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         while (1) {
 805                 struct bkey_s_c k;
 806                 unsigned bytes, sectors, offset_into_extent;
 807                 enum btree_id data_btree = BTREE_ID_extents;
 808
 809                 bch2_btree_iter_set_pos(iter,
 810                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 811
 812                 k = bch2_btree_iter_peek_slot(iter);
 813                 ret = bkey_err(k);
 814                 if (ret)
 815                         break;
 816
 817                 offset_into_extent = iter->pos.offset -
 818                         bkey_start_offset(k.k);
 819                 sectors = k.k->size - offset_into_extent;
 820
 821                 bch2_bkey_buf_reassemble(&sk, c, k);
 822
 823                 ret = bch2_read_indirect_extent(trans, &data_btree,
 824                                         &offset_into_extent, &sk);
 825                 if (ret)
 826                         break;
 827
 828                 k = bkey_i_to_s_c(sk.k);
 829
 830                 sectors = min(sectors, k.k->size - offset_into_extent);
 831
 832                 bch2_trans_unlock(trans);
 833
 834                 if (readpages_iter)
 835                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 836                                             extent_partial_reads_expensive(k));
 837
 838                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 839                 swap(rbio->bio.bi_iter.bi_size, bytes);
 840
 841                 if (rbio->bio.bi_iter.bi_size == bytes)
 842                         flags |= BCH_READ_LAST_FRAGMENT;
 843
 844                 if (bkey_extent_is_allocation(k.k))
 845                         bch2_add_page_sectors(&rbio->bio, k);
 846
 847                 bch2_read_extent(trans, rbio, iter->pos,
 848                                  data_btree, k, offset_into_extent, flags);
 849
 850                 if (flags & BCH_READ_LAST_FRAGMENT)
 851                         break;
 852
 853                 swap(rbio->bio.bi_iter.bi_size, bytes);
 854                 bio_advance(&rbio->bio, bytes);
 855         }
 856
 857         if (ret == -EINTR)
 858                 goto retry;
 859
 860         if (ret) {
 861                 bch_err_inum_ratelimited(c, inum,
 862                                 "read error %i from btree lookup", ret);
 863                 rbio->bio.bi_status = BLK_STS_IOERR;
 864                 bio_endio(&rbio->bio);
 865         }
 866
 867         bch2_bkey_buf_exit(&sk, c);
 868 }
 869
 870 void bch2_readahead(struct readahead_control *ractl)
 871 {
 872         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 873         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 874         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 875         struct btree_trans trans;
 876         struct btree_iter *iter;
 877         struct page *page;
 878         struct readpages_iter readpages_iter;
 879         int ret;
 880
 881         ret = readpages_iter_init(&readpages_iter, ractl);
 882         BUG_ON(ret);
 883
 884         bch2_trans_init(&trans, c, 0, 0);
 885         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 886                                    BTREE_ITER_SLOTS);
 887
 888         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 889
 890         while ((page = readpage_iter_next(&readpages_iter))) {
 891                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 892                 unsigned n = min_t(unsigned,
 893                                    readpages_iter.nr_pages -
 894                                    readpages_iter.idx,
 895                                    BIO_MAX_VECS);
 896                 struct bch_read_bio *rbio =
 897                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 898                                   opts);
 899
 900                 readpages_iter.idx++;
 901
 902                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 903                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 904                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 905                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 906
 907                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 908                            &readpages_iter);
 909         }
 910
 911         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 912
 913         bch2_trans_iter_put(&trans, iter);
 914         bch2_trans_exit(&trans);
 915         kfree(readpages_iter.pages);
 916 }
 917
 918 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 919                              u64 inum, struct page *page)
 920 {
 921         struct btree_trans trans;
 922         struct btree_iter *iter;
 923
 924         bch2_page_state_create(page, __GFP_NOFAIL);
 925
 926         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 927         rbio->bio.bi_iter.bi_sector =
 928                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 929         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 930
 931         bch2_trans_init(&trans, c, 0, 0);
 932         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 933                                    BTREE_ITER_SLOTS);
 934
 935         bchfs_read(&trans, iter, rbio, inum, NULL);
 936
 937         bch2_trans_iter_put(&trans, iter);
 938         bch2_trans_exit(&trans);
 939 }
 940
 941 int bch2_readpage(struct file *file, struct page *page)
 942 {
 943         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 944         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 945         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 946         struct bch_read_bio *rbio;
 947
 948         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 949         rbio->bio.bi_end_io = bch2_readpages_end_io;
 950
 951         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 952         return 0;
 953 }
 954
 955 static void bch2_read_single_page_end_io(struct bio *bio)
 956 {
 957         complete(bio->bi_private);
 958 }
 959
 960 static int bch2_read_single_page(struct page *page,
 961                                  struct address_space *mapping)
 962 {
 963         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 964         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 965         struct bch_read_bio *rbio;
 966         int ret;
 967         DECLARE_COMPLETION_ONSTACK(done);
 968
 969         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 970                          io_opts(c, &inode->ei_inode));
 971         rbio->bio.bi_private = &done;
 972         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 973
 974         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 975         wait_for_completion(&done);
 976
 977         ret = blk_status_to_errno(rbio->bio.bi_status);
 978         bio_put(&rbio->bio);
 979
 980         if (ret < 0)
 981                 return ret;
 982
 983         SetPageUptodate(page);
 984         return 0;
 985 }
 986
 987 /* writepages: */
 988
 989 struct bch_writepage_state {
 990         struct bch_writepage_io *io;
 991         struct bch_io_opts      opts;
 992 };
 993
 994 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 995                                                                   struct bch_inode_info *inode)
 996 {
 997         return (struct bch_writepage_state) {
 998                 .opts = io_opts(c, &inode->ei_inode)
 999         };
1000 }
1001
1002 static void bch2_writepage_io_free(struct closure *cl)
1003 {
1004         struct bch_writepage_io *io = container_of(cl,
1005                                         struct bch_writepage_io, cl);
1006
1007         bio_put(&io->op.wbio.bio);
1008 }
1009
1010 static void bch2_writepage_io_done(struct closure *cl)
1011 {
1012         struct bch_writepage_io *io = container_of(cl,
1013                                         struct bch_writepage_io, cl);
1014         struct bch_fs *c = io->op.c;
1015         struct bio *bio = &io->op.wbio.bio;
1016         struct bvec_iter_all iter;
1017         struct bio_vec *bvec;
1018         unsigned i;
1019
1020         up(&io->op.c->io_in_flight);
1021
1022         if (io->op.error) {
1023                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1024
1025                 bio_for_each_segment_all(bvec, bio, iter) {
1026                         struct bch_page_state *s;
1027
1028                         SetPageError(bvec->bv_page);
1029                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1030
1031                         s = __bch2_page_state(bvec->bv_page);
1032                         spin_lock(&s->lock);
1033                         for (i = 0; i < PAGE_SECTORS; i++)
1034                                 s->s[i].nr_replicas = 0;
1035                         spin_unlock(&s->lock);
1036                 }
1037         }
1038
1039         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1040                 bio_for_each_segment_all(bvec, bio, iter) {
1041                         struct bch_page_state *s;
1042
1043                         s = __bch2_page_state(bvec->bv_page);
1044                         spin_lock(&s->lock);
1045                         for (i = 0; i < PAGE_SECTORS; i++)
1046                                 s->s[i].nr_replicas = 0;
1047                         spin_unlock(&s->lock);
1048                 }
1049         }
1050
1051         /*
1052          * racing with fallocate can cause us to add fewer sectors than
1053          * expected - but we shouldn't add more sectors than expected:
1054          */
1055         BUG_ON(io->op.i_sectors_delta > 0);
1056
1057         /*
1058          * (error (due to going RO) halfway through a page can screw that up
1059          * slightly)
1060          * XXX wtf?
