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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 static inline struct address_space *faults_disabled_mapping(void)
  39 {
  40         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  41 }
  42
  43 static inline void set_fdm_dropped_locks(void)
  44 {
  45         current->faults_disabled_mapping =
  46                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  47 }
  48
  49 static inline bool fdm_dropped_locks(void)
  50 {
  51         return ((unsigned long) current->faults_disabled_mapping) & 1;
  52 }
  53
  54 struct quota_res {
  55         u64                             sectors;
  56 };
  57
  58 struct bch_writepage_io {
  59         struct closure                  cl;
  60         struct bch_inode_info           *inode;
  61
  62         /* must be last: */
  63         struct bch_write_op             op;
  64 };
  65
  66 struct dio_write {
  67         struct completion               done;
  68         struct kiocb                    *req;
  69         struct mm_struct                *mm;
  70         unsigned                        loop:1,
  71                                         sync:1,
  72                                         free_iov:1;
  73         struct quota_res                quota_res;
  74         u64                             written;
  75
  76         struct iov_iter                 iter;
  77         struct iovec                    inline_vecs[2];
  78
  79         /* must be last: */
  80         struct bch_write_op             op;
  81 };
  82
  83 struct dio_read {
  84         struct closure                  cl;
  85         struct kiocb                    *req;
  86         long                            ret;
  87         bool                            should_dirty;
  88         struct bch_read_bio             rbio;
  89 };
  90
  91 /* pagecache_block must be held */
  92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  93                                               loff_t start, loff_t end)
  94 {
  95         int ret;
  96
  97         /*
  98          * XXX: the way this is currently implemented, we can spin if a process
  99          * is continually redirtying a specific page
 100          */
 101         do {
 102                 if (!mapping->nrpages)
 103                         return 0;
 104
 105                 ret = filemap_write_and_wait_range(mapping, start, end);
 106                 if (ret)
 107                         break;
 108
 109                 if (!mapping->nrpages)
 110                         return 0;
 111
 112                 ret = invalidate_inode_pages2_range(mapping,
 113                                 start >> PAGE_SHIFT,
 114                                 end >> PAGE_SHIFT);
 115         } while (ret == -EBUSY);
 116
 117         return ret;
 118 }
 119
 120 /* quotas */
 121
 122 #ifdef CONFIG_BCACHEFS_QUOTA
 123
 124 static void bch2_quota_reservation_put(struct bch_fs *c,
 125                                        struct bch_inode_info *inode,
 126                                        struct quota_res *res)
 127 {
 128         if (!res->sectors)
 129                 return;
 130
 131         mutex_lock(&inode->ei_quota_lock);
 132         BUG_ON(res->sectors > inode->ei_quota_reserved);
 133
 134         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 135                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 136         inode->ei_quota_reserved -= res->sectors;
 137         mutex_unlock(&inode->ei_quota_lock);
 138
 139         res->sectors = 0;
 140 }
 141
 142 static int bch2_quota_reservation_add(struct bch_fs *c,
 143                                       struct bch_inode_info *inode,
 144                                       struct quota_res *res,
 145                                       unsigned sectors,
 146                                       bool check_enospc)
 147 {
 148         int ret;
 149
 150         mutex_lock(&inode->ei_quota_lock);
 151         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 152                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 153         if (likely(!ret)) {
 154                 inode->ei_quota_reserved += sectors;
 155                 res->sectors += sectors;
 156         }
 157         mutex_unlock(&inode->ei_quota_lock);
 158
 159         return ret;
 160 }
 161
 162 #else
 163
 164 static void bch2_quota_reservation_put(struct bch_fs *c,
 165                                        struct bch_inode_info *inode,
 166                                        struct quota_res *res)
 167 {
 168 }
 169
 170 static int bch2_quota_reservation_add(struct bch_fs *c,
 171                                       struct bch_inode_info *inode,
 172                                       struct quota_res *res,
 173                                       unsigned sectors,
 174                                       bool check_enospc)
 175 {
 176         return 0;
 177 }
 178
 179 #endif
 180
 181 /* i_size updates: */
 182
 183 struct inode_new_size {
 184         loff_t          new_size;
 185         u64             now;
 186         unsigned        fields;
 187 };
 188
 189 static int inode_set_size(struct bch_inode_info *inode,
 190                           struct bch_inode_unpacked *bi,
 191                           void *p)
 192 {
 193         struct inode_new_size *s = p;
 194
 195         bi->bi_size = s->new_size;
 196         if (s->fields & ATTR_ATIME)
 197                 bi->bi_atime = s->now;
 198         if (s->fields & ATTR_MTIME)
 199                 bi->bi_mtime = s->now;
 200         if (s->fields & ATTR_CTIME)
 201                 bi->bi_ctime = s->now;
 202
 203         return 0;
 204 }
 205
 206 int __must_check bch2_write_inode_size(struct bch_fs *c,
 207                                        struct bch_inode_info *inode,
 208                                        loff_t new_size, unsigned fields)
 209 {
 210         struct inode_new_size s = {
 211                 .new_size       = new_size,
 212                 .now            = bch2_current_time(c),
 213                 .fields         = fields,
 214         };
 215
 216         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 217 }
 218
 219 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 220                            struct quota_res *quota_res, s64 sectors)
 221 {
 222         if (!sectors)
 223                 return;
 224
 225         mutex_lock(&inode->ei_quota_lock);
 226 #ifdef CONFIG_BCACHEFS_QUOTA
 227         if (quota_res && sectors > 0) {
 228                 BUG_ON(sectors > quota_res->sectors);
 229                 BUG_ON(sectors > inode->ei_quota_reserved);
 230
 231                 quota_res->sectors -= sectors;
 232                 inode->ei_quota_reserved -= sectors;
 233         } else {
 234                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 235         }
 236 #endif
 237         inode->v.i_blocks += sectors;
 238         mutex_unlock(&inode->ei_quota_lock);
 239 }
 240
 241 /* page state: */
 242
 243 /* stored in page->private: */
 244
 245 struct bch_page_sector {
 246         /* Uncompressed, fully allocated replicas: */
 247         unsigned                nr_replicas:3;
 248
 249         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 250         unsigned                replicas_reserved:3;
 251
 252         /* i_sectors: */
 253         enum {
 254                 SECTOR_UNALLOCATED,
 255                 SECTOR_RESERVED,
 256                 SECTOR_DIRTY,
 257                 SECTOR_ALLOCATED,
 258         }                       state:2;
 259 };
 260
 261 struct bch_page_state {
 262         spinlock_t              lock;
 263         atomic_t                write_count;
 264         struct bch_page_sector  s[PAGE_SECTORS];
 265 };
 266
 267 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 268 {
 269         return page_has_private(page)
 270                 ? (struct bch_page_state *) page_private(page)
 271                 : NULL;
 272 }
 273
 274 static inline struct bch_page_state *bch2_page_state(struct page *page)
 275 {
 276         EBUG_ON(!PageLocked(page));
 277
 278         return __bch2_page_state(page);
 279 }
 280
 281 /* for newly allocated pages: */
 282 static void __bch2_page_state_release(struct page *page)
 283 {
 284         kfree(detach_page_private(page));
 285 }
 286
 287 static void bch2_page_state_release(struct page *page)
 288 {
 289         EBUG_ON(!PageLocked(page));
 290         __bch2_page_state_release(page);
 291 }
 292
 293 /* for newly allocated pages: */
 294 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 295                                                        gfp_t gfp)
 296 {
 297         struct bch_page_state *s;
 298
 299         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 300         if (!s)
 301                 return NULL;
 302
 303         spin_lock_init(&s->lock);
 304         attach_page_private(page, s);
 305         return s;
 306 }
 307
 308 static struct bch_page_state *bch2_page_state_create(struct page *page,
 309                                                      gfp_t gfp)
 310 {
 311         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 312 }
 313
 314 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 315 {
 316         /* XXX: this should not be open coded */
 317         return inode->ei_inode.bi_data_replicas
 318                 ? inode->ei_inode.bi_data_replicas - 1
 319                 : c->opts.data_replicas;
 320 }
 321
 322 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 323                                                   unsigned nr_replicas)
 324 {
 325         return max(0, (int) nr_replicas -
 326                    s->nr_replicas -
 327                    s->replicas_reserved);
 328 }
 329
 330 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 331                                 struct bch_inode_info *inode,
 332                                 struct page *page, bool check_enospc)
 333 {
 334         struct bch_page_state *s = bch2_page_state_create(page, 0);
 335         unsigned nr_replicas = inode_nr_replicas(c, inode);
 336         struct disk_reservation disk_res = { 0 };
 337         unsigned i, disk_res_sectors = 0;
 338         int ret;
 339
 340         if (!s)
 341                 return -ENOMEM;
 342
 343         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 344                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 345
 346         if (!disk_res_sectors)
 347                 return 0;
 348
 349         ret = bch2_disk_reservation_get(c, &disk_res,
 350                                         disk_res_sectors, 1,
 351                                         !check_enospc
 352                                         ? BCH_DISK_RESERVATION_NOFAIL
 353                                         : 0);
 354         if (unlikely(ret))
 355                 return ret;
 356
 357         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 358                 s->s[i].replicas_reserved +=
 359                         sectors_to_reserve(&s->s[i], nr_replicas);
 360
 361         return 0;
 362 }
 363
 364 struct bch2_page_reservation {
 365         struct disk_reservation disk;
 366         struct quota_res        quota;
 367 };
 368
 369 static void bch2_page_reservation_init(struct bch_fs *c,
 370                         struct bch_inode_info *inode,
 371                         struct bch2_page_reservation *res)
 372 {
 373         memset(res, 0, sizeof(*res));
 374
 375         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 376 }
 377
 378 static void bch2_page_reservation_put(struct bch_fs *c,
 379                         struct bch_inode_info *inode,
 380                         struct bch2_page_reservation *res)
 381 {
 382         bch2_disk_reservation_put(c, &res->disk);
 383         bch2_quota_reservation_put(c, inode, &res->quota);
 384 }
 385
 386 static int bch2_page_reservation_get(struct bch_fs *c,
 387                         struct bch_inode_info *inode, struct page *page,
 388                         struct bch2_page_reservation *res,
 389                         unsigned offset, unsigned len, bool check_enospc)
 390 {
 391         struct bch_page_state *s = bch2_page_state_create(page, 0);
 392         unsigned i, disk_sectors = 0, quota_sectors = 0;
 393         int ret;
 394
 395         if (!s)
 396                 return -ENOMEM;
 397
 398         for (i = round_down(offset, block_bytes(c)) >> 9;
 399              i < round_up(offset + len, block_bytes(c)) >> 9;
 400              i++) {
 401                 disk_sectors += sectors_to_reserve(&s->s[i],
 402                                                 res->disk.nr_replicas);
 403                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 404         }
 405
 406         if (disk_sectors) {
 407                 ret = bch2_disk_reservation_add(c, &res->disk,
 408                                                 disk_sectors,
 409                                                 !check_enospc
 410                                                 ? BCH_DISK_RESERVATION_NOFAIL
 411                                                 : 0);
 412                 if (unlikely(ret))
 413                         return ret;
 414         }
 415
 416         if (quota_sectors) {
 417                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 418                                                  quota_sectors,
 419                                                  check_enospc);
 420                 if (unlikely(ret)) {
 421                         struct disk_reservation tmp = {
 422                                 .sectors = disk_sectors
 423                         };
 424
