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_on_stack.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         struct bch_read_bio             rbio;
  88 };
  89
  90 /* pagecache_block must be held */
  91 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  92                                               loff_t start, loff_t end)
  93 {
  94         int ret;
  95
  96         /*
  97          * XXX: the way this is currently implemented, we can spin if a process
  98          * is continually redirtying a specific page
  99          */
 100         do {
 101                 if (!mapping->nrpages &&
 102                     !mapping->nrexceptional)
 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_on_stack 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         bkey_on_stack_init(&sk);
 803 retry:
 804         while (1) {
 805                 struct bkey_s_c k;
 806                 unsigned bytes, sectors, offset_into_extent;
 807
 808                 bch2_btree_iter_set_pos(iter,
 809                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 810
 811                 k = bch2_btree_iter_peek_slot(iter);
 812                 ret = bkey_err(k);
 813                 if (ret)
 814                         break;
 815
 816                 offset_into_extent = iter->pos.offset -
 817                         bkey_start_offset(k.k);
 818                 sectors = k.k->size - offset_into_extent;
 819
 820                 bkey_on_stack_reassemble(&sk, c, k);
 821
 822                 ret = bch2_read_indirect_extent(trans,
 823                                         &offset_into_extent, &sk);
 824                 if (ret)
 825                         break;
 826
 827                 k = bkey_i_to_s_c(sk.k);
 828
 829                 sectors = min(sectors, k.k->size - offset_into_extent);
 830
 831                 bch2_trans_unlock(trans);
 832
 833                 if (readpages_iter)
 834                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 835                                             extent_partial_reads_expensive(k));
 836
 837                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 838                 swap(rbio->bio.bi_iter.bi_size, bytes);
 839
 840                 if (rbio->bio.bi_iter.bi_size == bytes)
 841                         flags |= BCH_READ_LAST_FRAGMENT;
 842
 843                 if (bkey_extent_is_allocation(k.k))
 844                         bch2_add_page_sectors(&rbio->bio, k);
 845
 846                 bch2_read_extent(trans, rbio, k, offset_into_extent, flags);
 847
 848                 if (flags & BCH_READ_LAST_FRAGMENT)
 849                         break;
 850
 851                 swap(rbio->bio.bi_iter.bi_size, bytes);
 852                 bio_advance(&rbio->bio, bytes);
 853         }
 854
 855         if (ret == -EINTR)
 856                 goto retry;
 857
 858         if (ret) {
 859                 bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
 860                 bio_endio(&rbio->bio);
 861         }
 862
 863         bkey_on_stack_exit(&sk, c);
 864 }
 865
 866 void bch2_readahead(struct readahead_control *ractl)
 867 {
 868         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 869         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 870         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 871         struct btree_trans trans;
 872         struct btree_iter *iter;
 873         struct page *page;
 874         struct readpages_iter readpages_iter;
 875         int ret;
 876
 877         ret = readpages_iter_init(&readpages_iter, ractl);
 878         BUG_ON(ret);
 879
 880         bch2_trans_init(&trans, c, 0, 0);
 881
 882         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 883                                    BTREE_ITER_SLOTS);
 884
 885         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 886
 887         while ((page = readpage_iter_next(&readpages_iter))) {
 888                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 889                 unsigned n = min_t(unsigned,
 890                                    readpages_iter.nr_pages -
 891                                    readpages_iter.idx,
 892                                    BIO_MAX_PAGES);
 893                 struct bch_read_bio *rbio =
 894                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 895                                   opts);
 896
 897                 readpages_iter.idx++;
 898
 899                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 900                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 901                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 902                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 903
 904                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 905                            &readpages_iter);
 906         }
 907
 908         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 909
 910         bch2_trans_exit(&trans);
 911         kfree(readpages_iter.pages);
 912 }
 913
 914 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 915                              u64 inum, struct page *page)
 916 {
 917         struct btree_trans trans;
 918         struct btree_iter *iter;
 919
 920         bch2_page_state_create(page, __GFP_NOFAIL);
 921
 922         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 923         rbio->bio.bi_iter.bi_sector =
 924                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 925         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 926
 927         bch2_trans_init(&trans, c, 0, 0);
 928         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 929                                    BTREE_ITER_SLOTS);
 930
 931         bchfs_read(&trans, iter, rbio, inum, NULL);
 932
 933         bch2_trans_exit(&trans);
 934 }
 935
 936 int bch2_readpage(struct file *file, struct page *page)
 937 {
 938         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 939         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 940         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 941         struct bch_read_bio *rbio;
 942
 943         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 944         rbio->bio.bi_end_io = bch2_readpages_end_io;
 945
 946         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 947         return 0;
 948 }
 949
 950 static void bch2_read_single_page_end_io(struct bio *bio)
 951 {
 952         complete(bio->bi_private);
 953 }
 954
 955 static int bch2_read_single_page(struct page *page,
 956                                  struct address_space *mapping)
 957 {
 958         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 959         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 960         struct bch_read_bio *rbio;
 961         int ret;
 962         DECLARE_COMPLETION_ONSTACK(done);
 963
 964         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 965                          io_opts(c, &inode->ei_inode));
 966         rbio->bio.bi_private = &done;
 967         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 968
 969         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 970         wait_for_completion(&done);
 971
 972         ret = blk_status_to_errno(rbio->bio.bi_status);
 973         bio_put(&rbio->bio);
 974
 975         if (ret < 0)
 976                 return ret;
 977
 978         SetPageUptodate(page);
 979         return 0;
 980 }
 981
 982 /* writepages: */
 983
 984 struct bch_writepage_state {
 985         struct bch_writepage_io *io;
 986         struct bch_io_opts      opts;
 987 };
 988
 989 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 990                                                                   struct bch_inode_info *inode)
 991 {
 992         return (struct bch_writepage_state) {
 993                 .opts = io_opts(c, &inode->ei_inode)
 994         };
 995 }
 996
 997 static void bch2_writepage_io_free(struct closure *cl)
 998 {
 999         struct bch_writepage_io *io = container_of(cl,
1000                                         struct bch_writepage_io, cl);
1001
1002         bio_put(&io->op.wbio.bio);
1003 }
1004
1005 static void bch2_writepage_io_done(struct closure *cl)
1006 {
1007         struct bch_writepage_io *io = container_of(cl,
1008                                         struct bch_writepage_io, cl);
1009         struct bch_fs *c = io->op.c;
1010         struct bio *bio = &io->op.wbio.bio;
1011         struct bvec_iter_all iter;
1012         struct bio_vec *bvec;
1013         unsigned i;
1014
1015         if (io->op.error) {
1016                 bio_for_each_segment_all(bvec, bio, iter) {
1017                         struct bch_page_state *s;
1018
1019                         SetPageError(bvec->bv_page);
1020                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1021
1022                         s = __bch2_page_state(bvec->bv_page);
1023                         spin_lock(&s->lock);
1024                         for (i = 0; i < PAGE_SECTORS; i++)
1025                                 s->s[i].nr_replicas = 0;
1026                         spin_unlock(&s->lock);
1027                 }
1028         }
1029
1030         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1031                 bio_for_each_segment_all(bvec, bio, iter) {
1032                         struct bch_page_state *s;
1033
1034                         s = __bch2_page_state(bvec->bv_page);
1035                         spin_lock(&s->lock);
1036                         for (i = 0; i < PAGE_SECTORS; i++)
1037                                 s->s[i].nr_replicas = 0;
1038                         spin_unlock(&s->lock);
1039                 }
1040         }
1041
1042         /*
1043          * racing with fallocate can cause us to add fewer sectors than
1044          * expected - but we shouldn't add more sectors than expected:
1045          */
1046         BUG_ON(io->op.i_sectors_delta > 0);
1047
1048         /*
1049          * (error (due to going RO) halfway through a page can screw that up
1050          * slightly)
1051          * XXX wtf?