1061            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1062          */
1063
1064         /*
1065          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1066          * before calling end_page_writeback:
1067          */
1068         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1069
1070         bio_for_each_segment_all(bvec, bio, iter) {
1071                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1072
1073                 if (atomic_dec_and_test(&s->write_count))
1074                         end_page_writeback(bvec->bv_page);
1075         }
1076
1077         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1078 }
1079
1080 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1081 {
1082         struct bch_writepage_io *io = w->io;
1083
1084         down(&io->op.c->io_in_flight);
1085
1086         w->io = NULL;
1087         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1088         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1089 }
1090
1091 /*
1092  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1093  * possible, else allocating a new one:
1094  */
1095 static void bch2_writepage_io_alloc(struct bch_fs *c,
1096                                     struct writeback_control *wbc,
1097                                     struct bch_writepage_state *w,
1098                                     struct bch_inode_info *inode,
1099                                     u64 sector,
1100                                     unsigned nr_replicas)
1101 {
1102         struct bch_write_op *op;
1103
1104         w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1105                                               &c->writepage_bioset),
1106                              struct bch_writepage_io, op.wbio.bio);
1107
1108         closure_init(&w->io->cl, NULL);
1109         w->io->inode            = inode;
1110
1111         op                      = &w->io->op;
1112         bch2_write_op_init(op, c, w->opts);
1113         op->target              = w->opts.foreground_target;
1114         op_journal_seq_set(op, &inode->ei_journal_seq);
1115         op->nr_replicas         = nr_replicas;
1116         op->res.nr_replicas     = nr_replicas;
1117         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1118         op->pos                 = POS(inode->v.i_ino, sector);
1119         op->wbio.bio.bi_iter.bi_sector = sector;
1120         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1121 }
1122
1123 static int __bch2_writepage(struct page *page,
1124                             struct writeback_control *wbc,
1125                             void *data)
1126 {
1127         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1128         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1129         struct bch_writepage_state *w = data;
1130         struct bch_page_state *s, orig;
1131         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1132         loff_t i_size = i_size_read(&inode->v);
1133         pgoff_t end_index = i_size >> PAGE_SHIFT;
1134         int ret;
1135
1136         EBUG_ON(!PageUptodate(page));
1137
1138         /* Is the page fully inside i_size? */
1139         if (page->index < end_index)
1140                 goto do_io;
1141
1142         /* Is the page fully outside i_size? (truncate in progress) */
1143         offset = i_size & (PAGE_SIZE - 1);
1144         if (page->index > end_index || !offset) {
1145                 unlock_page(page);
1146                 return 0;
1147         }
1148
1149         /*
1150          * The page straddles i_size.  It must be zeroed out on each and every
1151          * writepage invocation because it may be mmapped.  "A file is mapped
1152          * in multiples of the page size.  For a file that is not a multiple of
1153          * the  page size, the remaining memory is zeroed when mapped, and
1154          * writes to that region are not written out to the file."
1155          */
1156         zero_user_segment(page, offset, PAGE_SIZE);
1157 do_io:
1158         s = bch2_page_state_create(page, __GFP_NOFAIL);
1159
1160         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1161         if (ret) {
1162                 SetPageError(page);
1163                 mapping_set_error(page->mapping, ret);
1164                 unlock_page(page);
1165                 return 0;
1166         }
1167
1168         /* Before unlocking the page, get copy of reservations: */
1169         orig = *s;
1170
1171         for (i = 0; i < PAGE_SECTORS; i++) {
1172                 if (s->s[i].state < SECTOR_DIRTY)
1173                         continue;
1174
1175                 nr_replicas_this_write =
1176                         min_t(unsigned, nr_replicas_this_write,
1177                               s->s[i].nr_replicas +
1178                               s->s[i].replicas_reserved);
1179         }
1180
1181         for (i = 0; i < PAGE_SECTORS; i++) {
1182                 if (s->s[i].state < SECTOR_DIRTY)
1183                         continue;
1184
1185                 s->s[i].nr_replicas = w->opts.compression
1186                         ? 0 : nr_replicas_this_write;
1187
1188                 s->s[i].replicas_reserved = 0;
1189                 s->s[i].state = SECTOR_ALLOCATED;
1190         }
1191
1192         BUG_ON(atomic_read(&s->write_count));
1193         atomic_set(&s->write_count, 1);
1194
1195         BUG_ON(PageWriteback(page));
1196         set_page_writeback(page);
1197
1198         unlock_page(page);
1199
1200         offset = 0;
1201         while (1) {
1202                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1203                 u64 sector;
1204
1205                 while (offset < PAGE_SECTORS &&
1206                        orig.s[offset].state < SECTOR_DIRTY)
1207                         offset++;
1208
1209                 if (offset == PAGE_SECTORS)
1210                         break;
1211
1212                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1213
1214                 while (offset + sectors < PAGE_SECTORS &&
1215                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1216                         sectors++;
1217
1218                 for (i = offset; i < offset + sectors; i++) {
1219                         reserved_sectors += orig.s[i].replicas_reserved;
1220                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1221                 }
1222
1223                 if (w->io &&
1224                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1225                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1226                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1227                      (BIO_MAX_VECS * PAGE_SIZE) ||
1228                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1229                         bch2_writepage_do_io(w);
1230
1231                 if (!w->io)
1232                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1233                                                 nr_replicas_this_write);
1234
1235                 atomic_inc(&s->write_count);
1236
1237                 BUG_ON(inode != w->io->inode);
1238                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1239                                      sectors << 9, offset << 9));
1240
1241                 /* Check for writing past i_size: */
1242                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1243                         round_up(i_size, block_bytes(c)));
1244
1245                 w->io->op.res.sectors += reserved_sectors;
1246                 w->io->op.i_sectors_delta -= dirty_sectors;
1247                 w->io->op.new_i_size = i_size;
1248
1249                 offset += sectors;
1250         }
1251
1252         if (atomic_dec_and_test(&s->write_count))
1253                 end_page_writeback(page);
1254
1255         return 0;
1256 }
1257
1258 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1259 {
1260         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1261         struct bch_writepage_state w =
1262                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1263         struct blk_plug plug;
1264         int ret;
1265
1266         blk_start_plug(&plug);
1267         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1268         if (w.io)
1269                 bch2_writepage_do_io(&w);
1270         blk_finish_plug(&plug);
1271         return ret;
1272 }
1273
1274 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1275 {
1276         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1277         struct bch_writepage_state w =
1278                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1279         int ret;
1280
1281         ret = __bch2_writepage(page, wbc, &w);
1282         if (w.io)
1283                 bch2_writepage_do_io(&w);
1284
1285         return ret;
1286 }
1287
1288 /* buffered writes: */
1289
1290 int bch2_write_begin(struct file *file, struct address_space *mapping,
1291                      loff_t pos, unsigned len, unsigned flags,
1292                      struct page **pagep, void **fsdata)
1293 {
1294         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1295         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1296         struct bch2_page_reservation *res;
1297         pgoff_t index = pos >> PAGE_SHIFT;
1298         unsigned offset = pos & (PAGE_SIZE - 1);
1299         struct page *page;
1300         int ret = -ENOMEM;
1301
1302         res = kmalloc(sizeof(*res), GFP_KERNEL);
1303         if (!res)
1304                 return -ENOMEM;
1305
1306         bch2_page_reservation_init(c, inode, res);
1307         *fsdata = res;
1308
1309         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1310
1311         page = grab_cache_page_write_begin(mapping, index, flags);
1312         if (!page)
1313                 goto err_unlock;
1314
1315         if (PageUptodate(page))
1316                 goto out;
1317
1318         /* If we're writing entire page, don't need to read it in first: */
1319         if (len == PAGE_SIZE)
1320                 goto out;
1321
1322         if (!offset && pos + len >= inode->v.i_size) {
1323                 zero_user_segment(page, len, PAGE_SIZE);
1324                 flush_dcache_page(page);
1325                 goto out;
1326         }
1327
1328         if (index > inode->v.i_size >> PAGE_SHIFT) {
1329                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1330                 flush_dcache_page(page);
1331                 goto out;
1332         }
1333 readpage:
1334         ret = bch2_read_single_page(page, mapping);
1335         if (ret)
1336                 goto err;
1337 out:
1338         ret = bch2_page_reservation_get(c, inode, page, res,
1339                                         offset, len, true);
1340         if (ret) {