 425                         bch2_disk_reservation_put(c, &tmp);
 426                         res->disk.sectors -= disk_sectors;
 427                         return ret;
 428                 }
 429         }
 430
 431         return 0;
 432 }
 433
 434 static void bch2_clear_page_bits(struct page *page)
 435 {
 436         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 437         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 438         struct bch_page_state *s = bch2_page_state(page);
 439         struct disk_reservation disk_res = { 0 };
 440         int i, dirty_sectors = 0;
 441
 442         if (!s)
 443                 return;
 444
 445         EBUG_ON(!PageLocked(page));
 446         EBUG_ON(PageWriteback(page));
 447
 448         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 449                 disk_res.sectors += s->s[i].replicas_reserved;
 450                 s->s[i].replicas_reserved = 0;
 451
 452                 if (s->s[i].state == SECTOR_DIRTY) {
 453                         dirty_sectors++;
 454                         s->s[i].state = SECTOR_UNALLOCATED;
 455                 }
 456         }
 457
 458         bch2_disk_reservation_put(c, &disk_res);
 459
 460         if (dirty_sectors)
 461                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 462
 463         bch2_page_state_release(page);
 464 }
 465
 466 static void bch2_set_page_dirty(struct bch_fs *c,
 467                         struct bch_inode_info *inode, struct page *page,
 468                         struct bch2_page_reservation *res,
 469                         unsigned offset, unsigned len)
 470 {
 471         struct bch_page_state *s = bch2_page_state(page);
 472         unsigned i, dirty_sectors = 0;
 473
 474         WARN_ON((u64) page_offset(page) + offset + len >
 475                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 476
 477         spin_lock(&s->lock);
 478
 479         for (i = round_down(offset, block_bytes(c)) >> 9;
 480              i < round_up(offset + len, block_bytes(c)) >> 9;
 481              i++) {
 482                 unsigned sectors = sectors_to_reserve(&s->s[i],
 483                                                 res->disk.nr_replicas);
 484
 485                 /*
 486                  * This can happen if we race with the error path in
 487                  * bch2_writepage_io_done():
 488                  */
 489                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 490
 491                 s->s[i].replicas_reserved += sectors;
 492                 res->disk.sectors -= sectors;
 493
 494                 if (s->s[i].state == SECTOR_UNALLOCATED)
 495                         dirty_sectors++;
 496
 497                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 498         }
 499
 500         spin_unlock(&s->lock);
 501
 502         if (dirty_sectors)
 503                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 504
 505         if (!PageDirty(page))
 506                 __set_page_dirty_nobuffers(page);
 507 }
 508
 509 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 510 {
 511         struct file *file = vmf->vma->vm_file;
 512         struct address_space *mapping = file->f_mapping;
 513         struct address_space *fdm = faults_disabled_mapping();
 514         struct bch_inode_info *inode = file_bch_inode(file);
 515         int ret;
 516
 517         if (fdm == mapping)
 518                 return VM_FAULT_SIGBUS;
 519
 520         /* Lock ordering: */
 521         if (fdm > mapping) {
 522                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 523
 524                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 525                         goto got_lock;
 526
 527                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 528
 529                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 530                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 531
 532                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 533
 534                 /* Signal that lock has been dropped: */
 535                 set_fdm_dropped_locks();
 536                 return VM_FAULT_SIGBUS;
 537         }
 538
 539         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 540 got_lock:
 541         ret = filemap_fault(vmf);
 542         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 543
 544         return ret;
 545 }
 546
 547 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 548 {
 549         struct page *page = vmf->page;
 550         struct file *file = vmf->vma->vm_file;
 551         struct bch_inode_info *inode = file_bch_inode(file);
 552         struct address_space *mapping = file->f_mapping;
 553         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 554         struct bch2_page_reservation res;
 555         unsigned len;
 556         loff_t isize;
 557         int ret = VM_FAULT_LOCKED;
 558
 559         bch2_page_reservation_init(c, inode, &res);
 560
 561         sb_start_pagefault(inode->v.i_sb);
 562         file_update_time(file);
 563
 564         /*
 565          * Not strictly necessary, but helps avoid dio writes livelocking in
 566          * write_invalidate_inode_pages_range() - can drop this if/when we get
 567          * a write_invalidate_inode_pages_range() that works without dropping
 568          * page lock before invalidating page
 569          */
 570         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 571
 572         lock_page(page);
 573         isize = i_size_read(&inode->v);
 574
 575         if (page->mapping != mapping || page_offset(page) >= isize) {
 576                 unlock_page(page);
 577                 ret = VM_FAULT_NOPAGE;
 578                 goto out;
 579         }
 580
 581         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 582
 583         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 584                 unlock_page(page);
 585                 ret = VM_FAULT_SIGBUS;
 586                 goto out;
 587         }
 588
 589         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 590         bch2_page_reservation_put(c, inode, &res);
 591
 592         wait_for_stable_page(page);
 593 out:
 594         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 595         sb_end_pagefault(inode->v.i_sb);
 596
 597         return ret;
 598 }
 599
 600 void bch2_invalidatepage(struct page *page, unsigned int offset,
 601                          unsigned int length)
 602 {
 603         if (offset || length < PAGE_SIZE)
 604                 return;
 605
 606         bch2_clear_page_bits(page);
 607 }
 608
 609 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 610 {
 611         if (PageDirty(page))
 612                 return 0;
 613
 614         bch2_clear_page_bits(page);
 615         return 1;
 616 }
 617
 618 #ifdef CONFIG_MIGRATION
 619 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 620                       struct page *page, enum migrate_mode mode)
 621 {
 622         int ret;
 623
 624         EBUG_ON(!PageLocked(page));
 625         EBUG_ON(!PageLocked(newpage));
 626
 627         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 628         if (ret != MIGRATEPAGE_SUCCESS)
 629                 return ret;
 630
 631         if (PagePrivate(page))
 632                 attach_page_private(newpage, detach_page_private(page));
 633
 634         if (mode != MIGRATE_SYNC_NO_COPY)
 635                 migrate_page_copy(newpage, page);
 636         else
 637                 migrate_page_states(newpage, page);
 638         return MIGRATEPAGE_SUCCESS;
 639 }
 640 #endif
 641
 642 /* readpage(s): */
 643
 644 static void bch2_readpages_end_io(struct bio *bio)
 645 {
 646         struct bvec_iter_all iter;
 647         struct bio_vec *bv;
 648
 649         bio_for_each_segment_all(bv, bio, iter) {
 650                 struct page *page = bv->bv_page;
 651
 652                 if (!bio->bi_status) {
 653                         SetPageUptodate(page);
 654                 } else {
 655                         ClearPageUptodate(page);
 656                         SetPageError(page);
 657                 }
 658                 unlock_page(page);
 659         }
 660
 661         bio_put(bio);
 662 }
 663
 664 struct readpages_iter {
 665         struct address_space    *mapping;
 666         struct page             **pages;
 667         unsigned                nr_pages;
 668         unsigned                idx;
 669         pgoff_t                 offset;
 670 };
 671
 672 static int readpages_iter_init(struct readpages_iter *iter,
 673                                struct readahead_control *ractl)
 674 {
 675         unsigned i, nr_pages = readahead_count(ractl);
 676
 677         memset(iter, 0, sizeof(*iter));
 678
 679         iter->mapping   = ractl->mapping;
 680         iter->offset    = readahead_index(ractl);
 681         iter->nr_pages  = nr_pages;
 682
 683         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 684         if (!iter->pages)
 685                 return -ENOMEM;
 686
 687         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 688         for (i = 0; i < nr_pages; i++) {
 689                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 690                 put_page(iter->pages[i]);
 691         }
 692
 693         return 0;
 694 }
 695
 696 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 697 {
 698         if (iter->idx >= iter->nr_pages)
 699                 return NULL;
 700
 701         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 702
 703         return iter->pages[iter->idx];
 704 }
 705
 706 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 707 {
 708         struct bvec_iter iter;
 709         struct bio_vec bv;
 710         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 711                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 712         unsigned state = k.k->type == KEY_TYPE_reservation
 713                 ? SECTOR_RESERVED
 714                 : SECTOR_ALLOCATED;
 715
 716         bio_for_each_segment(bv, bio, iter) {
 717                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 718                 unsigned i;
 719
 720                 for (i = bv.bv_offset >> 9;
 721                      i < (bv.bv_offset + bv.bv_len) >> 9;
 722                      i++) {
 723                         s->s[i].nr_replicas = nr_ptrs;
 724                         s->s[i].state = state;
 725                 }
 726         }
 727 }
 728
 729 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 730 {
 731         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 732         struct bch_extent_crc_unpacked crc;
 733         const union bch_extent_entry *i;
 734
 735         bkey_for_each_crc(k.k, ptrs, crc, i)
 736                 if (crc.csum_type || crc.compression_type)
 737                         return true;
 738         return false;
 739 }
 740
 741 static void readpage_bio_extend(struct readpages_iter *iter,
 742                                 struct bio *bio,
 743                                 unsigned sectors_this_extent,
 744                                 bool get_more)
 745 {
 746         while (bio_sectors(bio) < sectors_this_extent &&
 747                bio->bi_vcnt < bio->bi_max_vecs) {
 748                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 749                 struct page *page = readpage_iter_next(iter);
 750                 int ret;
 751
 752                 if (page) {
 753                         if (iter->offset + iter->idx != page_offset)
 754                                 break;
 755
 756                         iter->idx++;
 757                 } else {
 758                         if (!get_more)
 759                                 break;
 760
 761                         page = xa_load(&iter->mapping->i_pages, page_offset);
 762                         if (page && !xa_is_value(page))
 763                                 break;
 764
 765                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 766                         if (!page)
 767                                 break;
 768
 769                         if (!__bch2_page_state_create(page, 0)) {
 770                                 put_page(page);
 771                                 break;
 772                         }
 773
 774                         ret = add_to_page_cache_lru(page, iter->mapping,
 775                                                     page_offset, GFP_NOFS);
 776                         if (ret) {
 777                                 __bch2_page_state_release(page);
 778                                 put_page(page);
 779                                 break;
 780                         }
 781
 782                         put_page(page);
 783                 }
 784
 785                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 786         }
 787 }
 788
 789 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 790                        struct bch_read_bio *rbio, u64 inum,
 791                        struct readpages_iter *readpages_iter)
 792 {
 793         struct bch_fs *c = trans->c;
 794         struct bkey_buf sk;
 795         int flags = BCH_READ_RETRY_IF_STALE|
 796                 BCH_READ_MAY_PROMOTE;
 797         int ret = 0;
 798
 799         rbio->c = c;
 800         rbio->start_time = local_clock();
 801
 802         bch2_bkey_buf_init(&sk);
 803 retry:
 804         bch2_trans_begin(trans);
 805
 806         while (1) {
 807                 struct bkey_s_c k;
 808                 unsigned bytes, sectors, offset_into_extent;
 809                 enum btree_id data_btree = BTREE_ID_extents;
 810
 811                 /*
 812                  * read_extent -> io_time_reset may cause a transaction restart
