file *file = dio->req->ki_filp;
1380         struct address_space *mapping = file->f_mapping;
1381         struct inode *inode = file->f_inode;
1382         long ret = dio->error ?: dio->written;
1383
1384         bch_disk_reservation_put(dio->c, &dio->res);
1385
1386         __pagecache_block_put(&mapping->add_lock);
1387         inode_dio_end(inode);
1388
1389         if (dio->iovec && dio->iovec != dio->inline_vecs)
1390                 kfree(dio->iovec);
1391
1392         bio_put(&dio->bio.bio);
1393         return ret;
1394 }
1395
1396 static void bch_dio_write_complete(struct closure *cl)
1397 {
1398         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1399         struct kiocb *req = dio->req;
1400
1401         req->ki_complete(req, __bch_dio_write_complete(dio), 0);
1402 }
1403
1404 static void bch_dio_write_done(struct dio_write *dio)
1405 {
1406         struct bio_vec *bv;
1407         int i;
1408
1409         dio->written += dio->iop.op.written << 9;
1410
1411         if (dio->iop.op.error)
1412                 dio->error = dio->iop.op.error;
1413
1414         bio_for_each_segment_all(bv, &dio->bio.bio, i)
1415                 put_page(bv->bv_page);
1416
1417         if (dio->iter.count)
1418                 bio_reset(&dio->bio.bio);
1419 }
1420
1421 static void bch_do_direct_IO_write(struct dio_write *dio)
1422 {
1423         struct file *file = dio->req->ki_filp;
1424         struct inode *inode = file->f_inode;
1425         struct bch_inode_info *ei = to_bch_ei(inode);
1426         struct bio *bio = &dio->bio.bio;
1427         unsigned flags = 0;
1428         int ret;
1429
1430         if ((dio->req->ki_flags & IOCB_DSYNC) &&
1431             !dio->c->opts.journal_flush_disabled)
1432                 flags |= BCH_WRITE_FLUSH;
1433
1434         bio->bi_iter.bi_sector = (dio->offset + dio->written) >> 9;
1435
1436         ret = bio_get_user_pages(bio, &dio->iter, 0);
1437         if (ret < 0) {
1438                 /*
1439                  * these didn't get initialized, but bch_dio_write_done() will
1440                  * look at them:
1441                  */
1442                 dio->iop.op.error = 0;
1443                 dio->iop.op.written = 0;
1444                 dio->error = ret;
1445                 return;
1446         }
1447
1448         dio->iop.ei             = ei;
1449         dio->iop.sectors_added  = 0;
1450         dio->iop.is_dio         = true;
1451         dio->iop.new_i_size     = U64_MAX;
1452         bch_write_op_init(&dio->iop.op, dio->c, &dio->bio,
1453                           dio->res,
1454                           foreground_write_point(dio->c, inode->i_ino),
1455                           POS(inode->i_ino, bio->bi_iter.bi_sector),
1456                           &ei->journal_seq, flags);
1457         dio->iop.op.index_update_fn = bchfs_write_index_update;
1458
1459         dio->res.sectors -= bio_sectors(bio);
1460         dio->iop.op.res.sectors = bio_sectors(bio);
1461
1462         task_io_account_write(bio->bi_iter.bi_size);
1463
1464         closure_call(&dio->iop.op.cl, bch_write, NULL, &dio->cl);
1465 }
1466
1467 static void bch_dio_write_loop_async(struct closure *cl)
1468 {
1469         struct dio_write *dio =
1470                 container_of(cl, struct dio_write, cl);
1471         struct address_space *mapping = dio->req->ki_filp->f_mapping;
1472
1473         bch_dio_write_done(dio);
1474
1475         if (dio->iter.count && !dio->error) {
1476                 use_mm(dio->mm);
1477                 pagecache_block_get(&mapping->add_lock);
1478
1479                 bch_do_direct_IO_write(dio);
1480
1481                 pagecache_block_put(&mapping->add_lock);
1482                 unuse_mm(dio->mm);
1483
1484                 continue_at(&dio->cl, bch_dio_write_loop_async, NULL);
1485         } else {
1486 #if 0
1487                 closure_return_with_destructor(cl, bch_dio_write_complete);
1488 #else
1489                 closure_debug_destroy(cl);
1490                 bch_dio_write_complete(cl);
1491 #endif
1492         }
1493 }
1494
1495 static int bch_direct_IO_write(struct cache_set *c, struct kiocb *req,
1496                                struct file *file, struct inode *inode,
1497                                struct iov_iter *iter, loff_t offset)
1498 {
1499         struct address_space *mapping = file->f_mapping;
1500         struct dio_write *dio;
1501         struct bio *bio;
1502         ssize_t ret;
1503         bool sync = is_sync_kiocb(req);
1504
1505         lockdep_assert_held(&inode->i_rwsem);
1506
1507         if (unlikely(!iter->count))
1508                 return 0;
1509
1510         if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
1511                 return -EINVAL;
1512
1513         bio = bio_alloc_bioset(GFP_KERNEL,
1514                                iov_iter_npages(iter, BIO_MAX_PAGES),
1515                                bch_dio_write_bioset);
1516         dio = container_of(bio, struct dio_write, bio.bio);
1517         dio->req        = req;
1518         dio->c          = c;
1519         dio->written    = 0;
1520         dio->error      = 0;
1521         dio->offset     = offset;
1522         dio->iovec      = NULL;
1523         dio->iter       = *iter;
1524         dio->mm         = current->mm;
1525         closure_init(&dio->cl, NULL);
1526
1527         if (offset + iter->count > inode->i_size)
1528                 sync = true;
1529
1530         /*
1531          * XXX: we shouldn't return -ENOSPC if we're overwriting existing data -
1532          * if getting a reservation fails we should check if we are doing an
1533          * overwrite.