1052            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1053          */
1054
1055         /*
1056          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1057          * before calling end_page_writeback:
1058          */
1059         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1060
1061         bio_for_each_segment_all(bvec, bio, iter) {
1062                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1063
1064                 if (atomic_dec_and_test(&s->write_count))
1065                         end_page_writeback(bvec->bv_page);
1066         }
1067
1068         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1069 }
1070
1071 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1072 {
1073         struct bch_writepage_io *io = w->io;
1074
1075         w->io = NULL;
1076         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1077         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1078 }
1079
1080 /*
1081  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1082  * possible, else allocating a new one:
1083  */
1084 static void bch2_writepage_io_alloc(struct bch_fs *c,
1085                                     struct writeback_control *wbc,
1086                                     struct bch_writepage_state *w,
1087                                     struct bch_inode_info *inode,
1088                                     u64 sector,
1089                                     unsigned nr_replicas)
1090 {
1091         struct bch_write_op *op;
1092
1093         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1094                                               BIO_MAX_PAGES,
1095                                               &c->writepage_bioset),
1096                              struct bch_writepage_io, op.wbio.bio);
1097
1098         closure_init(&w->io->cl, NULL);
1099         w->io->inode            = inode;
1100
1101         op                      = &w->io->op;
1102         bch2_write_op_init(op, c, w->opts);
1103         op->target              = w->opts.foreground_target;
1104         op_journal_seq_set(op, &inode->ei_journal_seq);
1105         op->nr_replicas         = nr_replicas;
1106         op->res.nr_replicas     = nr_replicas;
1107         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1108         op->pos                 = POS(inode->v.i_ino, sector);
1109         op->wbio.bio.bi_iter.bi_sector = sector;
1110         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1111 }
1112
1113 static int __bch2_writepage(struct page *page,
1114                             struct writeback_control *wbc,
1115                             void *data)
1116 {
1117         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1118         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1119         struct bch_writepage_state *w = data;
1120         struct bch_page_state *s, orig;
1121         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1122         loff_t i_size = i_size_read(&inode->v);
1123         pgoff_t end_index = i_size >> PAGE_SHIFT;
1124         int ret;
1125
1126         EBUG_ON(!PageUptodate(page));
1127
1128         /* Is the page fully inside i_size? */
1129         if (page->index < end_index)
1130                 goto do_io;
1131
1132         /* Is the page fully outside i_size? (truncate in progress) */
1133         offset = i_size & (PAGE_SIZE - 1);
1134         if (page->index > end_index || !offset) {
1135                 unlock_page(page);
1136                 return 0;
1137         }
1138
1139         /*
1140          * The page straddles i_size.  It must be zeroed out on each and every
1141          * writepage invocation because it may be mmapped.  "A file is mapped
1142          * in multiples of the page size.  For a file that is not a multiple of
1143          * the  page size, the remaining memory is zeroed when mapped, and
1144          * writes to that region are not written out to the file."
1145          */
1146         zero_user_segment(page, offset, PAGE_SIZE);
1147 do_io:
1148         s = bch2_page_state_create(page, __GFP_NOFAIL);
1149
1150         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1151         if (ret) {
1152                 SetPageError(page);
1153                 mapping_set_error(page->mapping, ret);
1154                 unlock_page(page);
1155                 return 0;
1156         }
1157
1158         /* Before unlocking the page, get copy of reservations: */
1159         orig = *s;
1160
1161         for (i = 0; i < PAGE_SECTORS; i++) {
1162                 if (s->s[i].state < SECTOR_DIRTY)
1163                         continue;
1164
1165                 nr_replicas_this_write =
1166                         min_t(unsigned, nr_replicas_this_write,
1167                               s->s[i].nr_replicas +
1168                               s->s[i].replicas_reserved);
1169         }
1170
1171         for (i = 0; i < PAGE_SECTORS; i++) {
1172                 if (s->s[i].state < SECTOR_DIRTY)
1173                         continue;
1174
1175                 s->s[i].nr_replicas = w->opts.compression
1176                         ? 0 : nr_replicas_this_write;
1177
1178                 s->s[i].replicas_reserved = 0;
1179                 s->s[i].state = SECTOR_ALLOCATED;
1180         }
1181
1182         BUG_ON(atomic_read(&s->write_count));
1183         atomic_set(&s->write_count, 1);
1184
1185         BUG_ON(PageWriteback(page));
1186         set_page_writeback(page);
1187
1188         unlock_page(page);
1189
1190         offset = 0;
1191         while (1) {
1192                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1193                 u64 sector;
1194
1195                 while (offset < PAGE_SECTORS &&
1196                        orig.s[offset].state < SECTOR_DIRTY)
1197                         offset++;
1198
1199                 if (offset == PAGE_SECTORS)
1200                         break;
1201
1202                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1203
1204                 while (offset + sectors < PAGE_SECTORS &&
1205                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1206                         sectors++;
1207
1208                 for (i = offset; i < offset + sectors; i++) {
1209                         reserved_sectors += orig.s[i].replicas_reserved;
1210                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1211                 }
1212
1213                 if (w->io &&
1214                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1215                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1216                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1217                      (BIO_MAX_PAGES * PAGE_SIZE) ||
1218                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1219                         bch2_writepage_do_io(w);
1220
1221                 if (!w->io)
1222                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1223                                                 nr_replicas_this_write);
1224
1225                 atomic_inc(&s->write_count);
1226
1227                 BUG_ON(inode != w->io->inode);
1228                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1229                                      sectors << 9, offset << 9));
1230
1231                 /* Check for writing past i_size: */
1232                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1233                         round_up(i_size, block_bytes(c)));
1234
1235                 w->io->op.res.sectors += reserved_sectors;
1236                 w->io->op.i_sectors_delta -= dirty_sectors;
1237                 w->io->op.new_i_size = i_size;
1238
1239                 offset += sectors;
1240         }
1241
1242         if (atomic_dec_and_test(&s->write_count))
1243                 end_page_writeback(page);
1244
1245         return 0;
1246 }
1247
1248 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1249 {
1250         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1251         struct bch_writepage_state w =
1252                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1253         struct blk_plug plug;
1254         int ret;
1255
1256         blk_start_plug(&plug);
1257         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1258         if (w.io)
1259                 bch2_writepage_do_io(&w);
1260         blk_finish_plug(&plug);
1261         return ret;
1262 }
1263
1264 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1265 {
1266         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1267         struct bch_writepage_state w =
1268                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1269         int ret;
1270
1271         ret = __bch2_writepage(page, wbc, &w);
1272         if (w.io)
1273                 bch2_writepage_do_io(&w);
1274
1275         return ret;
1276 }
1277
1278 /* buffered writes: */
1279
1280 int bch2_write_begin(struct file *file, struct address_space *mapping,
1281                      loff_t pos, unsigned len, unsigned flags,
1282                      struct page **pagep, void **fsdata)
1283 {
1284         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1285         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1286         struct bch2_page_reservation *res;
1287         pgoff_t index = pos >> PAGE_SHIFT;
1288         unsigned offset = pos & (PAGE_SIZE - 1);
1289         struct page *page;
1290         int ret = -ENOMEM;
1291
1292         res = kmalloc(sizeof(*res), GFP_KERNEL);
1293         if (!res)
1294                 return -ENOMEM;
1295
1296         bch2_page_reservation_init(c, inode, res);
1297         *fsdata = res;
1298
1299         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1300
1301         page = grab_cache_page_write_begin(mapping, index, flags);
1302         if (!page)
1303                 goto err_unlock;
1304
1305         if (PageUptodate(page))
1306                 goto out;
1307
1308         /* If we're writing entire page, don't need to read it in first: */
1309         if (len == PAGE_SIZE)
1310                 goto out;
1311
1312         if (!offset && pos + len >= inode->v.i_size) {
1313                 zero_user_segment(page, len, PAGE_SIZE);
1314                 flush_dcache_page(page);
1315                 goto out;
1316         }
1317
1318         if (index > inode->v.i_size >> PAGE_SHIFT) {
1319                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1320                 flush_dcache_page(page);
1321                 goto out;
1322         }
1323 readpage:
1324         ret = bch2_read_single_page(page, mapping);
1325         if (ret)
1326                 goto err;
1327 out:
1328         ret = bch2_page_reservation_get(c, inode, page, res,
1329                                         offset, len, true);
