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

   1 #ifndef NO_BCACHEFS_FS
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_update.h"
   6 #include "buckets.h"
   7 #include "clock.h"
   8 #include "error.h"
   9 #include "extents.h"
  10 #include "fs.h"
  11 #include "fs-io.h"
  12 #include "fsck.h"
  13 #include "inode.h"
  14 #include "journal.h"
  15 #include "io.h"
  16 #include "keylist.h"
  17 #include "quota.h"
  18
  19 #include <linux/aio.h>
  20 #include <linux/backing-dev.h>
  21 #include <linux/falloc.h>
  22 #include <linux/migrate.h>
  23 #include <linux/mmu_context.h>
  24 #include <linux/pagevec.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/task_io_accounting_ops.h>
  27 #include <linux/uio.h>
  28 #include <linux/writeback.h>
  29
  30 #include <trace/events/bcachefs.h>
  31 #include <trace/events/writeback.h>
  32
  33 struct quota_res {
  34         u64                             sectors;
  35 };
  36
  37 struct bchfs_write_op {
  38         struct bch_inode_info           *inode;
  39         s64                             sectors_added;
  40         bool                            is_dio;
  41         bool                            unalloc;
  42         u64                             new_i_size;
  43
  44         /* must be last: */
  45         struct bch_write_op             op;
  46 };
  47
  48 struct bch_writepage_io {
  49         struct closure                  cl;
  50         u64                             new_sectors;
  51
  52         /* must be last: */
  53         struct bchfs_write_op           op;
  54 };
  55
  56 struct dio_write {
  57         struct closure                  cl;
  58         struct kiocb                    *req;
  59         struct mm_struct                *mm;
  60         unsigned                        loop:1,
  61                                         sync:1,
  62                                         free_iov:1;
  63         struct quota_res                quota_res;
  64
  65         struct iov_iter                 iter;
  66         struct iovec                    inline_vecs[2];
  67
  68         /* must be last: */
  69         struct bchfs_write_op           iop;
  70 };
  71
  72 struct dio_read {
  73         struct closure                  cl;
  74         struct kiocb                    *req;
  75         long                            ret;
  76         struct bch_read_bio             rbio;
  77 };
  78
  79 /* pagecache_block must be held */
  80 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  81                                               loff_t start, loff_t end)
  82 {
  83         int ret;
  84
  85         /*
  86          * XXX: the way this is currently implemented, we can spin if a process
  87          * is continually redirtying a specific page
  88          */
  89         do {
  90                 if (!mapping->nrpages &&
  91                     !mapping->nrexceptional)
  92                         return 0;
  93
  94                 ret = filemap_write_and_wait_range(mapping, start, end);
  95                 if (ret)
  96                         break;
  97
  98                 if (!mapping->nrpages)
  99                         return 0;
 100
 101                 ret = invalidate_inode_pages2_range(mapping,
 102                                 start >> PAGE_SHIFT,
 103                                 end >> PAGE_SHIFT);
 104         } while (ret == -EBUSY);
 105
 106         return ret;
 107 }
 108
 109 /* quotas */
 110
 111 #ifdef CONFIG_BCACHEFS_QUOTA
 112
 113 static void bch2_quota_reservation_put(struct bch_fs *c,
 114                                        struct bch_inode_info *inode,
 115                                        struct quota_res *res)
 116 {
 117         if (!res->sectors)
 118                 return;
 119
 120         mutex_lock(&inode->ei_quota_lock);
 121         BUG_ON(res->sectors > inode->ei_quota_reserved);
 122
 123         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 124                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 125         inode->ei_quota_reserved -= res->sectors;
 126         mutex_unlock(&inode->ei_quota_lock);
 127
 128         res->sectors = 0;
 129 }
 130
 131 static int bch2_quota_reservation_add(struct bch_fs *c,
 132                                       struct bch_inode_info *inode,
 133                                       struct quota_res *res,
 134                                       unsigned sectors,
 135                                       bool check_enospc)
 136 {
 137         int ret;
 138
 139         mutex_lock(&inode->ei_quota_lock);
 140         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 141                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 142         if (likely(!ret)) {
 143                 inode->ei_quota_reserved += sectors;
 144                 res->sectors += sectors;
 145         }
 146         mutex_unlock(&inode->ei_quota_lock);
 147
 148         return ret;
 149 }
 150
 151 #else
 152
 153 static void bch2_quota_reservation_put(struct bch_fs *c,
 154                                        struct bch_inode_info *inode,
 155                                        struct quota_res *res)
 156 {
 157 }
 158
 159 static int bch2_quota_reservation_add(struct bch_fs *c,
 160                                       struct bch_inode_info *inode,
 161                                       struct quota_res *res,
 162                                       unsigned sectors,
 163                                       bool check_enospc)
 164 {
 165         return 0;
 166 }
 167
 168 #endif
 169
 170 /* i_size updates: */
 171
 172 struct inode_new_size {
 173         loff_t          new_size;
 174         u64             now;
 175         unsigned        fields;
 176 };
 177
 178 static int inode_set_size(struct bch_inode_info *inode,
 179                           struct bch_inode_unpacked *bi,
 180                           void *p)
 181 {
 182         struct inode_new_size *s = p;
 183
 184         bi->bi_size = s->new_size;
 185         if (s->fields & ATTR_ATIME)
 186                 bi->bi_atime = s->now;
 187         if (s->fields & ATTR_MTIME)
 188                 bi->bi_mtime = s->now;
 189         if (s->fields & ATTR_CTIME)
 190                 bi->bi_ctime = s->now;
 191
 192         return 0;
 193 }
 194
 195 static int __must_check bch2_write_inode_size(struct bch_fs *c,
 196                                               struct bch_inode_info *inode,
 197                                               loff_t new_size, unsigned fields)
 198 {
 199         struct inode_new_size s = {
 200                 .new_size       = new_size,
 201                 .now            = bch2_current_time(c),
 202                 .fields         = fields,
 203         };
 204
 205         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 206 }
 207
 208 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 209                            struct quota_res *quota_res, s64 sectors)
 210 {
 211         if (!sectors)
 212                 return;
 213
 214         mutex_lock(&inode->ei_quota_lock);
 215 #ifdef CONFIG_BCACHEFS_QUOTA
 216         if (quota_res && sectors > 0) {
 217                 BUG_ON(sectors > quota_res->sectors);
 218                 BUG_ON(sectors > inode->ei_quota_reserved);
 219
 220                 quota_res->sectors -= sectors;
 221                 inode->ei_quota_reserved -= sectors;
 222         } else {
 223                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 224         }
 225 #endif
 226         inode->v.i_blocks += sectors;
 227         mutex_unlock(&inode->ei_quota_lock);
 228 }
 229
 230 /* normal i_size/i_sectors update machinery: */
 231
 232 static s64 sum_sector_overwrites(struct bkey_i *new, struct btree_iter *_iter,
 233                                  bool *allocating)
 234 {
 235         struct btree_iter iter;
 236         struct bkey_s_c old;
 237         s64 delta = 0;
 238
 239         bch2_btree_iter_init(&iter, _iter->c, BTREE_ID_EXTENTS, POS_MIN,
 240                              BTREE_ITER_SLOTS);
 241
 242         bch2_btree_iter_link(_iter, &iter);
 243         bch2_btree_iter_copy(&iter, _iter);
 244
 245         old = bch2_btree_iter_peek_slot(&iter);
 246
 247         while (1) {
 248                 /*
 249                  * should not be possible to get an error here, since we're
 250                  * carefully not advancing past @new and thus whatever leaf node
 251                  * @_iter currently points to:
 252                  */
 253                 BUG_ON(btree_iter_err(old));
 254
 255                 if (allocating &&
 256                     !bch2_extent_is_fully_allocated(old))
 257                         *allocating = true;
 258
 259                 delta += (min(new->k.p.offset,
 260                               old.k->p.offset) -
 261                           max(bkey_start_offset(&new->k),
 262                               bkey_start_offset(old.k))) *
 263                         (bkey_extent_is_allocation(&new->k) -
 264                          bkey_extent_is_allocation(old.k));
 265
 266                 if (bkey_cmp(old.k->p, new->k.p) >= 0)
 267                         break;
 268
 269                 old = bch2_btree_iter_next_slot(&iter);
 270         }
 271
 272         bch2_btree_iter_unlink(&iter);
 273
 274         return delta;
 275 }
 276
 277 static int bch2_extent_update(struct btree_trans *trans,
 278                               struct bch_inode_info *inode,
 279                               struct disk_reservation *disk_res,
 280                               struct quota_res *quota_res,
 281                               struct btree_iter *extent_iter,
 282                               struct bkey_i *k,
 283                               u64 new_i_size,
 284                               bool may_allocate,
 285                               bool direct,
 286                               s64 *total_delta)
 287 {
 288         struct btree_iter *inode_iter = NULL;
 289         struct bch_inode_unpacked inode_u;
 290         struct bkey_inode_buf inode_p;
 291         bool allocating = false;
 292         bool extended = false;
 293         s64 i_sectors_delta;
 294         int ret;
 295
 296         bch2_trans_begin_updates(trans);
 297
 298         ret = bch2_btree_iter_traverse(extent_iter);
 299         if (ret)
 300                 return ret;
 301
 302         bch2_extent_trim_atomic(k, extent_iter);
 303
 304         i_sectors_delta = sum_sector_overwrites(k, extent_iter, &allocating);
 305         if (!may_allocate && allocating)
 306                 return -ENOSPC;
 307
 308         bch2_trans_update(trans, BTREE_INSERT_ENTRY(extent_iter, k));
 309
 310         new_i_size = min(k->k.p.offset << 9, new_i_size);
 311
 312         /* XXX: inode->i_size locking */
 313         if (i_sectors_delta ||
 314             new_i_size > inode->ei_inode.bi_size) {
 315                 inode_iter = bch2_trans_get_iter(trans,
 316                         BTREE_ID_INODES,
 317                         POS(k->k.p.inode, 0),
 318                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 319                 if (IS_ERR(inode_iter))
 320                         return PTR_ERR(inode_iter);
 321
 322                 ret = bch2_btree_iter_traverse(inode_iter);
 323                 if (ret)
 324                         goto err;
 325
 326                 inode_u = inode->ei_inode;
 327                 inode_u.bi_sectors += i_sectors_delta;
 328
 329                 /* XXX: this is slightly suspect */
 330                 if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
 331                     new_i_size > inode_u.bi_size) {
 332                         inode_u.bi_size = new_i_size;
 333                         extended = true;
 334                 }
 335
 336                 bch2_inode_pack(&inode_p, &inode_u);
 337                 bch2_trans_update(trans,
 338                         BTREE_INSERT_ENTRY(inode_iter, &inode_p.inode.k_i));
 339         }
 340
 341         ret = bch2_trans_commit(trans, disk_res,
 342                                 &inode->ei_journal_seq,
 343                                 BTREE_INSERT_NOFAIL|
 344                                 BTREE_INSERT_ATOMIC|
 345                                 BTREE_INSERT_NOUNLOCK|
 346                                 BTREE_INSERT_USE_RESERVE);
 347         if (ret)
 348                 goto err;
 349
 350         inode->ei_inode.bi_sectors += i_sectors_delta;
 351
 352         EBUG_ON(i_sectors_delta &&
 353                 inode->ei_inode.bi_sectors != inode_u.bi_sectors);
 354
 355         if (extended) {
 356                 inode->ei_inode.bi_size = new_i_size;
 357
 358                 if (direct) {
 359                         spin_lock(&inode->v.i_lock);
 360                         if (new_i_size > inode->v.i_size)
 361                                 i_size_write(&inode->v, new_i_size);
 362                         spin_unlock(&inode->v.i_lock);
 363                 }
 364         }
 365
 366         if (direct)
 367                 i_sectors_acct(trans->c, inode, quota_res, i_sectors_delta);
 368
 369         if (total_delta)
 370                 *total_delta += i_sectors_delta;
 371 err:
 372         if (!IS_ERR_OR_NULL(inode_iter))
 373                 bch2_trans_iter_put(trans, inode_iter);
 374         return ret;
 375 }
 376
 377 static int bchfs_write_index_update(struct bch_write_op *wop)
 378 {
 379         struct bchfs_write_op *op = container_of(wop,
 380                                 struct bchfs_write_op, op);
 381         struct quota_res *quota_res = op->is_dio
 382                 ? &container_of(op, struct dio_write, iop)->quota_res
 383                 : NULL;
 384         struct bch_inode_info *inode = op->inode;
 385         struct keylist *keys = &op->op.insert_keys;
 386         struct bkey_i *k = bch2_keylist_front(keys);
 387         struct btree_trans trans;
 388         struct btree_iter *iter;
 389         int ret;
 390
 391         BUG_ON(k->k.p.inode != inode->v.i_ino);
 392
 393         bch2_trans_init(&trans, wop->c);
 394         bch2_trans_preload_iters(&trans);