1534          *
1535          * Have to then guard against racing with truncate (deleting data that
1536          * we would have been overwriting)
1537          */
1538         ret = bch_disk_reservation_get(c, &dio->res, iter->count >> 9, 0);
1539         if (unlikely(ret)) {
1540                 closure_debug_destroy(&dio->cl);
1541                 bio_put(bio);
1542                 return ret;
1543         }
1544
1545         inode_dio_begin(inode);
1546         __pagecache_block_get(&mapping->add_lock);
1547
1548         if (sync) {
1549                 do {
1550                         bch_do_direct_IO_write(dio);
1551
1552                         closure_sync(&dio->cl);
1553                         bch_dio_write_done(dio);
1554                 } while (dio->iter.count && !dio->error);
1555
1556                 closure_debug_destroy(&dio->cl);
1557                 return __bch_dio_write_complete(dio);
1558         } else {
1559                 bch_do_direct_IO_write(dio);
1560
1561                 if (dio->iter.count && !dio->error) {
1562                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1563                                 dio->iovec = kmalloc(dio->iter.nr_segs *
1564                                                      sizeof(struct iovec),
1565                                                      GFP_KERNEL);
1566                                 if (!dio->iovec)
1567                                         dio->error = -ENOMEM;
1568                         } else {
1569                                 dio->iovec = dio->inline_vecs;
1570                         }
1571
1572                         memcpy(dio->iovec,
1573                                dio->iter.iov,
1574                                dio->iter.nr_segs * sizeof(struct iovec));
1575                         dio->iter.iov = dio->iovec;
1576                 }
1577
1578                 continue_at_noreturn(&dio->cl, bch_dio_write_loop_async, NULL);
1579                 return -EIOCBQUEUED;
1580         }
1581 }
1582
1583 ssize_t bch_direct_IO(struct kiocb *req, struct iov_iter *iter)
1584 {
1585         struct file *file = req->ki_filp;
1586         struct inode *inode = file->f_inode;
1587         struct cache_set *c = inode->i_sb->s_fs_info;
1588
1589         return ((iov_iter_rw(iter) == WRITE)
1590                 ? bch_direct_IO_write
1591                 : bch_direct_IO_read)(c, req, file, inode, iter, req->ki_pos);
1592 }
1593
1594 static ssize_t
1595 bch_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1596 {
1597         struct file *file = iocb->ki_filp;
1598         struct inode *inode = file->f_inode;
1599         struct cache_set *c = inode->i_sb->s_fs_info;
1600         struct address_space *mapping = file->f_mapping;
1601         loff_t pos = iocb->ki_pos;
1602         ssize_t ret;
1603
1604         pagecache_block_get(&mapping->add_lock);
1605
1606         /* Write and invalidate pagecache range that we're writing to: */
1607         ret = write_invalidate_inode_pages_range(file->f_mapping, pos,
1608                                         pos + iov_iter_count(iter) - 1);
1609         if (unlikely(ret))
1610                 goto err;
1611
1612         ret = bch_direct_IO_write(c, iocb, file, inode, iter, pos);
1613 err:
1614         pagecache_block_put(&mapping->add_lock);
1615
1616         return ret;
1617 }
1618
1619 static ssize_t __bch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1620 {
1621         struct file *file = iocb->ki_filp;
1622         struct address_space *mapping = file->f_mapping;
1623         struct inode *inode = mapping->host;
1624         ssize_t ret;
1625
1626         /* We can write back this queue in page reclaim */
1627         current->backing_dev_info = inode_to_bdi(inode);
1628         ret = file_remove_privs(file);
1629         if (ret)
1630                 goto out;
1631
1632         ret = file_update_time(file);
1633         if (ret)
1634                 goto out;
1635
1636         ret = iocb->ki_flags & IOCB_DIRECT
1637                 ? bch_direct_write(iocb, from)
1638                 : generic_perform_write(file, from, iocb->ki_pos);
1639
1640         if (likely(ret > 0))
1641                 iocb->ki_pos += ret;
1642 out:
1643         current->backing_dev_info = NULL;
1644         return ret;
1645 }
1646
1647 ssize_t bch_write_iter(struct kiocb *iocb, struct iov_iter *from)
1648 {
1649         struct file *file = iocb->ki_filp;
1650         struct inode *inode = file->f_mapping->host;
1651         bool direct = iocb->ki_flags & IOCB_DIRECT;
1652         ssize_t ret;
1653
1654         inode_lock(inode);
1655         ret = generic_write_checks(iocb, from);
1656         if (ret > 0)
1657                 ret = __bch_write_iter(iocb, from);
1658         inode_unlock(inode);
1659
1660         if (ret > 0 && !direct)
1661                 ret = generic_write_sync(iocb, ret);
1662
1663         return ret;
1664 }
1665
1666 int bch_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1667 {
1668         struct page *page = vmf->page;