1330         if (ret) {
1331                 if (!PageUptodate(page)) {
1332                         /*
1333                          * If the page hasn't been read in, we won't know if we
1334                          * actually need a reservation - we don't actually need
1335                          * to read here, we just need to check if the page is
1336                          * fully backed by uncompressed data:
1337                          */
1338                         goto readpage;
1339                 }
1340
1341                 goto err;
1342         }
1343
1344         *pagep = page;
1345         return 0;
1346 err:
1347         unlock_page(page);
1348         put_page(page);
1349         *pagep = NULL;
1350 err_unlock:
1351         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1352         kfree(res);
1353         *fsdata = NULL;
1354         return ret;
1355 }
1356
1357 int bch2_write_end(struct file *file, struct address_space *mapping,
1358                    loff_t pos, unsigned len, unsigned copied,
1359                    struct page *page, void *fsdata)
1360 {
1361         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1362         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1363         struct bch2_page_reservation *res = fsdata;
1364         unsigned offset = pos & (PAGE_SIZE - 1);
1365
1366         lockdep_assert_held(&inode->v.i_rwsem);
1367
1368         if (unlikely(copied < len && !PageUptodate(page))) {
1369                 /*
1370                  * The page needs to be read in, but that would destroy
1371                  * our partial write - simplest thing is to just force
1372                  * userspace to redo the write:
1373                  */
1374                 zero_user(page, 0, PAGE_SIZE);
1375                 flush_dcache_page(page);
1376                 copied = 0;
1377         }
1378
1379         spin_lock(&inode->v.i_lock);
1380         if (pos + copied > inode->v.i_size)
1381                 i_size_write(&inode->v, pos + copied);
1382         spin_unlock(&inode->v.i_lock);
1383
1384         if (copied) {
1385                 if (!PageUptodate(page))
1386                         SetPageUptodate(page);
1387
1388                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1389
1390                 inode->ei_last_dirtied = (unsigned long) current;
1391         }
1392
1393         unlock_page(page);
1394         put_page(page);
1395         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1396
1397         bch2_page_reservation_put(c, inode, res);
1398         kfree(res);
1399
1400         return copied;
1401 }
1402
1403 #define WRITE_BATCH_PAGES       32
1404
1405 static int __bch2_buffered_write(struct bch_inode_info *inode,
1406                                  struct address_space *mapping,
1407                                  struct iov_iter *iter,
1408                                  loff_t pos, unsigned len)
1409 {
1410         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1411         struct page *pages[WRITE_BATCH_PAGES];
1412         struct bch2_page_reservation res;
1413         unsigned long index = pos >> PAGE_SHIFT;
1414         unsigned offset = pos & (PAGE_SIZE - 1);
1415         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1416         unsigned i, reserved = 0, set_dirty = 0;
1417         unsigned copied = 0, nr_pages_copied = 0;
1418         int ret = 0;
1419
1420         BUG_ON(!len);
1421         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1422
1423         bch2_page_reservation_init(c, inode, &res);
1424
1425         for (i = 0; i < nr_pages; i++) {
1426                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1427                 if (!pages[i]) {
1428                         nr_pages = i;
1429                         if (!i) {
1430                                 ret = -ENOMEM;
1431                                 goto out;
1432                         }
1433                         len = min_t(unsigned, len,
1434                                     nr_pages * PAGE_SIZE - offset);
1435                         break;
1436                 }
1437         }
1438
1439         if (offset && !PageUptodate(pages[0])) {
1440                 ret = bch2_read_single_page(pages[0], mapping);
1441                 if (ret)
1442                         goto out;
1443         }
1444
1445         if ((pos + len) & (PAGE_SIZE - 1) &&
1446             !PageUptodate(pages[nr_pages - 1])) {
1447                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1448                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1449                 } else {
1450                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1451                         if (ret)
1452                                 goto out;
1453                 }
1454         }
1455
1456         while (reserved < len) {
1457                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1458                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1459                 unsigned pg_len = min_t(unsigned, len - reserved,
1460                                         PAGE_SIZE - pg_offset);
1461 retry_reservation:
1462                 ret = bch2_page_reservation_get(c, inode, page, &res,
1463                                                 pg_offset, pg_len, true);
1464
1465                 if (ret && !PageUptodate(page)) {
1466                         ret = bch2_read_single_page(page, mapping);
1467                         if (!ret)
1468                                 goto retry_reservation;
1469                 }
1470
1471                 if (ret)
1472                         goto out;
1473
1474                 reserved += pg_len;
1475         }
1476
1477         if (mapping_writably_mapped(mapping))
1478                 for (i = 0; i < nr_pages; i++)
1479                         flush_dcache_page(pages[i]);
1480
1481         while (copied < len) {
1482                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1483                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1484                 unsigned pg_len = min_t(unsigned, len - copied,
1485                                         PAGE_SIZE - pg_offset);
1486                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1487                                                 iter, pg_offset, pg_len);
1488
1489                 if (!pg_copied)
1490                         break;
1491
1492                 if (!PageUptodate(page) &&
1493                     pg_copied != PAGE_SIZE &&
1494                     pos + copied + pg_copied < inode->v.i_size) {
1495                         zero_user(page, 0, PAGE_SIZE);
1496                         break;
1497                 }
1498
1499                 flush_dcache_page(page);
1500                 iov_iter_advance(iter, pg_copied);
1501                 copied += pg_copied;
1502
1503                 if (pg_copied != pg_len)
1504                         break;
1505         }
1506
1507         if (!copied)
1508                 goto out;
1509
1510         spin_lock(&inode->v.i_lock);
1511         if (pos + copied > inode->v.i_size)
1512                 i_size_write(&inode->v, pos + copied);
1513         spin_unlock(&inode->v.i_lock);
1514
1515         while (set_dirty < copied) {
1516                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1517                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1518                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1519                                         PAGE_SIZE - pg_offset);
1520
1521                 if (!PageUptodate(page))
1522                         SetPageUptodate(page);
1523
1524                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1525                 unlock_page(page);
1526                 put_page(page);
1527
1528                 set_dirty += pg_len;
1529         }
1530
1531         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1532         inode->ei_last_dirtied = (unsigned long) current;
1533 out:
1534         for (i = nr_pages_copied; i < nr_pages; i++) {
1535                 unlock_page(pages[i]);
1536                 put_page(pages[i]);
1537         }
1538
1539         bch2_page_reservation_put(c, inode, &res);
1540
1541         return copied ?: ret;
1542 }
1543
1544 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1545 {
1546         struct file *file = iocb->ki_filp;
1547         struct address_space *mapping = file->f_mapping;
1548         struct bch_inode_info *inode = file_bch_inode(file);
1549         loff_t pos = iocb->ki_pos;
1550         ssize_t written = 0;
1551         int ret = 0;
1552
1553         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1554
1555         do {
1556                 unsigned offset = pos & (PAGE_SIZE - 1);
1557                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1558                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1559 again:
1560                 /*
1561                  * Bring in the user page that we will copy from _first_.
1562                  * Otherwise there's a nasty deadlock on copying from the
1563                  * same page as we're writing to, without it being marked
1564                  * up-to-date.
1565                  *
1566                  * Not only is this an optimisation, but it is also required
1567                  * to check that the address is actually valid, when atomic
1568                  * usercopies are used, below.
1569                  */
1570                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1571                         bytes = min_t(unsigned long, iov_iter_count(iter),
1572                                       PAGE_SIZE - offset);
1573
1574                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1575                                 ret = -EFAULT;
1576                                 break;
1577                         }
1578                 }
1579
1580                 if (unlikely(fatal_signal_pending(current))) {
1581                         ret = -EINTR;
1582                         break;
1583                 }
1584
1585                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1586                 if (unlikely(ret < 0))
1587                         break;
1588
1589                 cond_resched();
1590
1591                 if (unlikely(ret == 0)) {
1592                         /*
1593                          * If we were unable to copy any data at all, we must
1594                          * fall back to a single segment length write.
1595                          *
1596                          * If we didn't fallback here, we could livelock
1597                          * because not all segments in the iov can be copied at
1598                          * once without a pagefault.
