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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 static inline struct address_space *faults_disabled_mapping(void)
  39 {
  40         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  41 }
  42
  43 static inline void set_fdm_dropped_locks(void)
  44 {
  45         current->faults_disabled_mapping =
  46                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  47 }
  48
  49 static inline bool fdm_dropped_locks(void)
  50 {
  51         return ((unsigned long) current->faults_disabled_mapping) & 1;
  52 }
  53
  54 struct quota_res {
  55         u64                             sectors;
  56 };
  57
  58 struct bch_writepage_io {
  59         struct closure                  cl;
  60         struct bch_inode_info           *inode;
  61
  62         /* must be last: */
  63         struct bch_write_op             op;
  64 };
  65
  66 struct dio_write {
  67         struct completion               done;
  68         struct kiocb                    *req;
  69         struct mm_struct                *mm;
  70         unsigned                        loop:1,
  71                                         sync:1,
  72                                         free_iov:1;
  73         struct quota_res                quota_res;
  74         u64                             written;
  75
  76         struct iov_iter                 iter;
  77         struct iovec                    inline_vecs[2];
  78
  79         /* must be last: */
  80         struct bch_write_op             op;
  81 };
  82
  83 struct dio_read {
  84         struct closure                  cl;
  85         struct kiocb                    *req;
  86         long                            ret;
  87         bool                            should_dirty;
  88         struct bch_read_bio             rbio;
  89 };
  90
  91 /* pagecache_block must be held */
  92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  93                                               loff_t start, loff_t end)
  94 {
  95         int ret;
  96
  97         /*
  98          * XXX: the way this is currently implemented, we can spin if a process
  99          * is continually redirtying a specific page
 100          */
 101         do {
 102                 if (!mapping->nrpages &&
 103                     !mapping->nrexceptional)
 104                         return 0;
 105
 106                 ret = filemap_write_and_wait_range(mapping, start, end);
 107                 if (ret)
 108                         break;
 109
 110                 if (!mapping->nrpages)
 111                         return 0;
 112
 113                 ret = invalidate_inode_pages2_range(mapping,
 114                                 start >> PAGE_SHIFT,
 115                                 end >> PAGE_SHIFT);
 116         } while (ret == -EBUSY);
 117
 118         return ret;
 119 }
 120
 121 /* quotas */
 122
 123 #ifdef CONFIG_BCACHEFS_QUOTA
 124
 125 static void bch2_quota_reservation_put(struct bch_fs *c,
 126                                        struct bch_inode_info *inode,
 127                                        struct quota_res *res)
 128 {
 129         if (!res->sectors)
 130                 return;
 131
 132         mutex_lock(&inode->ei_quota_lock);
 133         BUG_ON(res->sectors > inode->ei_quota_reserved);
 134
 135         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 136                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 137         inode->ei_quota_reserved -= res->sectors;
 138         mutex_unlock(&inode->ei_quota_lock);
 139
 140         res->sectors = 0;
 141 }
 142
 143 static int bch2_quota_reservation_add(struct bch_fs *c,
 144                                       struct bch_inode_info *inode,
 145                                       struct quota_res *res,
 146                                       unsigned sectors,
 147                                       bool check_enospc)
 148 {
 149         int ret;
 150
 151         mutex_lock(&inode->ei_quota_lock);
 152         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 153                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 154         if (likely(!ret)) {
 155                 inode->ei_quota_reserved += sectors;
 156                 res->sectors += sectors;
 157         }
 158         mutex_unlock(&inode->ei_quota_lock);
 159
 160         return ret;
 161 }
 162
 163 #else
 164
 165 static void bch2_quota_reservation_put(struct bch_fs *c,
 166                                        struct bch_inode_info *inode,
 167                                        struct quota_res *res)
 168 {
 169 }
 170
 171 static int bch2_quota_reservation_add(struct bch_fs *c,
 172                                       struct bch_inode_info *inode,
 173                                       struct quota_res *res,
 174                                       unsigned sectors,
 175                                       bool check_enospc)
 176 {
 177         return 0;
 178 }
 179
 180 #endif
 181
 182 /* i_size updates: */
 183
 184 struct inode_new_size {
 185         loff_t          new_size;
 186         u64             now;
 187         unsigned        fields;
 188 };
 189
 190 static int inode_set_size(struct bch_inode_info *inode,
 191                           struct bch_inode_unpacked *bi,
 192                           void *p)
 193 {
 194         struct inode_new_size *s = p;
 195
 196         bi->bi_size = s->new_size;
 197         if (s->fields & ATTR_ATIME)
 198                 bi->bi_atime = s->now;
 199         if (s->fields & ATTR_MTIME)
 200                 bi->bi_mtime = s->now;
 201         if (s->fields & ATTR_CTIME)
 202                 bi->bi_ctime = s->now;
 203
 204         return 0;
 205 }
 206
 207 int __must_check bch2_write_inode_size(struct bch_fs *c,
 208                                        struct bch_inode_info *inode,
 209                                        loff_t new_size, unsigned fields)
 210 {
 211         struct inode_new_size s = {
 212                 .new_size       = new_size,
 213                 .now            = bch2_current_time(c),
 214                 .fields         = fields,
 215         };
 216
 217         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 218 }
 219
 220 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 221                            struct quota_res *quota_res, s64 sectors)
 222 {
 223         if (!sectors)
 224                 return;
 225
 226         mutex_lock(&inode->ei_quota_lock);
 227 #ifdef CONFIG_BCACHEFS_QUOTA
 228         if (quota_res && sectors > 0) {
 229                 BUG_ON(sectors > quota_res->sectors);
 230                 BUG_ON(sectors > inode->ei_quota_reserved);
 231
 232                 quota_res->sectors -= sectors;
 233                 inode->ei_quota_reserved -= sectors;
 234         } else {
 235                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 236         }
 237 #endif
 238         inode->v.i_blocks += sectors;
 239         mutex_unlock(&inode->ei_quota_lock);
 240 }
 241
 242 /* page state: */
 243
 244 /* stored in page->private: */
 245
 246 struct bch_page_sector {
 247         /* Uncompressed, fully allocated replicas: */
 248         unsigned                nr_replicas:3;
 249
 250         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 251         unsigned                replicas_reserved:3;
 252
 253         /* i_sectors: */
 254         enum {
 255                 SECTOR_UNALLOCATED,
 256                 SECTOR_RESERVED,
 257                 SECTOR_DIRTY,
 258                 SECTOR_ALLOCATED,
 259         }                       state:2;
 260 };
 261
 262 struct bch_page_state {
 263         spinlock_t              lock;
 264         atomic_t                write_count;
 265         struct bch_page_sector  s[PAGE_SECTORS];
 266 };
 267
 268 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 269 {
 270         return page_has_private(page)
 271                 ? (struct bch_page_state *) page_private(page)
 272                 : NULL;
 273 }
 274
 275 static inline struct bch_page_state *bch2_page_state(struct page *page)
 276 {
 277         EBUG_ON(!PageLocked(page));