 395
 396         iter = bch2_trans_get_iter(&trans,
 397                                 BTREE_ID_EXTENTS,
 398                                 bkey_start_pos(&k->k),
 399                                 BTREE_ITER_INTENT);
 400
 401         do {
 402                 BKEY_PADDED(k) tmp;
 403
 404                 bkey_copy(&tmp.k, bch2_keylist_front(keys));
 405
 406                 ret = bch2_extent_update(&trans, inode,
 407                                 &wop->res, quota_res,
 408                                 iter, &tmp.k,
 409                                 op->new_i_size,
 410                                 !op->unalloc,
 411                                 op->is_dio,
 412                                 &op->sectors_added);
 413                 if (ret == -EINTR)
 414                         continue;
 415                 if (ret)
 416                         break;
 417
 418                 if (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) < 0)
 419                         bch2_cut_front(iter->pos, bch2_keylist_front(keys));
 420                 else
 421                         bch2_keylist_pop_front(keys);
 422         } while (!bch2_keylist_empty(keys));
 423
 424         bch2_trans_exit(&trans);
 425
 426         return ret;
 427 }
 428
 429 static inline void bch2_fswrite_op_init(struct bchfs_write_op *op,
 430                                         struct bch_fs *c,
 431                                         struct bch_inode_info *inode,
 432                                         struct bch_io_opts opts,
 433                                         bool is_dio)
 434 {
 435         op->inode               = inode;
 436         op->sectors_added       = 0;
 437         op->is_dio              = is_dio;
 438         op->unalloc             = false;
 439         op->new_i_size          = U64_MAX;
 440
 441         bch2_write_op_init(&op->op, c, opts);
 442         op->op.target           = opts.foreground_target;
 443         op->op.index_update_fn  = bchfs_write_index_update;
 444         op_journal_seq_set(&op->op, &inode->ei_journal_seq);
 445 }
 446
 447 static inline struct bch_io_opts io_opts(struct bch_fs *c, struct bch_inode_info *inode)
 448 {
 449         struct bch_io_opts opts = bch2_opts_to_inode_opts(c->opts);
 450
 451         bch2_io_opts_apply(&opts, bch2_inode_opts_get(&inode->ei_inode));
 452         return opts;
 453 }
 454
 455 /* page state: */
 456
 457 /* stored in page->private: */
 458
 459 /*
 460  * bch_page_state has to (unfortunately) be manipulated with cmpxchg - we could
 461  * almost protected it with the page lock, except that bch2_writepage_io_done has
 462  * to update the sector counts (and from interrupt/bottom half context).
 463  */
 464 struct bch_page_state {
 465 union { struct {
 466         /* existing data: */
 467         unsigned                sectors:PAGE_SECTOR_SHIFT + 1;
 468
 469         /* Uncompressed, fully allocated replicas: */
 470         unsigned                nr_replicas:4;
 471
 472         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 473         unsigned                replicas_reserved:4;
 474
 475         /* Owns PAGE_SECTORS sized quota reservation: */
 476         unsigned                quota_reserved:1;
 477
 478         /*
 479          * Number of sectors on disk - for i_blocks
 480          * Uncompressed size, not compressed size:
 481          */
 482         unsigned                dirty_sectors:PAGE_SECTOR_SHIFT + 1;
 483 };
 484         /* for cmpxchg: */
 485         unsigned long           v;
 486 };
 487 };
 488
 489 #define page_state_cmpxchg(_ptr, _new, _expr)                           \
 490 ({                                                                      \
 491         unsigned long _v = READ_ONCE((_ptr)->v);                        \
 492         struct bch_page_state _old;                                     \
 493                                                                         \
 494         do {                                                            \
 495                 _old.v = _new.v = _v;                                   \
 496                 _expr;                                                  \
 497                                                                         \
 498                 EBUG_ON(_new.sectors + _new.dirty_sectors > PAGE_SECTORS);\
 499         } while (_old.v != _new.v &&                                    \
 500                  (_v = cmpxchg(&(_ptr)->v, _old.v, _new.v)) != _old.v); \
 501                                                                         \
 502         _old;                                                           \
 503 })
 504
 505 static inline struct bch_page_state *page_state(struct page *page)
 506 {
 507         struct bch_page_state *s = (void *) &page->private;
 508
 509         BUILD_BUG_ON(sizeof(*s) > sizeof(page->private));
 510
 511         if (!PagePrivate(page))
 512                 SetPagePrivate(page);
 513
 514         return s;
 515 }
 516
 517 static inline unsigned page_res_sectors(struct bch_page_state s)
 518 {
 519
 520         return s.replicas_reserved * PAGE_SECTORS;
 521 }
 522
 523 static void __bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 524                                         struct bch_page_state s)
 525 {
 526         struct disk_reservation res = { .sectors = page_res_sectors(s) };
 527         struct quota_res quota_res = { .sectors = s.quota_reserved ? PAGE_SECTORS : 0 };
 528
 529         bch2_quota_reservation_put(c, inode, &quota_res);
 530         bch2_disk_reservation_put(c, &res);
 531 }
 532
 533 static void bch2_put_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 534                                       struct page *page)
 535 {
 536         struct bch_page_state s;
 537
 538         EBUG_ON(!PageLocked(page));
 539
 540         s = page_state_cmpxchg(page_state(page), s, {
 541                 s.replicas_reserved     = 0;
 542                 s.quota_reserved        = 0;
 543         });
 544
 545         __bch2_put_page_reservation(c, inode, s);
 546 }
 547
 548 static int bch2_get_page_reservation(struct bch_fs *c, struct bch_inode_info *inode,
 549                                      struct page *page, bool check_enospc)
 550 {
 551         struct bch_page_state *s = page_state(page), new;
 552
 553         /* XXX: this should not be open coded */
 554         unsigned nr_replicas = inode->ei_inode.bi_data_replicas
 555                 ? inode->ei_inode.bi_data_replicas - 1
 556                 : c->opts.data_replicas;
 557         struct disk_reservation disk_res;
 558         struct quota_res quota_res = { 0 };
 559         int ret;
 560
 561         EBUG_ON(!PageLocked(page));
 562
 563         if (s->replicas_reserved < nr_replicas) {
 564                 ret = bch2_disk_reservation_get(c, &disk_res, PAGE_SECTORS,
 565                                 nr_replicas - s->replicas_reserved,
 566                                 !check_enospc ? BCH_DISK_RESERVATION_NOFAIL : 0);
 567                 if (unlikely(ret))
 568                         return ret;
 569
 570                 page_state_cmpxchg(s, new, ({
 571                         BUG_ON(new.replicas_reserved +
 572                                disk_res.nr_replicas != nr_replicas);
 573                         new.replicas_reserved += disk_res.nr_replicas;
 574                 }));
 575         }
 576
 577         if (!s->quota_reserved &&
 578             s->sectors + s->dirty_sectors < PAGE_SECTORS) {
 579                 ret = bch2_quota_reservation_add(c, inode, &quota_res,
 580                                                  PAGE_SECTORS,
 581                                                  check_enospc);
 582                 if (unlikely(ret))
 583                         return ret;
 584
 585                 page_state_cmpxchg(s, new, ({
 586                         BUG_ON(new.quota_reserved);
 587                         new.quota_reserved = 1;
 588                 }));
 589         }
 590
 591         return ret;
 592 }
 593
 594 static void bch2_clear_page_bits(struct page *page)
 595 {
 596         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 597         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 598         struct bch_page_state s;
 599
 600         EBUG_ON(!PageLocked(page));
 601
 602         if (!PagePrivate(page))
 603                 return;
 604
 605         s.v = xchg(&page_state(page)->v, 0);
 606         ClearPagePrivate(page);
 607
 608         if (s.dirty_sectors)
 609                 i_sectors_acct(c, inode, NULL, -s.dirty_sectors);
 610
 611         __bch2_put_page_reservation(c, inode, s);
 612 }
 613
 614 int bch2_set_page_dirty(struct page *page)
 615 {
 616         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 617         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 618         struct quota_res quota_res = { 0 };
 619         struct bch_page_state old, new;
 620
 621         old = page_state_cmpxchg(page_state(page), new,
 622                 new.dirty_sectors = PAGE_SECTORS - new.sectors;
 623                 new.quota_reserved = 0;
 624         );
 625
 626         quota_res.sectors += old.quota_reserved * PAGE_SECTORS;
 627
 628         if (old.dirty_sectors != new.dirty_sectors)
 629                 i_sectors_acct(c, inode, &quota_res,
 630                                new.dirty_sectors - old.dirty_sectors);
 631         bch2_quota_reservation_put(c, inode, &quota_res);
 632
 633         return __set_page_dirty_nobuffers(page);
 634 }
 635
 636 int bch2_page_mkwrite(struct vm_fault *vmf)
 637 {
 638         struct page *page = vmf->page;
 639         struct file *file = vmf->vma->vm_file;
 640         struct bch_inode_info *inode = file_bch_inode(file);
 641         struct address_space *mapping = inode->v.i_mapping;
 642         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 643         int ret = VM_FAULT_LOCKED;
 644
 645         sb_start_pagefault(inode->v.i_sb);
 646         file_update_time(file);
 647
 648         /*
 649          * Not strictly necessary, but helps avoid dio writes livelocking in
 650          * write_invalidate_inode_pages_range() - can drop this if/when we get
 651          * a write_invalidate_inode_pages_range() that works without dropping
 652          * page lock before invalidating page
 653          */
 654         if (current->pagecache_lock != &mapping->add_lock)
 655                 pagecache_add_get(&mapping->add_lock);
 656
 657         lock_page(page);
 658         if (page->mapping != mapping ||
 659             page_offset(page) > i_size_read(&inode->v)) {
 660                 unlock_page(page);
 661                 ret = VM_FAULT_NOPAGE;
 662                 goto out;
 663         }
 664
 665         if (bch2_get_page_reservation(c, inode, page, true)) {
 666                 unlock_page(page);
 667                 ret = VM_FAULT_SIGBUS;
 668                 goto out;
 669         }
 670
 671         if (!PageDirty(page))
 672                 set_page_dirty(page);
 673         wait_for_stable_page(page);
 674 out:
 675         if (current->pagecache_lock != &mapping->add_lock)
 676                 pagecache_add_put(&mapping->add_lock);
 677         sb_end_pagefault(inode->v.i_sb);
 678         return ret;
 679 }
 680
 681 void bch2_invalidatepage(struct page *page, unsigned int offset,
 682                          unsigned int length)
 683 {
 684         EBUG_ON(!PageLocked(page));
 685         EBUG_ON(PageWriteback(page));
 686
 687         if (offset || length < PAGE_SIZE)
 688                 return;
 689
 690         bch2_clear_page_bits(page);
 691 }
 692
 693 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 694 {
 695         /* XXX: this can't take locks that are held while we allocate memory */
 696         EBUG_ON(!PageLocked(page));
 697         EBUG_ON(PageWriteback(page));
 698
 699         if (PageDirty(page))
 700                 return 0;
 701
 702         bch2_clear_page_bits(page);
 703         return 1;
 704 }
 705
 706 #ifdef CONFIG_MIGRATION
 707 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 708                       struct page *page, enum migrate_mode mode)
 709 {
 710         int ret;
 711
 712         EBUG_ON(!PageLocked(page));
 713         EBUG_ON(!PageLocked(newpage));
 714
 715         ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
 716         if (ret != MIGRATEPAGE_SUCCESS)
 717                 return ret;
 718
 719         if (PagePrivate(page)) {
 720                 *page_state(newpage) = *page_state(page);
 721                 ClearPagePrivate(page);
 722         }
 723
 724         migrate_page_copy(newpage, page);
 725         return MIGRATEPAGE_SUCCESS;
 726 }
 727 #endif
 728
 729 /* readpages/writepages: */
 730
 731 static bool bio_can_add_page_contig(struct bio *bio, struct page *page)
 732 {
 733         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 734
 735         return bio->bi_vcnt < bio->bi_max_vecs &&
 736                 bio_end_sector(bio) == offset;
 737 }
 738
 739 static int bio_add_page_contig(struct bio *bio, struct page *page)
 740 {
 741         sector_t offset = (sector_t) page->index << PAGE_SECTOR_SHIFT;
 742
 743         EBUG_ON(!bio->bi_max_vecs);
 744
 745         if (!bio->bi_vcnt)
 746                 bio->bi_iter.bi_sector = offset;
 747         else if (!bio_can_add_page_contig(bio, page))
 748                 return -1;
 749
 750         __bio_add_page(bio, page, PAGE_SIZE, 0);
 751         return 0;
 752 }
 753
 754 /* readpage(s): */
 755
 756 static void bch2_readpages_end_io(struct bio *bio)
 757 {
 758         struct bio_vec *bv;
 759         int i;
 760
 761         bio_for_each_segment_all(bv, bio, i) {
 762                 struct page *page = bv->bv_page;
 763
 764                 if (!bio->bi_status) {
 765                         SetPageUptodate(page);
 766                 } else {
 767                         ClearPageUptodate(page);
 768                         SetPageError(page);
 769                 }
 770                 unlock_page(page);
 771         }
 772
 773         bio_put(bio);
 774 }
 775