1669         struct inode *inode = file_inode(vma->vm_file);
1670         struct address_space *mapping = inode->i_mapping;
1671         struct cache_set *c = inode->i_sb->s_fs_info;
1672         int ret = VM_FAULT_LOCKED;
1673
1674         sb_start_pagefault(inode->i_sb);
1675         file_update_time(vma->vm_file);
1676
1677         /*
1678          * Not strictly necessary, but helps avoid dio writes livelocking in
1679          * write_invalidate_inode_pages_range() - can drop this if/when we get
1680          * a write_invalidate_inode_pages_range() that works without dropping
1681          * page lock before invalidating page
1682          */
1683         if (current->pagecache_lock != &mapping->add_lock)
1684                 pagecache_add_get(&mapping->add_lock);
1685
1686         lock_page(page);
1687         if (page->mapping != mapping ||
1688             page_offset(page) > i_size_read(inode)) {
1689                 unlock_page(page);
1690                 ret = VM_FAULT_NOPAGE;
1691                 goto out;
1692         }
1693
1694         if (bch_get_page_reservation(c, page, true)) {
1695                 unlock_page(page);
1696                 ret = VM_FAULT_SIGBUS;
1697                 goto out;
1698         }
1699
1700         if (!PageDirty(page))
1701                 set_page_dirty(page);
1702         wait_for_stable_page(page);
1703 out:
1704         if (current->pagecache_lock != &mapping->add_lock)
1705                 pagecache_add_put(&mapping->add_lock);
1706         sb_end_pagefault(inode->i_sb);
1707         return ret;
1708 }
1709
1710 void bch_invalidatepage(struct page *page, unsigned int offset,
1711                         unsigned int length)
1712 {
1713         EBUG_ON(!PageLocked(page));
1714         EBUG_ON(PageWriteback(page));
1715
1716         if (offset || length < PAGE_SIZE)
1717                 return;
1718
1719         bch_clear_page_bits(page);
1720 }
1721
1722 int bch_releasepage(struct page *page, gfp_t gfp_mask)
1723 {
1724         EBUG_ON(!PageLocked(page));
1725         EBUG_ON(PageWriteback(page));
1726
1727         if (PageDirty(page))
1728                 return 0;
1729
1730         bch_clear_page_bits(page);
1731         return 1;
1732 }
1733
1734 #ifdef CONFIG_MIGRATION
1735 int bch_migrate_page(struct address_space *mapping, struct page *newpage,
1736                      struct page *page, enum migrate_mode mode)
1737 {
1738         int ret;
1739
1740         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
1741         if (ret != MIGRATEPAGE_SUCCESS)
1742                 return ret;
1743
1744         if (PagePrivate(page)) {
1745                 *page_state(newpage) = *page_state(page);
1746                 ClearPagePrivate(page);
1747         }
1748
1749         migrate_page_copy(newpage, page);
1750         return MIGRATEPAGE_SUCCESS;
1751 }
1752 #endif
1753
1754 int bch_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1755 {
1756         struct inode *inode = file->f_mapping->host;
1757         struct bch_inode_info *ei = to_bch_ei(inode);
1758         struct cache_set *c = inode->i_sb->s_fs_info;
1759         int ret;
1760
1761         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1762         if (ret)
1763                 return ret;
1764
1765         if (c->opts.journal_flush_disabled)
1766                 return 0;
1767
1768         return bch_journal_flush_seq(&c->journal, ei->journal_seq);
1769 }
1770
1771 static int __bch_truncate_page(struct address_space *mapping,
1772                                pgoff_t index, loff_t start, loff_t end)
1773 {
1774         struct inode *inode = mapping->host;
1775         struct cache_set *c = inode->i_sb->s_fs_info;
1776         unsigned start_offset = start & (PAGE_SIZE - 1);
1777         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
1778         struct page *page;
1779         int ret = 0;
1780
1781         /* Page boundary? Nothing to do */
1782         if (!((index == start >> PAGE_SHIFT && start_offset) ||
1783               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
1784                 return 0;
1785
1786         /* Above i_size? */
1787         if (index << PAGE_SHIFT >= inode->i_size)
1788                 return 0;
1789
1790         page = find_lock_page(mapping, index);
1791         if (!page) {
1792                 struct btree_iter iter;
1793                 struct bkey_s_c k = bkey_s_c_null;
1794
1795                 /*
1796                  * XXX: we're doing two index lookups when we end up reading the
1797                  * page
1798                  */
1799                 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
1800                                    POS(inode->i_ino,
1801                                        index << (PAGE_SHIFT - 9)), k) {
1802                         if (bkey_cmp(bkey_start_pos(k.k),
1803                                      POS(inode->i_ino,
1804                                          (index + 1) << (PAGE_SHIFT - 9))) >= 0)
1805                                 break;
1806