 278
 279         return __bch2_page_state(page);
 280 }
 281
 282 /* for newly allocated pages: */
 283 static void __bch2_page_state_release(struct page *page)
 284 {
 285         kfree(detach_page_private(page));
 286 }
 287
 288 static void bch2_page_state_release(struct page *page)
 289 {
 290         EBUG_ON(!PageLocked(page));
 291         __bch2_page_state_release(page);
 292 }
 293
 294 /* for newly allocated pages: */
 295 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 296                                                        gfp_t gfp)
 297 {
 298         struct bch_page_state *s;
 299
 300         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 301         if (!s)
 302                 return NULL;
 303
 304         spin_lock_init(&s->lock);
 305         attach_page_private(page, s);
 306         return s;
 307 }
 308
 309 static struct bch_page_state *bch2_page_state_create(struct page *page,
 310                                                      gfp_t gfp)
 311 {
 312         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 313 }
 314
 315 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 316 {
 317         /* XXX: this should not be open coded */
 318         return inode->ei_inode.bi_data_replicas
 319                 ? inode->ei_inode.bi_data_replicas - 1
 320                 : c->opts.data_replicas;
 321 }
 322
 323 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 324                                                   unsigned nr_replicas)
 325 {
 326         return max(0, (int) nr_replicas -
 327                    s->nr_replicas -
 328                    s->replicas_reserved);
 329 }
 330
 331 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 332                                 struct bch_inode_info *inode,
 333                                 struct page *page, bool check_enospc)
 334 {
 335         struct bch_page_state *s = bch2_page_state_create(page, 0);
 336         unsigned nr_replicas = inode_nr_replicas(c, inode);
 337         struct disk_reservation disk_res = { 0 };
 338         unsigned i, disk_res_sectors = 0;
 339         int ret;
 340
 341         if (!s)
 342                 return -ENOMEM;
 343
 344         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 345                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 346
 347         if (!disk_res_sectors)
 348                 return 0;
 349
 350         ret = bch2_disk_reservation_get(c, &disk_res,
 351                                         disk_res_sectors, 1,
 352                                         !check_enospc
 353                                         ? BCH_DISK_RESERVATION_NOFAIL
 354                                         : 0);
 355         if (unlikely(ret))
 356                 return ret;
 357
 358         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 359                 s->s[i].replicas_reserved +=
 360                         sectors_to_reserve(&s->s[i], nr_replicas);
 361
 362         return 0;
 363 }
 364
 365 struct bch2_page_reservation {
 366         struct disk_reservation disk;
 367         struct quota_res        quota;
 368 };
 369
 370 static void bch2_page_reservation_init(struct bch_fs *c,
 371                         struct bch_inode_info *inode,
 372                         struct bch2_page_reservation *res)
 373 {
 374         memset(res, 0, sizeof(*res));
 375
 376         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 377 }
 378
 379 static void bch2_page_reservation_put(struct bch_fs *c,
 380                         struct bch_inode_info *inode,
 381                         struct bch2_page_reservation *res)
 382 {
 383         bch2_disk_reservation_put(c, &res->disk);
 384         bch2_quota_reservation_put(c, inode, &res->quota);
 385 }
 386
 387 static int bch2_page_reservation_get(struct bch_fs *c,
 388                         struct bch_inode_info *inode, struct page *page,
 389                         struct bch2_page_reservation *res,
 390                         unsigned offset, unsigned len, bool check_enospc)
 391 {
 392         struct bch_page_state *s = bch2_page_state_create(page, 0);
 393         unsigned i, disk_sectors = 0, quota_sectors = 0;
 394         int ret;
 395
 396         if (!s)
 397                 return -ENOMEM;
 398
 399         for (i = round_down(offset, block_bytes(c)) >> 9;
 400              i < round_up(offset + len, block_bytes(c)) >> 9;
 401              i++) {
 402                 disk_sectors += sectors_to_reserve(&s->s[i],
 403                                                 res->disk.nr_replicas);
 404                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 405         }
 406
 407         if (disk_sectors) {
 408                 ret = bch2_disk_reservation_add(c, &res->disk,
 409                                                 disk_sectors,
 410                                                 !check_enospc
 411                                                 ? BCH_DISK_RESERVATION_NOFAIL
 412                                                 : 0);
 413                 if (unlikely(ret))
 414                         return ret;
 415         }
 416
 417         if (quota_sectors) {
 418                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 419                                                  quota_sectors,
 420                                                  check_enospc);
 421                 if (unlikely(ret)) {
 422                         struct disk_reservation tmp = {
 423                                 .sectors = disk_sectors
 424                         };
 425
 426                         bch2_disk_reservation_put(c, &tmp);
 427                         res->disk.sectors -= disk_sectors;
 428                         return ret;
 429                 }
 430         }
 431
 432         return 0;
 433 }
 434
 435 static void bch2_clear_page_bits(struct page *page)
 436 {
 437         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 438         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 439         struct bch_page_state *s = bch2_page_state(page);
 440         struct disk_reservation disk_res = { 0 };
 441         int i, dirty_sectors = 0;
 442
 443         if (!s)
 444                 return;
 445
 446         EBUG_ON(!PageLocked(page));
 447         EBUG_ON(PageWriteback(page));
 448
 449         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 450                 disk_res.sectors += s->s[i].replicas_reserved;
 451                 s->s[i].replicas_reserved = 0;
 452
 453                 if (s->s[i].state == SECTOR_DIRTY) {
 454                         dirty_sectors++;
 455                         s->s[i].state = SECTOR_UNALLOCATED;
 456                 }
 457         }
 458
 459         bch2_disk_reservation_put(c, &disk_res);
 460
 461         if (dirty_sectors)
 462                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 463
 464         bch2_page_state_release(page);
 465 }
 466
 467 static void bch2_set_page_dirty(struct bch_fs *c,
 468                         struct bch_inode_info *inode, struct page *page,
 469                         struct bch2_page_reservation *res,
 470                         unsigned offset, unsigned len)
 471 {
 472         struct bch_page_state *s = bch2_page_state(page);
 473         unsigned i, dirty_sectors = 0;
 474
 475         WARN_ON((u64) page_offset(page) + offset + len >
 476                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 477
 478         spin_lock(&s->lock);
 479
 480         for (i = round_down(offset, block_bytes(c)) >> 9;
 481              i < round_up(offset + len, block_bytes(c)) >> 9;
 482              i++) {
 483                 unsigned sectors = sectors_to_reserve(&s->s[i],
 484                                                 res->disk.nr_replicas);
 485
 486                 /*
 487                  * This can happen if we race with the error path in
 488                  * bch2_writepage_io_done():
 489                  */
 490                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 491
 492                 s->s[i].replicas_reserved += sectors;
 493                 res->disk.sectors -= sectors;