 776 static inline void page_state_init_for_read(struct page *page)
 777 {
 778         SetPagePrivate(page);
 779         page->private = 0;
 780 }
 781
 782 struct readpages_iter {
 783         struct address_space    *mapping;
 784         struct page             **pages;
 785         unsigned                nr_pages;
 786         unsigned                nr_added;
 787         unsigned                idx;
 788         pgoff_t                 offset;
 789 };
 790
 791 static int readpages_iter_init(struct readpages_iter *iter,
 792                                struct address_space *mapping,
 793                                struct list_head *pages, unsigned nr_pages)
 794 {
 795         memset(iter, 0, sizeof(*iter));
 796
 797         iter->mapping   = mapping;
 798         iter->offset    = list_last_entry(pages, struct page, lru)->index;
 799
 800         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 801         if (!iter->pages)
 802                 return -ENOMEM;
 803
 804         while (!list_empty(pages)) {
 805                 struct page *page = list_last_entry(pages, struct page, lru);
 806
 807                 prefetchw(&page->flags);
 808                 iter->pages[iter->nr_pages++] = page;
 809                 list_del(&page->lru);
 810         }
 811
 812         return 0;
 813 }
 814
 815 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 816 {
 817         struct page *page;
 818         unsigned i;
 819         int ret;
 820
 821         BUG_ON(iter->idx > iter->nr_added);
 822         BUG_ON(iter->nr_added > iter->nr_pages);
 823
 824         if (iter->idx < iter->nr_added)
 825                 goto out;
 826
 827         while (1) {
 828                 if (iter->idx == iter->nr_pages)
 829                         return NULL;
 830
 831                 ret = add_to_page_cache_lru_vec(iter->mapping,
 832                                 iter->pages     + iter->nr_added,
 833                                 iter->nr_pages  - iter->nr_added,
 834                                 iter->offset    + iter->nr_added,
 835                                 GFP_NOFS);
 836                 if (ret > 0)
 837                         break;
 838
 839                 page = iter->pages[iter->nr_added];
 840                 iter->idx++;
 841                 iter->nr_added++;
 842
 843                 put_page(page);
 844         }
 845
 846         iter->nr_added += ret;
 847
 848         for (i = iter->idx; i < iter->nr_added; i++)
 849                 put_page(iter->pages[i]);
 850 out:
 851         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 852
 853         page_state_init_for_read(iter->pages[iter->idx]);
 854         return iter->pages[iter->idx];
 855 }
 856
 857 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 858 {
 859         struct bvec_iter iter;
 860         struct bio_vec bv;
 861         unsigned nr_ptrs = !bch2_extent_is_compressed(k)
 862                 ? bch2_bkey_nr_dirty_ptrs(k)
 863                 : 0;
 864
 865         bio_for_each_segment(bv, bio, iter) {
 866                 /* brand new pages, don't need to be locked: */
 867
 868                 struct bch_page_state *s = page_state(bv.bv_page);
 869
 870                 /* sectors in @k from the start of this page: */
 871                 unsigned k_sectors = k.k->size - (iter.bi_sector - k.k->p.offset);
 872
 873                 unsigned page_sectors = min(bv.bv_len >> 9, k_sectors);
 874
 875                 s->nr_replicas = page_sectors == PAGE_SECTORS
 876                         ? nr_ptrs : 0;
 877
 878                 BUG_ON(s->sectors + page_sectors > PAGE_SECTORS);
 879                 s->sectors += page_sectors;
 880         }
 881 }
 882
 883 static void readpage_bio_extend(struct readpages_iter *iter,
 884                                 struct bio *bio, u64 offset,
 885                                 bool get_more)
 886 {
 887         while (bio_end_sector(bio) < offset &&
 888                bio->bi_vcnt < bio->bi_max_vecs) {
 889                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 890                 struct page *page = readpage_iter_next(iter);
 891                 int ret;
 892
 893                 if (page) {
 894                         if (iter->offset + iter->idx != page_offset)
 895                                 break;
 896
 897                         iter->idx++;
 898                 } else {
 899                         if (!get_more)
 900                                 break;
 901
 902                         page = xa_load(&iter->mapping->i_pages, page_offset);
 903                         if (page && !xa_is_value(page))
 904                                 break;
 905
 906                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 907                         if (!page)
 908                                 break;
 909
 910                         page_state_init_for_read(page);
 911
 912                         ret = add_to_page_cache_lru(page, iter->mapping,
 913                                                     page_offset, GFP_NOFS);
 914                         if (ret) {
 915                                 ClearPagePrivate(page);
 916                                 put_page(page);
 917                                 break;
 918                         }
 919
 920                         put_page(page);
 921                 }
 922
 923                 __bio_add_page(bio, page, PAGE_SIZE, 0);
 924         }
 925 }
 926
 927 static void bchfs_read(struct bch_fs *c, struct btree_iter *iter,
 928                        struct bch_read_bio *rbio, u64 inum,
 929                        struct readpages_iter *readpages_iter)
 930 {
 931         struct bio *bio = &rbio->bio;
 932         int flags = BCH_READ_RETRY_IF_STALE|
 933                 BCH_READ_MAY_PROMOTE;
 934
 935         rbio->c = c;
 936         rbio->start_time = local_clock();
 937
 938         while (1) {
 939                 BKEY_PADDED(k) tmp;
 940                 struct bkey_s_c k;
 941                 unsigned bytes;
 942
 943                 bch2_btree_iter_set_pos(iter, POS(inum, bio->bi_iter.bi_sector));
 944
 945                 k = bch2_btree_iter_peek_slot(iter);
 946                 BUG_ON(!k.k);
 947
 948                 if (IS_ERR(k.k)) {
 949                         int ret = bch2_btree_iter_unlock(iter);
 950                         BUG_ON(!ret);
 951                         bcache_io_error(c, bio, "btree IO error %i", ret);
 952                         bio_endio(bio);
 953                         return;
 954                 }
 955
 956                 bkey_reassemble(&tmp.k, k);
 957                 bch2_btree_iter_unlock(iter);
 958                 k = bkey_i_to_s_c(&tmp.k);
 959
 960                 if (readpages_iter) {
 961                         bool want_full_extent = false;
 962
 963                         if (bkey_extent_is_data(k.k)) {
 964                                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 965                                 const union bch_extent_entry *i;
 966                                 struct extent_ptr_decoded p;
 967
 968                                 extent_for_each_ptr_decode(e, p, i)
 969                                         want_full_extent |= ((p.crc.csum_type != 0) |
 970                                                              (p.crc.compression_type != 0));
 971                         }
 972
 973                         readpage_bio_extend(readpages_iter,
 974                                             bio, k.k->p.offset,
 975                                             want_full_extent);
 976                 }
 977
 978                 bytes = (min_t(u64, k.k->p.offset, bio_end_sector(bio)) -
 979                          bio->bi_iter.bi_sector) << 9;
 980                 swap(bio->bi_iter.bi_size, bytes);
 981
 982                 if (bytes == bio->bi_iter.bi_size)
 983                         flags |= BCH_READ_LAST_FRAGMENT;
 984
 985                 if (bkey_extent_is_allocation(k.k))
 986                         bch2_add_page_sectors(bio, k);
 987
 988                 bch2_read_extent(c, rbio, k, flags);
 989
 990                 if (flags & BCH_READ_LAST_FRAGMENT)
 991                         return;
 992
 993                 swap(bio->bi_iter.bi_size, bytes);
 994                 bio_advance(bio, bytes);
 995         }
 996 }
 997
 998 int bch2_readpages(struct file *file, struct address_space *mapping,
 999                    struct list_head *pages, unsigned nr_pages)
1000 {
1001         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1002         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1003         struct bch_io_opts opts = io_opts(c, inode);
1004         struct btree_iter iter;
1005         struct page *page;
1006         struct readpages_iter readpages_iter;
1007         int ret;
1008
1009         ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
1010         BUG_ON(ret);
1011
1012         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1013                              BTREE_ITER_SLOTS);
1014
1015         if (current->pagecache_lock != &mapping->add_lock)
1016                 pagecache_add_get(&mapping->add_lock);
1017
1018         while ((page = readpage_iter_next(&readpages_iter))) {
1019                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1020                 unsigned n = min_t(unsigned,
1021                                    readpages_iter.nr_pages -
1022                                    readpages_iter.idx,
1023                                    BIO_MAX_PAGES);
1024                 struct bch_read_bio *rbio =
1025                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
1026                                   opts);
1027
1028                 readpages_iter.idx++;
1029
1030                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
1031                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
1032                 rbio->bio.bi_end_io = bch2_readpages_end_io;
1033                 __bio_add_page(&rbio->bio, page, PAGE_SIZE, 0);
1034
1035                 bchfs_read(c, &iter, rbio, inode->v.i_ino, &readpages_iter);
1036         }
1037
1038         if (current->pagecache_lock != &mapping->add_lock)
1039                 pagecache_add_put(&mapping->add_lock);
1040
1041         kfree(readpages_iter.pages);
1042
1043         return 0;
1044 }
1045
1046 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1047                              u64 inum, struct page *page)
1048 {
1049         struct btree_iter iter;
1050
1051         page_state_init_for_read(page);
1052
1053         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1054         bio_add_page_contig(&rbio->bio, page);
1055
1056         bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, POS_MIN,
1057                              BTREE_ITER_SLOTS);
1058         bchfs_read(c, &iter, rbio, inum, NULL);
1059 }
1060
1061 int bch2_readpage(struct file *file, struct page *page)
1062 {
1063         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1064         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1065         struct bch_io_opts opts = io_opts(c, inode);
1066         struct bch_read_bio *rbio;
1067
1068         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
1069         rbio->bio.bi_end_io = bch2_readpages_end_io;
1070
1071         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1072         return 0;
1073 }
1074
1075 static void bch2_read_single_page_end_io(struct bio *bio)
1076 {
1077         complete(bio->bi_private);
1078 }
1079
1080 static int bch2_read_single_page(struct page *page,
1081                                  struct address_space *mapping)
1082 {
1083         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1084         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1085         struct bch_read_bio *rbio;
1086         int ret;
1087         DECLARE_COMPLETION_ONSTACK(done);
1088
1089         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1090                          io_opts(c, inode));
1091         rbio->bio.bi_private = &done;
1092         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1093
1094         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
1095         wait_for_completion(&done);
1096
1097         ret = blk_status_to_errno(rbio->bio.bi_status);
1098         bio_put(&rbio->bio);
1099
1100         if (ret < 0)
1101                 return ret;
1102
1103         SetPageUptodate(page);
1104         return 0;
1105 }
1106
1107 /* writepages: */
1108
1109 struct bch_writepage_state {
1110         struct bch_writepage_io *io;
1111         struct bch_io_opts      opts;
1112 };
1113
1114 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1115                                                                   struct bch_inode_info *inode)
1116 {
1117         return (struct bch_writepage_state) { .opts = io_opts(c, inode) };
1118 }
1119
1120 static void bch2_writepage_io_free(struct closure *cl)
1121 {
1122         struct bch_writepage_io *io = container_of(cl,
1123                                         struct bch_writepage_io, cl);
1124
1125         bio_put(&io->op.op.wbio.bio);
1126 }
1127
1128 static void bch2_writepage_io_done(struct closure *cl)
1129 {
1130         struct bch_writepage_io *io = container_of(cl,
1131                                         struct bch_writepage_io, cl);
1132         struct bch_fs *c = io->op.op.c;
1133         struct bio *bio = &io->op.op.wbio.bio;
1134         struct bio_vec *bvec;
1135         unsigned i;
1136
1137         if (io->op.op.error) {
1138                 bio_for_each_segment_all(bvec, bio, i)
1139                         SetPageError(bvec->bv_page);
1140                 set_bit(AS_EIO, &io->op.inode->v.i_mapping->flags);
1141         }
1142
1143         /*
1144          * racing with fallocate can cause us to add fewer sectors than
1145          * expected - but we shouldn't add more sectors than expected:
1146          */
1147         BUG_ON(io->op.sectors_added > (s64) io->new_sectors);
1148
1149         /*
1150          * (error (due to going RO) halfway through a page can screw that up
1151          * slightly)
1152          * XXX wtf?