1807                         if (k.k->type != KEY_TYPE_DISCARD &&
1808                             k.k->type != BCH_RESERVATION) {
1809                                 bch_btree_iter_unlock(&iter);
1810                                 goto create;
1811                         }
1812                 }
1813                 bch_btree_iter_unlock(&iter);
1814                 return 0;
1815 create:
1816                 page = find_or_create_page(mapping, index, GFP_KERNEL);
1817                 if (unlikely(!page)) {
1818                         ret = -ENOMEM;
1819                         goto out;
1820                 }
1821         }
1822
1823         if (!PageUptodate(page)) {
1824                 ret = bch_read_single_page(page, mapping);
1825                 if (ret)
1826                         goto unlock;
1827         }
1828
1829         /*
1830          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
1831          *
1832          * XXX: because we aren't currently tracking whether the page has actual
1833          * data in it (vs. just 0s, or only partially written) this wrong. ick.
1834          */
1835         ret = bch_get_page_reservation(c, page, false);
1836         BUG_ON(ret);
1837
1838         if (index == start >> PAGE_SHIFT &&
1839             index == end >> PAGE_SHIFT)
1840                 zero_user_segment(page, start_offset, end_offset);
1841         else if (index == start >> PAGE_SHIFT)
1842                 zero_user_segment(page, start_offset, PAGE_SIZE);
1843         else if (index == end >> PAGE_SHIFT)
1844                 zero_user_segment(page, 0, end_offset);
1845
1846         if (!PageDirty(page))
1847                 set_page_dirty(page);
1848 unlock:
1849         unlock_page(page);
1850         put_page(page);
1851 out:
1852         return ret;
1853 }
1854
1855 static int bch_truncate_page(struct address_space *mapping, loff_t from)
1856 {
1857         return __bch_truncate_page(mapping, from >> PAGE_SHIFT,
1858                                    from, from + PAGE_SIZE);
1859 }
1860
1861 int bch_truncate(struct inode *inode, struct iattr *iattr)
1862 {
1863         struct address_space *mapping = inode->i_mapping;
1864         struct bch_inode_info *ei = to_bch_ei(inode);
1865         struct cache_set *c = inode->i_sb->s_fs_info;
1866         bool shrink = iattr->ia_size <= inode->i_size;
1867         int ret = 0;
1868
1869         inode_dio_wait(inode);
1870         pagecache_block_get(&mapping->add_lock);
1871
1872         truncate_setsize(inode, iattr->ia_size);
1873
1874         /* sync appends.. */
1875         /* XXX what protects ei->i_size? */
1876         if (iattr->ia_size > ei->i_size)
1877                 ret = filemap_write_and_wait_range(mapping, ei->i_size, S64_MAX);
1878         if (ret)
1879                 goto err_put_pagecache;
1880
1881         mutex_lock(&ei->update_lock);
1882         i_size_dirty_get(ei);
1883         ret = bch_write_inode_size(c, ei, inode->i_size);
1884         mutex_unlock(&ei->update_lock);
1885
1886         if (unlikely(ret))
1887                 goto err;
1888
1889         /*
1890          * There might be persistent reservations (from fallocate())
1891          * above i_size, which bch_inode_truncate() will discard - we're
1892          * only supposed to discard them if we're doing a real truncate
1893          * here (new i_size < current i_size):
1894          */
1895         if (shrink) {
1896                 struct i_sectors_hook i_sectors_hook;
1897                 int ret;
1898
1899                 ret = i_sectors_dirty_get(ei, &i_sectors_hook);
1900                 if (unlikely(ret))
1901                         goto err;
1902
1903                 ret = bch_truncate_page(inode->i_mapping, iattr->ia_size);
1904                 if (unlikely(ret)) {
1905                         i_sectors_dirty_put(ei, &i_sectors_hook);
1906                         goto err;
1907                 }
1908
1909                 ret = bch_inode_truncate(c, inode->i_ino,
1910                                          round_up(iattr->ia_size, PAGE_SIZE) >> 9,
1911                                          &i_sectors_hook.hook,
1912                                          &ei->journal_seq);
1913
1914                 i_sectors_dirty_put(ei, &i_sectors_hook);
1915
1916                 if (unlikely(ret))
1917                         goto err;
1918         }
1919
1920         mutex_lock(&ei->update_lock);
1921         setattr_copy(inode, iattr);
1922         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1923
1924         /* clear I_SIZE_DIRTY: */
1925         i_size_dirty_put(ei);
1926         ret = bch_write_inode_size(c, ei, inode->i_size);
1927         mutex_unlock(&ei->update_lock);
1928
1929         pagecache_block_put(&mapping->add_lock);
1930
1931         return 0;
1932 err:
1933         i_size_dirty_put(ei);
1934 err_put_pagecache:
1935         pagecache_block_put(&mapping->add_lock);
1936         return ret;
1937 }
1938
1939 static long bch_fpunch(struct inode *inode, loff_t offset, loff_t len)
1940 {