1599                          */
1600                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1601                                       iov_iter_single_seg_count(iter));
1602                         goto again;
1603                 }
1604                 pos += ret;
1605                 written += ret;
1606                 ret = 0;
1607
1608                 balance_dirty_pages_ratelimited(mapping);
1609         } while (iov_iter_count(iter));
1610
1611         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1612
1613         return written ? written : ret;
1614 }
1615
1616 /* O_DIRECT reads */
1617
1618 static void bch2_dio_read_complete(struct closure *cl)
1619 {
1620         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1621
1622         dio->req->ki_complete(dio->req, dio->ret, 0);
1623         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1624 }
1625
1626 static void bch2_direct_IO_read_endio(struct bio *bio)
1627 {
1628         struct dio_read *dio = bio->bi_private;
1629
1630         if (bio->bi_status)
1631                 dio->ret = blk_status_to_errno(bio->bi_status);
1632
1633         closure_put(&dio->cl);
1634 }
1635
1636 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1637 {
1638         bch2_direct_IO_read_endio(bio);
1639         bio_check_pages_dirty(bio);     /* transfers ownership */
1640 }
1641
1642 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1643 {
1644         struct file *file = req->ki_filp;
1645         struct bch_inode_info *inode = file_bch_inode(file);
1646         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1647         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1648         struct dio_read *dio;
1649         struct bio *bio;
1650         loff_t offset = req->ki_pos;
1651         bool sync = is_sync_kiocb(req);
1652         size_t shorten;
1653         ssize_t ret;
1654
1655         if ((offset|iter->count) & (block_bytes(c) - 1))
1656                 return -EINVAL;
1657
1658         ret = min_t(loff_t, iter->count,
1659                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1660
1661         if (!ret)
1662                 return ret;
1663
1664         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1665         iter->count -= shorten;
1666
1667         bio = bio_alloc_bioset(GFP_KERNEL,
1668                                iov_iter_npages(iter, BIO_MAX_PAGES),
1669                                &c->dio_read_bioset);
1670
1671         bio->bi_end_io = bch2_direct_IO_read_endio;
1672
1673         dio = container_of(bio, struct dio_read, rbio.bio);
1674         closure_init(&dio->cl, NULL);
1675
1676         /*
1677          * this is a _really_ horrible hack just to avoid an atomic sub at the
1678          * end:
1679          */
1680         if (!sync) {
1681                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1682                 atomic_set(&dio->cl.remaining,
1683                            CLOSURE_REMAINING_INITIALIZER -
1684                            CLOSURE_RUNNING +
1685                            CLOSURE_DESTRUCTOR);
1686         } else {
1687                 atomic_set(&dio->cl.remaining,
1688                            CLOSURE_REMAINING_INITIALIZER + 1);
1689         }
1690
1691         dio->req        = req;
1692         dio->ret        = ret;
1693
1694         goto start;
1695         while (iter->count) {
1696                 bio = bio_alloc_bioset(GFP_KERNEL,
1697                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1698                                        &c->bio_read);
1699                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1700 start:
1701                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1702                 bio->bi_iter.bi_sector  = offset >> 9;
1703                 bio->bi_private         = dio;
1704
1705                 ret = bio_iov_iter_get_pages(bio, iter);
1706                 if (ret < 0) {
1707                         /* XXX: fault inject this path */
1708                         bio->bi_status = BLK_STS_RESOURCE;
1709                         bio_endio(bio);
1710                         break;
1711                 }
1712
1713                 offset += bio->bi_iter.bi_size;
1714                 bio_set_pages_dirty(bio);
1715
1716                 if (iter->count)
1717                         closure_get(&dio->cl);
1718
1719                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1720         }
1721
1722         iter->count += shorten;
1723
1724         if (sync) {
1725                 closure_sync(&dio->cl);
1726                 closure_debug_destroy(&dio->cl);
1727                 ret = dio->ret;
1728                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1729                 return ret;
1730         } else {
1731                 return -EIOCBQUEUED;
1732         }
1733 }
1734
1735 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1736 {
1737         struct file *file = iocb->ki_filp;
1738         struct bch_inode_info *inode = file_bch_inode(file);
1739         struct address_space *mapping = file->f_mapping;
1740         size_t count = iov_iter_count(iter);
1741         ssize_t ret;
1742
1743         if (!count)
1744                 return 0; /* skip atime */
1745
1746         if (iocb->ki_flags & IOCB_DIRECT) {
1747                 struct blk_plug plug;
1748
1749                 ret = filemap_write_and_wait_range(mapping,
1750                                         iocb->ki_pos,
1751                                         iocb->ki_pos + count - 1);
1752                 if (ret < 0)
1753                         return ret;
1754
1755                 file_accessed(file);
1756
1757                 blk_start_plug(&plug);
1758                 ret = bch2_direct_IO_read(iocb, iter);
1759                 blk_finish_plug(&plug);
1760
1761                 if (ret >= 0)
1762                         iocb->ki_pos += ret;
1763         } else {
1764                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1765                 ret = generic_file_read_iter(iocb, iter);
1766                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1767         }
1768
1769         return ret;
1770 }
1771
1772 /* O_DIRECT writes */
1773
1774 static void bch2_dio_write_loop_async(struct bch_write_op *);
1775
1776 static long bch2_dio_write_loop(struct dio_write *dio)
1777 {
1778         bool kthread = (current->flags & PF_KTHREAD) != 0;
1779         struct kiocb *req = dio->req;
1780         struct address_space *mapping = req->ki_filp->f_mapping;
1781         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1782         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1783         struct bio *bio = &dio->op.wbio.bio;
1784         struct bvec_iter_all iter;
1785         struct bio_vec *bv;
1786         unsigned unaligned, iter_count;
1787         bool sync = dio->sync, dropped_locks;
1788         long ret;
1789
1790         if (dio->loop)
1791                 goto loop;
1792
1793         while (1) {
1794                 iter_count = dio->iter.count;
1795
1796                 if (kthread)
1797                         kthread_use_mm(dio->mm);
1798                 BUG_ON(current->faults_disabled_mapping);
1799                 current->faults_disabled_mapping = mapping;
1800
1801                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1802
1803                 dropped_locks = fdm_dropped_locks();
1804
1805                 current->faults_disabled_mapping = NULL;
1806                 if (kthread)
1807                         kthread_unuse_mm(dio->mm);
1808
1809                 /*
1810                  * If the fault handler returned an error but also signalled
1811                  * that it dropped & retook ei_pagecache_lock, we just need to
1812                  * re-shoot down the page cache and retry:
1813                  */
1814                 if (dropped_locks && ret)
1815                         ret = 0;
1816
1817                 if (unlikely(ret < 0))
1818                         goto err;
1819
1820                 if (unlikely(dropped_locks)) {
1821                         ret = write_invalidate_inode_pages_range(mapping,
1822                                         req->ki_pos,
1823                                         req->ki_pos + iter_count - 1);
1824                         if (unlikely(ret))
1825                                 goto err;
1826
1827                         if (!bio->bi_iter.bi_size)
1828                                 continue;
1829                 }
1830
1831                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1832                 bio->bi_iter.bi_size -= unaligned;
1833                 iov_iter_revert(&dio->iter, unaligned);
1834
1835                 if (!bio->bi_iter.bi_size) {
1836                         /*
1837                          * bio_iov_iter_get_pages was only able to get <
1838                          * blocksize worth of pages:
1839                          */
1840                         bio_for_each_segment_all(bv, bio, iter)
1841                                 put_page(bv->bv_page);
1842                         ret = -EFAULT;
1843                         goto err;
1844                 }
1845
1846                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1847                 dio->op.end_io          = bch2_dio_write_loop_async;
1848                 dio->op.target          = dio->op.opts.foreground_target;
1849                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1850                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1851                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1852                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1853
1854                 if ((req->ki_flags & IOCB_DSYNC) &&
1855                     !c->opts.journal_flush_disabled)
1856                         dio->op.flags |= BCH_WRITE_FLUSH;
1857
1858                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1859                                                 dio->op.opts.data_replicas, 0);
1860                 if (unlikely(ret) &&
1861                     !bch2_check_range_allocated(c, dio->op.pos,
1862                                 bio_sectors(bio), dio->op.opts.data_replicas))
1863                         goto err;
1864
1865                 task_io_account_write(bio->bi_iter.bi_size);
1866
1867                 if (!dio->sync && !dio->loop && dio->iter.count) {
1868                         struct iovec *iov = dio->inline_vecs;
1869
1870                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1871                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1872                                               GFP_KERNEL);
1873                                 if (unlikely(!iov)) {
1874                                         dio->sync = sync = true;
1875                                         goto do_io;
1876                                 }
1877
1878                                 dio->free_iov = true;
1879                         }
1880
1881                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1882                         dio->iter.iov = iov;
1883                 }
1884 do_io:
1885                 dio->loop = true;
1886                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1887
1888                 if (sync)
1889                         wait_for_completion(&dio->done);
1890                 else
1891                         return -EIOCBQUEUED;
1892 loop:
1893                 i_sectors_acct(c, inode, &dio->quota_res,
1894                                dio->op.i_sectors_delta);
1895                 req->ki_pos += (u64) dio->op.written << 9;
1896                 dio->written += dio->op.written;
1897
1898                 spin_lock(&inode->v.i_lock);
1899                 if (req->ki_pos > inode->v.i_size)
1900                         i_size_write(&inode->v, req->ki_pos);
1901                 spin_unlock(&inode->v.i_lock);
1902
1903                 bio_for_each_segment_all(bv, bio, iter)
1904                         put_page(bv->bv_page);
1905                 if (!dio->iter.count || dio->op.error)
1906                         break;
1907
1908                 bio_reset(bio);
1909                 reinit_completion(&dio->done);
1910         }
1911
1912         ret = dio->op.error ?: ((long) dio->written << 9);
1913 err:
1914         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1915         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1916
1917         if (dio->free_iov)
1918                 kfree(dio->iter.iov);
1919
1920         bio_put(bio);
1921
1922         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1923         inode_dio_end(&inode->v);
1924
1925         if (!sync) {