 494
 495                 if (s->s[i].state == SECTOR_UNALLOCATED)
 496                         dirty_sectors++;
 497
 498                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 499         }
 500
 501         spin_unlock(&s->lock);
 502
 503         if (dirty_sectors)
 504                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 505
 506         if (!PageDirty(page))
 507                 __set_page_dirty_nobuffers(page);
 508 }
 509
 510 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 511 {
 512         struct file *file = vmf->vma->vm_file;
 513         struct address_space *mapping = file->f_mapping;
 514         struct address_space *fdm = faults_disabled_mapping();
 515         struct bch_inode_info *inode = file_bch_inode(file);
 516         int ret;
 517
 518         if (fdm == mapping)
 519                 return VM_FAULT_SIGBUS;
 520
 521         /* Lock ordering: */
 522         if (fdm > mapping) {
 523                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 524
 525                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 526                         goto got_lock;
 527
 528                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 529
 530                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 531                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 532
 533                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 534
 535                 /* Signal that lock has been dropped: */
 536                 set_fdm_dropped_locks();
 537                 return VM_FAULT_SIGBUS;
 538         }
 539
 540         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 541 got_lock:
 542         ret = filemap_fault(vmf);
 543         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 544
 545         return ret;
 546 }
 547
 548 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 549 {
 550         struct page *page = vmf->page;
 551         struct file *file = vmf->vma->vm_file;
 552         struct bch_inode_info *inode = file_bch_inode(file);
 553         struct address_space *mapping = file->f_mapping;
 554         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 555         struct bch2_page_reservation res;
 556         unsigned len;
 557         loff_t isize;
 558         int ret = VM_FAULT_LOCKED;
 559
 560         bch2_page_reservation_init(c, inode, &res);
 561
 562         sb_start_pagefault(inode->v.i_sb);
 563         file_update_time(file);
 564
 565         /*
 566          * Not strictly necessary, but helps avoid dio writes livelocking in
 567          * write_invalidate_inode_pages_range() - can drop this if/when we get
 568          * a write_invalidate_inode_pages_range() that works without dropping
 569          * page lock before invalidating page
 570          */
 571         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 572
 573         lock_page(page);
 574         isize = i_size_read(&inode->v);
 575
 576         if (page->mapping != mapping || page_offset(page) >= isize) {
 577                 unlock_page(page);
 578                 ret = VM_FAULT_NOPAGE;
 579                 goto out;
 580         }
 581
 582         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 583
 584         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 585                 unlock_page(page);
 586                 ret = VM_FAULT_SIGBUS;
 587                 goto out;
 588         }
 589
 590         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 591         bch2_page_reservation_put(c, inode, &res);
 592
 593         wait_for_stable_page(page);
 594 out:
 595         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 596         sb_end_pagefault(inode->v.i_sb);
 597
 598         return ret;
 599 }
 600
 601 void bch2_invalidatepage(struct page *page, unsigned int offset,
 602                          unsigned int length)
 603 {
 604         if (offset || length < PAGE_SIZE)
 605                 return;
 606
 607         bch2_clear_page_bits(page);
 608 }
 609
 610 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 611 {
 612         if (PageDirty(page))
 613                 return 0;
 614
 615         bch2_clear_page_bits(page);
 616         return 1;
 617 }
 618
 619 #ifdef CONFIG_MIGRATION
 620 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 621                       struct page *page, enum migrate_mode mode)
 622 {
 623         int ret;
 624
 625         EBUG_ON(!PageLocked(page));
 626         EBUG_ON(!PageLocked(newpage));
 627
 628         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 629         if (ret != MIGRATEPAGE_SUCCESS)
 630                 return ret;
 631
 632         if (PagePrivate(page))
 633                 attach_page_private(newpage, detach_page_private(page));
 634
 635         if (mode != MIGRATE_SYNC_NO_COPY)
 636                 migrate_page_copy(newpage, page);
 637         else
 638                 migrate_page_states(newpage, page);
 639         return MIGRATEPAGE_SUCCESS;
 640 }
 641 #endif
 642
 643 /* readpage(s): */
 644
 645 static void bch2_readpages_end_io(struct bio *bio)
 646 {
 647         struct bvec_iter_all iter;
 648         struct bio_vec *bv;
 649
 650         bio_for_each_segment_all(bv, bio, iter) {
 651                 struct page *page = bv->bv_page;
 652
 653                 if (!bio->bi_status) {
 654                         SetPageUptodate(page);
 655                 } else {
 656                         ClearPageUptodate(page);
 657                         SetPageError(page);
 658                 }
 659                 unlock_page(page);
 660         }
 661
 662         bio_put(bio);
 663 }
 664
 665 struct readpages_iter {
 666         struct address_space    *mapping;
 667         struct page             **pages;
 668         unsigned                nr_pages;
 669         unsigned                idx;
 670         pgoff_t                 offset;
 671 };
 672
 673 static int readpages_iter_init(struct readpages_iter *iter,
 674                                struct readahead_control *ractl)
 675 {
 676         unsigned i, nr_pages = readahead_count(ractl);
 677
 678         memset(iter, 0, sizeof(*iter));
 679
 680         iter->mapping   = ractl->mapping;
 681         iter->offset    = readahead_index(ractl);
 682         iter->nr_pages  = nr_pages;
 683
 684         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 685         if (!iter->pages)
 686                 return -ENOMEM;
 687
 688         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 689         for (i = 0; i < nr_pages; i++) {
 690                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 691                 put_page(iter->pages[i]);
 692         }
 693
 694         return 0;
 695 }
 696
 697 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 698 {
 699         if (iter->idx >= iter->nr_pages)
 700                 return NULL;
 701
 702         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 703
 704         return iter->pages[iter->idx];
 705 }
 706
 707 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 708 {
 709         struct bvec_iter iter;
 710         struct bio_vec bv;
 711         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 712                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 713         unsigned state = k.k->type == KEY_TYPE_reservation
 714                 ? SECTOR_RESERVED
 715                 : SECTOR_ALLOCATED;
 716
 717         bio_for_each_segment(bv, bio, iter) {
 718                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 719                 unsigned i;
 720
 721                 for (i = bv.bv_offset >> 9;
 722                      i < (bv.bv_offset + bv.bv_len) >> 9;
 723                      i++) {
 724                         s->s[i].nr_replicas = nr_ptrs;
 725                         s->s[i].state = state;
 726                 }
 727         }
 728 }
 729
 730 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 731 {
 732         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 733         struct bch_extent_crc_unpacked crc;
 734         const union bch_extent_entry *i;
 735