1941         struct address_space *mapping = inode->i_mapping;
1942         struct bch_inode_info *ei = to_bch_ei(inode);
1943         struct cache_set *c = inode->i_sb->s_fs_info;
1944         u64 ino = inode->i_ino;
1945         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
1946         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
1947         int ret = 0;
1948
1949         inode_lock(inode);
1950         inode_dio_wait(inode);
1951         pagecache_block_get(&mapping->add_lock);
1952
1953         ret = __bch_truncate_page(inode->i_mapping,
1954                                   offset >> PAGE_SHIFT,
1955                                   offset, offset + len);
1956         if (unlikely(ret))
1957                 goto out;
1958
1959         if (offset >> PAGE_SHIFT !=
1960             (offset + len) >> PAGE_SHIFT) {
1961                 ret = __bch_truncate_page(inode->i_mapping,
1962                                           (offset + len) >> PAGE_SHIFT,
1963                                           offset, offset + len);
1964                 if (unlikely(ret))
1965                         goto out;
1966         }
1967
1968         truncate_pagecache_range(inode, offset, offset + len - 1);
1969
1970         if (discard_start < discard_end) {
1971                 struct disk_reservation disk_res;
1972                 struct i_sectors_hook i_sectors_hook;
1973                 int ret;
1974
1975                 BUG_ON(bch_disk_reservation_get(c, &disk_res, 0, 0));
1976
1977                 ret = i_sectors_dirty_get(ei, &i_sectors_hook);
1978                 if (unlikely(ret))
1979                         goto out;
1980
1981                 ret = bch_discard(c,
1982                                   POS(ino, discard_start),
1983                                   POS(ino, discard_end),
1984                                   0,
1985                                   &disk_res,
1986                                   &i_sectors_hook.hook,
1987                                   &ei->journal_seq);
1988
1989                 i_sectors_dirty_put(ei, &i_sectors_hook);
1990                 bch_disk_reservation_put(c, &disk_res);
1991         }
1992 out:
1993         pagecache_block_put(&mapping->add_lock);
1994         inode_unlock(inode);
1995
1996         return ret;
1997 }
1998
1999 static long bch_fcollapse(struct inode *inode, loff_t offset, loff_t len)
2000 {
2001         struct address_space *mapping = inode->i_mapping;
2002         struct bch_inode_info *ei = to_bch_ei(inode);
2003         struct cache_set *c = inode->i_sb->s_fs_info;
2004         struct btree_iter src;
2005         struct btree_iter dst;
2006         BKEY_PADDED(k) copy;
2007         struct bkey_s_c k;
2008         struct i_sectors_hook i_sectors_hook;
2009         loff_t new_size;
2010         int ret;
2011
2012         if ((offset | len) & (PAGE_SIZE - 1))
2013                 return -EINVAL;
2014
2015         bch_btree_iter_init_intent(&dst, c, BTREE_ID_EXTENTS,
2016                                    POS(inode->i_ino, offset >> 9));
2017         /* position will be set from dst iter's position: */
2018         bch_btree_iter_init(&src, c, BTREE_ID_EXTENTS, POS_MIN);
2019         bch_btree_iter_link(&src, &dst);
2020
2021         /*
2022          * We need i_mutex to keep the page cache consistent with the extents
2023          * btree, and the btree consistent with i_size - we don't need outside
2024          * locking for the extents btree itself, because we're using linked
2025          * iterators
2026          */
2027         inode_lock(inode);
2028         inode_dio_wait(inode);
2029         pagecache_block_get(&mapping->add_lock);
2030
2031         ret = -EINVAL;
2032         if (offset + len >= inode->i_size)
2033                 goto err;
2034
2035         if (inode->i_size < len)
2036                 goto err;
2037
2038         new_size = inode->i_size - len;
2039
2040         ret = write_invalidate_inode_pages_range(inode->i_mapping,
2041                                                  offset, LLONG_MAX);
2042         if (ret)
2043                 goto err;
2044
2045         ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2046         if (ret)
2047                 goto err;
2048
2049         while (bkey_cmp(dst.pos,
2050                         POS(inode->i_ino,
2051                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2052                 struct disk_reservation disk_res;
2053
2054                 bch_btree_iter_set_pos(&src,
2055                         POS(dst.pos.inode, dst.pos.offset + (len >> 9)));
2056
2057                 ret = bch_btree_iter_traverse(&dst);
2058                 if (ret)
2059                         goto btree_iter_err;
2060
2061                 k = bch_btree_iter_peek_with_holes(&src);
2062                 if ((ret = btree_iter_err(k)))
2063                         goto btree_iter_err;
2064
2065                 bkey_reassemble(&copy.k, k);
2066
2067                 if (bkey_deleted(&copy.k.k))
2068                         copy.k.k.type = KEY_TYPE_DISCARD;