1926                 req->ki_complete(req, ret, 0);
1927                 ret = -EIOCBQUEUED;
1928         }
1929         return ret;
1930 }
1931
1932 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1933 {
1934         struct dio_write *dio = container_of(op, struct dio_write, op);
1935
1936         if (dio->sync)
1937                 complete(&dio->done);
1938         else
1939                 bch2_dio_write_loop(dio);
1940 }
1941
1942 static noinline
1943 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1944 {
1945         struct file *file = req->ki_filp;
1946         struct address_space *mapping = file->f_mapping;
1947         struct bch_inode_info *inode = file_bch_inode(file);
1948         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1949         struct dio_write *dio;
1950         struct bio *bio;
1951         bool locked = true, extending;
1952         ssize_t ret;
1953
1954         prefetch(&c->opts);
1955         prefetch((void *) &c->opts + 64);
1956         prefetch(&inode->ei_inode);
1957         prefetch((void *) &inode->ei_inode + 64);
1958
1959         inode_lock(&inode->v);
1960
1961         ret = generic_write_checks(req, iter);
1962         if (unlikely(ret <= 0))
1963                 goto err;
1964
1965         ret = file_remove_privs(file);
1966         if (unlikely(ret))
1967                 goto err;
1968
1969         ret = file_update_time(file);
1970         if (unlikely(ret))
1971                 goto err;
1972
1973         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1974                 goto err;
1975
1976         inode_dio_begin(&inode->v);
1977         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
1978
1979         extending = req->ki_pos + iter->count > inode->v.i_size;
1980         if (!extending) {
1981                 inode_unlock(&inode->v);
1982                 locked = false;
1983         }
1984
1985         bio = bio_alloc_bioset(GFP_KERNEL,
1986                                iov_iter_npages(iter, BIO_MAX_PAGES),
1987                                &c->dio_write_bioset);
1988         dio = container_of(bio, struct dio_write, op.wbio.bio);
1989         init_completion(&dio->done);
1990         dio->req                = req;
1991         dio->mm                 = current->mm;
1992         dio->loop               = false;
1993         dio->sync               = is_sync_kiocb(req) || extending;
1994         dio->free_iov           = false;
1995         dio->quota_res.sectors  = 0;
1996         dio->written            = 0;
1997         dio->iter               = *iter;
1998
1999         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2000                                          iter->count >> 9, true);
2001         if (unlikely(ret))
2002                 goto err_put_bio;
2003
2004         ret = write_invalidate_inode_pages_range(mapping,
2005                                         req->ki_pos,
2006                                         req->ki_pos + iter->count - 1);
2007         if (unlikely(ret))
2008                 goto err_put_bio;
2009
2010         ret = bch2_dio_write_loop(dio);
2011 err:
2012         if (locked)
2013                 inode_unlock(&inode->v);
2014         return ret;
2015 err_put_bio:
2016         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2017         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2018         bio_put(bio);
2019         inode_dio_end(&inode->v);
2020         goto err;
2021 }
2022
2023 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2024 {
2025         struct file *file = iocb->ki_filp;
2026         struct bch_inode_info *inode = file_bch_inode(file);
2027         ssize_t ret;
2028
2029         if (iocb->ki_flags & IOCB_DIRECT)
2030                 return bch2_direct_write(iocb, from);
2031
2032         /* We can write back this queue in page reclaim */
2033         current->backing_dev_info = inode_to_bdi(&inode->v);
2034         inode_lock(&inode->v);
2035
2036         ret = generic_write_checks(iocb, from);
2037         if (ret <= 0)
2038                 goto unlock;
2039
2040         ret = file_remove_privs(file);
2041         if (ret)
2042                 goto unlock;
2043
2044         ret = file_update_time(file);
2045         if (ret)
2046                 goto unlock;
2047
2048         ret = bch2_buffered_write(iocb, from);
2049         if (likely(ret > 0))
2050                 iocb->ki_pos += ret;
2051 unlock:
2052         inode_unlock(&inode->v);
2053         current->backing_dev_info = NULL;
2054
2055         if (ret > 0)
2056                 ret = generic_write_sync(iocb, ret);
2057
2058         return ret;
2059 }
2060
2061 /* fsync: */
2062
2063 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2064 {
2065         struct bch_inode_info *inode = file_bch_inode(file);
2066         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2067         int ret, ret2;
2068
2069         ret = file_write_and_wait_range(file, start, end);
2070         if (ret)
2071                 return ret;
2072
2073         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2074                 goto out;
2075
2076         ret = sync_inode_metadata(&inode->v, 1);
2077         if (ret)
2078                 return ret;
2079 out:
2080         if (!c->opts.journal_flush_disabled)
2081                 ret = bch2_journal_flush_seq(&c->journal,
2082                                              inode->ei_journal_seq);
2083         ret2 = file_check_and_advance_wb_err(file);
2084
2085         return ret ?: ret2;
2086 }
2087
2088 /* truncate: */
2089
2090 static inline int range_has_data(struct bch_fs *c,
2091                                   struct bpos start,
2092                                   struct bpos end)
2093 {
2094         struct btree_trans trans;
2095         struct btree_iter *iter;
2096         struct bkey_s_c k;
2097         int ret = 0;
2098
2099         bch2_trans_init(&trans, c, 0, 0);
2100
2101         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2102                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2103                         break;
2104
2105                 if (bkey_extent_is_data(k.k)) {
2106                         ret = 1;
2107                         break;
2108                 }
2109         }
2110
2111         return bch2_trans_exit(&trans) ?: ret;
2112 }
2113
2114 static int __bch2_truncate_page(struct bch_inode_info *inode,
2115                                 pgoff_t index, loff_t start, loff_t end)
2116 {
2117         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2118         struct address_space *mapping = inode->v.i_mapping;
2119         struct bch_page_state *s;
2120         unsigned start_offset = start & (PAGE_SIZE - 1);
2121         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2122         unsigned i;
2123         struct page *page;
2124         int ret = 0;
2125
2126         /* Page boundary? Nothing to do */
2127         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2128               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2129                 return 0;
2130
2131         /* Above i_size? */
2132         if (index << PAGE_SHIFT >= inode->v.i_size)
2133                 return 0;
2134
2135         page = find_lock_page(mapping, index);
2136         if (!page) {
2137                 /*
2138                  * XXX: we're doing two index lookups when we end up reading the
2139                  * page
2140                  */
2141                 ret = range_has_data(c,
2142                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2143                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2144                 if (ret <= 0)
2145                         return ret;
2146
2147                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2148                 if (unlikely(!page)) {
2149                         ret = -ENOMEM;
2150                         goto out;
2151                 }
2152         }
2153
2154         s = bch2_page_state_create(page, 0);
2155         if (!s) {
2156                 ret = -ENOMEM;
2157                 goto unlock;
2158         }
2159
2160         if (!PageUptodate(page)) {
2161                 ret = bch2_read_single_page(page, mapping);
2162                 if (ret)
2163                         goto unlock;
2164         }
2165
2166         if (index != start >> PAGE_SHIFT)
2167                 start_offset = 0;
2168         if (index != end >> PAGE_SHIFT)
2169                 end_offset = PAGE_SIZE;
2170
2171         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2172              i < round_down(end_offset, block_bytes(c)) >> 9;
2173              i++) {
2174                 s->s[i].nr_replicas     = 0;
2175                 s->s[i].state           = SECTOR_UNALLOCATED;
2176         }
2177
2178         zero_user_segment(page, start_offset, end_offset);
2179
2180         /*
2181          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2182          *
2183          * XXX: because we aren't currently tracking whether the page has actual
2184          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2185          */
2186         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2187         BUG_ON(ret);
2188
2189         /*
2190          * This removes any writeable userspace mappings; we need to force
2191          * .page_mkwrite to be called again before any mmapped writes, to
2192          * redirty the full page:
2193          */
2194         page_mkclean(page);
2195         __set_page_dirty_nobuffers(page);
2196 unlock:
2197         unlock_page(page);
2198         put_page(page);
2199 out:
2200         return ret;
2201 }
2202
2203 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2204 {
2205         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2206                                     from, round_up(from, PAGE_SIZE));
2207 }
2208
2209 static int bch2_extend(struct bch_inode_info *inode,
2210                        struct bch_inode_unpacked *inode_u,
2211                        struct iattr *iattr)
2212 {
2213         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2214         struct address_space *mapping = inode->v.i_mapping;
2215         int ret;
2216
2217         /*
2218          * sync appends:
2219          *
2220          * this has to be done _before_ extending i_size:
2221          */
2222         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2223         if (ret)
2224                 return ret;
2225
2226         truncate_setsize(&inode->v, iattr->ia_size);
2227         setattr_copy(&inode->v, iattr);
2228
2229         mutex_lock(&inode->ei_update_lock);
2230         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2231                                     ATTR_MTIME|ATTR_CTIME);
2232         mutex_unlock(&inode->ei_update_lock);
2233
2234         return ret;
2235 }
2236
2237 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2238                                    struct bch_inode_unpacked *bi,
2239                                    void *p)
2240 {
2241         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2242
2243         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2244         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2245         return 0;
2246 }
2247
2248 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2249                                   struct bch_inode_unpacked *bi, void *p)
2250 {
2251         u64 *new_i_size = p;
2252
2253         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2254         bi->bi_size = *new_i_size;
2255         return 0;
2256 }
2257
2258 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2259 {
2260         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2261         struct address_space *mapping = inode->v.i_mapping;
2262         struct bch_inode_unpacked inode_u;
2263         struct btree_trans trans;
2264         struct btree_iter *iter;
2265         u64 new_i_size = iattr->ia_size;
2266         s64 i_sectors_delta = 0;
2267         int ret = 0;
2268
2269         inode_dio_wait(&inode->v);
2270         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2271
2272         /*
2273          * fetch current on disk i_size: inode is locked, i_size can only
2274          * increase underneath us:
2275          */
2276         bch2_trans_init(&trans, c, 0, 0);
2277         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2278         ret = PTR_ERR_OR_ZERO(iter);
2279         bch2_trans_exit(&trans);
2280
2281         if (ret)
2282                 goto err;
2283
2284         /*
2285          * check this before next assertion; on filesystem error our normal
2286          * invariants are a bit broken (truncate has to truncate the page cache
2287          * before the inode).