1341                 if (!PageUptodate(page)) {
1342                         /*
1343                          * If the page hasn't been read in, we won't know if we
1344                          * actually need a reservation - we don't actually need
1345                          * to read here, we just need to check if the page is
1346                          * fully backed by uncompressed data:
1347                          */
1348                         goto readpage;
1349                 }
1350
1351                 goto err;
1352         }
1353
1354         *pagep = page;
1355         return 0;
1356 err:
1357         unlock_page(page);
1358         put_page(page);
1359         *pagep = NULL;
1360 err_unlock:
1361         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1362         kfree(res);
1363         *fsdata = NULL;
1364         return ret;
1365 }
1366
1367 int bch2_write_end(struct file *file, struct address_space *mapping,
1368                    loff_t pos, unsigned len, unsigned copied,
1369                    struct page *page, void *fsdata)
1370 {
1371         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1372         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1373         struct bch2_page_reservation *res = fsdata;
1374         unsigned offset = pos & (PAGE_SIZE - 1);
1375
1376         lockdep_assert_held(&inode->v.i_rwsem);
1377
1378         if (unlikely(copied < len && !PageUptodate(page))) {
1379                 /*
1380                  * The page needs to be read in, but that would destroy
1381                  * our partial write - simplest thing is to just force
1382                  * userspace to redo the write:
1383                  */
1384                 zero_user(page, 0, PAGE_SIZE);
1385                 flush_dcache_page(page);
1386                 copied = 0;
1387         }
1388
1389         spin_lock(&inode->v.i_lock);
1390         if (pos + copied > inode->v.i_size)
1391                 i_size_write(&inode->v, pos + copied);
1392         spin_unlock(&inode->v.i_lock);
1393
1394         if (copied) {
1395                 if (!PageUptodate(page))
1396                         SetPageUptodate(page);
1397
1398                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1399
1400                 inode->ei_last_dirtied = (unsigned long) current;
1401         }
1402
1403         unlock_page(page);
1404         put_page(page);
1405         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1406
1407         bch2_page_reservation_put(c, inode, res);
1408         kfree(res);
1409
1410         return copied;
1411 }
1412
1413 #define WRITE_BATCH_PAGES       32
1414
1415 static int __bch2_buffered_write(struct bch_inode_info *inode,
1416                                  struct address_space *mapping,
1417                                  struct iov_iter *iter,
1418                                  loff_t pos, unsigned len)
1419 {
1420         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1421         struct page *pages[WRITE_BATCH_PAGES];
1422         struct bch2_page_reservation res;
1423         unsigned long index = pos >> PAGE_SHIFT;
1424         unsigned offset = pos & (PAGE_SIZE - 1);
1425         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1426         unsigned i, reserved = 0, set_dirty = 0;
1427         unsigned copied = 0, nr_pages_copied = 0;
1428         int ret = 0;
1429
1430         BUG_ON(!len);
1431         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1432
1433         bch2_page_reservation_init(c, inode, &res);
1434
1435         for (i = 0; i < nr_pages; i++) {
1436                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1437                 if (!pages[i]) {
1438                         nr_pages = i;
1439                         if (!i) {
1440                                 ret = -ENOMEM;
1441                                 goto out;
1442                         }
1443                         len = min_t(unsigned, len,
1444                                     nr_pages * PAGE_SIZE - offset);
1445                         break;
1446                 }
1447         }
1448
1449         if (offset && !PageUptodate(pages[0])) {
1450                 ret = bch2_read_single_page(pages[0], mapping);
1451                 if (ret)
1452                         goto out;
1453         }
1454
1455         if ((pos + len) & (PAGE_SIZE - 1) &&
1456             !PageUptodate(pages[nr_pages - 1])) {
1457                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1458                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1459                 } else {
1460                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1461                         if (ret)
1462                                 goto out;
1463                 }
1464         }
1465
1466         while (reserved < len) {
1467                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1468                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1469                 unsigned pg_len = min_t(unsigned, len - reserved,
1470                                         PAGE_SIZE - pg_offset);
1471 retry_reservation:
1472                 ret = bch2_page_reservation_get(c, inode, page, &res,
1473                                                 pg_offset, pg_len, true);
1474
1475                 if (ret && !PageUptodate(page)) {
1476                         ret = bch2_read_single_page(page, mapping);
1477                         if (!ret)
1478                                 goto retry_reservation;
1479                 }
1480
1481                 if (ret)
1482                         goto out;
1483
1484                 reserved += pg_len;
1485         }
1486
1487         if (mapping_writably_mapped(mapping))
1488                 for (i = 0; i < nr_pages; i++)
1489                         flush_dcache_page(pages[i]);
1490
1491         while (copied < len) {
1492                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1493                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1494                 unsigned pg_len = min_t(unsigned, len - copied,
1495                                         PAGE_SIZE - pg_offset);
1496                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1497                                                 iter, pg_offset, pg_len);
1498
1499                 if (!pg_copied)
1500                         break;
1501
1502                 if (!PageUptodate(page) &&
1503                     pg_copied != PAGE_SIZE &&
1504                     pos + copied + pg_copied < inode->v.i_size) {
1505                         zero_user(page, 0, PAGE_SIZE);
1506                         break;
1507                 }
1508
1509                 flush_dcache_page(page);
1510                 iov_iter_advance(iter, pg_copied);
1511                 copied += pg_copied;
1512
1513                 if (pg_copied != pg_len)
1514                         break;
1515         }
1516
1517         if (!copied)
1518                 goto out;
1519
1520         spin_lock(&inode->v.i_lock);
1521         if (pos + copied > inode->v.i_size)
1522                 i_size_write(&inode->v, pos + copied);
1523         spin_unlock(&inode->v.i_lock);
1524
1525         while (set_dirty < copied) {
1526                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1527                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1528                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1529                                         PAGE_SIZE - pg_offset);
1530
1531                 if (!PageUptodate(page))
1532                         SetPageUptodate(page);
1533
1534                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1535                 unlock_page(page);
1536                 put_page(page);
1537
1538                 set_dirty += pg_len;
1539         }
1540
1541         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1542         inode->ei_last_dirtied = (unsigned long) current;
1543 out:
1544         for (i = nr_pages_copied; i < nr_pages; i++) {
1545                 unlock_page(pages[i]);
1546                 put_page(pages[i]);
1547         }
1548
1549         bch2_page_reservation_put(c, inode, &res);
1550
1551         return copied ?: ret;
1552 }
1553
1554 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1555 {
1556         struct file *file = iocb->ki_filp;
1557         struct address_space *mapping = file->f_mapping;
1558         struct bch_inode_info *inode = file_bch_inode(file);
1559         loff_t pos = iocb->ki_pos;
1560         ssize_t written = 0;
1561         int ret = 0;
1562
1563         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1564
1565         do {
1566                 unsigned offset = pos & (PAGE_SIZE - 1);
1567                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1568                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1569 again:
1570                 /*
1571                  * Bring in the user page that we will copy from _first_.
1572                  * Otherwise there's a nasty deadlock on copying from the
1573                  * same page as we're writing to, without it being marked
1574                  * up-to-date.
1575                  *
1576                  * Not only is this an optimisation, but it is also required
1577                  * to check that the address is actually valid, when atomic
1578                  * usercopies are used, below.
1579                  */
1580                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1581                         bytes = min_t(unsigned long, iov_iter_count(iter),
1582                                       PAGE_SIZE - offset);
1583
1584                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1585                                 ret = -EFAULT;
1586                                 break;
1587                         }
1588                 }
1589
1590                 if (unlikely(fatal_signal_pending(current))) {
1591                         ret = -EINTR;
1592                         break;
1593                 }
1594
1595                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1596                 if (unlikely(ret < 0))
1597                         break;
1598
1599                 cond_resched();
1600
1601                 if (unlikely(ret == 0)) {
1602                         /*
1603                          * If we were unable to copy any data at all, we must
1604                          * fall back to a single segment length write.
1605                          *
1606                          * If we didn't fallback here, we could livelock
1607                          * because not all segments in the iov can be copied at
1608                          * once without a pagefault.