 813                  * without returning an error, we need to check for that here:
 814                  */
 815                 if (!bch2_trans_relock(trans)) {
 816                         ret = -EINTR;
 817                         break;
 818                 }
 819
 820                 bch2_btree_iter_set_pos(iter,
 821                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 822
 823                 k = bch2_btree_iter_peek_slot(iter);
 824                 ret = bkey_err(k);
 825                 if (ret)
 826                         break;
 827
 828                 offset_into_extent = iter->pos.offset -
 829                         bkey_start_offset(k.k);
 830                 sectors = k.k->size - offset_into_extent;
 831
 832                 bch2_bkey_buf_reassemble(&sk, c, k);
 833
 834                 ret = bch2_read_indirect_extent(trans, &data_btree,
 835                                         &offset_into_extent, &sk);
 836                 if (ret)
 837                         break;
 838
 839                 k = bkey_i_to_s_c(sk.k);
 840
 841                 sectors = min(sectors, k.k->size - offset_into_extent);
 842
 843                 bch2_trans_unlock(trans);
 844
 845                 if (readpages_iter)
 846                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 847                                             extent_partial_reads_expensive(k));
 848
 849                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 850                 swap(rbio->bio.bi_iter.bi_size, bytes);
 851
 852                 if (rbio->bio.bi_iter.bi_size == bytes)
 853                         flags |= BCH_READ_LAST_FRAGMENT;
 854
 855                 if (bkey_extent_is_allocation(k.k))
 856                         bch2_add_page_sectors(&rbio->bio, k);
 857
 858                 bch2_read_extent(trans, rbio, iter->pos,
 859                                  data_btree, k, offset_into_extent, flags);
 860
 861                 if (flags & BCH_READ_LAST_FRAGMENT)
 862                         break;
 863
 864                 swap(rbio->bio.bi_iter.bi_size, bytes);
 865                 bio_advance(&rbio->bio, bytes);
 866         }
 867
 868         if (ret == -EINTR)
 869                 goto retry;
 870
 871         if (ret) {
 872                 bch_err_inum_ratelimited(c, inum,
 873                                 "read error %i from btree lookup", ret);
 874                 rbio->bio.bi_status = BLK_STS_IOERR;
 875                 bio_endio(&rbio->bio);
 876         }
 877
 878         bch2_bkey_buf_exit(&sk, c);
 879 }
 880
 881 void bch2_readahead(struct readahead_control *ractl)
 882 {
 883         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 884         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 885         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 886         struct btree_trans trans;
 887         struct btree_iter iter;
 888         struct page *page;
 889         struct readpages_iter readpages_iter;
 890         int ret;
 891
 892         ret = readpages_iter_init(&readpages_iter, ractl);
 893         BUG_ON(ret);
 894
 895         bch2_trans_init(&trans, c, 0, 0);
 896         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, POS_MIN,
 897                              BTREE_ITER_SLOTS);
 898
 899         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 900
 901         while ((page = readpage_iter_next(&readpages_iter))) {
 902                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 903                 unsigned n = min_t(unsigned,
 904                                    readpages_iter.nr_pages -
 905                                    readpages_iter.idx,
 906                                    BIO_MAX_VECS);
 907                 struct bch_read_bio *rbio =
 908                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 909                                   opts);
 910
 911                 readpages_iter.idx++;
 912
 913                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 914                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 915                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 916                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 917
 918                 bchfs_read(&trans, &iter, rbio, inode->v.i_ino,
 919                            &readpages_iter);
 920         }
 921
 922         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 923
 924         bch2_trans_iter_exit(&trans, &iter);
 925         bch2_trans_exit(&trans);
 926         kfree(readpages_iter.pages);
 927 }
 928
 929 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 930                              u64 inum, struct page *page)
 931 {
 932         struct btree_trans trans;
 933         struct btree_iter iter;
 934
 935         bch2_page_state_create(page, __GFP_NOFAIL);
 936
 937         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 938         rbio->bio.bi_iter.bi_sector =
 939                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 940         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 941
 942         bch2_trans_init(&trans, c, 0, 0);
 943         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, POS_MIN,
 944                              BTREE_ITER_SLOTS);
 945
 946         bchfs_read(&trans, &iter, rbio, inum, NULL);
 947
 948         bch2_trans_iter_exit(&trans, &iter);
 949         bch2_trans_exit(&trans);
 950 }
 951
 952 int bch2_readpage(struct file *file, struct page *page)
 953 {
 954         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 955         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 956         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 957         struct bch_read_bio *rbio;
 958
 959         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 960         rbio->bio.bi_end_io = bch2_readpages_end_io;
 961
 962         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 963         return 0;
 964 }
 965
 966 static void bch2_read_single_page_end_io(struct bio *bio)
 967 {
 968         complete(bio->bi_private);
 969 }
 970
 971 static int bch2_read_single_page(struct page *page,
 972                                  struct address_space *mapping)
 973 {
 974         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 975         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 976         struct bch_read_bio *rbio;
 977         int ret;
 978         DECLARE_COMPLETION_ONSTACK(done);
 979
 980         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 981                          io_opts(c, &inode->ei_inode));
 982         rbio->bio.bi_private = &done;
 983         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 984
 985         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 986         wait_for_completion(&done);
 987
 988         ret = blk_status_to_errno(rbio->bio.bi_status);
 989         bio_put(&rbio->bio);
 990
 991         if (ret < 0)
 992                 return ret;
 993
 994         SetPageUptodate(page);
 995         return 0;
 996 }
 997
 998 /* writepages: */
 999
1000 struct bch_writepage_state {
1001         struct bch_writepage_io *io;
1002         struct bch_io_opts      opts;
1003 };
1004
1005 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1006                                                                   struct bch_inode_info *inode)
1007 {
1008         return (struct bch_writepage_state) {
1009                 .opts = io_opts(c, &inode->ei_inode)
1010         };
1011 }
1012
1013 static void bch2_writepage_io_free(struct closure *cl)
1014 {
1015         struct bch_writepage_io *io = container_of(cl,
1016                                         struct bch_writepage_io, cl);
1017
1018         bio_put(&io->op.wbio.bio);
1019 }
1020
1021 static void bch2_writepage_io_done(struct closure *cl)
1022 {
1023         struct bch_writepage_io *io = container_of(cl,
1024                                         struct bch_writepage_io, cl);
1025         struct bch_fs *c = io->op.c;
1026         struct bio *bio = &io->op.wbio.bio;
1027         struct bvec_iter_all iter;
1028         struct bio_vec *bvec;
1029         unsigned i;
1030
1031         up(&io->op.c->io_in_flight);
1032
1033         if (io->op.error) {
1034                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1035
1036                 bio_for_each_segment_all(bvec, bio, iter) {
1037                         struct bch_page_state *s;
1038
1039                         SetPageError(bvec->bv_page);
1040                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1041
1042                         s = __bch2_page_state(bvec->bv_page);
1043                         spin_lock(&s->lock);
1044                         for (i = 0; i < PAGE_SECTORS; i++)
1045                                 s->s[i].nr_replicas = 0;
1046                         spin_unlock(&s->lock);
1047                 }
1048         }
1049
1050         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1051                 bio_for_each_segment_all(bvec, bio, iter) {
1052                         struct bch_page_state *s;
1053
1054                         s = __bch2_page_state(bvec->bv_page);
1055                         spin_lock(&s->lock);
1056                         for (i = 0; i < PAGE_SECTORS; i++)
1057                                 s->s[i].nr_replicas = 0;
1058                         spin_unlock(&s->lock);
1059                 }
1060         }
1061
1062         /*
1063          * racing with fallocate can cause us to add fewer sectors than
1064          * expected - but we shouldn't add more sectors than expected:
1065          */
1066         BUG_ON(io->op.i_sectors_delta > 0);
1067
1068         /*
1069          * (error (due to going RO) halfway through a page can screw that up
1070          * slightly)
1071          * XXX wtf?
1072            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1073          */
1074
1075         /*
1076          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1077          * before calling end_page_writeback:
1078          */
1079         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1080
1081         bio_for_each_segment_all(bvec, bio, iter) {
1082                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1083
1084                 if (atomic_dec_and_test(&s->write_count))
1085                         end_page_writeback(bvec->bv_page);
1086         }
1087
1088         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1089 }
1090
1091 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1092 {
1093         struct bch_writepage_io *io = w->io;
1094
1095         down(&io->op.c->io_in_flight);
1096
1097         w->io = NULL;
1098         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1099         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1100 }
1101
1102 /*
1103  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1104  * possible, else allocating a new one:
1105  */
1106 static void bch2_writepage_io_alloc(struct bch_fs *c,
1107                                     struct writeback_control *wbc,
1108                                     struct bch_writepage_state *w,
1109                                     struct bch_inode_info *inode,
1110                                     u64 sector,
1111                                     unsigned nr_replicas)
1112 {
1113         struct bch_write_op *op;
1114
1115         w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1116                                               &c->writepage_bioset),
1117                              struct bch_writepage_io, op.wbio.bio);
1118
1119         closure_init(&w->io->cl, NULL);
1120         w->io->inode            = inode;
1121
1122         op                      = &w->io->op;
1123         bch2_write_op_init(op, c, w->opts);
1124         op->target              = w->opts.foreground_target;
1125         op_journal_seq_set(op, &inode->ei_journal_seq);
1126         op->nr_replicas         = nr_replicas;
1127         op->res.nr_replicas     = nr_replicas;
1128         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1129         op->pos                 = POS(inode->v.i_ino, sector);
1130         op->wbio.bio.bi_iter.bi_sector = sector;
1131         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1132 }
1133
1134 static int __bch2_writepage(struct page *page,
1135                             struct writeback_control *wbc,
1136                             void *data)
1137 {
1138         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1139         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1140         struct bch_writepage_state *w = data;
1141         struct bch_page_state *s, orig;
1142         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1143         loff_t i_size = i_size_read(&inode->v);
1144         pgoff_t end_index = i_size >> PAGE_SHIFT;
1145         int ret;
1146
1147         EBUG_ON(!PageUptodate(page));
1148
1149         /* Is the page fully inside i_size? */
1150         if (page->index < end_index)
1151                 goto do_io;
1152
1153         /* Is the page fully outside i_size? (truncate in progress) */
1154         offset = i_size & (PAGE_SIZE - 1);
1155         if (page->index > end_index || !offset) {
1156                 unlock_page(page);
1157                 return 0;
1158         }
1159
1160         /*
1161          * The page straddles i_size.  It must be zeroed out on each and every
1162          * writepage invocation because it may be mmapped.  "A file is mapped
1163          * in multiples of the page size.  For a file that is not a multiple of
1164          * the  page size, the remaining memory is zeroed when mapped, and
1165          * writes to that region are not written out to the file."