2288          */
2289         ret = bch2_journal_error(&c->journal);
2290         if (ret)
2291                 goto err;
2292
2293         BUG_ON(inode->v.i_size < inode_u.bi_size);
2294
2295         if (iattr->ia_size > inode->v.i_size) {
2296                 ret = bch2_extend(inode, &inode_u, iattr);
2297                 goto err;
2298         }
2299
2300         ret = bch2_truncate_page(inode, iattr->ia_size);
2301         if (unlikely(ret))
2302                 goto err;
2303
2304         /*
2305          * When extending, we're going to write the new i_size to disk
2306          * immediately so we need to flush anything above the current on disk
2307          * i_size first:
2308          *
2309          * Also, when extending we need to flush the page that i_size currently
2310          * straddles - if it's mapped to userspace, we need to ensure that
2311          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2312          * again to allocate the part of the page that was extended.
2313          */
2314         if (iattr->ia_size > inode_u.bi_size)
2315                 ret = filemap_write_and_wait_range(mapping,
2316                                 inode_u.bi_size,
2317                                 iattr->ia_size - 1);
2318         else if (iattr->ia_size & (PAGE_SIZE - 1))
2319                 ret = filemap_write_and_wait_range(mapping,
2320                                 round_down(iattr->ia_size, PAGE_SIZE),
2321                                 iattr->ia_size - 1);
2322         if (ret)
2323                 goto err;
2324
2325         mutex_lock(&inode->ei_update_lock);
2326         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2327                                &new_i_size, 0);
2328         mutex_unlock(&inode->ei_update_lock);
2329
2330         if (unlikely(ret))
2331                 goto err;
2332
2333         truncate_setsize(&inode->v, iattr->ia_size);
2334
2335         ret = bch2_fpunch(c, inode->v.i_ino,
2336                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2337                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2338         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2339
2340         if (unlikely(ret))
2341                 goto err;
2342
2343         setattr_copy(&inode->v, iattr);
2344
2345         mutex_lock(&inode->ei_update_lock);
2346         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2347                                ATTR_MTIME|ATTR_CTIME);
2348         mutex_unlock(&inode->ei_update_lock);
2349 err:
2350         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2351         return ret;
2352 }
2353
2354 /* fallocate: */
2355
2356 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2357 {
2358         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2359         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2360         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2361         int ret = 0;
2362
2363         inode_lock(&inode->v);
2364         inode_dio_wait(&inode->v);
2365         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2366
2367         ret = __bch2_truncate_page(inode,
2368                                    offset >> PAGE_SHIFT,
2369                                    offset, offset + len);
2370         if (unlikely(ret))
2371                 goto err;
2372
2373         if (offset >> PAGE_SHIFT !=
2374             (offset + len) >> PAGE_SHIFT) {
2375                 ret = __bch2_truncate_page(inode,
2376                                            (offset + len) >> PAGE_SHIFT,
2377                                            offset, offset + len);
2378                 if (unlikely(ret))
2379                         goto err;
2380         }
2381
2382         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2383
2384         if (discard_start < discard_end) {
2385                 s64 i_sectors_delta = 0;
2386
2387                 ret = bch2_fpunch(c, inode->v.i_ino,
2388                                   discard_start, discard_end,
2389                                   &inode->ei_journal_seq,
2390                                   &i_sectors_delta);
2391                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2392         }
2393 err:
2394         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2395         inode_unlock(&inode->v);
2396
2397         return ret;
2398 }
2399
2400 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2401                                    loff_t offset, loff_t len,
2402                                    bool insert)
2403 {
2404         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2405         struct address_space *mapping = inode->v.i_mapping;
2406         struct bkey_on_stack copy;
2407         struct btree_trans trans;
2408         struct btree_iter *src, *dst;
2409         loff_t shift, new_size;
2410         u64 src_start;
2411         int ret;
2412
2413         if ((offset | len) & (block_bytes(c) - 1))
2414                 return -EINVAL;
2415
2416         bkey_on_stack_init(&copy);
2417         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2418
2419         /*
2420          * We need i_mutex to keep the page cache consistent with the extents
2421          * btree, and the btree consistent with i_size - we don't need outside
2422          * locking for the extents btree itself, because we're using linked
2423          * iterators
2424          */
2425         inode_lock(&inode->v);
2426         inode_dio_wait(&inode->v);
2427         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2428
2429         if (insert) {
2430                 ret = -EFBIG;
2431                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2432                         goto err;
2433
2434                 ret = -EINVAL;
2435                 if (offset >= inode->v.i_size)
2436                         goto err;
2437
2438                 src_start       = U64_MAX;
2439                 shift           = len;
2440         } else {
2441                 ret = -EINVAL;
2442                 if (offset + len >= inode->v.i_size)
2443                         goto err;
2444
2445                 src_start       = offset + len;
2446                 shift           = -len;
2447         }
2448
2449         new_size = inode->v.i_size + shift;
2450
2451         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2452         if (ret)
2453                 goto err;
2454
2455         if (insert) {
2456                 i_size_write(&inode->v, new_size);
2457                 mutex_lock(&inode->ei_update_lock);
2458                 ret = bch2_write_inode_size(c, inode, new_size,
2459                                             ATTR_MTIME|ATTR_CTIME);
2460                 mutex_unlock(&inode->ei_update_lock);
2461         } else {
2462                 s64 i_sectors_delta = 0;
2463
2464                 ret = bch2_fpunch(c, inode->v.i_ino,
2465                                   offset >> 9, (offset + len) >> 9,
2466                                   &inode->ei_journal_seq,
2467                                   &i_sectors_delta);
2468                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2469
2470                 if (ret)
2471                         goto err;
2472         }
2473
2474         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2475                         POS(inode->v.i_ino, src_start >> 9),
2476                         BTREE_ITER_INTENT);
2477         dst = bch2_trans_copy_iter(&trans, src);
2478
2479         while (1) {
2480                 struct disk_reservation disk_res =
2481                         bch2_disk_reservation_init(c, 0);
2482                 struct bkey_i delete;
2483                 struct bkey_s_c k;
2484                 struct bpos next_pos;
2485                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2486                 struct bpos atomic_end;
2487                 unsigned trigger_flags = 0;
2488
2489                 k = insert
2490                         ? bch2_btree_iter_peek_prev(src)
2491                         : bch2_btree_iter_peek(src);
2492                 if ((ret = bkey_err(k)))
2493                         goto bkey_err;
2494
2495                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2496                         break;
2497
2498                 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2499
2500                 if (insert &&
2501                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2502                         break;
2503 reassemble:
2504                 bkey_on_stack_reassemble(&copy, c, k);
2505
2506                 if (insert &&
2507                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2508                         bch2_cut_front(move_pos, copy.k);
2509
2510                 copy.k->k.p.offset += shift >> 9;
2511                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2512
2513                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2514                 if (ret)
2515                         goto bkey_err;
2516
2517                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2518                         if (insert) {
2519                                 move_pos = atomic_end;
2520                                 move_pos.offset -= shift >> 9;
2521                                 goto reassemble;
2522                         } else {
2523                                 bch2_cut_back(atomic_end, copy.k);
2524                         }
2525                 }
2526
2527                 bkey_init(&delete.k);
2528                 delete.k.p = copy.k->k.p;
2529                 delete.k.size = copy.k->k.size;
2530                 delete.k.p.offset -= shift >> 9;
2531
2532                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2533
2534                 if (copy.k->k.size == k.k->size) {
2535                         /*
2536                          * If we're moving the entire extent, we can skip
2537                          * running triggers:
2538                          */
2539                         trigger_flags |= BTREE_TRIGGER_NORUN;
2540                 } else {
2541                         /* We might end up splitting compressed extents: */
2542                         unsigned nr_ptrs =
2543                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2544
2545                         ret = bch2_disk_reservation_get(c, &disk_res,
2546                                         copy.k->k.size, nr_ptrs,
2547                                         BCH_DISK_RESERVATION_NOFAIL);
2548                         BUG_ON(ret);
2549                 }
2550
2551                 bch2_btree_iter_set_pos(src, bkey_start_pos(&delete.k));
2552
2553                 ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
2554                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2555                         bch2_trans_commit(&trans, &disk_res,
2556                                           &inode->ei_journal_seq,
2557                                           BTREE_INSERT_NOFAIL);