1609                          */
1610                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1611                                       iov_iter_single_seg_count(iter));
1612                         goto again;
1613                 }
1614                 pos += ret;
1615                 written += ret;
1616                 ret = 0;
1617
1618                 balance_dirty_pages_ratelimited(mapping);
1619         } while (iov_iter_count(iter));
1620
1621         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1622
1623         return written ? written : ret;
1624 }
1625
1626 /* O_DIRECT reads */
1627
1628 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1629 {
1630         if (check_dirty) {
1631                 bio_check_pages_dirty(bio);
1632         } else {
1633                 bio_release_pages(bio, false);
1634                 bio_put(bio);
1635         }
1636 }
1637
1638 static void bch2_dio_read_complete(struct closure *cl)
1639 {
1640         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1641
1642         dio->req->ki_complete(dio->req, dio->ret, 0);
1643         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1644 }
1645
1646 static void bch2_direct_IO_read_endio(struct bio *bio)
1647 {
1648         struct dio_read *dio = bio->bi_private;
1649
1650         if (bio->bi_status)
1651                 dio->ret = blk_status_to_errno(bio->bi_status);
1652
1653         closure_put(&dio->cl);
1654 }
1655
1656 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1657 {
1658         struct dio_read *dio = bio->bi_private;
1659         bool should_dirty = dio->should_dirty;
1660
1661         bch2_direct_IO_read_endio(bio);
1662         bio_check_or_release(bio, should_dirty);
1663 }
1664
1665 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1666 {
1667         struct file *file = req->ki_filp;
1668         struct bch_inode_info *inode = file_bch_inode(file);
1669         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1670         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1671         struct dio_read *dio;
1672         struct bio *bio;
1673         loff_t offset = req->ki_pos;
1674         bool sync = is_sync_kiocb(req);
1675         size_t shorten;
1676         ssize_t ret;
1677
1678         if ((offset|iter->count) & (block_bytes(c) - 1))
1679                 return -EINVAL;
1680
1681         ret = min_t(loff_t, iter->count,
1682                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1683
1684         if (!ret)
1685                 return ret;
1686
1687         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1688         iter->count -= shorten;
1689
1690         bio = bio_alloc_bioset(GFP_KERNEL,
1691                                iov_iter_npages(iter, BIO_MAX_VECS),
1692                                &c->dio_read_bioset);
1693
1694         bio->bi_end_io = bch2_direct_IO_read_endio;
1695
1696         dio = container_of(bio, struct dio_read, rbio.bio);
1697         closure_init(&dio->cl, NULL);
1698
1699         /*
1700          * this is a _really_ horrible hack just to avoid an atomic sub at the
1701          * end:
1702          */
1703         if (!sync) {
1704                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1705                 atomic_set(&dio->cl.remaining,
1706                            CLOSURE_REMAINING_INITIALIZER -
1707                            CLOSURE_RUNNING +
1708                            CLOSURE_DESTRUCTOR);
1709         } else {
1710                 atomic_set(&dio->cl.remaining,
1711                            CLOSURE_REMAINING_INITIALIZER + 1);
1712         }
1713
1714         dio->req        = req;
1715         dio->ret        = ret;
1716         /*
1717          * This is one of the sketchier things I've encountered: we have to skip
1718          * the dirtying of requests that are internal from the kernel (i.e. from
1719          * loopback), because we'll deadlock on page_lock.
1720          */
1721         dio->should_dirty = iter_is_iovec(iter);
1722
1723         goto start;
1724         while (iter->count) {
1725                 bio = bio_alloc_bioset(GFP_KERNEL,
1726                                        iov_iter_npages(iter, BIO_MAX_VECS),
1727                                        &c->bio_read);
1728                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1729 start:
1730                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1731                 bio->bi_iter.bi_sector  = offset >> 9;
1732                 bio->bi_private         = dio;
1733
1734                 ret = bio_iov_iter_get_pages(bio, iter);
1735                 if (ret < 0) {
1736                         /* XXX: fault inject this path */
1737                         bio->bi_status = BLK_STS_RESOURCE;
1738                         bio_endio(bio);
1739                         break;
1740                 }
1741
1742                 offset += bio->bi_iter.bi_size;
1743
1744                 if (dio->should_dirty)
1745                         bio_set_pages_dirty(bio);
1746
1747                 if (iter->count)
1748                         closure_get(&dio->cl);
1749
1750                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1751         }
1752
1753         iter->count += shorten;
1754
1755         if (sync) {
1756                 closure_sync(&dio->cl);
1757                 closure_debug_destroy(&dio->cl);
1758                 ret = dio->ret;
1759                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1760                 return ret;
1761         } else {
1762                 return -EIOCBQUEUED;
1763         }
1764 }
1765
1766 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1767 {
1768         struct file *file = iocb->ki_filp;
1769         struct bch_inode_info *inode = file_bch_inode(file);
1770         struct address_space *mapping = file->f_mapping;
1771         size_t count = iov_iter_count(iter);
1772         ssize_t ret;
1773
1774         if (!count)
1775                 return 0; /* skip atime */
1776
1777         if (iocb->ki_flags & IOCB_DIRECT) {
1778                 struct blk_plug plug;
1779
1780                 ret = filemap_write_and_wait_range(mapping,
1781                                         iocb->ki_pos,
1782                                         iocb->ki_pos + count - 1);
1783                 if (ret < 0)
1784                         return ret;
1785
1786                 file_accessed(file);
1787
1788                 blk_start_plug(&plug);
1789                 ret = bch2_direct_IO_read(iocb, iter);
1790                 blk_finish_plug(&plug);
1791
1792                 if (ret >= 0)
1793                         iocb->ki_pos += ret;
1794         } else {
1795                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1796                 ret = generic_file_read_iter(iocb, iter);
1797                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1798         }
1799
1800         return ret;
1801 }
1802
1803 /* O_DIRECT writes */
1804
1805 static void bch2_dio_write_loop_async(struct bch_write_op *);
1806
1807 static long bch2_dio_write_loop(struct dio_write *dio)
1808 {
1809         bool kthread = (current->flags & PF_KTHREAD) != 0;
1810         struct kiocb *req = dio->req;
1811         struct address_space *mapping = req->ki_filp->f_mapping;
1812         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1813         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1814         struct bio *bio = &dio->op.wbio.bio;
1815         struct bvec_iter_all iter;
1816         struct bio_vec *bv;
1817         unsigned unaligned, iter_count;
1818         bool sync = dio->sync, dropped_locks;
1819         long ret;
1820
1821         if (dio->loop)
1822                 goto loop;
1823
1824         down(&c->io_in_flight);
1825
1826         while (1) {
1827                 iter_count = dio->iter.count;
1828
1829                 if (kthread)
1830                         kthread_use_mm(dio->mm);
1831                 BUG_ON(current->faults_disabled_mapping);
1832                 current->faults_disabled_mapping = mapping;
1833
1834                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1835
1836                 dropped_locks = fdm_dropped_locks();
1837
1838                 current->faults_disabled_mapping = NULL;
1839                 if (kthread)
1840                         kthread_unuse_mm(dio->mm);
1841
1842                 /*
1843                  * If the fault handler returned an error but also signalled
1844                  * that it dropped & retook ei_pagecache_lock, we just need to
1845                  * re-shoot down the page cache and retry:
1846                  */
1847                 if (dropped_locks && ret)
1848                         ret = 0;
1849
1850                 if (unlikely(ret < 0))
1851                         goto err;
1852
1853                 if (unlikely(dropped_locks)) {
1854                         ret = write_invalidate_inode_pages_range(mapping,
1855                                         req->ki_pos,
1856                                         req->ki_pos + iter_count - 1);
1857                         if (unlikely(ret))
1858                                 goto err;
1859
1860                         if (!bio->bi_iter.bi_size)
1861                                 continue;
1862                 }
1863
1864                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1865                 bio->bi_iter.bi_size -= unaligned;
1866                 iov_iter_revert(&dio->iter, unaligned);
1867
1868                 if (!bio->bi_iter.bi_size) {
1869                         /*
1870                          * bio_iov_iter_get_pages was only able to get <
1871                          * blocksize worth of pages:
1872                          */
1873                         ret = -EFAULT;
1874                         goto err;
1875                 }
1876
1877                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1878                 dio->op.end_io          = bch2_dio_write_loop_async;
1879                 dio->op.target          = dio->op.opts.foreground_target;
1880                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1881                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1882                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1883                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1884
1885                 if ((req->ki_flags & IOCB_DSYNC) &&
1886                     !c->opts.journal_flush_disabled)
1887                         dio->op.flags |= BCH_WRITE_FLUSH;