2069
2070                 bch_cut_front(src.pos, &copy.k);
2071                 copy.k.k.p.offset -= len >> 9;
2072
2073                 BUG_ON(bkey_cmp(dst.pos, bkey_start_pos(&copy.k.k)));
2074
2075                 ret = bch_disk_reservation_get(c, &disk_res, copy.k.k.size,
2076                                                BCH_DISK_RESERVATION_NOFAIL);
2077                 BUG_ON(ret);
2078
2079                 ret = bch_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2080                                           &ei->journal_seq,
2081                                           BTREE_INSERT_ATOMIC|
2082                                           BTREE_INSERT_NOFAIL,
2083                                           BTREE_INSERT_ENTRY(&dst, &copy.k));
2084                 bch_disk_reservation_put(c, &disk_res);
2085 btree_iter_err:
2086                 if (ret < 0 && ret != -EINTR)
2087                         goto err_unwind;
2088
2089                 bch_btree_iter_cond_resched(&src);
2090         }
2091
2092         bch_btree_iter_unlock(&src);
2093         bch_btree_iter_unlock(&dst);
2094
2095         ret = bch_inode_truncate(c, inode->i_ino,
2096                                  round_up(new_size, PAGE_SIZE) >> 9,
2097                                  &i_sectors_hook.hook,
2098                                  &ei->journal_seq);
2099         if (ret)
2100                 goto err_unwind;
2101
2102         i_sectors_dirty_put(ei, &i_sectors_hook);
2103
2104         mutex_lock(&ei->update_lock);
2105         i_size_write(inode, new_size);
2106         ret = bch_write_inode_size(c, ei, inode->i_size);
2107         mutex_unlock(&ei->update_lock);
2108
2109         pagecache_block_put(&mapping->add_lock);
2110         inode_unlock(inode);
2111
2112         return ret;
2113 err_unwind:
2114         /*
2115          * XXX: we've left data with multiple pointers... which isn't a _super_
2116          * serious problem...
2117          */
2118         i_sectors_dirty_put(ei, &i_sectors_hook);
2119 err:
2120         bch_btree_iter_unlock(&src);
2121         bch_btree_iter_unlock(&dst);
2122         pagecache_block_put(&mapping->add_lock);
2123         inode_unlock(inode);
2124         return ret;
2125 }
2126
2127 static long bch_fallocate(struct inode *inode, int mode,
2128                           loff_t offset, loff_t len)
2129 {
2130         struct address_space *mapping = inode->i_mapping;
2131         struct bch_inode_info *ei = to_bch_ei(inode);
2132         struct cache_set *c = inode->i_sb->s_fs_info;
2133         struct i_sectors_hook i_sectors_hook;
2134         struct btree_iter iter;
2135         struct bkey_i reservation;
2136         struct bkey_s_c k;
2137         struct bpos end;
2138         loff_t block_start, block_end;
2139         loff_t new_size = offset + len;
2140         unsigned sectors;
2141         int ret;
2142
2143         bch_btree_iter_init_intent(&iter, c, BTREE_ID_EXTENTS, POS_MIN);
2144
2145         inode_lock(inode);
2146         inode_dio_wait(inode);
2147         pagecache_block_get(&mapping->add_lock);
2148
2149         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2150             new_size > inode->i_size) {
2151                 ret = inode_newsize_ok(inode, new_size);
2152                 if (ret)
2153                         goto err;
2154         }
2155
2156         if (mode & FALLOC_FL_ZERO_RANGE) {
2157                 ret = __bch_truncate_page(inode->i_mapping,
2158                                           offset >> PAGE_SHIFT,
2159                                           offset, offset + len);
2160
2161                 if (!ret &&
2162                     offset >> PAGE_SHIFT !=
2163                     (offset + len) >> PAGE_SHIFT)
2164                         ret = __bch_truncate_page(inode->i_mapping,
2165                                                   (offset + len) >> PAGE_SHIFT,
2166                                                   offset, offset + len);
2167
2168                 if (unlikely(ret))
2169                         goto err;
2170
2171                 truncate_pagecache_range(inode, offset, offset + len - 1);
2172
2173                 block_start     = round_up(offset, PAGE_SIZE);
2174                 block_end       = round_down(offset + len, PAGE_SIZE);
2175         } else {
2176                 block_start     = round_down(offset, PAGE_SIZE);
2177                 block_end       = round_up(offset + len, PAGE_SIZE);
2178         }
2179
2180         bch_btree_iter_set_pos(&iter, POS(inode->i_ino, block_start >> 9));
2181         end = POS(inode->i_ino, block_end >> 9);
2182
2183         ret = i_sectors_dirty_get(ei, &i_sectors_hook);
2184         if (unlikely(ret))
2185                 goto err;
2186
2187         while (bkey_cmp(iter.pos, end) < 0) {
2188                 struct disk_reservation disk_res = { 0 };
2189
2190                 k = bch_btree_iter_peek_with_holes(&iter);
2191                 if ((ret = btree_iter_err(k)))
2192                         goto btree_iter_err;
2193