1166          */
1167         zero_user_segment(page, offset, PAGE_SIZE);
1168 do_io:
1169         s = bch2_page_state_create(page, __GFP_NOFAIL);
1170
1171         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1172         if (ret) {
1173                 SetPageError(page);
1174                 mapping_set_error(page->mapping, ret);
1175                 unlock_page(page);
1176                 return 0;
1177         }
1178
1179         /* Before unlocking the page, get copy of reservations: */
1180         orig = *s;
1181
1182         for (i = 0; i < PAGE_SECTORS; i++) {
1183                 if (s->s[i].state < SECTOR_DIRTY)
1184                         continue;
1185
1186                 nr_replicas_this_write =
1187                         min_t(unsigned, nr_replicas_this_write,
1188                               s->s[i].nr_replicas +
1189                               s->s[i].replicas_reserved);
1190         }
1191
1192         for (i = 0; i < PAGE_SECTORS; i++) {
1193                 if (s->s[i].state < SECTOR_DIRTY)
1194                         continue;
1195
1196                 s->s[i].nr_replicas = w->opts.compression
1197                         ? 0 : nr_replicas_this_write;
1198
1199                 s->s[i].replicas_reserved = 0;
1200                 s->s[i].state = SECTOR_ALLOCATED;
1201         }
1202
1203         BUG_ON(atomic_read(&s->write_count));
1204         atomic_set(&s->write_count, 1);
1205
1206         BUG_ON(PageWriteback(page));
1207         set_page_writeback(page);
1208
1209         unlock_page(page);
1210
1211         offset = 0;
1212         while (1) {
1213                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1214                 u64 sector;
1215
1216                 while (offset < PAGE_SECTORS &&
1217                        orig.s[offset].state < SECTOR_DIRTY)
1218                         offset++;
1219
1220                 if (offset == PAGE_SECTORS)
1221                         break;
1222
1223                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1224
1225                 while (offset + sectors < PAGE_SECTORS &&
1226                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1227                         sectors++;
1228
1229                 for (i = offset; i < offset + sectors; i++) {
1230                         reserved_sectors += orig.s[i].replicas_reserved;
1231                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1232                 }
1233
1234                 if (w->io &&
1235                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1236                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1237                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1238                      (BIO_MAX_VECS * PAGE_SIZE) ||
1239                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1240                         bch2_writepage_do_io(w);
1241
1242                 if (!w->io)
1243                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1244                                                 nr_replicas_this_write);
1245
1246                 atomic_inc(&s->write_count);
1247
1248                 BUG_ON(inode != w->io->inode);
1249                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1250                                      sectors << 9, offset << 9));
1251
1252                 /* Check for writing past i_size: */
1253                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1254                         round_up(i_size, block_bytes(c)));
1255
1256                 w->io->op.res.sectors += reserved_sectors;
1257                 w->io->op.i_sectors_delta -= dirty_sectors;
1258                 w->io->op.new_i_size = i_size;
1259
1260                 offset += sectors;
1261         }
1262
1263         if (atomic_dec_and_test(&s->write_count))
1264                 end_page_writeback(page);
1265
1266         return 0;
1267 }
1268
1269 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1270 {
1271         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1272         struct bch_writepage_state w =
1273                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1274         struct blk_plug plug;
1275         int ret;
1276
1277         blk_start_plug(&plug);
1278         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1279         if (w.io)
1280                 bch2_writepage_do_io(&w);
1281         blk_finish_plug(&plug);
1282         return ret;
1283 }
1284
1285 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1286 {
1287         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1288         struct bch_writepage_state w =
1289                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1290         int ret;
1291
1292         ret = __bch2_writepage(page, wbc, &w);
1293         if (w.io)
1294                 bch2_writepage_do_io(&w);
1295
1296         return ret;
1297 }
1298
1299 /* buffered writes: */
1300
1301 int bch2_write_begin(struct file *file, struct address_space *mapping,
1302                      loff_t pos, unsigned len, unsigned flags,
1303                      struct page **pagep, void **fsdata)
1304 {
1305         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1306         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1307         struct bch2_page_reservation *res;
1308         pgoff_t index = pos >> PAGE_SHIFT;
1309         unsigned offset = pos & (PAGE_SIZE - 1);
1310         struct page *page;
1311         int ret = -ENOMEM;
1312
1313         res = kmalloc(sizeof(*res), GFP_KERNEL);
1314         if (!res)
1315                 return -ENOMEM;
1316
1317         bch2_page_reservation_init(c, inode, res);
1318         *fsdata = res;
1319
1320         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1321
1322         page = grab_cache_page_write_begin(mapping, index, flags);
1323         if (!page)
1324                 goto err_unlock;
1325
1326         if (PageUptodate(page))
1327                 goto out;
1328
1329         /* If we're writing entire page, don't need to read it in first: */
1330         if (len == PAGE_SIZE)
1331                 goto out;
1332
1333         if (!offset && pos + len >= inode->v.i_size) {
1334                 zero_user_segment(page, len, PAGE_SIZE);
1335                 flush_dcache_page(page);
1336                 goto out;
1337         }
1338
1339         if (index > inode->v.i_size >> PAGE_SHIFT) {
1340                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1341                 flush_dcache_page(page);
1342                 goto out;
1343         }
1344 readpage:
1345         ret = bch2_read_single_page(page, mapping);
1346         if (ret)
1347                 goto err;
1348 out:
1349         ret = bch2_page_reservation_get(c, inode, page, res,
1350                                         offset, len, true);
1351         if (ret) {
1352                 if (!PageUptodate(page)) {
1353                         /*
1354                          * If the page hasn't been read in, we won't know if we
1355                          * actually need a reservation - we don't actually need
1356                          * to read here, we just need to check if the page is
1357                          * fully backed by uncompressed data:
1358                          */
1359                         goto readpage;
1360                 }
1361
1362                 goto err;
1363         }
1364
1365         *pagep = page;
1366         return 0;
1367 err:
1368         unlock_page(page);
1369         put_page(page);
1370         *pagep = NULL;
1371 err_unlock:
1372         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1373         kfree(res);
1374         *fsdata = NULL;
1375         return ret;
1376 }
1377
1378 int bch2_write_end(struct file *file, struct address_space *mapping,
1379                    loff_t pos, unsigned len, unsigned copied,
1380                    struct page *page, void *fsdata)
1381 {
1382         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1383         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1384         struct bch2_page_reservation *res = fsdata;
1385         unsigned offset = pos & (PAGE_SIZE - 1);
1386
1387         lockdep_assert_held(&inode->v.i_rwsem);
1388
1389         if (unlikely(copied < len && !PageUptodate(page))) {
1390                 /*
1391                  * The page needs to be read in, but that would destroy
1392                  * our partial write - simplest thing is to just force
1393                  * userspace to redo the write:
1394                  */
1395                 zero_user(page, 0, PAGE_SIZE);
1396                 flush_dcache_page(page);
1397                 copied = 0;
1398         }
1399
1400         spin_lock(&inode->v.i_lock);
1401         if (pos + copied > inode->v.i_size)
1402                 i_size_write(&inode->v, pos + copied);
1403         spin_unlock(&inode->v.i_lock);
1404
1405         if (copied) {
1406                 if (!PageUptodate(page))
1407                         SetPageUptodate(page);
1408
1409                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1410
1411                 inode->ei_last_dirtied = (unsigned long) current;
1412         }
1413
1414         unlock_page(page);
1415         put_page(page);
1416         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1417
1418         bch2_page_reservation_put(c, inode, res);
1419         kfree(res);
1420
1421         return copied;
1422 }
1423
1424 #define WRITE_BATCH_PAGES       32
1425
1426 static int __bch2_buffered_write(struct bch_inode_info *inode,
1427                                  struct address_space *mapping,
1428                                  struct iov_iter *iter,
1429                                  loff_t pos, unsigned len)
1430 {
1431         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1432         struct page *pages[WRITE_BATCH_PAGES];
1433         struct bch2_page_reservation res;
1434         unsigned long index = pos >> PAGE_SHIFT;
1435         unsigned offset = pos & (PAGE_SIZE - 1);
1436         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1437         unsigned i, reserved = 0, set_dirty = 0;
1438         unsigned copied = 0, nr_pages_copied = 0;
1439         int ret = 0;
1440
1441         BUG_ON(!len);
1442         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1443
1444         bch2_page_reservation_init(c, inode, &res);
1445
1446         for (i = 0; i < nr_pages; i++) {
1447                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1448                 if (!pages[i]) {
1449                         nr_pages = i;
1450                         if (!i) {
1451                                 ret = -ENOMEM;
1452                                 goto out;
1453                         }
1454                         len = min_t(unsigned, len,
1455                                     nr_pages * PAGE_SIZE - offset);
1456                         break;
1457                 }
1458         }
1459
1460         if (offset && !PageUptodate(pages[0])) {
1461                 ret = bch2_read_single_page(pages[0], mapping);
1462                 if (ret)
1463                         goto out;
1464         }
1465
1466         if ((pos + len) & (PAGE_SIZE - 1) &&
1467             !PageUptodate(pages[nr_pages - 1])) {
1468                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1469                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1470                 } else {
1471                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1472                         if (ret)
1473                                 goto out;
1474                 }
1475         }
1476
1477         while (reserved < len) {
1478                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1479                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1480                 unsigned pg_len = min_t(unsigned, len - reserved,
1481                                         PAGE_SIZE - pg_offset);
1482 retry_reservation:
1483                 ret = bch2_page_reservation_get(c, inode, page, &res,
1484                                                 pg_offset, pg_len, true);
1485
1486                 if (ret && !PageUptodate(page)) {
1487                         ret = bch2_read_single_page(page, mapping);
1488                         if (!ret)
1489                                 goto retry_reservation;
1490                 }
1491
1492                 if (ret)
1493                         goto out;
1494
1495                 reserved += pg_len;
1496         }
1497
1498         if (mapping_writably_mapped(mapping))
1499                 for (i = 0; i < nr_pages; i++)
1500                         flush_dcache_page(pages[i]);
1501
1502         while (copied < len) {
1503                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1504                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1505                 unsigned pg_len = min_t(unsigned, len - copied,
1506                                         PAGE_SIZE - pg_offset);
1507                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1508                                                 iter, pg_offset, pg_len);
1509
1510                 if (!pg_copied)
1511                         break;
1512
1513                 if (!PageUptodate(page) &&
1514                     pg_copied != PAGE_SIZE &&
1515                     pos + copied + pg_copied < inode->v.i_size) {
1516                         zero_user(page, 0, PAGE_SIZE);
1517                         break;
1518                 }
1519
1520                 flush_dcache_page(page);
1521                 iov_iter_advance(iter, pg_copied);
1522                 copied += pg_copied;
1523
1524                 if (pg_copied != pg_len)
1525                         break;
1526         }
1527
1528         if (!copied)
1529                 goto out;
1530
1531         spin_lock(&inode->v.i_lock);
1532         if (pos + copied > inode->v.i_size)
1533                 i_size_write(&inode->v, pos + copied);
1534         spin_unlock(&inode->v.i_lock);
1535
1536         while (set_dirty < copied) {
1537                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1538                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1539                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1540                                         PAGE_SIZE - pg_offset);
1541
1542                 if (!PageUptodate(page))
1543                         SetPageUptodate(page);
1544
1545                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1546                 unlock_page(page);
1547                 put_page(page);
1548
1549                 set_dirty += pg_len;
1550         }
1551
1552         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1553         inode->ei_last_dirtied = (unsigned long) current;
1554 out:
1555         for (i = nr_pages_copied; i < nr_pages; i++) {
1556                 unlock_page(pages[i]);
1557                 put_page(pages[i]);
1558         }
1559
1560         bch2_page_reservation_put(c, inode, &res);
1561
1562         return copied ?: ret;
1563 }
1564
1565 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1566 {
1567         struct file *file = iocb->ki_filp;
1568         struct address_space *mapping = file->f_mapping;
1569         struct bch_inode_info *inode = file_bch_inode(file);
1570         loff_t pos = iocb->ki_pos;
1571         ssize_t written = 0;
1572         int ret = 0;
1573
1574         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1575
1576         do {
1577                 unsigned offset = pos & (PAGE_SIZE - 1);
1578                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1579                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1580 again:
1581                 /*
1582                  * Bring in the user page that we will copy from _first_.
1583                  * Otherwise there's a nasty deadlock on copying from the
1584                  * same page as we're writing to, without it being marked
1585                  * up-to-date.
1586                  *
1587                  * Not only is this an optimisation, but it is also required
1588                  * to check that the address is actually valid, when atomic
1589                  * usercopies are used, below.
1590                  */
1591                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1592                         bytes = min_t(unsigned long, iov_iter_count(iter),
1593                                       PAGE_SIZE - offset);
1594
1595                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1596                                 ret = -EFAULT;
1597                                 break;
1598                         }
1599                 }
1600
1601                 if (unlikely(fatal_signal_pending(current))) {
1602                         ret = -EINTR;
1603                         break;
1604                 }
1605
1606                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1607                 if (unlikely(ret < 0))
1608                         break;
1609
1610                 cond_resched();
1611
1612                 if (unlikely(ret == 0)) {
1613                         /*
1614                          * If we were unable to copy any data at all, we must
1615                          * fall back to a single segment length write.
1616                          *
1617                          * If we didn't fallback here, we could livelock
1618                          * because not all segments in the iov can be copied at
1619                          * once without a pagefault.