1888                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
1889
1890                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1891                                                 dio->op.opts.data_replicas, 0);
1892                 if (unlikely(ret) &&
1893                     !bch2_check_range_allocated(c, dio->op.pos,
1894                                 bio_sectors(bio),
1895                                 dio->op.opts.data_replicas,
1896                                 dio->op.opts.compression != 0))
1897                         goto err;
1898
1899                 task_io_account_write(bio->bi_iter.bi_size);
1900
1901                 if (!dio->sync && !dio->loop && dio->iter.count) {
1902                         struct iovec *iov = dio->inline_vecs;
1903
1904                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1905                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1906                                               GFP_KERNEL);
1907                                 if (unlikely(!iov)) {
1908                                         dio->sync = sync = true;
1909                                         goto do_io;
1910                                 }
1911
1912                                 dio->free_iov = true;
1913                         }
1914
1915                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1916                         dio->iter.iov = iov;
1917                 }
1918 do_io:
1919                 dio->loop = true;
1920                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1921
1922                 if (sync)
1923                         wait_for_completion(&dio->done);
1924                 else
1925                         return -EIOCBQUEUED;
1926 loop:
1927                 i_sectors_acct(c, inode, &dio->quota_res,
1928                                dio->op.i_sectors_delta);
1929                 req->ki_pos += (u64) dio->op.written << 9;
1930                 dio->written += dio->op.written;
1931
1932                 spin_lock(&inode->v.i_lock);
1933                 if (req->ki_pos > inode->v.i_size)
1934                         i_size_write(&inode->v, req->ki_pos);
1935                 spin_unlock(&inode->v.i_lock);
1936
1937                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1938                         bio_for_each_segment_all(bv, bio, iter)
1939                                 put_page(bv->bv_page);
1940                 bio->bi_vcnt = 0;
1941
1942                 if (dio->op.error) {
1943                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
1944                         break;
1945                 }
1946
1947                 if (!dio->iter.count)
1948                         break;
1949
1950                 bio_reset(bio);
1951                 reinit_completion(&dio->done);
1952         }
1953
1954         ret = dio->op.error ?: ((long) dio->written << 9);
1955 err:
1956         up(&c->io_in_flight);
1957         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1958         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1959
1960         if (dio->free_iov)
1961                 kfree(dio->iter.iov);
1962
1963         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1964                 bio_for_each_segment_all(bv, bio, iter)
1965                         put_page(bv->bv_page);
1966         bio_put(bio);
1967
1968         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1969         inode_dio_end(&inode->v);
1970
1971         if (!sync) {
1972                 req->ki_complete(req, ret, 0);
1973                 ret = -EIOCBQUEUED;
1974         }
1975         return ret;
1976 }
1977
1978 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1979 {
1980         struct dio_write *dio = container_of(op, struct dio_write, op);
1981
1982         if (dio->sync)
1983                 complete(&dio->done);
1984         else
1985                 bch2_dio_write_loop(dio);
1986 }
1987
1988 static noinline
1989 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1990 {
1991         struct file *file = req->ki_filp;
1992         struct address_space *mapping = file->f_mapping;
1993         struct bch_inode_info *inode = file_bch_inode(file);
1994         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1995         struct dio_write *dio;
1996         struct bio *bio;
1997         bool locked = true, extending;
1998         ssize_t ret;
1999
2000         prefetch(&c->opts);
2001         prefetch((void *) &c->opts + 64);
2002         prefetch(&inode->ei_inode);
2003         prefetch((void *) &inode->ei_inode + 64);
2004
2005         inode_lock(&inode->v);
2006
2007         ret = generic_write_checks(req, iter);
2008         if (unlikely(ret <= 0))
2009                 goto err;
2010
2011         ret = file_remove_privs(file);
2012         if (unlikely(ret))
2013                 goto err;
2014
2015         ret = file_update_time(file);
2016         if (unlikely(ret))
2017                 goto err;
2018
2019         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2020                 goto err;
2021
2022         inode_dio_begin(&inode->v);
2023         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2024
2025         extending = req->ki_pos + iter->count > inode->v.i_size;
2026         if (!extending) {
2027                 inode_unlock(&inode->v);
2028                 locked = false;
2029         }
2030
2031         bio = bio_alloc_bioset(GFP_KERNEL,
2032                                iov_iter_is_bvec(iter)
2033                                ? 0
2034                                : iov_iter_npages(iter, BIO_MAX_VECS),
2035                                &c->dio_write_bioset);
2036         dio = container_of(bio, struct dio_write, op.wbio.bio);
2037         init_completion(&dio->done);
2038         dio->req                = req;
2039         dio->mm                 = current->mm;
2040         dio->loop               = false;
2041         dio->sync               = is_sync_kiocb(req) || extending;
2042         dio->free_iov           = false;
2043         dio->quota_res.sectors  = 0;
2044         dio->written            = 0;
2045         dio->iter               = *iter;
2046
2047         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2048                                          iter->count >> 9, true);
2049         if (unlikely(ret))
2050                 goto err_put_bio;
2051
2052         ret = write_invalidate_inode_pages_range(mapping,
2053                                         req->ki_pos,
2054                                         req->ki_pos + iter->count - 1);
2055         if (unlikely(ret))
2056                 goto err_put_bio;
2057
2058         ret = bch2_dio_write_loop(dio);
2059 err:
2060         if (locked)
2061                 inode_unlock(&inode->v);
2062         return ret;
2063 err_put_bio:
2064         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2065         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2066         bio_put(bio);
2067         inode_dio_end(&inode->v);
2068         goto err;
2069 }
2070
2071 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2072 {
2073         struct file *file = iocb->ki_filp;
2074         struct bch_inode_info *inode = file_bch_inode(file);
2075         ssize_t ret;
2076
2077         if (iocb->ki_flags & IOCB_DIRECT)
2078                 return bch2_direct_write(iocb, from);
2079
2080         /* We can write back this queue in page reclaim */
2081         current->backing_dev_info = inode_to_bdi(&inode->v);
2082         inode_lock(&inode->v);
2083
2084         ret = generic_write_checks(iocb, from);
2085         if (ret <= 0)
2086                 goto unlock;
2087
2088         ret = file_remove_privs(file);
2089         if (ret)
2090                 goto unlock;
2091
2092         ret = file_update_time(file);
2093         if (ret)
2094                 goto unlock;
2095
2096         ret = bch2_buffered_write(iocb, from);
2097         if (likely(ret > 0))
2098                 iocb->ki_pos += ret;
2099 unlock:
2100         inode_unlock(&inode->v);
2101         current->backing_dev_info = NULL;
2102
2103         if (ret > 0)
2104                 ret = generic_write_sync(iocb, ret);
2105
2106         return ret;
2107 }
2108
2109 /* fsync: */
2110
2111 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2112 {
2113         struct bch_inode_info *inode = file_bch_inode(file);
2114         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2115         int ret, ret2;
2116
2117         ret = file_write_and_wait_range(file, start, end);
2118         if (ret)
2119                 return ret;
2120
2121         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2122                 goto out;
2123
2124         ret = sync_inode_metadata(&inode->v, 1);
2125         if (ret)
2126                 return ret;
2127 out:
2128         if (!c->opts.journal_flush_disabled)
2129                 ret = bch2_journal_flush_seq(&c->journal,
2130                                              inode->ei_journal_seq);
2131         ret2 = file_check_and_advance_wb_err(file);
2132
2133         return ret ?: ret2;
2134 }
2135
2136 /* truncate: */
2137
2138 static inline int range_has_data(struct bch_fs *c,
2139                                   struct bpos start,
2140                                   struct bpos end)
2141 {
2142         struct btree_trans trans;
2143         struct btree_iter *iter;
2144         struct bkey_s_c k;
2145         int ret = 0;
2146
2147         bch2_trans_init(&trans, c, 0, 0);
2148
2149         for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2150                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2151                         break;
2152
2153                 if (bkey_extent_is_data(k.k)) {
2154                         ret = 1;
2155                         break;
2156                 }
2157         }
2158         bch2_trans_iter_put(&trans, iter);
2159
2160         return bch2_trans_exit(&trans) ?: ret;
2161 }
2162
2163 static int __bch2_truncate_page(struct bch_inode_info *inode,
2164                                 pgoff_t index, loff_t start, loff_t end)
2165 {
2166         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2167         struct address_space *mapping = inode->v.i_mapping;
2168         struct bch_page_state *s;
2169         unsigned start_offset = start & (PAGE_SIZE - 1);
2170         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2171         unsigned i;
2172         struct page *page;
2173         int ret = 0;