 736         bkey_for_each_crc(k.k, ptrs, crc, i)
 737                 if (crc.csum_type || crc.compression_type)
 738                         return true;
 739         return false;
 740 }
 741
 742 static void readpage_bio_extend(struct readpages_iter *iter,
 743                                 struct bio *bio,
 744                                 unsigned sectors_this_extent,
 745                                 bool get_more)
 746 {
 747         while (bio_sectors(bio) < sectors_this_extent &&
 748                bio->bi_vcnt < bio->bi_max_vecs) {
 749                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 750                 struct page *page = readpage_iter_next(iter);
 751                 int ret;
 752
 753                 if (page) {
 754                         if (iter->offset + iter->idx != page_offset)
 755                                 break;
 756
 757                         iter->idx++;
 758                 } else {
 759                         if (!get_more)
 760                                 break;
 761
 762                         page = xa_load(&iter->mapping->i_pages, page_offset);
 763                         if (page && !xa_is_value(page))
 764                                 break;
 765
 766                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 767                         if (!page)
 768                                 break;
 769
 770                         if (!__bch2_page_state_create(page, 0)) {
 771                                 put_page(page);
 772                                 break;
 773                         }
 774
 775                         ret = add_to_page_cache_lru(page, iter->mapping,
 776                                                     page_offset, GFP_NOFS);
 777                         if (ret) {
 778                                 __bch2_page_state_release(page);
 779                                 put_page(page);
 780                                 break;
 781                         }
 782
 783                         put_page(page);
 784                 }
 785
 786                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 787         }
 788 }
 789
 790 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 791                        struct bch_read_bio *rbio, u64 inum,
 792                        struct readpages_iter *readpages_iter)
 793 {
 794         struct bch_fs *c = trans->c;
 795         struct bkey_buf sk;
 796         int flags = BCH_READ_RETRY_IF_STALE|
 797                 BCH_READ_MAY_PROMOTE;
 798         int ret = 0;
 799
 800         rbio->c = c;
 801         rbio->start_time = local_clock();
 802
 803         bch2_bkey_buf_init(&sk);
 804 retry:
 805         while (1) {
 806                 struct bkey_s_c k;
 807                 unsigned bytes, sectors, offset_into_extent;
 808                 enum btree_id data_btree = BTREE_ID_extents;
 809
 810                 bch2_btree_iter_set_pos(iter,
 811                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 812
 813                 k = bch2_btree_iter_peek_slot(iter);
 814                 ret = bkey_err(k);
 815                 if (ret)
 816                         break;
 817
 818                 offset_into_extent = iter->pos.offset -
 819                         bkey_start_offset(k.k);
 820                 sectors = k.k->size - offset_into_extent;
 821
 822                 bch2_bkey_buf_reassemble(&sk, c, k);
 823
 824                 ret = bch2_read_indirect_extent(trans, &data_btree,
 825                                         &offset_into_extent, &sk);
 826                 if (ret)
 827                         break;
 828
 829                 k = bkey_i_to_s_c(sk.k);
 830
 831                 sectors = min(sectors, k.k->size - offset_into_extent);
 832
 833                 bch2_trans_unlock(trans);
 834
 835                 if (readpages_iter)
 836                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 837                                             extent_partial_reads_expensive(k));
 838
 839                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 840                 swap(rbio->bio.bi_iter.bi_size, bytes);
 841
 842                 if (rbio->bio.bi_iter.bi_size == bytes)
 843                         flags |= BCH_READ_LAST_FRAGMENT;
 844
 845                 if (bkey_extent_is_allocation(k.k))
 846                         bch2_add_page_sectors(&rbio->bio, k);
 847
 848                 bch2_read_extent(trans, rbio, iter->pos,
 849                                  data_btree, k, offset_into_extent, flags);
 850
 851                 if (flags & BCH_READ_LAST_FRAGMENT)
 852                         break;
 853
 854                 swap(rbio->bio.bi_iter.bi_size, bytes);
 855                 bio_advance(&rbio->bio, bytes);
 856         }
 857
 858         if (ret == -EINTR)
 859                 goto retry;
 860
 861         if (ret) {
 862                 bch_err_inum_ratelimited(c, inum,
 863                                 "read error %i from btree lookup", ret);
 864                 rbio->bio.bi_status = BLK_STS_IOERR;
 865                 bio_endio(&rbio->bio);
 866         }
 867
 868         bch2_bkey_buf_exit(&sk, c);
 869 }
 870
 871 void bch2_readahead(struct readahead_control *ractl)
 872 {
 873         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 874         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 875         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 876         struct btree_trans trans;
 877         struct btree_iter *iter;
 878         struct page *page;
 879         struct readpages_iter readpages_iter;
 880         int ret;
 881
 882         ret = readpages_iter_init(&readpages_iter, ractl);
 883         BUG_ON(ret);
 884
 885         bch2_trans_init(&trans, c, 0, 0);
 886         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 887                                    BTREE_ITER_SLOTS);
 888
 889         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 890
 891         while ((page = readpage_iter_next(&readpages_iter))) {
 892                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 893                 unsigned n = min_t(unsigned,
 894                                    readpages_iter.nr_pages -
 895                                    readpages_iter.idx,
 896                                    BIO_MAX_PAGES);
 897                 struct bch_read_bio *rbio =
 898                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 899                                   opts);
 900
 901                 readpages_iter.idx++;
 902
 903                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 904                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 905                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 906                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 907
 908                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 909                            &readpages_iter);
 910         }
 911
 912         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 913
 914         bch2_trans_iter_put(&trans, iter);
 915         bch2_trans_exit(&trans);
 916         kfree(readpages_iter.pages);
 917 }
 918
 919 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 920                              u64 inum, struct page *page)
 921 {
 922         struct btree_trans trans;
 923         struct btree_iter *iter;
 924
 925         bch2_page_state_create(page, __GFP_NOFAIL);
 926
 927         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 928         rbio->bio.bi_iter.bi_sector =
 929                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 930         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 931
 932         bch2_trans_init(&trans, c, 0, 0);
 933         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 934                                    BTREE_ITER_SLOTS);
 935
 936         bchfs_read(&trans, iter, rbio, inum, NULL);
 937
 938         bch2_trans_iter_put(&trans, iter);
 939         bch2_trans_exit(&trans);
 940 }
 941
 942 int bch2_readpage(struct file *file, struct page *page)
 943 {