1153            BUG_ON(io->op.sectors_added - io->new_sectors >= (s64) PAGE_SECTORS);
1154          */
1155
1156         /*
1157          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1158          * before calling end_page_writeback:
1159          */
1160         if (io->op.sectors_added != io->new_sectors)
1161                 i_sectors_acct(c, io->op.inode, NULL,
1162                                io->op.sectors_added - (s64) io->new_sectors);
1163
1164         bio_for_each_segment_all(bvec, bio, i)
1165                 end_page_writeback(bvec->bv_page);
1166
1167         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1168 }
1169
1170 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1171 {
1172         struct bch_writepage_io *io = w->io;
1173
1174         w->io = NULL;
1175         closure_call(&io->op.op.cl, bch2_write, NULL, &io->cl);
1176         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1177 }
1178
1179 /*
1180  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1181  * possible, else allocating a new one:
1182  */
1183 static void bch2_writepage_io_alloc(struct bch_fs *c,
1184                                     struct bch_writepage_state *w,
1185                                     struct bch_inode_info *inode,
1186                                     struct page *page,
1187                                     unsigned nr_replicas)
1188 {
1189         struct bch_write_op *op;
1190         u64 offset = (u64) page->index << PAGE_SECTOR_SHIFT;
1191
1192         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1193                                               BIO_MAX_PAGES,
1194                                               &c->writepage_bioset),
1195                              struct bch_writepage_io, op.op.wbio.bio);
1196
1197         closure_init(&w->io->cl, NULL);
1198         w->io->new_sectors      = 0;
1199         bch2_fswrite_op_init(&w->io->op, c, inode, w->opts, false);
1200         op                      = &w->io->op.op;
1201         op->nr_replicas         = nr_replicas;
1202         op->res.nr_replicas     = nr_replicas;
1203         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1204         op->pos                 = POS(inode->v.i_ino, offset);
1205         op->wbio.bio.bi_iter.bi_sector = offset;
1206 }
1207
1208 static int __bch2_writepage(struct page *page,
1209                             struct writeback_control *wbc,
1210                             void *data)
1211 {
1212         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1213         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1214         struct bch_writepage_state *w = data;
1215         struct bch_page_state new, old;
1216         unsigned offset, nr_replicas_this_write;
1217         loff_t i_size = i_size_read(&inode->v);
1218         pgoff_t end_index = i_size >> PAGE_SHIFT;
1219
1220         EBUG_ON(!PageUptodate(page));
1221
1222         /* Is the page fully inside i_size? */
1223         if (page->index < end_index)
1224                 goto do_io;
1225
1226         /* Is the page fully outside i_size? (truncate in progress) */
1227         offset = i_size & (PAGE_SIZE - 1);
1228         if (page->index > end_index || !offset) {
1229                 unlock_page(page);
1230                 return 0;
1231         }
1232
1233         /*
1234          * The page straddles i_size.  It must be zeroed out on each and every
1235          * writepage invocation because it may be mmapped.  "A file is mapped
1236          * in multiples of the page size.  For a file that is not a multiple of
1237          * the  page size, the remaining memory is zeroed when mapped, and
1238          * writes to that region are not written out to the file."
1239          */
1240         zero_user_segment(page, offset, PAGE_SIZE);
1241 do_io:
1242         EBUG_ON(!PageLocked(page));
1243
1244         /* Before unlocking the page, transfer reservation to w->io: */
1245         old = page_state_cmpxchg(page_state(page), new, {
1246                 /*
1247                  * If we didn't get a reservation, we can only write out the
1248                  * number of (fully allocated) replicas that currently exist,
1249                  * and only if the entire page has been written:
1250                  */
1251                 nr_replicas_this_write =
1252                         max_t(unsigned,
1253                               new.replicas_reserved,
1254                               (new.sectors == PAGE_SECTORS
1255                                ? new.nr_replicas : 0));
1256
1257                 BUG_ON(!nr_replicas_this_write);
1258
1259                 new.nr_replicas = w->opts.compression
1260                         ? 0
1261                         : nr_replicas_this_write;
1262
1263                 new.replicas_reserved = 0;
1264
1265                 new.sectors += new.dirty_sectors;
1266                 BUG_ON(new.sectors != PAGE_SECTORS);
1267                 new.dirty_sectors = 0;
1268         });
1269
1270         BUG_ON(PageWriteback(page));
1271         set_page_writeback(page);
1272         unlock_page(page);
1273
1274         if (w->io &&
1275             (w->io->op.op.res.nr_replicas != nr_replicas_this_write ||
1276              !bio_can_add_page_contig(&w->io->op.op.wbio.bio, page)))
1277                 bch2_writepage_do_io(w);
1278
1279         if (!w->io)
1280                 bch2_writepage_io_alloc(c, w, inode, page,
1281                                         nr_replicas_this_write);
1282
1283         w->io->new_sectors += new.sectors - old.sectors;
1284
1285         BUG_ON(inode != w->io->op.inode);
1286         BUG_ON(bio_add_page_contig(&w->io->op.op.wbio.bio, page));
1287
1288         w->io->op.op.res.sectors += old.replicas_reserved * PAGE_SECTORS;
1289         w->io->op.new_i_size = i_size;
1290
1291         if (wbc->sync_mode == WB_SYNC_ALL)
1292                 w->io->op.op.wbio.bio.bi_opf |= REQ_SYNC;
1293
1294         return 0;
1295 }
1296
1297 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1298 {
1299         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1300         struct bch_writepage_state w =
1301                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1302         struct blk_plug plug;
1303         int ret;
1304
1305         blk_start_plug(&plug);
1306         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1307         if (w.io)
1308                 bch2_writepage_do_io(&w);
1309         blk_finish_plug(&plug);
1310         return ret;
1311 }
1312
1313 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1314 {
1315         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1316         struct bch_writepage_state w =
1317                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1318         int ret;
1319
1320         ret = __bch2_writepage(page, wbc, &w);
1321         if (w.io)
1322                 bch2_writepage_do_io(&w);
1323
1324         return ret;
1325 }
1326
1327 /* buffered writes: */
1328
1329 int bch2_write_begin(struct file *file, struct address_space *mapping,
1330                      loff_t pos, unsigned len, unsigned flags,
1331                      struct page **pagep, void **fsdata)
1332 {
1333         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1334         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1335         pgoff_t index = pos >> PAGE_SHIFT;
1336         unsigned offset = pos & (PAGE_SIZE - 1);
1337         struct page *page;
1338         int ret = -ENOMEM;
1339
1340         BUG_ON(inode_unhashed(&inode->v));
1341
1342         /* Not strictly necessary - same reason as mkwrite(): */
1343         pagecache_add_get(&mapping->add_lock);
1344
1345         page = grab_cache_page_write_begin(mapping, index, flags);
1346         if (!page)
1347                 goto err_unlock;
1348
1349         if (PageUptodate(page))
1350                 goto out;
1351
1352         /* If we're writing entire page, don't need to read it in first: */
1353         if (len == PAGE_SIZE)
1354                 goto out;
1355
1356         if (!offset && pos + len >= inode->v.i_size) {
1357                 zero_user_segment(page, len, PAGE_SIZE);
1358                 flush_dcache_page(page);
1359                 goto out;
1360         }
1361
1362         if (index > inode->v.i_size >> PAGE_SHIFT) {
1363                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1364                 flush_dcache_page(page);
1365                 goto out;
1366         }
1367 readpage:
1368         ret = bch2_read_single_page(page, mapping);
1369         if (ret)
1370                 goto err;
1371 out:
1372         ret = bch2_get_page_reservation(c, inode, page, true);
1373         if (ret) {
1374                 if (!PageUptodate(page)) {
1375                         /*
1376                          * If the page hasn't been read in, we won't know if we
1377                          * actually need a reservation - we don't actually need
1378                          * to read here, we just need to check if the page is
1379                          * fully backed by uncompressed data:
1380                          */
1381                         goto readpage;
1382                 }
1383
1384                 goto err;
1385         }
1386
1387         *pagep = page;
1388         return 0;
1389 err:
1390         unlock_page(page);
1391         put_page(page);
1392         *pagep = NULL;
1393 err_unlock:
1394         pagecache_add_put(&mapping->add_lock);
1395         return ret;
1396 }
1397
1398 int bch2_write_end(struct file *file, struct address_space *mapping,
1399                    loff_t pos, unsigned len, unsigned copied,
1400                    struct page *page, void *fsdata)
1401 {
1402         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1403         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1404
1405         lockdep_assert_held(&inode->v.i_rwsem);
1406
1407         if (unlikely(copied < len && !PageUptodate(page))) {
1408                 /*
1409                  * The page needs to be read in, but that would destroy
1410                  * our partial write - simplest thing is to just force
1411                  * userspace to redo the write:
1412                  */
1413                 zero_user(page, 0, PAGE_SIZE);
1414                 flush_dcache_page(page);
1415                 copied = 0;
1416         }
1417
1418         spin_lock(&inode->v.i_lock);
1419         if (pos + copied > inode->v.i_size)
1420                 i_size_write(&inode->v, pos + copied);
1421         spin_unlock(&inode->v.i_lock);
1422
1423         if (copied) {
1424                 if (!PageUptodate(page))
1425                         SetPageUptodate(page);
1426                 if (!PageDirty(page))
1427                         set_page_dirty(page);
1428
1429                 inode->ei_last_dirtied = (unsigned long) current;
1430         } else {
1431                 bch2_put_page_reservation(c, inode, page);
1432         }
1433
1434         unlock_page(page);
1435         put_page(page);
1436         pagecache_add_put(&mapping->add_lock);
1437
1438         return copied;
1439 }
1440
1441 #define WRITE_BATCH_PAGES       32
1442
1443 static int __bch2_buffered_write(struct bch_inode_info *inode,
1444                                  struct address_space *mapping,
1445                                  struct iov_iter *iter,
1446                                  loff_t pos, unsigned len)
1447 {
1448         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1449         struct page *pages[WRITE_BATCH_PAGES];
1450         unsigned long index = pos >> PAGE_SHIFT;
1451         unsigned offset = pos & (PAGE_SIZE - 1);
1452         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1453         unsigned i, copied = 0, nr_pages_copied = 0;
1454         int ret = 0;
1455
1456         BUG_ON(!len);
1457         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1458
1459         for (i = 0; i < nr_pages; i++) {
1460                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1461                 if (!pages[i]) {
1462                         nr_pages = i;