1620                          */
1621                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1622                                       iov_iter_single_seg_count(iter));
1623                         goto again;
1624                 }
1625                 pos += ret;
1626                 written += ret;
1627                 ret = 0;
1628
1629                 balance_dirty_pages_ratelimited(mapping);
1630         } while (iov_iter_count(iter));
1631
1632         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1633
1634         return written ? written : ret;
1635 }
1636
1637 /* O_DIRECT reads */
1638
1639 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1640 {
1641         if (check_dirty) {
1642                 bio_check_pages_dirty(bio);
1643         } else {
1644                 bio_release_pages(bio, false);
1645                 bio_put(bio);
1646         }
1647 }
1648
1649 static void bch2_dio_read_complete(struct closure *cl)
1650 {
1651         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1652
1653         dio->req->ki_complete(dio->req, dio->ret, 0);
1654         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1655 }
1656
1657 static void bch2_direct_IO_read_endio(struct bio *bio)
1658 {
1659         struct dio_read *dio = bio->bi_private;
1660
1661         if (bio->bi_status)
1662                 dio->ret = blk_status_to_errno(bio->bi_status);
1663
1664         closure_put(&dio->cl);
1665 }
1666
1667 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1668 {
1669         struct dio_read *dio = bio->bi_private;
1670         bool should_dirty = dio->should_dirty;
1671
1672         bch2_direct_IO_read_endio(bio);
1673         bio_check_or_release(bio, should_dirty);
1674 }
1675
1676 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1677 {
1678         struct file *file = req->ki_filp;
1679         struct bch_inode_info *inode = file_bch_inode(file);
1680         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1681         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1682         struct dio_read *dio;
1683         struct bio *bio;
1684         loff_t offset = req->ki_pos;
1685         bool sync = is_sync_kiocb(req);
1686         size_t shorten;
1687         ssize_t ret;
1688
1689         if ((offset|iter->count) & (block_bytes(c) - 1))
1690                 return -EINVAL;
1691
1692         ret = min_t(loff_t, iter->count,
1693                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1694
1695         if (!ret)
1696                 return ret;
1697
1698         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1699         iter->count -= shorten;
1700
1701         bio = bio_alloc_bioset(GFP_KERNEL,
1702                                iov_iter_npages(iter, BIO_MAX_VECS),
1703                                &c->dio_read_bioset);
1704
1705         bio->bi_end_io = bch2_direct_IO_read_endio;
1706
1707         dio = container_of(bio, struct dio_read, rbio.bio);
1708         closure_init(&dio->cl, NULL);
1709
1710         /*
1711          * this is a _really_ horrible hack just to avoid an atomic sub at the
1712          * end:
1713          */
1714         if (!sync) {
1715                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1716                 atomic_set(&dio->cl.remaining,
1717                            CLOSURE_REMAINING_INITIALIZER -
1718                            CLOSURE_RUNNING +
1719                            CLOSURE_DESTRUCTOR);
1720         } else {
1721                 atomic_set(&dio->cl.remaining,
1722                            CLOSURE_REMAINING_INITIALIZER + 1);
1723         }
1724
1725         dio->req        = req;
1726         dio->ret        = ret;
1727         /*
1728          * This is one of the sketchier things I've encountered: we have to skip
1729          * the dirtying of requests that are internal from the kernel (i.e. from
1730          * loopback), because we'll deadlock on page_lock.
1731          */
1732         dio->should_dirty = iter_is_iovec(iter);
1733
1734         goto start;
1735         while (iter->count) {
1736                 bio = bio_alloc_bioset(GFP_KERNEL,
1737                                        iov_iter_npages(iter, BIO_MAX_VECS),
1738                                        &c->bio_read);
1739                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1740 start:
1741                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1742                 bio->bi_iter.bi_sector  = offset >> 9;
1743                 bio->bi_private         = dio;
1744
1745                 ret = bio_iov_iter_get_pages(bio, iter);
1746                 if (ret < 0) {
1747                         /* XXX: fault inject this path */
1748                         bio->bi_status = BLK_STS_RESOURCE;
1749                         bio_endio(bio);
1750                         break;
1751                 }
1752
1753                 offset += bio->bi_iter.bi_size;
1754
1755                 if (dio->should_dirty)
1756                         bio_set_pages_dirty(bio);
1757
1758                 if (iter->count)
1759                         closure_get(&dio->cl);
1760
1761                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1762         }
1763
1764         iter->count += shorten;
1765
1766         if (sync) {
1767                 closure_sync(&dio->cl);
1768                 closure_debug_destroy(&dio->cl);
1769                 ret = dio->ret;
1770                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1771                 return ret;
1772         } else {
1773                 return -EIOCBQUEUED;
1774         }
1775 }
1776
1777 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1778 {
1779         struct file *file = iocb->ki_filp;
1780         struct bch_inode_info *inode = file_bch_inode(file);
1781         struct address_space *mapping = file->f_mapping;
1782         size_t count = iov_iter_count(iter);
1783         ssize_t ret;
1784
1785         if (!count)
1786                 return 0; /* skip atime */
1787
1788         if (iocb->ki_flags & IOCB_DIRECT) {
1789                 struct blk_plug plug;
1790
1791                 ret = filemap_write_and_wait_range(mapping,
1792                                         iocb->ki_pos,
1793                                         iocb->ki_pos + count - 1);
1794                 if (ret < 0)
1795                         return ret;
1796
1797                 file_accessed(file);
1798
1799                 blk_start_plug(&plug);
1800                 ret = bch2_direct_IO_read(iocb, iter);
1801                 blk_finish_plug(&plug);
1802
1803                 if (ret >= 0)
1804                         iocb->ki_pos += ret;
1805         } else {
1806                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1807                 ret = generic_file_read_iter(iocb, iter);
1808                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1809         }
1810
1811         return ret;
1812 }
1813
1814 /* O_DIRECT writes */
1815
1816 static void bch2_dio_write_loop_async(struct bch_write_op *);
1817
1818 static long bch2_dio_write_loop(struct dio_write *dio)
1819 {
1820         bool kthread = (current->flags & PF_KTHREAD) != 0;
1821         struct kiocb *req = dio->req;
1822         struct address_space *mapping = req->ki_filp->f_mapping;
1823         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1824         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1825         struct bio *bio = &dio->op.wbio.bio;
1826         struct bvec_iter_all iter;
1827         struct bio_vec *bv;
1828         unsigned unaligned, iter_count;
1829         bool sync = dio->sync, dropped_locks;
1830         long ret;
1831
1832         if (dio->loop)
1833                 goto loop;
1834
1835         down(&c->io_in_flight);
1836
1837         while (1) {
1838                 iter_count = dio->iter.count;
1839
1840                 if (kthread)
1841                         kthread_use_mm(dio->mm);
1842                 BUG_ON(current->faults_disabled_mapping);
1843                 current->faults_disabled_mapping = mapping;
1844
1845                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1846
1847                 dropped_locks = fdm_dropped_locks();
1848
1849                 current->faults_disabled_mapping = NULL;
1850                 if (kthread)
1851                         kthread_unuse_mm(dio->mm);
1852
1853                 /*
1854                  * If the fault handler returned an error but also signalled
1855                  * that it dropped & retook ei_pagecache_lock, we just need to
1856                  * re-shoot down the page cache and retry:
1857                  */
1858                 if (dropped_locks && ret)
1859                         ret = 0;
1860
1861                 if (unlikely(ret < 0))
1862                         goto err;
1863
1864                 if (unlikely(dropped_locks)) {
1865                         ret = write_invalidate_inode_pages_range(mapping,
1866                                         req->ki_pos,
1867                                         req->ki_pos + iter_count - 1);
1868                         if (unlikely(ret))
1869                                 goto err;
1870
1871                         if (!bio->bi_iter.bi_size)
1872                                 continue;
1873                 }
1874
1875                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1876                 bio->bi_iter.bi_size -= unaligned;
1877                 iov_iter_revert(&dio->iter, unaligned);
1878
1879                 if (!bio->bi_iter.bi_size) {
1880                         /*
1881                          * bio_iov_iter_get_pages was only able to get <
1882                          * blocksize worth of pages:
1883                          */
1884                         ret = -EFAULT;
1885                         goto err;
1886                 }
1887
1888                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1889                 dio->op.end_io          = bch2_dio_write_loop_async;
1890                 dio->op.target          = dio->op.opts.foreground_target;
1891                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1892                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1893                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1894                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1895
1896                 if ((req->ki_flags & IOCB_DSYNC) &&
1897                     !c->opts.journal_flush_disabled)
1898                         dio->op.flags |= BCH_WRITE_FLUSH;
1899                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
1900
1901                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1902                                                 dio->op.opts.data_replicas, 0);
1903                 if (unlikely(ret) &&
1904                     !bch2_check_range_allocated(c, dio->op.pos,
1905                                 bio_sectors(bio),
1906                                 dio->op.opts.data_replicas,
1907                                 dio->op.opts.compression != 0))
1908                         goto err;
1909
1910                 task_io_account_write(bio->bi_iter.bi_size);
1911
1912                 if (!dio->sync && !dio->loop && dio->iter.count) {
1913                         struct iovec *iov = dio->inline_vecs;
1914
1915                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1916                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1917                                               GFP_KERNEL);
1918                                 if (unlikely(!iov)) {
1919                                         dio->sync = sync = true;
1920                                         goto do_io;
1921                                 }
1922
1923                                 dio->free_iov = true;
1924                         }
1925
1926                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1927                         dio->iter.iov = iov;
1928                 }
1929 do_io:
1930                 dio->loop = true;
1931                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1932
1933                 if (sync)
1934                         wait_for_completion(&dio->done);
1935                 else
1936                         return -EIOCBQUEUED;
1937 loop:
1938                 i_sectors_acct(c, inode, &dio->quota_res,
1939                                dio->op.i_sectors_delta);
1940                 req->ki_pos += (u64) dio->op.written << 9;
1941                 dio->written += dio->op.written;
1942
1943                 spin_lock(&inode->v.i_lock);
1944                 if (req->ki_pos > inode->v.i_size)
1945                         i_size_write(&inode->v, req->ki_pos);
1946                 spin_unlock(&inode->v.i_lock);
1947
1948                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1949                         bio_for_each_segment_all(bv, bio, iter)
1950                                 put_page(bv->bv_page);
1951                 bio->bi_vcnt = 0;
1952
1953                 if (dio->op.error) {
1954                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
1955                         break;
1956                 }
1957
1958                 if (!dio->iter.count)
1959                         break;
1960
1961                 bio_reset(bio);
1962                 reinit_completion(&dio->done);
1963         }
1964
1965         ret = dio->op.error ?: ((long) dio->written << 9);
1966 err:
1967         up(&c->io_in_flight);
1968         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1969         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1970
1971         if (dio->free_iov)
1972                 kfree(dio->iter.iov);
1973
1974         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1975                 bio_for_each_segment_all(bv, bio, iter)
1976                         put_page(bv->bv_page);
1977         bio_put(bio);
1978
1979         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1980         inode_dio_end(&inode->v);