 944         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 945         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 946         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 947         struct bch_read_bio *rbio;
 948
 949         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 950         rbio->bio.bi_end_io = bch2_readpages_end_io;
 951
 952         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 953         return 0;
 954 }
 955
 956 static void bch2_read_single_page_end_io(struct bio *bio)
 957 {
 958         complete(bio->bi_private);
 959 }
 960
 961 static int bch2_read_single_page(struct page *page,
 962                                  struct address_space *mapping)
 963 {
 964         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 965         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 966         struct bch_read_bio *rbio;
 967         int ret;
 968         DECLARE_COMPLETION_ONSTACK(done);
 969
 970         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 971                          io_opts(c, &inode->ei_inode));
 972         rbio->bio.bi_private = &done;
 973         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 974
 975         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 976         wait_for_completion(&done);
 977
 978         ret = blk_status_to_errno(rbio->bio.bi_status);
 979         bio_put(&rbio->bio);
 980
 981         if (ret < 0)
 982                 return ret;
 983
 984         SetPageUptodate(page);
 985         return 0;
 986 }
 987
 988 /* writepages: */
 989
 990 struct bch_writepage_state {
 991         struct bch_writepage_io *io;
 992         struct bch_io_opts      opts;
 993 };
 994
 995 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 996                                                                   struct bch_inode_info *inode)
 997 {
 998         return (struct bch_writepage_state) {
 999                 .opts = io_opts(c, &inode->ei_inode)
1000         };
1001 }
1002
1003 static void bch2_writepage_io_free(struct closure *cl)
1004 {
1005         struct bch_writepage_io *io = container_of(cl,
1006                                         struct bch_writepage_io, cl);
1007
1008         bio_put(&io->op.wbio.bio);
1009 }
1010
1011 static void bch2_writepage_io_done(struct closure *cl)
1012 {
1013         struct bch_writepage_io *io = container_of(cl,
1014                                         struct bch_writepage_io, cl);
1015         struct bch_fs *c = io->op.c;
1016         struct bio *bio = &io->op.wbio.bio;
1017         struct bvec_iter_all iter;
1018         struct bio_vec *bvec;
1019         unsigned i;
1020
1021         if (io->op.error) {
1022                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1023
1024                 bio_for_each_segment_all(bvec, bio, iter) {
1025                         struct bch_page_state *s;
1026
1027                         SetPageError(bvec->bv_page);
1028                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1029
1030                         s = __bch2_page_state(bvec->bv_page);
1031                         spin_lock(&s->lock);
1032                         for (i = 0; i < PAGE_SECTORS; i++)
1033                                 s->s[i].nr_replicas = 0;
1034                         spin_unlock(&s->lock);
1035                 }
1036         }
1037
1038         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1039                 bio_for_each_segment_all(bvec, bio, iter) {
1040                         struct bch_page_state *s;
1041
1042                         s = __bch2_page_state(bvec->bv_page);
1043                         spin_lock(&s->lock);
1044                         for (i = 0; i < PAGE_SECTORS; i++)
1045                                 s->s[i].nr_replicas = 0;
1046                         spin_unlock(&s->lock);
1047                 }
1048         }
1049
1050         /*
1051          * racing with fallocate can cause us to add fewer sectors than
1052          * expected - but we shouldn't add more sectors than expected:
1053          */
1054         BUG_ON(io->op.i_sectors_delta > 0);
1055
1056         /*
1057          * (error (due to going RO) halfway through a page can screw that up
1058          * slightly)
1059          * XXX wtf?
1060            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1061          */
1062
1063         /*
1064          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1065          * before calling end_page_writeback:
1066          */
1067         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1068
1069         bio_for_each_segment_all(bvec, bio, iter) {
1070                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1071
1072                 if (atomic_dec_and_test(&s->write_count))
1073                         end_page_writeback(bvec->bv_page);
1074         }
1075
1076         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1077 }
1078
1079 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1080 {
1081         struct bch_writepage_io *io = w->io;
1082
1083         w->io = NULL;
1084         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1085         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1086 }
1087
1088 /*
1089  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1090  * possible, else allocating a new one:
1091  */
1092 static void bch2_writepage_io_alloc(struct bch_fs *c,
1093                                     struct writeback_control *wbc,
1094                                     struct bch_writepage_state *w,
1095                                     struct bch_inode_info *inode,
1096                                     u64 sector,
1097                                     unsigned nr_replicas)
1098 {
1099         struct bch_write_op *op;
1100
1101         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1102                                               BIO_MAX_PAGES,
1103                                               &c->writepage_bioset),
1104                              struct bch_writepage_io, op.wbio.bio);
1105
1106         closure_init(&w->io->cl, NULL);
1107         w->io->inode            = inode;
1108
1109         op                      = &w->io->op;
1110         bch2_write_op_init(op, c, w->opts);
1111         op->target              = w->opts.foreground_target;
1112         op_journal_seq_set(op, &inode->ei_journal_seq);
1113         op->nr_replicas         = nr_replicas;
1114         op->res.nr_replicas     = nr_replicas;
1115         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1116         op->pos                 = POS(inode->v.i_ino, sector);
1117         op->wbio.bio.bi_iter.bi_sector = sector;
1118         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1119 }
1120
1121 static int __bch2_writepage(struct page *page,
1122                             struct writeback_control *wbc,
1123                             void *data)
1124 {
1125         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1126         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1127         struct bch_writepage_state *w = data;
1128         struct bch_page_state *s, orig;
1129         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1130         loff_t i_size = i_size_read(&inode->v);
1131         pgoff_t end_index = i_size >> PAGE_SHIFT;
1132         int ret;
1133
1134         EBUG_ON(!PageUptodate(page));
1135
1136         /* Is the page fully inside i_size? */
1137         if (page->index < end_index)
1138                 goto do_io;
1139
1140         /* Is the page fully outside i_size? (truncate in progress) */
1141         offset = i_size & (PAGE_SIZE - 1);
1142         if (page->index > end_index || !offset) {
1143                 unlock_page(page);
1144                 return 0;
1145         }
1146
1147         /*
1148          * The page straddles i_size.  It must be zeroed out on each and every
1149          * writepage invocation because it may be mmapped.  "A file is mapped
1150          * in multiples of the page size.  For a file that is not a multiple of
1151          * the  page size, the remaining memory is zeroed when mapped, and
1152          * writes to that region are not written out to the file."