1463                         ret = -ENOMEM;
1464                         goto out;
1465                 }
1466         }
1467
1468         if (offset && !PageUptodate(pages[0])) {
1469                 ret = bch2_read_single_page(pages[0], mapping);
1470                 if (ret)
1471                         goto out;
1472         }
1473
1474         if ((pos + len) & (PAGE_SIZE - 1) &&
1475             !PageUptodate(pages[nr_pages - 1])) {
1476                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1477                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1478                 } else {
1479                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1480                         if (ret)
1481                                 goto out;
1482                 }
1483         }
1484
1485         for (i = 0; i < nr_pages; i++) {
1486                 ret = bch2_get_page_reservation(c, inode, pages[i], true);
1487
1488                 if (ret && !PageUptodate(pages[i])) {
1489                         ret = bch2_read_single_page(pages[i], mapping);
1490                         if (ret)
1491                                 goto out;
1492
1493                         ret = bch2_get_page_reservation(c, inode, pages[i], true);
1494                 }
1495
1496                 if (ret)
1497                         goto out;
1498         }
1499
1500         if (mapping_writably_mapped(mapping))
1501                 for (i = 0; i < nr_pages; i++)
1502                         flush_dcache_page(pages[i]);
1503
1504         while (copied < len) {
1505                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1506                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1507                 unsigned pg_bytes = min_t(unsigned, len - copied,
1508                                           PAGE_SIZE - pg_offset);
1509                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1510                                                 iter, pg_offset, pg_bytes);
1511
1512                 if (!pg_copied)
1513                         break;
1514
1515                 flush_dcache_page(page);
1516                 iov_iter_advance(iter, pg_copied);
1517                 copied += pg_copied;
1518         }
1519
1520         if (!copied)
1521                 goto out;
1522
1523         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1524         inode->ei_last_dirtied = (unsigned long) current;
1525
1526         spin_lock(&inode->v.i_lock);
1527         if (pos + copied > inode->v.i_size)
1528                 i_size_write(&inode->v, pos + copied);
1529         spin_unlock(&inode->v.i_lock);
1530
1531         if (copied < len &&
1532             ((offset + copied) & (PAGE_SIZE - 1))) {
1533                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1534
1535                 if (!PageUptodate(page)) {
1536                         zero_user(page, 0, PAGE_SIZE);
1537                         copied -= (offset + copied) & (PAGE_SIZE - 1);
1538                 }
1539         }
1540 out:
1541         for (i = 0; i < nr_pages_copied; i++) {
1542                 if (!PageUptodate(pages[i]))
1543                         SetPageUptodate(pages[i]);
1544                 if (!PageDirty(pages[i]))
1545                         set_page_dirty(pages[i]);
1546                 unlock_page(pages[i]);
1547                 put_page(pages[i]);
1548         }
1549
1550         for (i = nr_pages_copied; i < nr_pages; i++) {
1551                 if (!PageDirty(pages[i]))
1552                         bch2_put_page_reservation(c, inode, pages[i]);
1553                 unlock_page(pages[i]);
1554                 put_page(pages[i]);
1555         }
1556
1557         return copied ?: ret;
1558 }
1559
1560 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1561 {
1562         struct file *file = iocb->ki_filp;
1563         struct address_space *mapping = file->f_mapping;
1564         struct bch_inode_info *inode = file_bch_inode(file);
1565         loff_t pos = iocb->ki_pos;
1566         ssize_t written = 0;
1567         int ret = 0;
1568
1569         pagecache_add_get(&mapping->add_lock);
1570
1571         do {
1572                 unsigned offset = pos & (PAGE_SIZE - 1);
1573                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1574                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1575 again:
1576                 /*
1577                  * Bring in the user page that we will copy from _first_.
1578                  * Otherwise there's a nasty deadlock on copying from the
1579                  * same page as we're writing to, without it being marked
1580                  * up-to-date.
1581                  *
1582                  * Not only is this an optimisation, but it is also required
1583                  * to check that the address is actually valid, when atomic
1584                  * usercopies are used, below.
1585                  */
1586                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1587                         bytes = min_t(unsigned long, iov_iter_count(iter),
1588                                       PAGE_SIZE - offset);
1589
1590                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1591                                 ret = -EFAULT;
1592                                 break;
1593                         }
1594                 }
1595
1596                 if (unlikely(fatal_signal_pending(current))) {
1597                         ret = -EINTR;
1598                         break;
1599                 }
1600
1601                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1602                 if (unlikely(ret < 0))
1603                         break;
1604
1605                 cond_resched();
1606
1607                 if (unlikely(ret == 0)) {
1608                         /*
1609                          * If we were unable to copy any data at all, we must
1610                          * fall back to a single segment length write.
1611                          *
1612                          * If we didn't fallback here, we could livelock
1613                          * because not all segments in the iov can be copied at
1614                          * once without a pagefault.
1615                          */
1616                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1617                                       iov_iter_single_seg_count(iter));
1618                         goto again;
1619                 }
1620                 pos += ret;
1621                 written += ret;
1622
1623                 balance_dirty_pages_ratelimited(mapping);
1624         } while (iov_iter_count(iter));
1625
1626         pagecache_add_put(&mapping->add_lock);
1627
1628         return written ? written : ret;
1629 }
1630
1631 /* O_DIRECT reads */
1632
1633 static void bch2_dio_read_complete(struct closure *cl)
1634 {
1635         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1636
1637         dio->req->ki_complete(dio->req, dio->ret, 0);
1638         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1639 }
1640
1641 static void bch2_direct_IO_read_endio(struct bio *bio)
1642 {
1643         struct dio_read *dio = bio->bi_private;
1644
1645         if (bio->bi_status)
1646                 dio->ret = blk_status_to_errno(bio->bi_status);
1647
1648         closure_put(&dio->cl);
1649 }
1650
1651 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1652 {
1653         bch2_direct_IO_read_endio(bio);
1654         bio_check_pages_dirty(bio);     /* transfers ownership */
1655 }
1656
1657 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1658 {
1659         struct file *file = req->ki_filp;
1660         struct bch_inode_info *inode = file_bch_inode(file);
1661         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1662         struct bch_io_opts opts = io_opts(c, inode);
1663         struct dio_read *dio;
1664         struct bio *bio;
1665         loff_t offset = req->ki_pos;
1666         bool sync = is_sync_kiocb(req);
1667         size_t shorten;
1668         ssize_t ret;
1669
1670         if ((offset|iter->count) & (block_bytes(c) - 1))
1671                 return -EINVAL;
1672
1673         ret = min_t(loff_t, iter->count,
1674                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1675
1676         if (!ret)
1677                 return ret;
1678
1679         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1680         iter->count -= shorten;
1681
1682         bio = bio_alloc_bioset(GFP_KERNEL,
1683                                iov_iter_npages(iter, BIO_MAX_PAGES),
1684                                &c->dio_read_bioset);
1685
1686         bio->bi_end_io = bch2_direct_IO_read_endio;
1687
1688         dio = container_of(bio, struct dio_read, rbio.bio);
1689         closure_init(&dio->cl, NULL);
1690
1691         /*
1692          * this is a _really_ horrible hack just to avoid an atomic sub at the
1693          * end:
1694          */
1695         if (!sync) {
1696                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1697                 atomic_set(&dio->cl.remaining,
1698                            CLOSURE_REMAINING_INITIALIZER -
1699                            CLOSURE_RUNNING +
1700                            CLOSURE_DESTRUCTOR);
1701         } else {
1702                 atomic_set(&dio->cl.remaining,
1703                            CLOSURE_REMAINING_INITIALIZER + 1);
1704         }
1705
1706         dio->req        = req;
1707         dio->ret        = ret;
1708
1709         goto start;
1710         while (iter->count) {
1711                 bio = bio_alloc_bioset(GFP_KERNEL,
1712                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1713                                        &c->bio_read);
1714                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1715 start:
1716                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1717                 bio->bi_iter.bi_sector  = offset >> 9;
1718                 bio->bi_private         = dio;
1719
1720                 ret = bio_iov_iter_get_pages(bio, iter);
1721                 if (ret < 0) {
1722                         /* XXX: fault inject this path */
1723                         bio->bi_status = BLK_STS_RESOURCE;
1724                         bio_endio(bio);
1725                         break;
1726                 }
1727
1728                 offset += bio->bi_iter.bi_size;
1729                 bio_set_pages_dirty(bio);
1730
1731                 if (iter->count)
1732                         closure_get(&dio->cl);
1733
1734                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1735         }
1736
1737         iter->count += shorten;
1738
1739         if (sync) {
1740                 closure_sync(&dio->cl);
1741                 closure_debug_destroy(&dio->cl);
1742                 ret = dio->ret;
1743                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1744                 return ret;
1745         } else {
1746                 return -EIOCBQUEUED;
1747         }
1748 }
1749
1750 /* O_DIRECT writes */
1751
1752 static void bch2_dio_write_loop_async(struct closure *);
1753
1754 static long bch2_dio_write_loop(struct dio_write *dio)
1755 {
1756         bool kthread = (current->flags & PF_KTHREAD) != 0;
1757         struct kiocb *req = dio->req;
1758         struct address_space *mapping = req->ki_filp->f_mapping;
1759         struct bch_inode_info *inode = dio->iop.inode;
1760         struct bio *bio = &dio->iop.op.wbio.bio;
1761         struct bio_vec *bv;
1762         bool sync;
1763         long ret;
1764         int i;
1765
1766         if (dio->loop)
1767                 goto loop;
1768
1769         inode_dio_begin(&inode->v);
1770         __pagecache_block_get(&mapping->add_lock);
1771
1772         /* Write and invalidate pagecache range that we're writing to: */
1773         ret = write_invalidate_inode_pages_range(mapping, req->ki_pos,
1774                                 req->ki_pos + iov_iter_count(&dio->iter) - 1);
1775         if (unlikely(ret))
1776                 goto err;
1777
1778         while (1) {
1779                 BUG_ON(current->pagecache_lock);