2174
2175         /* Page boundary? Nothing to do */
2176         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2177               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2178                 return 0;
2179
2180         /* Above i_size? */
2181         if (index << PAGE_SHIFT >= inode->v.i_size)
2182                 return 0;
2183
2184         page = find_lock_page(mapping, index);
2185         if (!page) {
2186                 /*
2187                  * XXX: we're doing two index lookups when we end up reading the
2188                  * page
2189                  */
2190                 ret = range_has_data(c,
2191                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2192                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2193                 if (ret <= 0)
2194                         return ret;
2195
2196                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2197                 if (unlikely(!page)) {
2198                         ret = -ENOMEM;
2199                         goto out;
2200                 }
2201         }
2202
2203         s = bch2_page_state_create(page, 0);
2204         if (!s) {
2205                 ret = -ENOMEM;
2206                 goto unlock;
2207         }
2208
2209         if (!PageUptodate(page)) {
2210                 ret = bch2_read_single_page(page, mapping);
2211                 if (ret)
2212                         goto unlock;
2213         }
2214
2215         if (index != start >> PAGE_SHIFT)
2216                 start_offset = 0;
2217         if (index != end >> PAGE_SHIFT)
2218                 end_offset = PAGE_SIZE;
2219
2220         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2221              i < round_down(end_offset, block_bytes(c)) >> 9;
2222              i++) {
2223                 s->s[i].nr_replicas     = 0;
2224                 s->s[i].state           = SECTOR_UNALLOCATED;
2225         }
2226
2227         zero_user_segment(page, start_offset, end_offset);
2228
2229         /*
2230          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2231          *
2232          * XXX: because we aren't currently tracking whether the page has actual
2233          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2234          */
2235         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2236         BUG_ON(ret);
2237
2238         /*
2239          * This removes any writeable userspace mappings; we need to force
2240          * .page_mkwrite to be called again before any mmapped writes, to
2241          * redirty the full page:
2242          */
2243         page_mkclean(page);
2244         __set_page_dirty_nobuffers(page);
2245 unlock:
2246         unlock_page(page);
2247         put_page(page);
2248 out:
2249         return ret;
2250 }
2251
2252 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2253 {
2254         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2255                                     from, round_up(from, PAGE_SIZE));
2256 }
2257
2258 static int bch2_extend(struct user_namespace *mnt_userns,
2259                        struct bch_inode_info *inode,
2260                        struct bch_inode_unpacked *inode_u,
2261                        struct iattr *iattr)
2262 {
2263         struct address_space *mapping = inode->v.i_mapping;
2264         int ret;
2265
2266         /*
2267          * sync appends:
2268          *
2269          * this has to be done _before_ extending i_size:
2270          */
2271         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2272         if (ret)
2273                 return ret;
2274
2275         truncate_setsize(&inode->v, iattr->ia_size);
2276
2277         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2278 }
2279
2280 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2281                                    struct bch_inode_unpacked *bi,
2282                                    void *p)
2283 {
2284         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2285         return 0;
2286 }
2287
2288 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2289                                   struct bch_inode_unpacked *bi, void *p)
2290 {
2291         u64 *new_i_size = p;
2292
2293         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2294         bi->bi_size = *new_i_size;
2295         return 0;
2296 }
2297
2298 int bch2_truncate(struct user_namespace *mnt_userns,
2299                   struct bch_inode_info *inode, struct iattr *iattr)
2300 {
2301         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2302         struct address_space *mapping = inode->v.i_mapping;
2303         struct bch_inode_unpacked inode_u;
2304         struct btree_trans trans;
2305         struct btree_iter *iter;
2306         u64 new_i_size = iattr->ia_size;
2307         s64 i_sectors_delta = 0;
2308         int ret = 0;
2309
2310         /*
2311          * If the truncate call with change the size of the file, the
2312          * cmtimes should be updated. If the size will not change, we
2313          * do not need to update the cmtimes.
2314          */
2315         if (iattr->ia_size != inode->v.i_size) {
2316                 if (!(iattr->ia_valid & ATTR_MTIME))
2317                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2318                 if (!(iattr->ia_valid & ATTR_CTIME))
2319                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2320                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2321         }
2322
2323         inode_dio_wait(&inode->v);
2324         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2325
2326         /*
2327          * fetch current on disk i_size: inode is locked, i_size can only
2328          * increase underneath us:
2329          */
2330         bch2_trans_init(&trans, c, 0, 0);
2331         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2332         ret = PTR_ERR_OR_ZERO(iter);
2333         bch2_trans_iter_put(&trans, iter);
2334         bch2_trans_exit(&trans);
2335
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         src = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2545                         POS(inode->v.i_ino, src_start >> 9),
2546                         BTREE_ITER_INTENT);
2547         dst = bch2_trans_copy_iter(&trans, src);
2548         del = bch2_trans_copy_iter(&trans, 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                 k = insert
2561                         ? bch2_btree_iter_peek_prev(src)
2562                         : bch2_btree_iter_peek(src);
2563                 if ((ret = bkey_err(k)))
2564                         continue;
2565
2566                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2567                         break;
2568
2569                 if (insert &&
2570                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2571                         break;
2572 reassemble:
2573                 bch2_bkey_buf_reassemble(&copy, c, k);
2574
2575                 if (insert &&
2576                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2577                         bch2_cut_front(move_pos, copy.k);
2578
2579                 copy.k->k.p.offset += shift >> 9;
2580                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2581
2582                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2583                 if (ret)
2584                         continue;
2585
2586                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2587                         if (insert) {
2588                                 move_pos = atomic_end;
2589                                 move_pos.offset -= shift >> 9;
2590                                 goto reassemble;
2591                         } else {
2592                                 bch2_cut_back(atomic_end, copy.k);
2593                         }
2594                 }
2595
2596                 bkey_init(&delete.k);
2597                 delete.k.p = copy.k->k.p;
2598                 delete.k.size = copy.k->k.size;
2599                 delete.k.p.offset -= shift >> 9;
2600                 bch2_btree_iter_set_pos(del, bkey_start_pos(&delete.k));
2601
2602                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2603
2604                 if (copy.k->k.size == k.k->size) {
2605                         /*
2606                          * If we're moving the entire extent, we can skip
2607                          * running triggers:
2608                          */
2609                         trigger_flags |= BTREE_TRIGGER_NORUN;
2610                 } else {
2611                         /* We might end up splitting compressed extents: */
2612                         unsigned nr_ptrs =
2613                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2614
2615                         ret = bch2_disk_reservation_get(c, &disk_res,
2616                                         copy.k->k.size, nr_ptrs,
2617                                         BCH_DISK_RESERVATION_NOFAIL);
2618                         BUG_ON(ret);
2619                 }
2620
2621                 ret =   bch2_btree_iter_traverse(del) ?:
2622                         bch2_trans_update(&trans, del, &delete, trigger_flags) ?:
2623                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2624                         bch2_trans_commit(&trans, &disk_res,
2625                                           &inode->ei_journal_seq,
2626                                           BTREE_INSERT_NOFAIL);
2627                 bch2_disk_reservation_put(c, &disk_res);
2628
2629                 if (!ret)
2630                         bch2_btree_iter_set_pos(src, next_pos);
2631         }
2632         bch2_trans_iter_put(&trans, del);
2633         bch2_trans_iter_put(&trans, dst);
2634         bch2_trans_iter_put(&trans, src);
2635         bch2_trans_exit(&trans);
2636         bch2_bkey_buf_exit(&copy, c);
2637
2638         if (ret)
2639                 goto err;
2640
2641         if (!insert) {
2642                 i_size_write(&inode->v, new_size);
2643                 mutex_lock(&inode->ei_update_lock);
2644                 ret = bch2_write_inode_size(c, inode, new_size,
2645                                             ATTR_MTIME|ATTR_CTIME);
2646                 mutex_unlock(&inode->ei_update_lock);
2647         }
2648 err:
2649         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2650         inode_unlock(&inode->v);
2651         return ret;
2652 }
2653
2654 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2655                              u64 start_sector, u64 end_sector)