1153          */
1154         zero_user_segment(page, offset, PAGE_SIZE);
1155 do_io:
1156         s = bch2_page_state_create(page, __GFP_NOFAIL);
1157
1158         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1159         if (ret) {
1160                 SetPageError(page);
1161                 mapping_set_error(page->mapping, ret);
1162                 unlock_page(page);
1163                 return 0;
1164         }
1165
1166         /* Before unlocking the page, get copy of reservations: */
1167         orig = *s;
1168
1169         for (i = 0; i < PAGE_SECTORS; i++) {
1170                 if (s->s[i].state < SECTOR_DIRTY)
1171                         continue;
1172
1173                 nr_replicas_this_write =
1174                         min_t(unsigned, nr_replicas_this_write,
1175                               s->s[i].nr_replicas +
1176                               s->s[i].replicas_reserved);
1177         }
1178
1179         for (i = 0; i < PAGE_SECTORS; i++) {
1180                 if (s->s[i].state < SECTOR_DIRTY)
1181                         continue;
1182
1183                 s->s[i].nr_replicas = w->opts.compression
1184                         ? 0 : nr_replicas_this_write;
1185
1186                 s->s[i].replicas_reserved = 0;
1187                 s->s[i].state = SECTOR_ALLOCATED;
1188         }
1189
1190         BUG_ON(atomic_read(&s->write_count));
1191         atomic_set(&s->write_count, 1);
1192
1193         BUG_ON(PageWriteback(page));
1194         set_page_writeback(page);
1195
1196         unlock_page(page);
1197
1198         offset = 0;
1199         while (1) {
1200                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1201                 u64 sector;
1202
1203                 while (offset < PAGE_SECTORS &&
1204                        orig.s[offset].state < SECTOR_DIRTY)
1205                         offset++;
1206
1207                 if (offset == PAGE_SECTORS)
1208                         break;
1209
1210                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1211
1212                 while (offset + sectors < PAGE_SECTORS &&
1213                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1214                         sectors++;
1215
1216                 for (i = offset; i < offset + sectors; i++) {
1217                         reserved_sectors += orig.s[i].replicas_reserved;
1218                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1219                 }
1220
1221                 if (w->io &&
1222                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1223                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1224                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1225                      (BIO_MAX_PAGES * PAGE_SIZE) ||
1226                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1227                         bch2_writepage_do_io(w);
1228
1229                 if (!w->io)
1230                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1231                                                 nr_replicas_this_write);
1232
1233                 atomic_inc(&s->write_count);
1234
1235                 BUG_ON(inode != w->io->inode);
1236                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1237                                      sectors << 9, offset << 9));
1238
1239                 /* Check for writing past i_size: */
1240                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1241                         round_up(i_size, block_bytes(c)));
1242
1243                 w->io->op.res.sectors += reserved_sectors;
1244                 w->io->op.i_sectors_delta -= dirty_sectors;
1245                 w->io->op.new_i_size = i_size;
1246
1247                 offset += sectors;
1248         }
1249
1250         if (atomic_dec_and_test(&s->write_count))
1251                 end_page_writeback(page);
1252
1253         return 0;
1254 }
1255
1256 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1257 {
1258         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1259         struct bch_writepage_state w =
1260                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1261         struct blk_plug plug;
1262         int ret;
1263
1264         blk_start_plug(&plug);
1265         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1266         if (w.io)
1267                 bch2_writepage_do_io(&w);
1268         blk_finish_plug(&plug);
1269         return ret;
1270 }
1271
1272 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1273 {
1274         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1275         struct bch_writepage_state w =
1276                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1277         int ret;
1278
1279         ret = __bch2_writepage(page, wbc, &w);
1280         if (w.io)
1281                 bch2_writepage_do_io(&w);
1282
1283         return ret;
1284 }
1285
1286 /* buffered writes: */
1287
1288 int bch2_write_begin(struct file *file, struct address_space *mapping,
1289                      loff_t pos, unsigned len, unsigned flags,
1290                      struct page **pagep, void **fsdata)
1291 {
1292         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1293         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1294         struct bch2_page_reservation *res;
1295         pgoff_t index = pos >> PAGE_SHIFT;
1296         unsigned offset = pos & (PAGE_SIZE - 1);
1297         struct page *page;
1298         int ret = -ENOMEM;
1299
1300         res = kmalloc(sizeof(*res), GFP_KERNEL);
1301         if (!res)
1302                 return -ENOMEM;
1303
1304         bch2_page_reservation_init(c, inode, res);
1305         *fsdata = res;
1306
1307         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1308
1309         page = grab_cache_page_write_begin(mapping, index, flags);
1310         if (!page)
1311                 goto err_unlock;
1312
1313         if (PageUptodate(page))
1314                 goto out;
1315
1316         /* If we're writing entire page, don't need to read it in first: */
1317         if (len == PAGE_SIZE)
1318                 goto out;
1319
1320         if (!offset && pos + len >= inode->v.i_size) {
1321                 zero_user_segment(page, len, PAGE_SIZE);
1322                 flush_dcache_page(page);
1323                 goto out;
1324         }
1325
1326         if (index > inode->v.i_size >> PAGE_SHIFT) {
1327                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1328                 flush_dcache_page(page);
1329                 goto out;
1330         }
1331 readpage:
1332         ret = bch2_read_single_page(page, mapping);
1333         if (ret)
1334                 goto err;
1335 out:
1336         ret = bch2_page_reservation_get(c, inode, page, res,
1337                                         offset, len, true);
1338         if (ret) {
1339                 if (!PageUptodate(page)) {
1340                         /*
1341                          * If the page hasn't been read in, we won't know if we
1342                          * actually need a reservation - we don't actually need
1343                          * to read here, we just need to check if the page is
1344                          * fully backed by uncompressed data:
1345                          */
1346                         goto readpage;
1347                 }
1348
1349                 goto err;
1350         }
1351
1352         *pagep = page;
1353         return 0;
1354 err:
1355         unlock_page(page);
1356         put_page(page);
1357         *pagep = NULL;
1358 err_unlock:
1359         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1360         kfree(res);
1361         *fsdata = NULL;
1362         return ret;
1363 }
1364
1365 int bch2_write_end(struct file *file, struct address_space *mapping,
1366                    loff_t pos, unsigned len, unsigned copied,
1367                    struct page *page, void *fsdata)
1368 {
1369         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1370         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1371         struct bch2_page_reservation *res = fsdata;
1372         unsigned offset = pos & (PAGE_SIZE - 1);
1373
1374         lockdep_assert_held(&inode->v.i_rwsem);
1375
1376         if (unlikely(copied < len && !PageUptodate(page))) {
1377                 /*
1378                  * The page needs to be read in, but that would destroy
1379                  * our partial write - simplest thing is to just force
1380                  * userspace to redo the write:
1381                  */
1382                 zero_user(page, 0, PAGE_SIZE);
1383                 flush_dcache_page(page);
1384                 copied = 0;
1385         }
1386
1387         spin_lock(&inode->v.i_lock);
1388         if (pos + copied > inode->v.i_size)
1389                 i_size_write(&inode->v, pos + copied);
1390         spin_unlock(&inode->v.i_lock);
1391
1392         if (copied) {
1393                 if (!PageUptodate(page))
1394                         SetPageUptodate(page);
1395
1396                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1397
1398                 inode->ei_last_dirtied = (unsigned long) current;
1399         }
1400
1401         unlock_page(page);
1402         put_page(page);
1403         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1404
1405         bch2_page_reservation_put(c, inode, res);
1406         kfree(res);
1407
1408         return copied;
1409 }
1410
1411 #define WRITE_BATCH_PAGES       32
1412
1413 static int __bch2_buffered_write(struct bch_inode_info *inode,
1414                                  struct address_space *mapping,
1415                                  struct iov_iter *iter,
1416                                  loff_t pos, unsigned len)
1417 {
1418         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1419         struct page *pages[WRITE_BATCH_PAGES];
1420         struct bch2_page_reservation res;
1421         unsigned long index = pos >> PAGE_SHIFT;
1422         unsigned offset = pos & (PAGE_SIZE - 1);
1423         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1424         unsigned i, reserved = 0, set_dirty = 0;
1425         unsigned copied = 0, nr_pages_copied = 0;
1426         int ret = 0;
1427
1428         BUG_ON(!len);
1429         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1430
1431         bch2_page_reservation_init(c, inode, &res);
1432
1433         for (i = 0; i < nr_pages; i++) {
1434                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1435                 if (!pages[i]) {
1436                         nr_pages = i;
1437                         if (!i) {
1438                                 ret = -ENOMEM;
1439                                 goto out;
1440                         }