1780                 current->pagecache_lock = &mapping->add_lock;
1781                 if (kthread)
1782                         use_mm(dio->mm);
1783
1784                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1785
1786                 if (kthread)
1787                         unuse_mm(dio->mm);
1788                 current->pagecache_lock = NULL;
1789
1790                 if (unlikely(ret < 0))
1791                         goto err;
1792
1793                 /* gup might have faulted pages back in: */
1794                 ret = write_invalidate_inode_pages_range(mapping,
1795                                 req->ki_pos + (dio->iop.op.written << 9),
1796                                 req->ki_pos + iov_iter_count(&dio->iter) - 1);
1797                 if (unlikely(ret))
1798                         goto err;
1799
1800                 dio->iop.op.pos = POS(inode->v.i_ino,
1801                                 (req->ki_pos >> 9) + dio->iop.op.written);
1802
1803                 task_io_account_write(bio->bi_iter.bi_size);
1804
1805                 closure_call(&dio->iop.op.cl, bch2_write, NULL, &dio->cl);
1806
1807                 if (!dio->sync && !dio->loop && dio->iter.count) {
1808                         struct iovec *iov = dio->inline_vecs;
1809
1810                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1811                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1812                                               GFP_KERNEL);
1813                                 if (unlikely(!iov)) {
1814                                         dio->iop.op.error = -ENOMEM;
1815                                         goto err_wait_io;
1816                                 }
1817
1818                                 dio->free_iov = true;
1819                         }
1820
1821                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1822                         dio->iter.iov = iov;
1823                 }
1824 err_wait_io:
1825                 dio->loop = true;
1826
1827                 if (!dio->sync) {
1828                         continue_at(&dio->cl, bch2_dio_write_loop_async, NULL);
1829                         return -EIOCBQUEUED;
1830                 }
1831
1832                 closure_sync(&dio->cl);
1833 loop:
1834                 bio_for_each_segment_all(bv, bio, i)
1835                         put_page(bv->bv_page);
1836                 if (!dio->iter.count || dio->iop.op.error)
1837                         break;
1838                 bio_reset(bio);
1839         }
1840
1841         ret = dio->iop.op.error ?: ((long) dio->iop.op.written << 9);
1842 err:
1843         __pagecache_block_put(&mapping->add_lock);
1844         bch2_disk_reservation_put(dio->iop.op.c, &dio->iop.op.res);
1845         bch2_quota_reservation_put(dio->iop.op.c, inode, &dio->quota_res);
1846
1847         if (dio->free_iov)
1848                 kfree(dio->iter.iov);
1849
1850         closure_debug_destroy(&dio->cl);
1851
1852         sync = dio->sync;
1853         bio_put(bio);
1854
1855         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1856         inode_dio_end(&inode->v);
1857
1858         if (!sync) {
1859                 req->ki_complete(req, ret, 0);
1860                 ret = -EIOCBQUEUED;
1861         }
1862         return ret;
1863 }
1864
1865 static void bch2_dio_write_loop_async(struct closure *cl)
1866 {
1867         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1868
1869         bch2_dio_write_loop(dio);
1870 }
1871
1872 static int bch2_direct_IO_write(struct kiocb *req,
1873                                 struct iov_iter *iter,
1874                                 bool swap)
1875 {
1876         struct file *file = req->ki_filp;
1877         struct bch_inode_info *inode = file_bch_inode(file);
1878         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1879         struct dio_write *dio;
1880         struct bio *bio;
1881         loff_t offset = req->ki_pos;
1882         ssize_t ret;
1883
1884         lockdep_assert_held(&inode->v.i_rwsem);
1885
1886         if (unlikely(!iter->count))
1887                 return 0;
1888
1889         if (unlikely((offset|iter->count) & (block_bytes(c) - 1)))
1890                 return -EINVAL;
1891
1892         bio = bio_alloc_bioset(GFP_KERNEL,
1893                                iov_iter_npages(iter, BIO_MAX_PAGES),
1894                                &c->dio_write_bioset);
1895         dio = container_of(bio, struct dio_write, iop.op.wbio.bio);
1896         closure_init(&dio->cl, NULL);
1897         dio->req                = req;
1898         dio->mm                 = current->mm;
1899         dio->loop               = false;
1900         dio->sync               = is_sync_kiocb(req) ||
1901                 offset + iter->count > inode->v.i_size;
1902         dio->free_iov           = false;
1903         dio->quota_res.sectors  = 0;
1904         dio->iter               = *iter;
1905         bch2_fswrite_op_init(&dio->iop, c, inode, io_opts(c, inode), true);
1906         dio->iop.op.write_point = writepoint_hashed((unsigned long) current);
1907         dio->iop.op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1908
1909         if ((req->ki_flags & IOCB_DSYNC) &&
1910             !c->opts.journal_flush_disabled)
1911                 dio->iop.op.flags |= BCH_WRITE_FLUSH;
1912
1913         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
1914                                          iter->count >> 9, true);
1915         if (unlikely(ret))
1916                 goto err;
1917
1918         ret = bch2_disk_reservation_get(c, &dio->iop.op.res, iter->count >> 9,
1919                                         dio->iop.op.opts.data_replicas, 0);
1920         if (unlikely(ret)) {
1921                 if (bch2_check_range_allocated(c, POS(inode->v.i_ino,
1922                                                       offset >> 9),
1923                                                iter->count >> 9))
1924                         goto err;
1925
1926                 dio->iop.unalloc = true;
1927         }
1928
1929         dio->iop.op.nr_replicas = dio->iop.op.res.nr_replicas;
1930
1931         return bch2_dio_write_loop(dio);
1932 err:
1933         bch2_disk_reservation_put(c, &dio->iop.op.res);
1934         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1935         closure_debug_destroy(&dio->cl);
1936         bio_put(bio);
1937         return ret;
1938 }
1939
1940 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1941 {
1942         struct blk_plug plug;
1943         ssize_t ret;
1944
1945         blk_start_plug(&plug);
1946         ret = iov_iter_rw(iter) == WRITE
1947                 ? bch2_direct_IO_write(req, iter, false)
1948                 : bch2_direct_IO_read(req, iter);
1949         blk_finish_plug(&plug);
1950
1951         return ret;
1952 }
1953
1954 static ssize_t
1955 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1956 {
1957         return bch2_direct_IO_write(iocb, iter, true);
1958 }
1959
1960 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1961 {
1962         struct file *file = iocb->ki_filp;
1963         struct bch_inode_info *inode = file_bch_inode(file);
1964         ssize_t ret;
1965
1966         /* We can write back this queue in page reclaim */
1967         current->backing_dev_info = inode_to_bdi(&inode->v);
1968         ret = file_remove_privs(file);
1969         if (ret)
1970                 goto out;
1971
1972         ret = file_update_time(file);
1973         if (ret)
1974                 goto out;
1975
1976         ret = iocb->ki_flags & IOCB_DIRECT
1977                 ? bch2_direct_write(iocb, from)
1978                 : bch2_buffered_write(iocb, from);
1979
1980         if (likely(ret > 0))
1981                 iocb->ki_pos += ret;
1982 out:
1983         current->backing_dev_info = NULL;
1984         return ret;
1985 }
1986
1987 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1988 {
1989         struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp);
1990         bool direct = iocb->ki_flags & IOCB_DIRECT;
1991         ssize_t ret;
1992
1993         inode_lock(&inode->v);
1994         ret = generic_write_checks(iocb, from);
1995         if (ret > 0)
1996                 ret = __bch2_write_iter(iocb, from);
1997         inode_unlock(&inode->v);
1998
1999         if (ret > 0 && !direct)
2000                 ret = generic_write_sync(iocb, ret);
2001
2002         return ret;
2003 }
2004
2005 /* fsync: */
2006
2007 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2008 {
2009         struct bch_inode_info *inode = file_bch_inode(file);
2010         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2011         int ret, ret2;
2012
2013         ret = file_write_and_wait_range(file, start, end);
2014         if (ret)
2015                 return ret;
2016
2017         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2018                 goto out;
2019
2020         ret = sync_inode_metadata(&inode->v, 1);
2021         if (ret)
2022                 return ret;
2023 out:
2024         if (c->opts.journal_flush_disabled)
2025                 return 0;
2026
2027         ret = bch2_journal_flush_seq(&c->journal, inode->ei_journal_seq);
2028         ret2 = file_check_and_advance_wb_err(file);
2029
2030         return ret ?: ret2;
2031 }
2032
2033 /* truncate: */
2034
2035 static int __bch2_fpunch(struct bch_fs *c, struct bch_inode_info *inode,
2036                          u64 start_offset, u64 end_offset, u64 *journal_seq)
2037 {
2038         struct bpos start       = POS(inode->v.i_ino, start_offset);
2039         struct bpos end         = POS(inode->v.i_ino, end_offset);
2040         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
2041         struct btree_trans trans;
2042         struct btree_iter *iter;
2043         struct bkey_s_c k;
2044         int ret = 0;
2045
2046         bch2_trans_init(&trans, c);
2047         bch2_trans_preload_iters(&trans);
2048
2049         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, start,
2050                                    BTREE_ITER_INTENT);
2051
2052         while ((k = bch2_btree_iter_peek(iter)).k &&
2053                !(ret = btree_iter_err(k)) &&
2054                bkey_cmp(iter->pos, end) < 0) {
2055                 struct disk_reservation disk_res =
2056                         bch2_disk_reservation_init(c, 0);
2057                 struct bkey_i delete;
2058
2059                 bkey_init(&delete.k);
2060                 delete.k.p = iter->pos;
2061
2062                 /* create the biggest key we can */
2063                 bch2_key_resize(&delete.k, max_sectors);
2064                 bch2_cut_back(end, &delete.k);
2065
2066                 ret = bch2_extent_update(&trans, inode,
2067                                 &disk_res, NULL, iter, &delete,
2068                                 0, true, true, NULL);
2069                 bch2_disk_reservation_put(c, &disk_res);
2070
2071                 if (ret == -EINTR)
2072                         ret = 0;
2073                 if (ret)
2074                         break;
2075
2076                 bch2_btree_iter_cond_resched(iter);
2077         }
2078
2079         bch2_trans_exit(&trans);
2080
2081         return ret;
2082 }
2083
2084 static inline int range_has_data(struct bch_fs *c,
2085                                   struct bpos start,
2086                                   struct bpos end)
2087 {
2088
2089         struct btree_iter iter;
2090         struct bkey_s_c k;
2091         int ret = 0;
2092
2093         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2094                            start, 0, k) {
2095                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2096                         break;
2097
2098                 if (bkey_extent_is_data(k.k)) {
2099                         ret = 1;
2100                         break;
2101                 }
2102         }
2103
2104         return bch2_btree_iter_unlock(&iter) ?: ret;
2105 }
2106
2107 static int __bch2_truncate_page(struct bch_inode_info *inode,
2108                                 pgoff_t index, loff_t start, loff_t end)