2558                 bch2_disk_reservation_put(c, &disk_res);
2559 bkey_err:
2560                 if (!ret)
2561                         bch2_btree_iter_set_pos(src, next_pos);
2562
2563                 if (ret == -EINTR)
2564                         ret = 0;
2565                 if (ret)
2566                         goto err;
2567
2568                 bch2_trans_cond_resched(&trans);
2569         }
2570         bch2_trans_unlock(&trans);
2571
2572         if (!insert) {
2573                 i_size_write(&inode->v, new_size);
2574                 mutex_lock(&inode->ei_update_lock);
2575                 ret = bch2_write_inode_size(c, inode, new_size,
2576                                             ATTR_MTIME|ATTR_CTIME);
2577                 mutex_unlock(&inode->ei_update_lock);
2578         }
2579 err:
2580         bch2_trans_exit(&trans);
2581         bkey_on_stack_exit(&copy, c);
2582         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2583         inode_unlock(&inode->v);
2584         return ret;
2585 }
2586
2587 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2588                             loff_t offset, loff_t len)
2589 {
2590         struct address_space *mapping = inode->v.i_mapping;
2591         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2592         struct btree_trans trans;
2593         struct btree_iter *iter;
2594         struct bpos end_pos;
2595         loff_t end              = offset + len;
2596         loff_t block_start      = round_down(offset,    block_bytes(c));
2597         loff_t block_end        = round_up(end,         block_bytes(c));
2598         unsigned sectors;
2599         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2600         int ret;
2601
2602         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2603
2604         inode_lock(&inode->v);
2605         inode_dio_wait(&inode->v);
2606         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2607
2608         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2609                 ret = inode_newsize_ok(&inode->v, end);
2610                 if (ret)
2611                         goto err;
2612         }
2613
2614         if (mode & FALLOC_FL_ZERO_RANGE) {
2615                 ret = __bch2_truncate_page(inode,
2616                                            offset >> PAGE_SHIFT,
2617                                            offset, end);
2618
2619                 if (!ret &&
2620                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2621                         ret = __bch2_truncate_page(inode,
2622                                                    end >> PAGE_SHIFT,
2623                                                    offset, end);
2624
2625                 if (unlikely(ret))
2626                         goto err;
2627
2628                 truncate_pagecache_range(&inode->v, offset, end - 1);
2629         }
2630
2631         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2632                         POS(inode->v.i_ino, block_start >> 9),
2633                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2634         end_pos = POS(inode->v.i_ino, block_end >> 9);
2635
2636         while (bkey_cmp(iter->pos, end_pos) < 0) {
2637                 s64 i_sectors_delta = 0;
2638                 struct disk_reservation disk_res = { 0 };
2639                 struct quota_res quota_res = { 0 };
2640                 struct bkey_i_reservation reservation;
2641                 struct bkey_s_c k;
2642
2643                 bch2_trans_begin(&trans);
2644
2645                 k = bch2_btree_iter_peek_slot(iter);
2646                 if ((ret = bkey_err(k)))
2647                         goto bkey_err;
2648
2649                 /* already reserved */
2650                 if (k.k->type == KEY_TYPE_reservation &&
2651                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2652                         bch2_btree_iter_next_slot(iter);
2653                         continue;
2654                 }
2655
2656                 if (bkey_extent_is_data(k.k) &&
2657                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2658                         bch2_btree_iter_next_slot(iter);
2659                         continue;
2660                 }
2661
2662                 bkey_reservation_init(&reservation.k_i);
2663                 reservation.k.type      = KEY_TYPE_reservation;
2664                 reservation.k.p         = k.k->p;
2665                 reservation.k.size      = k.k->size;
2666
2667                 bch2_cut_front(iter->pos,       &reservation.k_i);
2668                 bch2_cut_back(end_pos,          &reservation.k_i);
2669
2670                 sectors = reservation.k.size;
2671                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2672
2673                 if (!bkey_extent_is_allocation(k.k)) {
2674                         ret = bch2_quota_reservation_add(c, inode,
2675                                         &quota_res,
2676                                         sectors, true);
2677                         if (unlikely(ret))
2678                                 goto bkey_err;
2679                 }
2680
2681                 if (reservation.v.nr_replicas < replicas ||
2682                     bch2_bkey_sectors_compressed(k)) {
2683                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2684                                                         replicas, 0);
2685                         if (unlikely(ret))
2686                                 goto bkey_err;
2687
2688                         reservation.v.nr_replicas = disk_res.nr_replicas;
2689                 }
2690
2691                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2692                                 &disk_res, &inode->ei_journal_seq,
2693                                 0, &i_sectors_delta);
2694                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2695 bkey_err:
2696                 bch2_quota_reservation_put(c, inode, &quota_res);
2697                 bch2_disk_reservation_put(c, &disk_res);
2698                 if (ret == -EINTR)
2699                         ret = 0;
2700                 if (ret)
2701                         goto err;
2702         }
2703
2704         /*
2705          * Do we need to extend the file?
2706          *
2707          * If we zeroed up to the end of the file, we dropped whatever writes
2708          * were going to write out the current i_size, so we have to extend
2709          * manually even if FL_KEEP_SIZE was set:
2710          */
2711         if (end >= inode->v.i_size &&
2712             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2713              (mode & FALLOC_FL_ZERO_RANGE))) {
2714                 struct btree_iter *inode_iter;
2715                 struct bch_inode_unpacked inode_u;
2716
2717                 do {
2718                         bch2_trans_begin(&trans);
2719                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2720                                                      inode->v.i_ino, 0);
2721                         ret = PTR_ERR_OR_ZERO(inode_iter);
2722                 } while (ret == -EINTR);
2723
2724                 bch2_trans_unlock(&trans);
2725
2726                 if (ret)
2727                         goto err;
2728
2729                 /*
2730                  * Sync existing appends before extending i_size,
2731                  * as in bch2_extend():
2732                  */
2733                 ret = filemap_write_and_wait_range(mapping,
2734                                         inode_u.bi_size, S64_MAX);
2735                 if (ret)
2736                         goto err;
2737
2738                 if (mode & FALLOC_FL_KEEP_SIZE)
2739                         end = inode->v.i_size;
2740                 else
2741                         i_size_write(&inode->v, end);
2742
2743                 mutex_lock(&inode->ei_update_lock);
2744                 ret = bch2_write_inode_size(c, inode, end, 0);
2745                 mutex_unlock(&inode->ei_update_lock);
2746         }
2747 err:
2748         bch2_trans_exit(&trans);
2749         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2750         inode_unlock(&inode->v);
2751         return ret;
2752 }
2753
2754 long bch2_fallocate_dispatch(struct file *file, int mode,
2755                              loff_t offset, loff_t len)
2756 {
2757         struct bch_inode_info *inode = file_bch_inode(file);
2758         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2759         long ret;
2760
2761         if (!percpu_ref_tryget(&c->writes))
2762                 return -EROFS;
2763
2764         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2765                 ret = bchfs_fallocate(inode, mode, offset, len);
2766         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2767                 ret = bchfs_fpunch(inode, offset, len);
2768         else if (mode == FALLOC_FL_INSERT_RANGE)
2769                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2770         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2771                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2772         else
2773                 ret = -EOPNOTSUPP;
2774
2775         percpu_ref_put(&c->writes);
2776
2777         return ret;
2778 }
2779
2780 static void mark_range_unallocated(struct bch_inode_info *inode,
2781                                    loff_t start, loff_t end)
2782 {
2783         pgoff_t index = start >> PAGE_SHIFT;
2784         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2785         struct pagevec pvec;
2786
2787         pagevec_init(&pvec);
2788
2789         do {
2790                 unsigned nr_pages, i, j;
2791
2792                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2793                                                 &index, end_index);
2794                 if (nr_pages == 0)
2795                         break;
2796
2797                 for (i = 0; i < nr_pages; i++) {
2798                         struct page *page = pvec.pages[i];
2799                         struct bch_page_state *s;
2800
2801                         lock_page(page);
2802                         s = bch2_page_state(page);
2803
2804                         if (s) {
2805                                 spin_lock(&s->lock);
2806                                 for (j = 0; j < PAGE_SECTORS; j++)
2807                                         s->s[j].nr_replicas = 0;
2808                                 spin_unlock(&s->lock);
2809                         }
2810
2811                         unlock_page(page);
2812                 }
2813                 pagevec_release(&pvec);
2814         } while (index <= end_index);
2815 }
2816
2817 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2818                              struct file *file_dst, loff_t pos_dst,
2819                              loff_t len, unsigned remap_flags)
2820 {
2821         struct bch_inode_info *src = file_bch_inode(file_src);
2822         struct bch_inode_info *dst = file_bch_inode(file_dst);
2823         struct bch_fs *c = src->v.i_sb->s_fs_info;