1441                         len = min_t(unsigned, len,
1442                                     nr_pages * PAGE_SIZE - offset);
1443                         break;
1444                 }
1445         }
1446
1447         if (offset && !PageUptodate(pages[0])) {
1448                 ret = bch2_read_single_page(pages[0], mapping);
1449                 if (ret)
1450                         goto out;
1451         }
1452
1453         if ((pos + len) & (PAGE_SIZE - 1) &&
1454             !PageUptodate(pages[nr_pages - 1])) {
1455                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1456                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1457                 } else {
1458                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1459                         if (ret)
1460                                 goto out;
1461                 }
1462         }
1463
1464         while (reserved < len) {
1465                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1466                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1467                 unsigned pg_len = min_t(unsigned, len - reserved,
1468                                         PAGE_SIZE - pg_offset);
1469 retry_reservation:
1470                 ret = bch2_page_reservation_get(c, inode, page, &res,
1471                                                 pg_offset, pg_len, true);
1472
1473                 if (ret && !PageUptodate(page)) {
1474                         ret = bch2_read_single_page(page, mapping);
1475                         if (!ret)
1476                                 goto retry_reservation;
1477                 }
1478
1479                 if (ret)
1480                         goto out;
1481
1482                 reserved += pg_len;
1483         }
1484
1485         if (mapping_writably_mapped(mapping))
1486                 for (i = 0; i < nr_pages; i++)
1487                         flush_dcache_page(pages[i]);
1488
1489         while (copied < len) {
1490                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1491                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1492                 unsigned pg_len = min_t(unsigned, len - copied,
1493                                         PAGE_SIZE - pg_offset);
1494                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1495                                                 iter, pg_offset, pg_len);
1496
1497                 if (!pg_copied)
1498                         break;
1499
1500                 if (!PageUptodate(page) &&
1501                     pg_copied != PAGE_SIZE &&
1502                     pos + copied + pg_copied < inode->v.i_size) {
1503                         zero_user(page, 0, PAGE_SIZE);
1504                         break;
1505                 }
1506
1507                 flush_dcache_page(page);
1508                 iov_iter_advance(iter, pg_copied);
1509                 copied += pg_copied;
1510
1511                 if (pg_copied != pg_len)
1512                         break;
1513         }
1514
1515         if (!copied)
1516                 goto out;
1517
1518         spin_lock(&inode->v.i_lock);
1519         if (pos + copied > inode->v.i_size)
1520                 i_size_write(&inode->v, pos + copied);
1521         spin_unlock(&inode->v.i_lock);
1522
1523         while (set_dirty < copied) {
1524                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1525                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1526                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1527                                         PAGE_SIZE - pg_offset);
1528
1529                 if (!PageUptodate(page))
1530                         SetPageUptodate(page);
1531
1532                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1533                 unlock_page(page);
1534                 put_page(page);
1535
1536                 set_dirty += pg_len;
1537         }
1538
1539         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1540         inode->ei_last_dirtied = (unsigned long) current;
1541 out:
1542         for (i = nr_pages_copied; i < nr_pages; i++) {
1543                 unlock_page(pages[i]);
1544                 put_page(pages[i]);
1545         }
1546
1547         bch2_page_reservation_put(c, inode, &res);
1548
1549         return copied ?: ret;
1550 }
1551
1552 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1553 {
1554         struct file *file = iocb->ki_filp;
1555         struct address_space *mapping = file->f_mapping;
1556         struct bch_inode_info *inode = file_bch_inode(file);
1557         loff_t pos = iocb->ki_pos;
1558         ssize_t written = 0;
1559         int ret = 0;
1560
1561         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1562
1563         do {
1564                 unsigned offset = pos & (PAGE_SIZE - 1);
1565                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1566                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1567 again:
1568                 /*
1569                  * Bring in the user page that we will copy from _first_.
1570                  * Otherwise there's a nasty deadlock on copying from the
1571                  * same page as we're writing to, without it being marked
1572                  * up-to-date.
1573                  *
1574                  * Not only is this an optimisation, but it is also required
1575                  * to check that the address is actually valid, when atomic
1576                  * usercopies are used, below.
1577                  */
1578                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1579                         bytes = min_t(unsigned long, iov_iter_count(iter),
1580                                       PAGE_SIZE - offset);
1581
1582                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1583                                 ret = -EFAULT;
1584                                 break;
1585                         }
1586                 }
1587
1588                 if (unlikely(fatal_signal_pending(current))) {
1589                         ret = -EINTR;
1590                         break;
1591                 }
1592
1593                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1594                 if (unlikely(ret < 0))
1595                         break;
1596
1597                 cond_resched();
1598
1599                 if (unlikely(ret == 0)) {
1600                         /*
1601                          * If we were unable to copy any data at all, we must
1602                          * fall back to a single segment length write.
1603                          *
1604                          * If we didn't fallback here, we could livelock
1605                          * because not all segments in the iov can be copied at
1606                          * once without a pagefault.
1607                          */
1608                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1609                                       iov_iter_single_seg_count(iter));
1610                         goto again;
1611                 }
1612                 pos += ret;
1613                 written += ret;
1614                 ret = 0;
1615
1616                 balance_dirty_pages_ratelimited(mapping);
1617         } while (iov_iter_count(iter));
1618
1619         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1620
1621         return written ? written : ret;
1622 }
1623
1624 /* O_DIRECT reads */
1625
1626 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1627 {
1628         if (check_dirty) {
1629                 bio_check_pages_dirty(bio);
1630         } else {
1631                 bio_release_pages(bio, false);
1632                 bio_put(bio);
1633         }
1634 }
1635
1636 static void bch2_dio_read_complete(struct closure *cl)
1637 {
1638         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1639
1640         dio->req->ki_complete(dio->req, dio->ret, 0);
1641         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1642 }
1643
1644 static void bch2_direct_IO_read_endio(struct bio *bio)
1645 {
1646         struct dio_read *dio = bio->bi_private;
1647
1648         if (bio->bi_status)
1649                 dio->ret = blk_status_to_errno(bio->bi_status);
1650
1651         closure_put(&dio->cl);
1652 }
1653
1654 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1655 {
1656         struct dio_read *dio = bio->bi_private;
1657         bool should_dirty = dio->should_dirty;
1658
1659         bch2_direct_IO_read_endio(bio);
1660         bio_check_or_release(bio, should_dirty);
1661 }
1662
1663 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1664 {
1665         struct file *file = req->ki_filp;
1666         struct bch_inode_info *inode = file_bch_inode(file);
1667         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1668         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1669         struct dio_read *dio;
1670         struct bio *bio;
1671         loff_t offset = req->ki_pos;
1672         bool sync = is_sync_kiocb(req);
1673         size_t shorten;
1674         ssize_t ret;
1675
1676         if ((offset|iter->count) & (block_bytes(c) - 1))
1677                 return -EINVAL;
1678
1679         ret = min_t(loff_t, iter->count,
1680                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1681
1682         if (!ret)
1683                 return ret;
1684
1685         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1686         iter->count -= shorten;
1687
1688         bio = bio_alloc_bioset(GFP_KERNEL,
1689                                iov_iter_npages(iter, BIO_MAX_PAGES),
1690                                &c->dio_read_bioset);
1691
1692         bio->bi_end_io = bch2_direct_IO_read_endio;
1693
1694         dio = container_of(bio, struct dio_read, rbio.bio);
1695         closure_init(&dio->cl, NULL);
1696
1697         /*
1698          * this is a _really_ horrible hack just to avoid an atomic sub at the
1699          * end:
1700          */
1701         if (!sync) {
1702                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1703                 atomic_set(&dio->cl.remaining,
1704                            CLOSURE_REMAINING_INITIALIZER -
1705                            CLOSURE_RUNNING +
1706                            CLOSURE_DESTRUCTOR);
1707         } else {
1708                 atomic_set(&dio->cl.remaining,
1709                            CLOSURE_REMAINING_INITIALIZER + 1);
1710         }
1711
1712         dio->req        = req;
1713         dio->ret        = ret;
1714         /*
1715          * This is one of the sketchier things I've encountered: we have to skip
1716          * the dirtying of requests that are internal from the kernel (i.e. from
1717          * loopback), because we'll deadlock on page_lock.