2656 {
2657         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2658         struct btree_trans trans;
2659         struct btree_iter *iter;
2660         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2661         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2662         int ret = 0;
2663
2664         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2665
2666         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2667                         POS(inode->v.i_ino, start_sector),
2668                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2669
2670         while (!ret && bkey_cmp(iter->pos, end_pos) < 0) {
2671                 s64 i_sectors_delta = 0;
2672                 struct disk_reservation disk_res = { 0 };
2673                 struct quota_res quota_res = { 0 };
2674                 struct bkey_i_reservation reservation;
2675                 struct bkey_s_c k;
2676                 unsigned sectors;
2677
2678                 bch2_trans_begin(&trans);
2679
2680                 k = bch2_btree_iter_peek_slot(iter);
2681                 if ((ret = bkey_err(k)))
2682                         goto bkey_err;
2683
2684                 /* already reserved */
2685                 if (k.k->type == KEY_TYPE_reservation &&
2686                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2687                         bch2_btree_iter_next_slot(iter);
2688                         continue;
2689                 }
2690
2691                 if (bkey_extent_is_data(k.k) &&
2692                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2693                         bch2_btree_iter_next_slot(iter);
2694                         continue;
2695                 }
2696
2697                 bkey_reservation_init(&reservation.k_i);
2698                 reservation.k.type      = KEY_TYPE_reservation;
2699                 reservation.k.p         = k.k->p;
2700                 reservation.k.size      = k.k->size;
2701
2702                 bch2_cut_front(iter->pos,       &reservation.k_i);
2703                 bch2_cut_back(end_pos,          &reservation.k_i);
2704
2705                 sectors = reservation.k.size;
2706                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2707
2708                 if (!bkey_extent_is_allocation(k.k)) {
2709                         ret = bch2_quota_reservation_add(c, inode,
2710                                         &quota_res,
2711                                         sectors, true);
2712                         if (unlikely(ret))
2713                                 goto bkey_err;
2714                 }
2715
2716                 if (reservation.v.nr_replicas < replicas ||
2717                     bch2_bkey_sectors_compressed(k)) {
2718                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2719                                                         replicas, 0);
2720                         if (unlikely(ret))
2721                                 goto bkey_err;
2722
2723                         reservation.v.nr_replicas = disk_res.nr_replicas;
2724                 }
2725
2726                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2727                                 &disk_res, &inode->ei_journal_seq,
2728                                 0, &i_sectors_delta, true);
2729                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2730 bkey_err:
2731                 bch2_quota_reservation_put(c, inode, &quota_res);
2732                 bch2_disk_reservation_put(c, &disk_res);
2733                 if (ret == -EINTR)
2734                         ret = 0;
2735         }
2736         bch2_trans_iter_put(&trans, iter);
2737         bch2_trans_exit(&trans);
2738         return ret;
2739 }
2740
2741 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2742                             loff_t offset, loff_t len)
2743 {
2744         struct address_space *mapping = inode->v.i_mapping;
2745         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2746         loff_t end              = offset + len;
2747         loff_t block_start      = round_down(offset,    block_bytes(c));
2748         loff_t block_end        = round_up(end,         block_bytes(c));
2749         int ret;
2750
2751         inode_lock(&inode->v);
2752         inode_dio_wait(&inode->v);
2753         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2754
2755         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2756                 ret = inode_newsize_ok(&inode->v, end);
2757                 if (ret)
2758                         goto err;
2759         }
2760
2761         if (mode & FALLOC_FL_ZERO_RANGE) {
2762                 ret = __bch2_truncate_page(inode,
2763                                            offset >> PAGE_SHIFT,
2764                                            offset, end);
2765
2766                 if (!ret &&
2767                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2768                         ret = __bch2_truncate_page(inode,
2769                                                    end >> PAGE_SHIFT,
2770                                                    offset, end);
2771
2772                 if (unlikely(ret))
2773                         goto err;
2774
2775                 truncate_pagecache_range(&inode->v, offset, end - 1);
2776         }
2777
2778         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
2779         if (ret)
2780                 goto err;
2781
2782         /*
2783          * Do we need to extend the file?
2784          *
2785          * If we zeroed up to the end of the file, we dropped whatever writes
2786          * were going to write out the current i_size, so we have to extend
2787          * manually even if FL_KEEP_SIZE was set:
2788          */
2789         if (end >= inode->v.i_size &&
2790             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2791              (mode & FALLOC_FL_ZERO_RANGE))) {
2792
2793                 /*
2794                  * Sync existing appends before extending i_size,
2795                  * as in bch2_extend():
2796                  */
2797                 ret = filemap_write_and_wait_range(mapping,
2798                                         inode->ei_inode.bi_size, S64_MAX);
2799                 if (ret)
2800                         goto err;
2801
2802                 if (mode & FALLOC_FL_KEEP_SIZE)
2803                         end = inode->v.i_size;
2804                 else
2805                         i_size_write(&inode->v, end);
2806
2807                 mutex_lock(&inode->ei_update_lock);
2808                 ret = bch2_write_inode_size(c, inode, end, 0);
2809                 mutex_unlock(&inode->ei_update_lock);
2810         }
2811 err:
2812         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2813         inode_unlock(&inode->v);
2814         return ret;
2815 }
2816
2817 long bch2_fallocate_dispatch(struct file *file, int mode,
2818                              loff_t offset, loff_t len)
2819 {
2820         struct bch_inode_info *inode = file_bch_inode(file);
2821         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2822         long ret;
2823
2824         if (!percpu_ref_tryget(&c->writes))
2825                 return -EROFS;
2826
2827         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2828                 ret = bchfs_fallocate(inode, mode, offset, len);
2829         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2830                 ret = bchfs_fpunch(inode, offset, len);
2831         else if (mode == FALLOC_FL_INSERT_RANGE)
2832                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2833         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2834                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2835         else
2836                 ret = -EOPNOTSUPP;
2837
2838         percpu_ref_put(&c->writes);
2839
2840         return ret;
2841 }
2842
2843 static void mark_range_unallocated(struct bch_inode_info *inode,
2844                                    loff_t start, loff_t end)
2845 {
2846         pgoff_t index = start >> PAGE_SHIFT;
2847         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2848         struct pagevec pvec;
2849
2850         pagevec_init(&pvec);
2851
2852         do {
2853                 unsigned nr_pages, i, j;
2854
2855                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2856                                                 &index, end_index);
2857                 if (nr_pages == 0)
2858                         break;
2859
2860                 for (i = 0; i < nr_pages; i++) {
2861                         struct page *page = pvec.pages[i];
2862                         struct bch_page_state *s;
2863
2864                         lock_page(page);
2865                         s = bch2_page_state(page);
2866
2867                         if (s) {
2868                                 spin_lock(&s->lock);
2869                                 for (j = 0; j < PAGE_SECTORS; j++)
2870                                         s->s[j].nr_replicas = 0;
2871                                 spin_unlock(&s->lock);
2872                         }
2873
2874                         unlock_page(page);
2875                 }
2876                 pagevec_release(&pvec);
2877         } while (index <= end_index);
2878 }
2879
2880 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2881                              struct file *file_dst, loff_t pos_dst,
2882                              loff_t len, unsigned remap_flags)
2883 {
2884         struct bch_inode_info *src = file_bch_inode(file_src);
2885         struct bch_inode_info *dst = file_bch_inode(file_dst);
2886         struct bch_fs *c = src->v.i_sb->s_fs_info;
2887         s64 i_sectors_delta = 0;
2888         u64 aligned_len;
2889         loff_t ret = 0;
2890
2891         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2892                 return -EINVAL;
2893
2894         if (remap_flags & REMAP_FILE_DEDUP)
2895                 return -EOPNOTSUPP;
2896
2897         if ((pos_src & (block_bytes(c) - 1)) ||
2898             (pos_dst & (block_bytes(c) - 1)))
2899                 return -EINVAL;
2900
2901         if (src == dst &&
2902             abs(pos_src - pos_dst) < len)
2903                 return -EINVAL;
2904
2905         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2906
2907         file_update_time(file_dst);
2908
2909         inode_dio_wait(&src->v);
2910         inode_dio_wait(&dst->v);
2911