2109 {
2110         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2111         struct address_space *mapping = inode->v.i_mapping;
2112         unsigned start_offset = start & (PAGE_SIZE - 1);
2113         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2114         struct page *page;
2115         int ret = 0;
2116
2117         /* Page boundary? Nothing to do */
2118         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2119               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2120                 return 0;
2121
2122         /* Above i_size? */
2123         if (index << PAGE_SHIFT >= inode->v.i_size)
2124                 return 0;
2125
2126         page = find_lock_page(mapping, index);
2127         if (!page) {
2128                 /*
2129                  * XXX: we're doing two index lookups when we end up reading the
2130                  * page
2131                  */
2132                 ret = range_has_data(c,
2133                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2134                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2135                 if (ret <= 0)
2136                         return ret;
2137
2138                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2139                 if (unlikely(!page)) {
2140                         ret = -ENOMEM;
2141                         goto out;
2142                 }
2143         }
2144
2145         if (!PageUptodate(page)) {
2146                 ret = bch2_read_single_page(page, mapping);
2147                 if (ret)
2148                         goto unlock;
2149         }
2150
2151         /*
2152          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2153          *
2154          * XXX: because we aren't currently tracking whether the page has actual
2155          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2156          */
2157         ret = bch2_get_page_reservation(c, inode, page, false);
2158         BUG_ON(ret);
2159
2160         if (index == start >> PAGE_SHIFT &&
2161             index == end >> PAGE_SHIFT)
2162                 zero_user_segment(page, start_offset, end_offset);
2163         else if (index == start >> PAGE_SHIFT)
2164                 zero_user_segment(page, start_offset, PAGE_SIZE);
2165         else if (index == end >> PAGE_SHIFT)
2166                 zero_user_segment(page, 0, end_offset);
2167
2168         if (!PageDirty(page))
2169                 set_page_dirty(page);
2170 unlock:
2171         unlock_page(page);
2172         put_page(page);
2173 out:
2174         return ret;
2175 }
2176
2177 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2178 {
2179         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2180                                     from, from + PAGE_SIZE);
2181 }
2182
2183 static int bch2_extend(struct bch_inode_info *inode, struct iattr *iattr)
2184 {
2185         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2186         struct address_space *mapping = inode->v.i_mapping;
2187         int ret;
2188
2189         ret = filemap_write_and_wait_range(mapping,
2190                         inode->ei_inode.bi_size, S64_MAX);
2191         if (ret)
2192                 return ret;
2193
2194         truncate_setsize(&inode->v, iattr->ia_size);
2195         setattr_copy(&inode->v, iattr);
2196
2197         mutex_lock(&inode->ei_update_lock);
2198         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2199                                     ATTR_MTIME|ATTR_CTIME);
2200         mutex_unlock(&inode->ei_update_lock);
2201
2202         return ret;
2203 }
2204
2205 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2206                                    struct bch_inode_unpacked *bi,
2207                                    void *p)
2208 {
2209         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2210
2211         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2212         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2213         return 0;
2214 }
2215
2216 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2217                                   struct bch_inode_unpacked *bi, void *p)
2218 {
2219         u64 *new_i_size = p;
2220
2221         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2222         bi->bi_size = *new_i_size;
2223         return 0;
2224 }
2225
2226 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2227 {
2228         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2229         struct address_space *mapping = inode->v.i_mapping;
2230         u64 new_i_size = iattr->ia_size;
2231         bool shrink;
2232         int ret = 0;
2233
2234         inode_dio_wait(&inode->v);
2235         pagecache_block_get(&mapping->add_lock);
2236
2237         BUG_ON(inode->v.i_size < inode->ei_inode.bi_size);
2238
2239         shrink = iattr->ia_size <= inode->v.i_size;
2240
2241         if (!shrink) {
2242                 ret = bch2_extend(inode, iattr);
2243                 goto err;
2244         }
2245
2246         ret = bch2_truncate_page(inode, iattr->ia_size);
2247         if (unlikely(ret))
2248                 goto err;
2249
2250         if (iattr->ia_size > inode->ei_inode.bi_size)
2251                 ret = filemap_write_and_wait_range(mapping,
2252                                 inode->ei_inode.bi_size,
2253                                 iattr->ia_size - 1);
2254         else if (iattr->ia_size & (PAGE_SIZE - 1))
2255                 ret = filemap_write_and_wait_range(mapping,
2256                                 round_down(iattr->ia_size, PAGE_SIZE),
2257                                 iattr->ia_size - 1);
2258         if (ret)
2259                 goto err;
2260
2261         mutex_lock(&inode->ei_update_lock);
2262         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2263                                &new_i_size, 0);
2264         mutex_unlock(&inode->ei_update_lock);
2265
2266         if (unlikely(ret))
2267                 goto err;
2268
2269         truncate_setsize(&inode->v, iattr->ia_size);
2270
2271         /*
2272          * XXX: need a comment explaining why PAGE_SIZE and not block_bytes()
2273          * here:
2274          */
2275         ret = __bch2_fpunch(c, inode,
2276                         round_up(iattr->ia_size, PAGE_SIZE) >> 9,
2277                         U64_MAX, &inode->ei_journal_seq);
2278         if (unlikely(ret))
2279                 goto err;
2280
2281         setattr_copy(&inode->v, iattr);
2282
2283         mutex_lock(&inode->ei_update_lock);
2284         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2285                                ATTR_MTIME|ATTR_CTIME);
2286         mutex_unlock(&inode->ei_update_lock);
2287 err:
2288         pagecache_block_put(&mapping->add_lock);
2289         return ret;
2290 }
2291
2292 /* fallocate: */
2293
2294 static long bch2_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2295 {
2296         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2297         struct address_space *mapping = inode->v.i_mapping;
2298         u64 discard_start = round_up(offset, PAGE_SIZE) >> 9;
2299         u64 discard_end = round_down(offset + len, PAGE_SIZE) >> 9;
2300         int ret = 0;
2301
2302         inode_lock(&inode->v);
2303         inode_dio_wait(&inode->v);
2304         pagecache_block_get(&mapping->add_lock);
2305
2306         ret = __bch2_truncate_page(inode,
2307                                    offset >> PAGE_SHIFT,
2308                                    offset, offset + len);
2309         if (unlikely(ret))
2310                 goto err;
2311
2312         if (offset >> PAGE_SHIFT !=
2313             (offset + len) >> PAGE_SHIFT) {
2314                 ret = __bch2_truncate_page(inode,
2315                                            (offset + len) >> PAGE_SHIFT,
2316                                            offset, offset + len);
2317                 if (unlikely(ret))
2318                         goto err;
2319         }
2320
2321         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2322
2323         if (discard_start < discard_end)
2324                 ret = __bch2_fpunch(c, inode, discard_start, discard_end,
2325                                     &inode->ei_journal_seq);
2326 err:
2327         pagecache_block_put(&mapping->add_lock);
2328         inode_unlock(&inode->v);
2329
2330         return ret;
2331 }
2332
2333 static long bch2_fcollapse(struct bch_inode_info *inode,
2334                            loff_t offset, loff_t len)
2335 {
2336         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2337         struct address_space *mapping = inode->v.i_mapping;
2338         struct btree_trans trans;
2339         struct btree_iter *src, *dst;
2340         BKEY_PADDED(k) copy;
2341         struct bkey_s_c k;
2342         loff_t new_size;
2343         int ret;
2344
2345         if ((offset | len) & (block_bytes(c) - 1))
2346                 return -EINVAL;
2347
2348         bch2_trans_init(&trans, c);
2349         bch2_trans_preload_iters(&trans);
2350
2351         /*
2352          * We need i_mutex to keep the page cache consistent with the extents
2353          * btree, and the btree consistent with i_size - we don't need outside
2354          * locking for the extents btree itself, because we're using linked
2355          * iterators
2356          */
2357         inode_lock(&inode->v);
2358         inode_dio_wait(&inode->v);
2359         pagecache_block_get(&mapping->add_lock);
2360
2361         ret = -EINVAL;
2362         if (offset + len >= inode->v.i_size)
2363                 goto err;
2364
2365         if (inode->v.i_size < len)
2366                 goto err;
2367
2368         new_size = inode->v.i_size - len;
2369
2370         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2371         if (ret)
2372                 goto err;
2373
2374         dst = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2375                         POS(inode->v.i_ino, offset >> 9),
2376                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2377         BUG_ON(IS_ERR_OR_NULL(dst));
2378
2379         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2380                         POS_MIN, BTREE_ITER_SLOTS);
2381         BUG_ON(IS_ERR_OR_NULL(src));
2382
2383         while (bkey_cmp(dst->pos,
2384                         POS(inode->v.i_ino,
2385                             round_up(new_size, PAGE_SIZE) >> 9)) < 0) {
2386                 struct disk_reservation disk_res;
2387
2388                 ret = bch2_btree_iter_traverse(dst);
2389                 if (ret)
2390                         goto btree_iter_err;
2391
2392                 bch2_btree_iter_set_pos(src,
2393                         POS(dst->pos.inode, dst->pos.offset + (len >> 9)));
2394
2395                 k = bch2_btree_iter_peek_slot(src);
2396                 if ((ret = btree_iter_err(k)))
2397                         goto btree_iter_err;
2398
2399                 bkey_reassemble(&copy.k, k);
2400
2401                 bch2_cut_front(src->pos, &copy.k);
2402                 copy.k.k.p.offset -= len >> 9;
2403
2404                 bch2_extent_trim_atomic(&copy.k, dst);
2405
2406                 BUG_ON(bkey_cmp(dst->pos, bkey_start_pos(&copy.k.k)));
2407
2408                 ret = bch2_disk_reservation_get(c, &disk_res, copy.k.k.size,
2409                                 bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&copy.k)),
2410                                 BCH_DISK_RESERVATION_NOFAIL);
2411                 BUG_ON(ret);
2412
2413                 ret = bch2_extent_update(&trans, inode,
2414                                 &disk_res, NULL,
2415                                 dst, &copy.k,
2416                                 0, true, true, NULL);
2417                 bch2_disk_reservation_put(c, &disk_res);
2418 btree_iter_err:
2419                 if (ret == -EINTR)
2420                         ret = 0;
2421                 if (ret)
2422                         goto err;
2423                 /*
2424                  * XXX: if we error here we've left data with multiple
2425                  * pointers... which isn't a _super_ serious problem...