2824         s64 i_sectors_delta = 0;
2825         u64 aligned_len;
2826         loff_t ret = 0;
2827
2828         if (!c->opts.reflink)
2829                 return -EOPNOTSUPP;
2830
2831         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2832                 return -EINVAL;
2833
2834         if (remap_flags & REMAP_FILE_DEDUP)
2835                 return -EOPNOTSUPP;
2836
2837         if ((pos_src & (block_bytes(c) - 1)) ||
2838             (pos_dst & (block_bytes(c) - 1)))
2839                 return -EINVAL;
2840
2841         if (src == dst &&
2842             abs(pos_src - pos_dst) < len)
2843                 return -EINVAL;
2844
2845         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2846
2847         file_update_time(file_dst);
2848
2849         inode_dio_wait(&src->v);
2850         inode_dio_wait(&dst->v);
2851
2852         ret = generic_remap_file_range_prep(file_src, pos_src,
2853                                             file_dst, pos_dst,
2854                                             &len, remap_flags);
2855         if (ret < 0 || len == 0)
2856                 goto err;
2857
2858         aligned_len = round_up((u64) len, block_bytes(c));
2859
2860         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2861                                 pos_dst, pos_dst + len - 1);
2862         if (ret)
2863                 goto err;
2864
2865         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2866
2867         ret = bch2_remap_range(c,
2868                                POS(dst->v.i_ino, pos_dst >> 9),
2869                                POS(src->v.i_ino, pos_src >> 9),
2870                                aligned_len >> 9,
2871                                &dst->ei_journal_seq,
2872                                pos_dst + len, &i_sectors_delta);
2873         if (ret < 0)
2874                 goto err;
2875
2876         /*
2877          * due to alignment, we might have remapped slightly more than requsted
2878          */
2879         ret = min((u64) ret << 9, (u64) len);
2880
2881         /* XXX get a quota reservation */
2882         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2883
2884         spin_lock(&dst->v.i_lock);
2885         if (pos_dst + ret > dst->v.i_size)
2886                 i_size_write(&dst->v, pos_dst + ret);
2887         spin_unlock(&dst->v.i_lock);
2888 err:
2889         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2890
2891         return ret;
2892 }
2893
2894 /* fseek: */
2895
2896 static int page_data_offset(struct page *page, unsigned offset)
2897 {
2898         struct bch_page_state *s = bch2_page_state(page);
2899         unsigned i;
2900
2901         if (s)
2902                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2903                         if (s->s[i].state >= SECTOR_DIRTY)
2904                                 return i << 9;
2905
2906         return -1;
2907 }
2908
2909 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2910                                        loff_t start_offset,
2911                                        loff_t end_offset)
2912 {
2913         struct address_space *mapping = vinode->i_mapping;
2914         struct page *page;
2915         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2916         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2917         pgoff_t index           = start_index;
2918         loff_t ret;
2919         int offset;
2920
2921         while (index <= end_index) {
2922                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2923                         lock_page(page);
2924
2925                         offset = page_data_offset(page,
2926                                         page->index == start_index
2927                                         ? start_offset & (PAGE_SIZE - 1)
2928                                         : 0);
2929                         if (offset >= 0) {
2930                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2931                                             offset,
2932                                             start_offset, end_offset);
2933                                 unlock_page(page);
2934                                 put_page(page);
2935                                 return ret;
2936                         }
2937
2938                         unlock_page(page);
2939                         put_page(page);
2940                 } else {
2941                         break;
2942                 }
2943         }
2944
2945         return end_offset;
2946 }
2947
2948 static loff_t bch2_seek_data(struct file *file, u64 offset)
2949 {
2950         struct bch_inode_info *inode = file_bch_inode(file);
2951         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2952         struct btree_trans trans;
2953         struct btree_iter *iter;
2954         struct bkey_s_c k;
2955         u64 isize, next_data = MAX_LFS_FILESIZE;
2956         int ret;
2957
2958         isize = i_size_read(&inode->v);
2959         if (offset >= isize)
2960                 return -ENXIO;
2961
2962         bch2_trans_init(&trans, c, 0, 0);
2963
2964         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2965                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2966                 if (k.k->p.inode != inode->v.i_ino) {
2967                         break;
2968                 } else if (bkey_extent_is_data(k.k)) {
2969                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2970                         break;
2971                 } else if (k.k->p.offset >> 9 > isize)
2972                         break;
2973         }
2974
2975         ret = bch2_trans_exit(&trans) ?: ret;
2976         if (ret)
2977                 return ret;
2978
2979         if (next_data > offset)
2980                 next_data = bch2_seek_pagecache_data(&inode->v,
2981                                                      offset, next_data);
2982
2983         if (next_data >= isize)
2984                 return -ENXIO;
2985
2986         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2987 }
2988
2989 static int __page_hole_offset(struct page *page, unsigned offset)
2990 {
2991         struct bch_page_state *s = bch2_page_state(page);
2992         unsigned i;
2993
2994         if (!s)
2995                 return 0;
2996
2997         for (i = offset >> 9; i < PAGE_SECTORS; i++)
2998                 if (s->s[i].state < SECTOR_DIRTY)
2999                         return i << 9;
3000
3001         return -1;
3002 }
3003
3004 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3005 {
3006         pgoff_t index = offset >> PAGE_SHIFT;
3007         struct page *page;
3008         int pg_offset;
3009         loff_t ret = -1;
3010
3011         page = find_lock_entry(mapping, index);
3012         if (!page || xa_is_value(page))
3013                 return offset;
3014
3015         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3016         if (pg_offset >= 0)
3017                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3018
3019         unlock_page(page);
3020
3021         return ret;
3022 }
3023
3024 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3025                                        loff_t start_offset,
3026                                        loff_t end_offset)
3027 {
3028         struct address_space *mapping = vinode->i_mapping;
3029         loff_t offset = start_offset, hole;
3030
3031         while (offset < end_offset) {
3032                 hole = page_hole_offset(mapping, offset);
3033                 if (hole >= 0 && hole <= end_offset)
3034                         return max(start_offset, hole);
3035
3036                 offset += PAGE_SIZE;
3037                 offset &= PAGE_MASK;
3038         }
3039
3040         return end_offset;
3041 }
3042
3043 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3044 {
3045         struct bch_inode_info *inode = file_bch_inode(file);
3046         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3047         struct btree_trans trans;
3048         struct btree_iter *iter;
3049         struct bkey_s_c k;
3050         u64 isize, next_hole = MAX_LFS_FILESIZE;
3051         int ret;
3052
3053         isize = i_size_read(&inode->v);
3054         if (offset >= isize)
3055                 return -ENXIO;
3056
3057         bch2_trans_init(&trans, c, 0, 0);
3058
3059         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3060                            POS(inode->v.i_ino, offset >> 9),
3061                            BTREE_ITER_SLOTS, k, ret) {
3062                 if (k.k->p.inode != inode->v.i_ino) {
3063                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3064                                         offset, MAX_LFS_FILESIZE);
3065                         break;
3066                 } else if (!bkey_extent_is_data(k.k)) {
3067                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3068                                         max(offset, bkey_start_offset(k.k) << 9),
3069                                         k.k->p.offset << 9);
3070
3071                         if (next_hole < k.k->p.offset << 9)
3072                                 break;
3073                 } else {
3074                         offset = max(offset, bkey_start_offset(k.k) << 9);
3075                 }
3076         }
3077
3078         ret = bch2_trans_exit(&trans) ?: ret;
3079         if (ret)
3080                 return ret;
3081
3082         if (next_hole > isize)
3083                 next_hole = isize;
3084
3085         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3086 }
3087
3088 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3089 {
3090         switch (whence) {
3091         case SEEK_SET:
3092         case SEEK_CUR:
3093         case SEEK_END:
3094                 return generic_file_llseek(file, offset, whence);
3095         case SEEK_DATA:
3096                 return bch2_seek_data(file, offset);
3097         case SEEK_HOLE:
3098                 return bch2_seek_hole(file, offset);
3099         }
3100
3101         return -EINVAL;
3102 }
3103
3104 void bch2_fs_fsio_exit(struct bch_fs *c)
3105 {
3106         bioset_exit(&c->dio_write_bioset);
3107         bioset_exit(&c->dio_read_bioset);
3108         bioset_exit(&c->writepage_bioset);
3109 }
3110
3111 int bch2_fs_fsio_init(struct bch_fs *c)
3112 {
3113         int ret = 0;
3114
3115         pr_verbose_init(c->opts, "");
3116
3117         if (bioset_init(&c->writepage_bioset,
3118                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3119                         BIOSET_NEED_BVECS) ||
3120             bioset_init(&c->dio_read_bioset,
3121                         4, offsetof(struct dio_read, rbio.bio),
3122                         BIOSET_NEED_BVECS) ||
3123             bioset_init(&c->dio_write_bioset,
3124                         4, offsetof(struct dio_write, op.wbio.bio),
3125                         BIOSET_NEED_BVECS))
3126                 ret = -ENOMEM;
3127
3128         pr_verbose_init(c->opts, "ret %i", ret);
3129         return ret;
3130 }
3131
3132 #endif /* NO_BCACHEFS_FS */