2912         ret = generic_remap_file_range_prep(file_src, pos_src,
2913                                             file_dst, pos_dst,
2914                                             &len, remap_flags);
2915         if (ret < 0 || len == 0)
2916                 goto err;
2917
2918         aligned_len = round_up((u64) len, block_bytes(c));
2919
2920         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2921                                 pos_dst, pos_dst + len - 1);
2922         if (ret)
2923                 goto err;
2924
2925         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2926
2927         ret = bch2_remap_range(c,
2928                                POS(dst->v.i_ino, pos_dst >> 9),
2929                                POS(src->v.i_ino, pos_src >> 9),
2930                                aligned_len >> 9,
2931                                &dst->ei_journal_seq,
2932                                pos_dst + len, &i_sectors_delta);
2933         if (ret < 0)
2934                 goto err;
2935
2936         /*
2937          * due to alignment, we might have remapped slightly more than requsted
2938          */
2939         ret = min((u64) ret << 9, (u64) len);
2940
2941         /* XXX get a quota reservation */
2942         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2943
2944         spin_lock(&dst->v.i_lock);
2945         if (pos_dst + ret > dst->v.i_size)
2946                 i_size_write(&dst->v, pos_dst + ret);
2947         spin_unlock(&dst->v.i_lock);
2948
2949         if (((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
2950              IS_SYNC(file_inode(file_dst))) &&
2951             !c->opts.journal_flush_disabled)
2952                 ret = bch2_journal_flush_seq(&c->journal, dst->ei_journal_seq);
2953 err:
2954         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2955
2956         return ret;
2957 }
2958
2959 /* fseek: */
2960
2961 static int page_data_offset(struct page *page, unsigned offset)
2962 {
2963         struct bch_page_state *s = bch2_page_state(page);
2964         unsigned i;
2965
2966         if (s)
2967                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2968                         if (s->s[i].state >= SECTOR_DIRTY)
2969                                 return i << 9;
2970
2971         return -1;
2972 }
2973
2974 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2975                                        loff_t start_offset,
2976                                        loff_t end_offset)
2977 {
2978         struct address_space *mapping = vinode->i_mapping;
2979         struct page *page;
2980         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2981         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2982         pgoff_t index           = start_index;
2983         loff_t ret;
2984         int offset;
2985
2986         while (index <= end_index) {
2987                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2988                         lock_page(page);
2989
2990                         offset = page_data_offset(page,
2991                                         page->index == start_index
2992                                         ? start_offset & (PAGE_SIZE - 1)
2993                                         : 0);
2994                         if (offset >= 0) {
2995                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2996                                             offset,
2997                                             start_offset, end_offset);
2998                                 unlock_page(page);
2999                                 put_page(page);
3000                                 return ret;
3001                         }
3002
3003                         unlock_page(page);
3004                         put_page(page);
3005                 } else {
3006                         break;
3007                 }
3008         }
3009
3010         return end_offset;
3011 }
3012
3013 static loff_t bch2_seek_data(struct file *file, u64 offset)
3014 {
3015         struct bch_inode_info *inode = file_bch_inode(file);
3016         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3017         struct btree_trans trans;
3018         struct btree_iter *iter;
3019         struct bkey_s_c k;
3020         u64 isize, next_data = MAX_LFS_FILESIZE;
3021         int ret;
3022
3023         isize = i_size_read(&inode->v);
3024         if (offset >= isize)
3025                 return -ENXIO;
3026
3027         bch2_trans_init(&trans, c, 0, 0);
3028
3029         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3030                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3031                 if (k.k->p.inode != inode->v.i_ino) {
3032                         break;
3033                 } else if (bkey_extent_is_data(k.k)) {
3034                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3035                         break;
3036                 } else if (k.k->p.offset >> 9 > isize)
3037                         break;
3038         }
3039         bch2_trans_iter_put(&trans, iter);
3040
3041         ret = bch2_trans_exit(&trans) ?: ret;
3042         if (ret)
3043                 return ret;
3044
3045         if (next_data > offset)
3046                 next_data = bch2_seek_pagecache_data(&inode->v,
3047                                                      offset, next_data);
3048
3049         if (next_data >= isize)
3050                 return -ENXIO;
3051
3052         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3053 }
3054
3055 static int __page_hole_offset(struct page *page, unsigned offset)
3056 {
3057         struct bch_page_state *s = bch2_page_state(page);
3058         unsigned i;
3059
3060         if (!s)
3061                 return 0;
3062
3063         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3064                 if (s->s[i].state < SECTOR_DIRTY)
3065                         return i << 9;
3066
3067         return -1;
3068 }
3069
3070 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3071 {
3072         pgoff_t index = offset >> PAGE_SHIFT;
3073         struct page *page;
3074         int pg_offset;
3075         loff_t ret = -1;
3076
3077         page = find_lock_page(mapping, index);
3078         if (!page)
3079                 return offset;
3080
3081         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3082         if (pg_offset >= 0)
3083                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3084
3085         unlock_page(page);
3086
3087         return ret;
3088 }
3089
3090 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3091                                        loff_t start_offset,
3092                                        loff_t end_offset)
3093 {
3094         struct address_space *mapping = vinode->i_mapping;
3095         loff_t offset = start_offset, hole;
3096
3097         while (offset < end_offset) {
3098                 hole = page_hole_offset(mapping, offset);
3099                 if (hole >= 0 && hole <= end_offset)
3100                         return max(start_offset, hole);
3101
3102                 offset += PAGE_SIZE;
3103                 offset &= PAGE_MASK;
3104         }
3105
3106         return end_offset;
3107 }
3108
3109 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3110 {
3111         struct bch_inode_info *inode = file_bch_inode(file);
3112         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3113         struct btree_trans trans;
3114         struct btree_iter *iter;
3115         struct bkey_s_c k;
3116         u64 isize, next_hole = MAX_LFS_FILESIZE;
3117         int ret;
3118
3119         isize = i_size_read(&inode->v);
3120         if (offset >= isize)
3121                 return -ENXIO;
3122
3123         bch2_trans_init(&trans, c, 0, 0);
3124
3125         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3126                            POS(inode->v.i_ino, offset >> 9),
3127                            BTREE_ITER_SLOTS, k, ret) {
3128                 if (k.k->p.inode != inode->v.i_ino) {
3129                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3130                                         offset, MAX_LFS_FILESIZE);
3131                         break;
3132                 } else if (!bkey_extent_is_data(k.k)) {
3133                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3134                                         max(offset, bkey_start_offset(k.k) << 9),
3135                                         k.k->p.offset << 9);
3136
3137                         if (next_hole < k.k->p.offset << 9)
3138                                 break;
3139                 } else {
3140                         offset = max(offset, bkey_start_offset(k.k) << 9);
3141                 }
3142         }
3143         bch2_trans_iter_put(&trans, iter);
3144
3145         ret = bch2_trans_exit(&trans) ?: ret;
3146         if (ret)
3147                 return ret;
3148
3149         if (next_hole > isize)
3150                 next_hole = isize;
3151
3152         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3153 }
3154
3155 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3156 {
3157         switch (whence) {
3158         case SEEK_SET:
3159         case SEEK_CUR:
3160         case SEEK_END:
3161                 return generic_file_llseek(file, offset, whence);
3162         case SEEK_DATA:
3163                 return bch2_seek_data(file, offset);
3164         case SEEK_HOLE:
3165                 return bch2_seek_hole(file, offset);
3166         }
3167
3168         return -EINVAL;
3169 }
3170
3171 void bch2_fs_fsio_exit(struct bch_fs *c)
3172 {
3173         bioset_exit(&c->dio_write_bioset);
3174         bioset_exit(&c->dio_read_bioset);
3175         bioset_exit(&c->writepage_bioset);
3176 }
3177
3178 int bch2_fs_fsio_init(struct bch_fs *c)
3179 {
3180         int ret = 0;
3181
3182         pr_verbose_init(c->opts, "");
3183
3184         if (bioset_init(&c->writepage_bioset,
3185                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3186                         BIOSET_NEED_BVECS) ||
3187             bioset_init(&c->dio_read_bioset,
3188                         4, offsetof(struct dio_read, rbio.bio),
3189                         BIOSET_NEED_BVECS) ||
3190             bioset_init(&c->dio_write_bioset,
3191                         4, offsetof(struct dio_write, op.wbio.bio),
3192                         BIOSET_NEED_BVECS))
3193                 ret = -ENOMEM;
3194
3195         pr_verbose_init(c->opts, "ret %i", ret);
3196         return ret;
3197 }
3198
3199 #endif /* NO_BCACHEFS_FS */