2194                 /* already reserved */
2195                 if (k.k->type == BCH_RESERVATION) {
2196                         bch_btree_iter_advance_pos(&iter);
2197                         continue;
2198                 }
2199
2200                 if (bkey_extent_is_data(k.k)) {
2201                         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
2202                                 bch_btree_iter_advance_pos(&iter);
2203                                 continue;
2204                         }
2205                 }
2206
2207                 bkey_init(&reservation.k);
2208                 reservation.k.type      = BCH_RESERVATION;
2209                 reservation.k.p         = k.k->p;
2210                 reservation.k.size      = k.k->size;
2211
2212                 bch_cut_front(iter.pos, &reservation);
2213                 bch_cut_back(end, &reservation.k);
2214
2215                 sectors = reservation.k.size;
2216
2217                 if (!bkey_extent_is_allocation(k.k) ||
2218                     bkey_extent_is_compressed(c, k)) {
2219                         ret = bch_disk_reservation_get(c, &disk_res,
2220                                                        sectors, 0);
2221                         if (ret)
2222                                 goto err_put_sectors_dirty;
2223                 }
2224
2225                 ret = bch_btree_insert_at(c, &disk_res, &i_sectors_hook.hook,
2226                                           &ei->journal_seq,
2227                                           BTREE_INSERT_ATOMIC|
2228                                           BTREE_INSERT_NOFAIL,
2229                                           BTREE_INSERT_ENTRY(&iter, &reservation));
2230                 bch_disk_reservation_put(c, &disk_res);
2231 btree_iter_err:
2232                 if (ret < 0 && ret != -EINTR)
2233                         goto err_put_sectors_dirty;
2234
2235         }
2236         bch_btree_iter_unlock(&iter);
2237
2238         i_sectors_dirty_put(ei, &i_sectors_hook);
2239
2240         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2241             new_size > inode->i_size) {
2242                 i_size_write(inode, new_size);
2243
2244                 mutex_lock(&ei->update_lock);
2245                 ret = bch_write_inode_size(c, ei, inode->i_size);
2246                 mutex_unlock(&ei->update_lock);
2247         }
2248
2249         /* blech */
2250         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2251             (mode & FALLOC_FL_ZERO_RANGE) &&
2252             ei->i_size != inode->i_size) {
2253                 /* sync appends.. */
2254                 ret = filemap_write_and_wait_range(mapping, ei->i_size, S64_MAX);
2255                 if (ret)
2256                         goto err;
2257
2258                 if (ei->i_size != inode->i_size) {
2259                         mutex_lock(&ei->update_lock);
2260                         ret = bch_write_inode_size(c, ei, inode->i_size);
2261                         mutex_unlock(&ei->update_lock);
2262                 }
2263         }
2264
2265         pagecache_block_put(&mapping->add_lock);
2266         inode_unlock(inode);
2267
2268         return 0;
2269 err_put_sectors_dirty:
2270         i_sectors_dirty_put(ei, &i_sectors_hook);
2271 err:
2272         bch_btree_iter_unlock(&iter);
2273         pagecache_block_put(&mapping->add_lock);
2274         inode_unlock(inode);
2275         return ret;
2276 }
2277
2278 long bch_fallocate_dispatch(struct file *file, int mode,
2279                             loff_t offset, loff_t len)
2280 {
2281         struct inode *inode = file_inode(file);
2282
2283         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2284                 return bch_fallocate(inode, mode, offset, len);
2285
2286         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2287                 return bch_fpunch(inode, offset, len);
2288
2289         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2290                 return bch_fcollapse(inode, offset, len);
2291
2292         return -EOPNOTSUPP;
2293 }
2294
2295 static bool page_is_data(struct page *page)
2296 {
2297         /* XXX: should only have to check PageDirty */
2298         return PagePrivate(page) &&
2299                 (page_state(page)->sectors ||
2300                  page_state(page)->dirty_sectors);
2301 }
2302
2303 static loff_t bch_next_pagecache_data(struct inode *inode,
2304                                       loff_t start_offset,
2305                                       loff_t end_offset)
2306 {
2307         struct address_space *mapping = inode->i_mapping;
2308         struct page *page;
2309         pgoff_t index;
2310
2311         for (index = start_offset >> PAGE_SHIFT;
2312              index < end_offset >> PAGE_SHIFT;
2313              index++) {
2314                 if (find_get_pages(mapping, index, 1, &page)) {
2315                         lock_page(page);
2316                         index = page->index;
2317
2318                         if (page_is_data(page))