2426                  */
2427
2428                 bch2_btree_iter_cond_resched(src);
2429         }
2430         bch2_trans_unlock(&trans);
2431
2432         ret = __bch2_fpunch(c, inode,
2433                         round_up(new_size, block_bytes(c)) >> 9,
2434                         U64_MAX, &inode->ei_journal_seq);
2435         if (ret)
2436                 goto err;
2437
2438         i_size_write(&inode->v, new_size);
2439         mutex_lock(&inode->ei_update_lock);
2440         ret = bch2_write_inode_size(c, inode, new_size,
2441                                     ATTR_MTIME|ATTR_CTIME);
2442         mutex_unlock(&inode->ei_update_lock);
2443 err:
2444         bch2_trans_exit(&trans);
2445         pagecache_block_put(&mapping->add_lock);
2446         inode_unlock(&inode->v);
2447         return ret;
2448 }
2449
2450 static long bch2_fallocate(struct bch_inode_info *inode, int mode,
2451                            loff_t offset, loff_t len)
2452 {
2453         struct address_space *mapping = inode->v.i_mapping;
2454         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2455         struct btree_trans trans;
2456         struct btree_iter *iter;
2457         struct bpos end_pos;
2458         loff_t block_start, block_end;
2459         loff_t end = offset + len;
2460         unsigned sectors;
2461         unsigned replicas = io_opts(c, inode).data_replicas;
2462         int ret;
2463
2464         bch2_trans_init(&trans, c);
2465         bch2_trans_preload_iters(&trans);
2466
2467         inode_lock(&inode->v);
2468         inode_dio_wait(&inode->v);
2469         pagecache_block_get(&mapping->add_lock);
2470
2471         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2472                 ret = inode_newsize_ok(&inode->v, end);
2473                 if (ret)
2474                         goto err;
2475         }
2476
2477         if (mode & FALLOC_FL_ZERO_RANGE) {
2478                 ret = __bch2_truncate_page(inode,
2479                                            offset >> PAGE_SHIFT,
2480                                            offset, end);
2481
2482                 if (!ret &&
2483                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2484                         ret = __bch2_truncate_page(inode,
2485                                                    end >> PAGE_SHIFT,
2486                                                    offset, end);
2487
2488                 if (unlikely(ret))
2489                         goto err;
2490
2491                 truncate_pagecache_range(&inode->v, offset, end - 1);
2492
2493                 block_start     = round_up(offset, PAGE_SIZE);
2494                 block_end       = round_down(end, PAGE_SIZE);
2495         } else {
2496                 block_start     = round_down(offset, PAGE_SIZE);
2497                 block_end       = round_up(end, PAGE_SIZE);
2498         }
2499
2500         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2501                         POS(inode->v.i_ino, block_start >> 9),
2502                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2503         end_pos = POS(inode->v.i_ino, block_end >> 9);
2504
2505         while (bkey_cmp(iter->pos, end_pos) < 0) {
2506                 struct disk_reservation disk_res = { 0 };
2507                 struct quota_res quota_res = { 0 };
2508                 struct bkey_i_reservation reservation;
2509                 struct bkey_s_c k;
2510
2511                 k = bch2_btree_iter_peek_slot(iter);
2512                 if ((ret = btree_iter_err(k)))
2513                         goto btree_iter_err;
2514
2515                 /* already reserved */
2516                 if (k.k->type == KEY_TYPE_reservation &&
2517                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2518                         bch2_btree_iter_next_slot(iter);
2519                         continue;
2520                 }
2521
2522                 if (bkey_extent_is_data(k.k) &&
2523                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2524                         bch2_btree_iter_next_slot(iter);
2525                         continue;
2526                 }
2527
2528                 bkey_reservation_init(&reservation.k_i);
2529                 reservation.k.type      = KEY_TYPE_reservation;
2530                 reservation.k.p         = k.k->p;
2531                 reservation.k.size      = k.k->size;
2532
2533                 bch2_cut_front(iter->pos, &reservation.k_i);
2534                 bch2_cut_back(end_pos, &reservation.k);
2535
2536                 sectors = reservation.k.size;
2537                 reservation.v.nr_replicas = bch2_bkey_nr_dirty_ptrs(k);
2538
2539                 if (!bkey_extent_is_allocation(k.k)) {
2540                         ret = bch2_quota_reservation_add(c, inode,
2541                                         &quota_res,
2542                                         sectors, true);
2543                         if (unlikely(ret))
2544                                 goto btree_iter_err;
2545                 }
2546
2547                 if (reservation.v.nr_replicas < replicas ||
2548                     bch2_extent_is_compressed(k)) {
2549                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2550                                                         replicas, 0);
2551                         if (unlikely(ret))
2552                                 goto btree_iter_err;
2553
2554                         reservation.v.nr_replicas = disk_res.nr_replicas;
2555                 }
2556
2557                 ret = bch2_extent_update(&trans, inode,
2558                                 &disk_res, &quota_res,
2559                                 iter, &reservation.k_i,
2560                                 0, true, true, NULL);
2561 btree_iter_err:
2562                 bch2_quota_reservation_put(c, inode, &quota_res);
2563                 bch2_disk_reservation_put(c, &disk_res);
2564                 if (ret == -EINTR)
2565                         ret = 0;
2566                 if (ret)
2567                         goto err;
2568         }
2569         bch2_trans_unlock(&trans);
2570
2571         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2572             end > inode->v.i_size) {
2573                 i_size_write(&inode->v, end);
2574
2575                 mutex_lock(&inode->ei_update_lock);
2576                 ret = bch2_write_inode_size(c, inode, inode->v.i_size, 0);
2577                 mutex_unlock(&inode->ei_update_lock);
2578         }
2579
2580         /* blech */
2581         if ((mode & FALLOC_FL_KEEP_SIZE) &&
2582             (mode & FALLOC_FL_ZERO_RANGE) &&
2583             inode->ei_inode.bi_size != inode->v.i_size) {
2584                 /* sync appends.. */
2585                 ret = filemap_write_and_wait_range(mapping,
2586                                         inode->ei_inode.bi_size, S64_MAX);
2587                 if (ret)
2588                         goto err;
2589
2590                 if (inode->ei_inode.bi_size != inode->v.i_size) {
2591                         mutex_lock(&inode->ei_update_lock);
2592                         ret = bch2_write_inode_size(c, inode,
2593                                                     inode->v.i_size, 0);
2594                         mutex_unlock(&inode->ei_update_lock);
2595                 }
2596         }
2597 err:
2598         bch2_trans_exit(&trans);
2599         pagecache_block_put(&mapping->add_lock);
2600         inode_unlock(&inode->v);
2601         return ret;
2602 }
2603
2604 long bch2_fallocate_dispatch(struct file *file, int mode,
2605                              loff_t offset, loff_t len)
2606 {
2607         struct bch_inode_info *inode = file_bch_inode(file);
2608
2609         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2610                 return bch2_fallocate(inode, mode, offset, len);
2611
2612         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2613                 return bch2_fpunch(inode, offset, len);
2614
2615         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2616                 return bch2_fcollapse(inode, offset, len);
2617
2618         return -EOPNOTSUPP;
2619 }
2620
2621 /* fseek: */
2622
2623 static bool page_is_data(struct page *page)
2624 {
2625         EBUG_ON(!PageLocked(page));
2626
2627         /* XXX: should only have to check PageDirty */
2628         return PagePrivate(page) &&
2629                 (page_state(page)->sectors ||
2630                  page_state(page)->dirty_sectors);
2631 }
2632
2633 static loff_t bch2_next_pagecache_data(struct inode *vinode,
2634                                        loff_t start_offset,
2635                                        loff_t end_offset)
2636 {
2637         struct address_space *mapping = vinode->i_mapping;
2638         struct page *page;
2639         pgoff_t index;
2640
2641         for (index = start_offset >> PAGE_SHIFT;
2642              index < end_offset >> PAGE_SHIFT;
2643              index++) {
2644                 if (find_get_pages(mapping, &index, 1, &page)) {
2645                         lock_page(page);
2646
2647                         if (page_is_data(page))
2648                                 end_offset =
2649                                         min(end_offset,
2650                                         max(start_offset,
2651                                             ((loff_t) index) << PAGE_SHIFT));
2652                         unlock_page(page);
2653                         put_page(page);
2654                 } else {
2655                         break;
2656                 }
2657         }
2658
2659         return end_offset;
2660 }
2661
2662 static loff_t bch2_seek_data(struct file *file, u64 offset)
2663 {
2664         struct bch_inode_info *inode = file_bch_inode(file);
2665         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2666         struct btree_iter iter;
2667         struct bkey_s_c k;
2668         u64 isize, next_data = MAX_LFS_FILESIZE;
2669         int ret;
2670
2671         isize = i_size_read(&inode->v);
2672         if (offset >= isize)
2673                 return -ENXIO;
2674
2675         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2676                            POS(inode->v.i_ino, offset >> 9), 0, k) {
2677                 if (k.k->p.inode != inode->v.i_ino) {
2678                         break;
2679                 } else if (bkey_extent_is_data(k.k)) {
2680                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2681                         break;
2682                 } else if (k.k->p.offset >> 9 > isize)
2683                         break;
2684         }
2685
2686         ret = bch2_btree_iter_unlock(&iter);
2687         if (ret)
2688                 return ret;
2689
2690         if (next_data > offset)
2691                 next_data = bch2_next_pagecache_data(&inode->v,
2692                                                      offset, next_data);
2693
2694         if (next_data > isize)
2695                 return -ENXIO;
2696
2697         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2698 }
2699
2700 static bool page_slot_is_data(struct address_space *mapping, pgoff_t index)
2701 {
2702         struct page *page;
2703         bool ret;
2704
2705         page = find_lock_entry(mapping, index);
2706         if (!page || xa_is_value(page))
2707                 return false;
2708
2709         ret = page_is_data(page);
2710         unlock_page(page);
2711
2712         return ret;
2713 }
2714
2715 static loff_t bch2_next_pagecache_hole(struct inode *vinode,
2716                                        loff_t start_offset,
2717                                        loff_t end_offset)
2718 {
2719         struct address_space *mapping = vinode->i_mapping;
2720         pgoff_t index;
2721
2722         for (index = start_offset >> PAGE_SHIFT;
2723              index < end_offset >> PAGE_SHIFT;
2724              index++)
2725                 if (!page_slot_is_data(mapping, index))
2726                         end_offset = max(start_offset,
2727                                          ((loff_t) index) << PAGE_SHIFT);
2728
2729         return end_offset;
2730 }
2731
2732 static loff_t bch2_seek_hole(struct file *file, u64 offset)
2733 {
2734         struct bch_inode_info *inode = file_bch_inode(file);
2735         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2736         struct btree_iter iter;
2737         struct bkey_s_c k;
2738         u64 isize, next_hole = MAX_LFS_FILESIZE;
2739         int ret;
2740
2741         isize = i_size_read(&inode->v);
2742         if (offset >= isize)
2743                 return -ENXIO;
2744
2745         for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
2746                            POS(inode->v.i_ino, offset >> 9),
2747                            BTREE_ITER_SLOTS, k) {
2748                 if (k.k->p.inode != inode->v.i_ino) {
2749                         next_hole = bch2_next_pagecache_hole(&inode->v,
2750                                         offset, MAX_LFS_FILESIZE);
2751                         break;
2752                 } else if (!bkey_extent_is_data(k.k)) {
2753                         next_hole = bch2_next_pagecache_hole(&inode->v,
2754                                         max(offset, bkey_start_offset(k.k) << 9),
2755                                         k.k->p.offset << 9);
2756
2757                         if (next_hole < k.k->p.offset << 9)
2758                                 break;
2759                 } else {
2760                         offset = max(offset, bkey_start_offset(k.k) << 9);
2761                 }
2762         }
2763
2764         ret = bch2_btree_iter_unlock(&iter);
2765         if (ret)
2766                 return ret;
2767
2768         if (next_hole > isize)
2769                 next_hole = isize;
2770
2771         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
2772 }
2773
2774 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
2775 {
2776         switch (whence) {
2777         case SEEK_SET:
2778         case SEEK_CUR:
2779         case SEEK_END:
2780                 return generic_file_llseek(file, offset, whence);
2781         case SEEK_DATA:
2782                 return bch2_seek_data(file, offset);
2783         case SEEK_HOLE:
2784                 return bch2_seek_hole(file, offset);
2785         }
2786
2787         return -EINVAL;
2788 }
2789
2790 void bch2_fs_fsio_exit(struct bch_fs *c)
2791 {
2792         bioset_exit(&c->dio_write_bioset);
2793         bioset_exit(&c->dio_read_bioset);
2794         bioset_exit(&c->writepage_bioset);
2795 }
2796
2797 int bch2_fs_fsio_init(struct bch_fs *c)
2798 {
2799         int ret = 0;
2800
2801         pr_verbose_init(c->opts, "");
2802
2803         if (bioset_init(&c->writepage_bioset,
2804                         4, offsetof(struct bch_writepage_io, op.op.wbio.bio),
2805                         BIOSET_NEED_BVECS) ||
2806             bioset_init(&c->dio_read_bioset,
2807                         4, offsetof(struct dio_read, rbio.bio),
2808                         BIOSET_NEED_BVECS) ||
2809             bioset_init(&c->dio_write_bioset,
2810                         4, offsetof(struct dio_write, iop.op.wbio.bio),
2811                         BIOSET_NEED_BVECS))
2812                 ret = -ENOMEM;
2813
2814         pr_verbose_init(c->opts, "ret %i", ret);
2815         return ret;
2816 }
2817
2818 #endif /* NO_BCACHEFS_FS */