1981
1982         if (!sync) {
1983                 req->ki_complete(req, ret, 0);
1984                 ret = -EIOCBQUEUED;
1985         }
1986         return ret;
1987 }
1988
1989 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1990 {
1991         struct dio_write *dio = container_of(op, struct dio_write, op);
1992
1993         if (dio->sync)
1994                 complete(&dio->done);
1995         else
1996                 bch2_dio_write_loop(dio);
1997 }
1998
1999 static noinline
2000 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2001 {
2002         struct file *file = req->ki_filp;
2003         struct address_space *mapping = file->f_mapping;
2004         struct bch_inode_info *inode = file_bch_inode(file);
2005         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2006         struct dio_write *dio;
2007         struct bio *bio;
2008         bool locked = true, extending;
2009         ssize_t ret;
2010
2011         prefetch(&c->opts);
2012         prefetch((void *) &c->opts + 64);
2013         prefetch(&inode->ei_inode);
2014         prefetch((void *) &inode->ei_inode + 64);
2015
2016         inode_lock(&inode->v);
2017
2018         ret = generic_write_checks(req, iter);
2019         if (unlikely(ret <= 0))
2020                 goto err;
2021
2022         ret = file_remove_privs(file);
2023         if (unlikely(ret))
2024                 goto err;
2025
2026         ret = file_update_time(file);
2027         if (unlikely(ret))
2028                 goto err;
2029
2030         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2031                 goto err;
2032
2033         inode_dio_begin(&inode->v);
2034         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2035
2036         extending = req->ki_pos + iter->count > inode->v.i_size;
2037         if (!extending) {
2038                 inode_unlock(&inode->v);
2039                 locked = false;
2040         }
2041
2042         bio = bio_alloc_bioset(GFP_KERNEL,
2043                                iov_iter_is_bvec(iter)
2044                                ? 0
2045                                : iov_iter_npages(iter, BIO_MAX_VECS),
2046                                &c->dio_write_bioset);
2047         dio = container_of(bio, struct dio_write, op.wbio.bio);
2048         init_completion(&dio->done);
2049         dio->req                = req;
2050         dio->mm                 = current->mm;
2051         dio->loop               = false;
2052         dio->sync               = is_sync_kiocb(req) || extending;
2053         dio->free_iov           = false;
2054         dio->quota_res.sectors  = 0;
2055         dio->written            = 0;
2056         dio->iter               = *iter;
2057
2058         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2059                                          iter->count >> 9, true);
2060         if (unlikely(ret))
2061                 goto err_put_bio;
2062
2063         ret = write_invalidate_inode_pages_range(mapping,
2064                                         req->ki_pos,
2065                                         req->ki_pos + iter->count - 1);
2066         if (unlikely(ret))
2067                 goto err_put_bio;
2068
2069         ret = bch2_dio_write_loop(dio);
2070 err:
2071         if (locked)
2072                 inode_unlock(&inode->v);
2073         return ret;
2074 err_put_bio:
2075         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2076         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2077         bio_put(bio);
2078         inode_dio_end(&inode->v);
2079         goto err;
2080 }
2081
2082 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2083 {
2084         struct file *file = iocb->ki_filp;
2085         struct bch_inode_info *inode = file_bch_inode(file);
2086         ssize_t ret;
2087
2088         if (iocb->ki_flags & IOCB_DIRECT)
2089                 return bch2_direct_write(iocb, from);
2090
2091         /* We can write back this queue in page reclaim */
2092         current->backing_dev_info = inode_to_bdi(&inode->v);
2093         inode_lock(&inode->v);
2094
2095         ret = generic_write_checks(iocb, from);
2096         if (ret <= 0)
2097                 goto unlock;
2098
2099         ret = file_remove_privs(file);
2100         if (ret)
2101                 goto unlock;
2102
2103         ret = file_update_time(file);
2104         if (ret)
2105                 goto unlock;
2106
2107         ret = bch2_buffered_write(iocb, from);
2108         if (likely(ret > 0))
2109                 iocb->ki_pos += ret;
2110 unlock:
2111         inode_unlock(&inode->v);
2112         current->backing_dev_info = NULL;
2113
2114         if (ret > 0)
2115                 ret = generic_write_sync(iocb, ret);
2116
2117         return ret;
2118 }
2119
2120 /* fsync: */
2121
2122 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2123 {
2124         struct bch_inode_info *inode = file_bch_inode(file);
2125         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2126         int ret, ret2;
2127
2128         ret = file_write_and_wait_range(file, start, end);
2129         if (ret)
2130                 return ret;
2131
2132         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2133                 goto out;
2134
2135         ret = sync_inode_metadata(&inode->v, 1);
2136         if (ret)
2137                 return ret;
2138 out:
2139         if (!c->opts.journal_flush_disabled)
2140                 ret = bch2_journal_flush_seq(&c->journal,
2141                                              inode->ei_journal_seq);
2142         ret2 = file_check_and_advance_wb_err(file);
2143
2144         return ret ?: ret2;
2145 }
2146
2147 /* truncate: */
2148
2149 static inline int range_has_data(struct bch_fs *c,
2150                                   struct bpos start,
2151                                   struct bpos end)
2152 {
2153         struct btree_trans trans;
2154         struct btree_iter iter;
2155         struct bkey_s_c k;
2156         int ret = 0;
2157
2158         bch2_trans_init(&trans, c, 0, 0);
2159
2160         for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2161                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2162                         break;
2163
2164                 if (bkey_extent_is_data(k.k)) {
2165                         ret = 1;
2166                         break;
2167                 }
2168         }
2169         bch2_trans_iter_exit(&trans, &iter);
2170
2171         return bch2_trans_exit(&trans) ?: ret;
2172 }
2173
2174 static int __bch2_truncate_page(struct bch_inode_info *inode,
2175                                 pgoff_t index, loff_t start, loff_t end)
2176 {
2177         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2178         struct address_space *mapping = inode->v.i_mapping;
2179         struct bch_page_state *s;
2180         unsigned start_offset = start & (PAGE_SIZE - 1);
2181         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2182         unsigned i;
2183         struct page *page;
2184         int ret = 0;
2185
2186         /* Page boundary? Nothing to do */
2187         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2188               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2189                 return 0;
2190
2191         /* Above i_size? */
2192         if (index << PAGE_SHIFT >= inode->v.i_size)
2193                 return 0;
2194
2195         page = find_lock_page(mapping, index);
2196         if (!page) {
2197                 /*
2198                  * XXX: we're doing two index lookups when we end up reading the
2199                  * page
2200                  */
2201                 ret = range_has_data(c,
2202                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2203                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2204                 if (ret <= 0)
2205                         return ret;
2206
2207                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2208                 if (unlikely(!page)) {
2209                         ret = -ENOMEM;
2210                         goto out;
2211                 }
2212         }
2213
2214         s = bch2_page_state_create(page, 0);
2215         if (!s) {
2216                 ret = -ENOMEM;
2217                 goto unlock;
2218         }
2219
2220         if (!PageUptodate(page)) {
2221                 ret = bch2_read_single_page(page, mapping);
2222                 if (ret)
2223                         goto unlock;
2224         }
2225
2226         if (index != start >> PAGE_SHIFT)
2227                 start_offset = 0;
2228         if (index != end >> PAGE_SHIFT)
2229                 end_offset = PAGE_SIZE;
2230
2231         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2232              i < round_down(end_offset, block_bytes(c)) >> 9;
2233              i++) {
2234                 s->s[i].nr_replicas     = 0;
2235                 s->s[i].state           = SECTOR_UNALLOCATED;
2236         }
2237
2238         zero_user_segment(page, start_offset, end_offset);
2239
2240         /*
2241          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2242          *
2243          * XXX: because we aren't currently tracking whether the page has actual
2244          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2245          */
2246         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2247         BUG_ON(ret);
2248
2249         /*
2250          * This removes any writeable userspace mappings; we need to force
2251          * .page_mkwrite to be called again before any mmapped writes, to
2252          * redirty the full page:
2253          */
2254         page_mkclean(page);
2255         __set_page_dirty_nobuffers(page);
2256 unlock:
2257         unlock_page(page);
2258         put_page(page);
2259 out:
2260         return ret;
2261 }
2262
2263 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2264 {
2265         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2266                                     from, round_up(from, PAGE_SIZE));
2267 }
2268
2269 static int bch2_extend(struct user_namespace *mnt_userns,
2270                        struct bch_inode_info *inode,
2271                        struct bch_inode_unpacked *inode_u,
2272                        struct iattr *iattr)
2273 {
2274         struct address_space *mapping = inode->v.i_mapping;
2275         int ret;
2276
2277         /*
2278          * sync appends:
2279          *
2280          * this has to be done _before_ extending i_size:
2281          */
2282         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2283         if (ret)
2284                 return ret;
2285
2286         truncate_setsize(&inode->v, iattr->ia_size);
2287
2288         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2289 }
2290
2291 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2292                                    struct bch_inode_unpacked *bi,
2293                                    void *p)
2294 {
2295         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2296         return 0;
2297 }
2298
2299 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2300                                   struct bch_inode_unpacked *bi, void *p)
2301 {
2302         u64 *new_i_size = p;
2303
2304         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2305         bi->bi_size = *new_i_size;
2306         return 0;
2307 }
2308
2309 int bch2_truncate(struct user_namespace *mnt_userns,
2310                   struct bch_inode_info *inode, struct iattr *iattr)
2311 {
2312         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2313         struct address_space *mapping = inode->v.i_mapping;
2314         struct bch_inode_unpacked inode_u;
2315         u64 new_i_size = iattr->ia_size;
2316         s64 i_sectors_delta = 0;
2317         int ret = 0;
2318
2319         /*
2320          * If the truncate call with change the size of the file, the
2321          * cmtimes should be updated. If the size will not change, we
2322          * do not need to update the cmtimes.
2323          */
2324         if (iattr->ia_size != inode->v.i_size) {
2325                 if (!(iattr->ia_valid & ATTR_MTIME))
2326                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2327                 if (!(iattr->ia_valid & ATTR_CTIME))
2328                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2329                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2330         }
2331
2332         inode_dio_wait(&inode->v);
2333         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2334
2335         ret = bch2_inode_find_by_inum(c, inode->v.i_ino, &inode_u);
2336         if (ret)
2337                 goto err;
2338
2339         /*
2340          * check this before next assertion; on filesystem error our normal
2341          * invariants are a bit broken (truncate has to truncate the page cache
2342          * before the inode).
2343          */
2344         ret = bch2_journal_error(&c->journal);
2345         if (ret)
2346                 goto err;
2347
2348         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2349                 inode->v.i_size < inode_u.bi_size);
2350
2351         if (iattr->ia_size > inode->v.i_size) {
2352                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2353                 goto err;
2354         }
2355
2356         iattr->ia_valid &= ~ATTR_SIZE;
2357
2358         ret = bch2_truncate_page(inode, iattr->ia_size);
2359         if (unlikely(ret))
2360                 goto err;
2361
2362         /*
2363          * When extending, we're going to write the new i_size to disk
2364          * immediately so we need to flush anything above the current on disk
2365          * i_size first:
2366          *
2367          * Also, when extending we need to flush the page that i_size currently
2368          * straddles - if it's mapped to userspace, we need to ensure that
2369          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2370          * again to allocate the part of the page that was extended.