2319                                 end_offset =
2320                                         min(end_offset,
2321                                         max(start_offset,
2322                                             ((loff_t) index) << PAGE_SHIFT));
2323                         unlock_page(page);
2324                         put_page(page);
2325                 } else {
2326                         break;
2327                 }
2328         }
2329
2330         return end_offset;
2331 }
2332
2333 static loff_t bch_seek_data(struct file *file, u64 offset)
2334 {
2335         struct inode *inode = file->f_mapping->host;
2336         struct cache_set *c = inode->i_sb->s_fs_info;
2337         struct btree_iter iter;
2338         struct bkey_s_c k;
2339         u64 isize, next_data = MAX_LFS_FILESIZE;
2340         int ret;
2341
2342         isize = i_size_read(inode);
2343         if (offset >= isize)
2344                 return -ENXIO;
2345
2346         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2347                            POS(inode->i_ino, offset >> 9), k) {
2348                 if (k.k->p.inode != inode->i_ino) {
2349                         break;
2350                 } else if (bkey_extent_is_data(k.k)) {
2351                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2352                         break;
2353                 } else if (k.k->p.offset >> 9 > isize)
2354                         break;
2355         }
2356
2357         ret = bch_btree_iter_unlock(&iter);
2358         if (ret)
2359                 return ret;
2360
2361         if (next_data > offset)
2362                 next_data = bch_next_pagecache_data(inode, offset, next_data);
2363
2364         if (next_data > isize)
2365                 return -ENXIO;
2366
2367         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2368 }
2369
2370 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2371 {
2372         struct page *page;
2373         bool ret;
2374
2375         page = find_lock_entry(mapping, index);
2376         if (!page || radix_tree_exception(page))
2377                 return false;
2378
2379         ret = page_is_data(page);
2380         unlock_page(page);
2381
2382         return ret;
2383 }
2384
2385 static loff_t bch_next_pagecache_hole(struct inode *inode,
2386                                       loff_t start_offset,
2387                                       loff_t end_offset)
2388 {
2389         struct address_space *mapping = inode->i_mapping;
2390         pgoff_t index;
2391
2392         for (index = start_offset >> PAGE_SHIFT;
2393              index < end_offset >> PAGE_SHIFT;
2394              index++)
2395                 if (!page_slot_is_data(mapping, index))
2396                         end_offset = max(start_offset,
2397                                          ((loff_t) index) << PAGE_SHIFT);
2398
2399         return end_offset;
2400 }
2401
2402 static loff_t bch_seek_hole(struct file *file, u64 offset)
2403 {
2404         struct inode *inode = file->f_mapping->host;
2405         struct cache_set *c = inode->i_sb->s_fs_info;
2406         struct btree_iter iter;
2407         struct bkey_s_c k;
2408         u64 isize, next_hole = MAX_LFS_FILESIZE;
2409         int ret;
2410
2411         isize = i_size_read(inode);
2412         if (offset >= isize)
2413                 return -ENXIO;
2414
2415         for_each_btree_key_with_holes(&iter, c, BTREE_ID_EXTENTS,
2416                                       POS(inode->i_ino, offset >> 9), k) {
2417                 if (k.k->p.inode != inode->i_ino) {
2418                         next_hole = bch_next_pagecache_hole(inode,
2419                                         offset, MAX_LFS_FILESIZE);
2420                         break;
2421                 } else if (!bkey_extent_is_data(k.k)) {
2422                         next_hole = bch_next_pagecache_hole(inode,
2423                                         max(offset, bkey_start_offset(k.k) << 9),
2424                                         k.k->p.offset << 9);
2425
2426                         if (next_hole < k.k->p.offset << 9)
2427                                 break;
2428                 } else {
2429                         offset = max(offset, bkey_start_offset(k.k) << 9);
2430                 }
2431         }
2432
2433         ret = bch_btree_iter_unlock(&iter);
2434         if (ret)
2435                 return ret;
2436
2437         if (next_hole > isize)
2438                 next_hole = isize;
2439
2440         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2441 }
2442
2443 loff_t bch_llseek(struct file *file, loff_t offset, int whence)
2444 {
2445         switch (whence) {
2446         case SEEK_SET:
2447         case SEEK_CUR:
2448         case SEEK_END:
2449                 return generic_file_llseek(file, offset, whence);
2450         case SEEK_DATA:
2451                 return bch_seek_data(file, offset);
2452         case SEEK_HOLE:
2453                 return bch_seek_hole(file, offset);
2454         }
2455
2456         return -EINVAL;
2457 }