2371          */
2372         if (iattr->ia_size > inode_u.bi_size)
2373                 ret = filemap_write_and_wait_range(mapping,
2374                                 inode_u.bi_size,
2375                                 iattr->ia_size - 1);
2376         else if (iattr->ia_size & (PAGE_SIZE - 1))
2377                 ret = filemap_write_and_wait_range(mapping,
2378                                 round_down(iattr->ia_size, PAGE_SIZE),
2379                                 iattr->ia_size - 1);
2380         if (ret)
2381                 goto err;
2382
2383         mutex_lock(&inode->ei_update_lock);
2384         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2385                                &new_i_size, 0);
2386         mutex_unlock(&inode->ei_update_lock);
2387
2388         if (unlikely(ret))
2389                 goto err;
2390
2391         truncate_setsize(&inode->v, iattr->ia_size);
2392
2393         ret = bch2_fpunch(c, inode->v.i_ino,
2394                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2395                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2396         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2397
2398         if (unlikely(ret))
2399                 goto err;
2400
2401         mutex_lock(&inode->ei_update_lock);
2402         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2403         mutex_unlock(&inode->ei_update_lock);
2404
2405         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2406 err:
2407         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2408         return ret;
2409 }
2410
2411 /* fallocate: */
2412
2413 static int inode_update_times_fn(struct bch_inode_info *inode,
2414                                  struct bch_inode_unpacked *bi, void *p)
2415 {
2416         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2417
2418         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2419         return 0;
2420 }
2421
2422 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2423 {
2424         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2425         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2426         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2427         int ret = 0;
2428
2429         inode_lock(&inode->v);
2430         inode_dio_wait(&inode->v);
2431         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2432
2433         ret = __bch2_truncate_page(inode,
2434                                    offset >> PAGE_SHIFT,
2435                                    offset, offset + len);
2436         if (unlikely(ret))
2437                 goto err;
2438
2439         if (offset >> PAGE_SHIFT !=
2440             (offset + len) >> PAGE_SHIFT) {
2441                 ret = __bch2_truncate_page(inode,
2442                                            (offset + len) >> PAGE_SHIFT,
2443                                            offset, offset + len);
2444                 if (unlikely(ret))
2445                         goto err;
2446         }
2447
2448         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2449
2450         if (discard_start < discard_end) {
2451                 s64 i_sectors_delta = 0;
2452
2453                 ret = bch2_fpunch(c, inode->v.i_ino,
2454                                   discard_start, discard_end,
2455                                   &inode->ei_journal_seq,
2456                                   &i_sectors_delta);
2457                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2458         }
2459
2460         mutex_lock(&inode->ei_update_lock);
2461         ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2462                                ATTR_MTIME|ATTR_CTIME) ?: ret;
2463         mutex_unlock(&inode->ei_update_lock);
2464 err:
2465         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2466         inode_unlock(&inode->v);
2467
2468         return ret;
2469 }
2470
2471 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2472                                    loff_t offset, loff_t len,
2473                                    bool insert)
2474 {
2475         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2476         struct address_space *mapping = inode->v.i_mapping;
2477         struct bkey_buf copy;
2478         struct btree_trans trans;
2479         struct btree_iter src, dst, del;
2480         loff_t shift, new_size;
2481         u64 src_start;
2482         int ret = 0;
2483
2484         if ((offset | len) & (block_bytes(c) - 1))
2485                 return -EINVAL;
2486
2487         /*
2488          * We need i_mutex to keep the page cache consistent with the extents
2489          * btree, and the btree consistent with i_size - we don't need outside
2490          * locking for the extents btree itself, because we're using linked
2491          * iterators
2492          */
2493         inode_lock(&inode->v);
2494         inode_dio_wait(&inode->v);
2495         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2496
2497         if (insert) {
2498                 ret = -EFBIG;
2499                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2500                         goto err;
2501
2502                 ret = -EINVAL;
2503                 if (offset >= inode->v.i_size)
2504                         goto err;
2505
2506                 src_start       = U64_MAX;
2507                 shift           = len;
2508         } else {
2509                 ret = -EINVAL;
2510                 if (offset + len >= inode->v.i_size)
2511                         goto err;
2512
2513                 src_start       = offset + len;
2514                 shift           = -len;
2515         }
2516
2517         new_size = inode->v.i_size + shift;
2518
2519         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2520         if (ret)
2521                 goto err;
2522
2523         if (insert) {
2524                 i_size_write(&inode->v, new_size);
2525                 mutex_lock(&inode->ei_update_lock);
2526                 ret = bch2_write_inode_size(c, inode, new_size,
2527                                             ATTR_MTIME|ATTR_CTIME);
2528                 mutex_unlock(&inode->ei_update_lock);
2529         } else {
2530                 s64 i_sectors_delta = 0;
2531
2532                 ret = bch2_fpunch(c, inode->v.i_ino,
2533                                   offset >> 9, (offset + len) >> 9,
2534                                   &inode->ei_journal_seq,
2535                                   &i_sectors_delta);
2536                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2537
2538                 if (ret)
2539                         goto err;
2540         }
2541
2542         bch2_bkey_buf_init(&copy);
2543         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2544         bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
2545                         POS(inode->v.i_ino, src_start >> 9),
2546                         BTREE_ITER_INTENT);
2547         bch2_trans_copy_iter(&dst, &src);
2548         bch2_trans_copy_iter(&del, &src);
2549
2550         while (ret == 0 || ret == -EINTR) {
2551                 struct disk_reservation disk_res =
2552                         bch2_disk_reservation_init(c, 0);
2553                 struct bkey_i delete;
2554                 struct bkey_s_c k;
2555                 struct bpos next_pos;
2556                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2557                 struct bpos atomic_end;
2558                 unsigned trigger_flags = 0;
2559
2560                 bch2_trans_begin(&trans);
2561
2562                 k = insert
2563                         ? bch2_btree_iter_peek_prev(&src)
2564                         : bch2_btree_iter_peek(&src);
2565                 if ((ret = bkey_err(k)))
2566                         continue;
2567
2568                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2569                         break;
2570
2571                 if (insert &&
2572                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2573                         break;
2574 reassemble:
2575                 bch2_bkey_buf_reassemble(&copy, c, k);
2576
2577                 if (insert &&
2578                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2579                         bch2_cut_front(move_pos, copy.k);
2580
2581                 copy.k->k.p.offset += shift >> 9;
2582                 bch2_btree_iter_set_pos(&dst, bkey_start_pos(&copy.k->k));
2583
2584                 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
2585                 if (ret)
2586                         continue;
2587
2588                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2589                         if (insert) {
2590                                 move_pos = atomic_end;
2591                                 move_pos.offset -= shift >> 9;
2592                                 goto reassemble;
2593                         } else {
2594                                 bch2_cut_back(atomic_end, copy.k);
2595                         }
2596                 }
2597
2598                 bkey_init(&delete.k);
2599                 delete.k.p = copy.k->k.p;
2600                 delete.k.size = copy.k->k.size;
2601                 delete.k.p.offset -= shift >> 9;
2602                 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
2603
2604                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2605
2606                 if (copy.k->k.size == k.k->size) {
2607                         /*
2608                          * If we're moving the entire extent, we can skip
2609                          * running triggers:
2610                          */
2611                         trigger_flags |= BTREE_TRIGGER_NORUN;
2612                 } else {
2613                         /* We might end up splitting compressed extents: */
2614                         unsigned nr_ptrs =
2615                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2616
2617                         ret = bch2_disk_reservation_get(c, &disk_res,
2618                                         copy.k->k.size, nr_ptrs,
2619                                         BCH_DISK_RESERVATION_NOFAIL);
2620                         BUG_ON(ret);
2621                 }
2622
2623                 ret =   bch2_btree_iter_traverse(&del) ?:
2624                         bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
2625                         bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
2626                         bch2_trans_commit(&trans, &disk_res,
2627                                           &inode->ei_journal_seq,
2628                                           BTREE_INSERT_NOFAIL);
2629                 bch2_disk_reservation_put(c, &disk_res);
2630
2631                 if (!ret)
2632                         bch2_btree_iter_set_pos(&src, next_pos);
2633         }
2634         bch2_trans_iter_exit(&trans, &del);
2635         bch2_trans_iter_exit(&trans, &dst);
2636         bch2_trans_iter_exit(&trans, &src);
2637         bch2_trans_exit(&trans);
2638         bch2_bkey_buf_exit(&copy, c);
2639
2640         if (ret)
2641                 goto err;
2642
2643         if (!insert) {
2644                 i_size_write(&inode->v, new_size);
2645                 mutex_lock(&inode->ei_update_lock);
2646                 ret = bch2_write_inode_size(c, inode, new_size,
2647                                             ATTR_MTIME|ATTR_CTIME);
2648                 mutex_unlock(&inode->ei_update_lock);
2649         }
2650 err:
2651         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2652         inode_unlock(&inode->v);
2653         return ret;
2654 }
2655
2656 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2657                              u64 start_sector, u64 end_sector)
2658 {
2659         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2660         struct btree_trans trans;
2661         struct btree_iter iter;
2662         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2663         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2664         int ret = 0;
2665
2666         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2667
2668         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2669                         POS(inode->v.i_ino, start_sector),
2670                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2671
2672         while (!ret && bkey_cmp(iter.pos, end_pos) < 0) {
2673                 s64 i_sectors_delta = 0;
2674                 struct disk_reservation disk_res = { 0 };
2675                 struct quota_res quota_res = { 0 };
2676                 struct bkey_i_reservation reservation;
2677                 struct bkey_s_c k;
2678                 unsigned sectors;
2679
2680                 bch2_trans_begin(&trans);
2681
2682                 k = bch2_btree_iter_peek_slot(&iter);
2683                 if ((ret = bkey_err(k)))
2684                         goto bkey_err;
2685
2686                 /* already reserved */
2687                 if (k.k->type == KEY_TYPE_reservation &&
2688                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2689                         bch2_btree_iter_advance(&iter);
2690                         continue;
2691                 }
2692
2693                 if (bkey_extent_is_data(k.k) &&
2694                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2695                         bch2_btree_iter_advance(&iter);
2696                         continue;
2697                 }
2698
2699                 bkey_reservation_init(&reservation.k_i);
2700                 reservation.k.type      = KEY_TYPE_reservation;
2701                 reservation.k.p         = k.k->p;
2702                 reservation.k.size      = k.k->size;
2703
2704                 bch2_cut_front(iter.pos,        &reservation.k_i);
2705                 bch2_cut_back(end_pos,          &reservation.k_i);
2706
2707                 sectors = reservation.k.size;
2708                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2709
2710                 if (!bkey_extent_is_allocation(k.k)) {
2711                         ret = bch2_quota_reservation_add(c, inode,
2712                                         &quota_res,
2713                                         sectors, true);
2714                         if (unlikely(ret))
2715                                 goto bkey_err;
2716                 }
2717
2718                 if (reservation.v.nr_replicas < replicas ||
2719                     bch2_bkey_sectors_compressed(k)) {
2720                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2721                                                         replicas, 0);
2722                         if (unlikely(ret))
2723                                 goto bkey_err;
2724
2725                         reservation.v.nr_replicas = disk_res.nr_replicas;
2726                 }
2727
2728                 ret = bch2_extent_update(&trans, &iter, &reservation.k_i,
2729                                 &disk_res, &inode->ei_journal_seq,
2730                                 0, &i_sectors_delta, true);
2731                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2732 bkey_err:
2733                 bch2_quota_reservation_put(c, inode, &quota_res);
2734                 bch2_disk_reservation_put(c, &disk_res);
2735                 if (ret == -EINTR)
2736                         ret = 0;
2737         }
2738         bch2_trans_iter_exit(&trans, &iter);
2739         bch2_trans_exit(&trans);
2740         return ret;
2741 }
2742
2743 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2744                             loff_t offset, loff_t len)
2745 {
2746         struct address_space *mapping = inode->v.i_mapping;
2747         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2748         loff_t end              = offset + len;
2749         loff_t block_start      = round_down(offset,    block_bytes(c));
2750         loff_t block_end        = round_up(end,         block_bytes(c));
2751         int ret;
2752
2753         inode_lock(&inode->v);
2754         inode_dio_wait(&inode->v);
2755         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2756
2757         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2758                 ret = inode_newsize_ok(&inode->v, end);
2759                 if (ret)
2760                         goto err;
2761         }
2762
2763         if (mode & FALLOC_FL_ZERO_RANGE) {
2764                 ret = __bch2_truncate_page(inode,
2765                                            offset >> PAGE_SHIFT,
2766                                            offset, end);
2767
2768                 if (!ret &&
2769                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2770                         ret = __bch2_truncate_page(inode,
2771                                                    end >> PAGE_SHIFT,
2772                                                    offset, end);
2773
2774                 if (unlikely(ret))
2775                         goto err;
2776
2777                 truncate_pagecache_range(&inode->v, offset, end - 1);
2778         }
2779
2780         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
2781         if (ret)
2782                 goto err;
2783
2784         /*
2785          * Do we need to extend the file?
2786          *
2787          * If we zeroed up to the end of the file, we dropped whatever writes
2788          * were going to write out the current i_size, so we have to extend
2789          * manually even if FL_KEEP_SIZE was set:
2790          */
2791         if (end >= inode->v.i_size &&
2792             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2793              (mode & FALLOC_FL_ZERO_RANGE))) {
2794
2795                 /*
2796                  * Sync existing appends before extending i_size,
2797                  * as in bch2_extend():
2798                  */
2799                 ret = filemap_write_and_wait_range(mapping,
2800                                         inode->ei_inode.bi_size, S64_MAX);
2801                 if (ret)
2802                         goto err;
2803
2804                 if (mode & FALLOC_FL_KEEP_SIZE)
2805                         end = inode->v.i_size;
2806                 else
2807                         i_size_write(&inode->v, end);
2808
2809                 mutex_lock(&inode->ei_update_lock);
2810                 ret = bch2_write_inode_size(c, inode, end, 0);
2811                 mutex_unlock(&inode->ei_update_lock);
2812         }
2813 err:
2814         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2815         inode_unlock(&inode->v);
2816         return ret;
2817 }
2818
2819 long bch2_fallocate_dispatch(struct file *file, int mode,
2820                              loff_t offset, loff_t len)
2821 {
2822         struct bch_inode_info *inode = file_bch_inode(file);
2823         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2824         long ret;
2825
2826         if (!percpu_ref_tryget(&c->writes))
2827                 return -EROFS;
2828
2829         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2830                 ret = bchfs_fallocate(inode, mode, offset, len);
2831         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2832                 ret = bchfs_fpunch(inode, offset, len);
2833         else if (mode == FALLOC_FL_INSERT_RANGE)
2834                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2835         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2836                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2837         else
2838                 ret = -EOPNOTSUPP;
2839
2840         percpu_ref_put(&c->writes);
2841
2842         return ret;
2843 }
2844
2845 static void mark_range_unallocated(struct bch_inode_info *inode,
2846                                    loff_t start, loff_t end)
2847 {
2848         pgoff_t index = start >> PAGE_SHIFT;
2849         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2850         struct pagevec pvec;
2851
2852         pagevec_init(&pvec);
2853
2854         do {
2855                 unsigned nr_pages, i, j;
2856
2857                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2858                                                 &index, end_index);
2859                 if (nr_pages == 0)
2860                         break;
2861
2862                 for (i = 0; i < nr_pages; i++) {
2863                         struct page *page = pvec.pages[i];
2864                         struct bch_page_state *s;
2865
2866                         lock_page(page);
2867                         s = bch2_page_state(page);
2868
2869                         if (s) {
2870                                 spin_lock(&s->lock);
2871                                 for (j = 0; j < PAGE_SECTORS; j++)
2872                                         s->s[j].nr_replicas = 0;
2873                                 spin_unlock(&s->lock);
2874                         }
2875
2876                         unlock_page(page);
2877                 }
2878                 pagevec_release(&pvec);
2879         } while (index <= end_index);
2880 }
2881
2882 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2883                              struct file *file_dst, loff_t pos_dst,
2884                              loff_t len, unsigned remap_flags)
2885 {
2886         struct bch_inode_info *src = file_bch_inode(file_src);
2887         struct bch_inode_info *dst = file_bch_inode(file_dst);
2888         struct bch_fs *c = src->v.i_sb->s_fs_info;
2889         s64 i_sectors_delta = 0;
2890         u64 aligned_len;
2891         loff_t ret = 0;
2892
2893         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2894                 return -EINVAL;
2895
2896         if (remap_flags & REMAP_FILE_DEDUP)
2897                 return -EOPNOTSUPP;
2898
2899         if ((pos_src & (block_bytes(c) - 1)) ||
2900             (pos_dst & (block_bytes(c) - 1)))
2901                 return -EINVAL;
2902
2903         if (src == dst &&
2904             abs(pos_src - pos_dst) < len)
2905                 return -EINVAL;
2906
2907         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2908
2909         file_update_time(file_dst);
2910
2911         inode_dio_wait(&src->v);
2912         inode_dio_wait(&dst->v);
2913
2914         ret = generic_remap_file_range_prep(file_src, pos_src,
2915                                             file_dst, pos_dst,
2916                                             &len, remap_flags);
2917         if (ret < 0 || len == 0)
2918                 goto err;
2919
2920         aligned_len = round_up((u64) len, block_bytes(c));
2921
2922         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2923                                 pos_dst, pos_dst + len - 1);
2924         if (ret)
2925                 goto err;
2926
2927         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2928
2929         ret = bch2_remap_range(c,
2930                                POS(dst->v.i_ino, pos_dst >> 9),
2931                                POS(src->v.i_ino, pos_src >> 9),
2932                                aligned_len >> 9,
2933                                &dst->ei_journal_seq,
2934                                pos_dst + len, &i_sectors_delta);
2935         if (ret < 0)
2936                 goto err;
2937
2938         /*
2939          * due to alignment, we might have remapped slightly more than requsted
2940          */
2941         ret = min((u64) ret << 9, (u64) len);
2942
2943         /* XXX get a quota reservation */
2944         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2945
2946         spin_lock(&dst->v.i_lock);
2947         if (pos_dst + ret > dst->v.i_size)
2948                 i_size_write(&dst->v, pos_dst + ret);
2949         spin_unlock(&dst->v.i_lock);
2950
2951         if (((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
2952              IS_SYNC(file_inode(file_dst))) &&
2953             !c->opts.journal_flush_disabled)
2954                 ret = bch2_journal_flush_seq(&c->journal, dst->ei_journal_seq);
2955 err:
2956         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2957
2958         return ret;
2959 }
2960
2961 /* fseek: */
2962
2963 static int page_data_offset(struct page *page, unsigned offset)
2964 {
2965         struct bch_page_state *s = bch2_page_state(page);
2966         unsigned i;
2967
2968         if (s)
2969                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2970                         if (s->s[i].state >= SECTOR_DIRTY)
2971                                 return i << 9;
2972
2973         return -1;
2974 }
2975
2976 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2977                                        loff_t start_offset,
2978                                        loff_t end_offset)
2979 {
2980         struct address_space *mapping = vinode->i_mapping;
2981         struct page *page;
2982         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2983         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2984         pgoff_t index           = start_index;
2985         loff_t ret;
2986         int offset;
2987
2988         while (index <= end_index) {
2989                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2990                         lock_page(page);
2991
2992                         offset = page_data_offset(page,
2993                                         page->index == start_index
2994                                         ? start_offset & (PAGE_SIZE - 1)
2995                                         : 0);
2996                         if (offset >= 0) {
2997                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2998                                             offset,
2999                                             start_offset, end_offset);
3000                                 unlock_page(page);
3001                                 put_page(page);
3002                                 return ret;
3003                         }
3004
3005                         unlock_page(page);
3006                         put_page(page);
3007                 } else {
3008                         break;
3009                 }
3010         }
3011
3012         return end_offset;
3013 }
3014
3015 static loff_t bch2_seek_data(struct file *file, u64 offset)
3016 {
3017         struct bch_inode_info *inode = file_bch_inode(file);
3018         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3019         struct btree_trans trans;
3020         struct btree_iter iter;
3021         struct bkey_s_c k;
3022         u64 isize, next_data = MAX_LFS_FILESIZE;
3023         int ret;
3024
3025         isize = i_size_read(&inode->v);
3026         if (offset >= isize)
3027                 return -ENXIO;
3028
3029         bch2_trans_init(&trans, c, 0, 0);
3030
3031         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3032                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3033                 if (k.k->p.inode != inode->v.i_ino) {
3034                         break;
3035                 } else if (bkey_extent_is_data(k.k)) {
3036                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3037                         break;
3038                 } else if (k.k->p.offset >> 9 > isize)
3039                         break;
3040         }
3041         bch2_trans_iter_exit(&trans, &iter);
3042
3043         ret = bch2_trans_exit(&trans) ?: ret;
3044         if (ret)
3045                 return ret;
3046
3047         if (next_data > offset)
3048                 next_data = bch2_seek_pagecache_data(&inode->v,
3049                                                      offset, next_data);
3050
3051         if (next_data >= isize)
3052                 return -ENXIO;
3053
3054         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3055 }
3056
3057 static int __page_hole_offset(struct page *page, unsigned offset)
3058 {
3059         struct bch_page_state *s = bch2_page_state(page);
3060         unsigned i;
3061
3062         if (!s)
3063                 return 0;
3064
3065         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3066                 if (s->s[i].state < SECTOR_DIRTY)
3067                         return i << 9;
3068
3069         return -1;
3070 }
3071
3072 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3073 {
3074         pgoff_t index = offset >> PAGE_SHIFT;
3075         struct page *page;
3076         int pg_offset;
3077         loff_t ret = -1;
3078
3079         page = find_lock_page(mapping, index);
3080         if (!page)
3081                 return offset;
3082
3083         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3084         if (pg_offset >= 0)
3085                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3086
3087         unlock_page(page);
3088
3089         return ret;
3090 }
3091
3092 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3093                                        loff_t start_offset,
3094                                        loff_t end_offset)
3095 {
3096         struct address_space *mapping = vinode->i_mapping;
3097         loff_t offset = start_offset, hole;
3098
3099         while (offset < end_offset) {
3100                 hole = page_hole_offset(mapping, offset);
3101                 if (hole >= 0 && hole <= end_offset)
3102                         return max(start_offset, hole);
3103
3104                 offset += PAGE_SIZE;
3105                 offset &= PAGE_MASK;
3106         }
3107
3108         return end_offset;
3109 }
3110
3111 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3112 {
3113         struct bch_inode_info *inode = file_bch_inode(file);
3114         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3115         struct btree_trans trans;
3116         struct btree_iter iter;
3117         struct bkey_s_c k;
3118         u64 isize, next_hole = MAX_LFS_FILESIZE;
3119         int ret;
3120
3121         isize = i_size_read(&inode->v);
3122         if (offset >= isize)
3123                 return -ENXIO;
3124
3125         bch2_trans_init(&trans, c, 0, 0);
3126
3127         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3128                            POS(inode->v.i_ino, offset >> 9),
3129                            BTREE_ITER_SLOTS, k, ret) {
3130                 if (k.k->p.inode != inode->v.i_ino) {
3131                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3132                                         offset, MAX_LFS_FILESIZE);
3133                         break;
3134                 } else if (!bkey_extent_is_data(k.k)) {
3135                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3136                                         max(offset, bkey_start_offset(k.k) << 9),
3137                                         k.k->p.offset << 9);
3138
3139                         if (next_hole < k.k->p.offset << 9)
3140                                 break;
3141                 } else {
3142                         offset = max(offset, bkey_start_offset(k.k) << 9);
3143                 }
3144         }
3145         bch2_trans_iter_exit(&trans, &iter);
3146
3147         ret = bch2_trans_exit(&trans) ?: ret;
3148         if (ret)
3149                 return ret;
3150
3151         if (next_hole > isize)
3152                 next_hole = isize;
3153
3154         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3155 }
3156
3157 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3158 {
3159         switch (whence) {
3160         case SEEK_SET:
3161         case SEEK_CUR:
3162         case SEEK_END:
3163                 return generic_file_llseek(file, offset, whence);
3164         case SEEK_DATA:
3165                 return bch2_seek_data(file, offset);
3166         case SEEK_HOLE:
3167                 return bch2_seek_hole(file, offset);
3168         }
3169
3170         return -EINVAL;
3171 }
3172
3173 void bch2_fs_fsio_exit(struct bch_fs *c)
3174 {
3175         bioset_exit(&c->dio_write_bioset);
3176         bioset_exit(&c->dio_read_bioset);
3177         bioset_exit(&c->writepage_bioset);
3178 }
3179
3180 int bch2_fs_fsio_init(struct bch_fs *c)
3181 {
3182         int ret = 0;
3183
3184         pr_verbose_init(c->opts, "");
3185
3186         if (bioset_init(&c->writepage_bioset,
3187                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3188                         BIOSET_NEED_BVECS) ||
3189             bioset_init(&c->dio_read_bioset,
3190                         4, offsetof(struct dio_read, rbio.bio),
3191                         BIOSET_NEED_BVECS) ||
3192             bioset_init(&c->dio_write_bioset,
3193                         4, offsetof(struct dio_write, op.wbio.bio),
3194                         BIOSET_NEED_BVECS))
3195                 ret = -ENOMEM;
3196
3197         pr_verbose_init(c->opts, "ret %i", ret);
3198         return ret;
3199 }
3200
3201 #endif /* NO_BCACHEFS_FS */