1718          */
1719         dio->should_dirty = iter_is_iovec(iter);
1720
1721         goto start;
1722         while (iter->count) {
1723                 bio = bio_alloc_bioset(GFP_KERNEL,
1724                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1725                                        &c->bio_read);
1726                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1727 start:
1728                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1729                 bio->bi_iter.bi_sector  = offset >> 9;
1730                 bio->bi_private         = dio;
1731
1732                 ret = bio_iov_iter_get_pages(bio, iter);
1733                 if (ret < 0) {
1734                         /* XXX: fault inject this path */
1735                         bio->bi_status = BLK_STS_RESOURCE;
1736                         bio_endio(bio);
1737                         break;
1738                 }
1739
1740                 offset += bio->bi_iter.bi_size;
1741
1742                 if (dio->should_dirty)
1743                         bio_set_pages_dirty(bio);
1744
1745                 if (iter->count)
1746                         closure_get(&dio->cl);
1747
1748                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1749         }
1750
1751         iter->count += shorten;
1752
1753         if (sync) {
1754                 closure_sync(&dio->cl);
1755                 closure_debug_destroy(&dio->cl);
1756                 ret = dio->ret;
1757                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1758                 return ret;
1759         } else {
1760                 return -EIOCBQUEUED;
1761         }
1762 }
1763
1764 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1765 {
1766         struct file *file = iocb->ki_filp;
1767         struct bch_inode_info *inode = file_bch_inode(file);
1768         struct address_space *mapping = file->f_mapping;
1769         size_t count = iov_iter_count(iter);
1770         ssize_t ret;
1771
1772         if (!count)
1773                 return 0; /* skip atime */
1774
1775         if (iocb->ki_flags & IOCB_DIRECT) {
1776                 struct blk_plug plug;
1777
1778                 ret = filemap_write_and_wait_range(mapping,
1779                                         iocb->ki_pos,
1780                                         iocb->ki_pos + count - 1);
1781                 if (ret < 0)
1782                         return ret;
1783
1784                 file_accessed(file);
1785
1786                 blk_start_plug(&plug);
1787                 ret = bch2_direct_IO_read(iocb, iter);
1788                 blk_finish_plug(&plug);
1789
1790                 if (ret >= 0)
1791                         iocb->ki_pos += ret;
1792         } else {
1793                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1794                 ret = generic_file_read_iter(iocb, iter);
1795                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1796         }
1797
1798         return ret;
1799 }
1800
1801 /* O_DIRECT writes */
1802
1803 static void bch2_dio_write_loop_async(struct bch_write_op *);
1804
1805 static long bch2_dio_write_loop(struct dio_write *dio)
1806 {
1807         bool kthread = (current->flags & PF_KTHREAD) != 0;
1808         struct kiocb *req = dio->req;
1809         struct address_space *mapping = req->ki_filp->f_mapping;
1810         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1811         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1812         struct bio *bio = &dio->op.wbio.bio;
1813         struct bvec_iter_all iter;
1814         struct bio_vec *bv;
1815         unsigned unaligned, iter_count;
1816         bool sync = dio->sync, dropped_locks;
1817         long ret;
1818
1819         if (dio->loop)
1820                 goto loop;
1821
1822         while (1) {
1823                 iter_count = dio->iter.count;
1824
1825                 if (kthread)
1826                         kthread_use_mm(dio->mm);
1827                 BUG_ON(current->faults_disabled_mapping);
1828                 current->faults_disabled_mapping = mapping;
1829
1830                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1831
1832                 dropped_locks = fdm_dropped_locks();
1833
1834                 current->faults_disabled_mapping = NULL;
1835                 if (kthread)
1836                         kthread_unuse_mm(dio->mm);
1837
1838                 /*
1839                  * If the fault handler returned an error but also signalled
1840                  * that it dropped & retook ei_pagecache_lock, we just need to
1841                  * re-shoot down the page cache and retry:
1842                  */
1843                 if (dropped_locks && ret)
1844                         ret = 0;
1845
1846                 if (unlikely(ret < 0))
1847                         goto err;
1848
1849                 if (unlikely(dropped_locks)) {
1850                         ret = write_invalidate_inode_pages_range(mapping,
1851                                         req->ki_pos,
1852                                         req->ki_pos + iter_count - 1);
1853                         if (unlikely(ret))
1854                                 goto err;
1855
1856                         if (!bio->bi_iter.bi_size)
1857                                 continue;
1858                 }
1859
1860                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1861                 bio->bi_iter.bi_size -= unaligned;
1862                 iov_iter_revert(&dio->iter, unaligned);
1863
1864                 if (!bio->bi_iter.bi_size) {
1865                         /*
1866                          * bio_iov_iter_get_pages was only able to get <
1867                          * blocksize worth of pages:
1868                          */
1869                         bio_for_each_segment_all(bv, bio, iter)
1870                                 put_page(bv->bv_page);
1871                         ret = -EFAULT;
1872                         goto err;
1873                 }
1874
1875                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1876                 dio->op.end_io          = bch2_dio_write_loop_async;
1877                 dio->op.target          = dio->op.opts.foreground_target;
1878                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1879                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1880                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1881                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1882
1883                 if ((req->ki_flags & IOCB_DSYNC) &&
1884                     !c->opts.journal_flush_disabled)
1885                         dio->op.flags |= BCH_WRITE_FLUSH;
1886
1887                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1888                                                 dio->op.opts.data_replicas, 0);
1889                 if (unlikely(ret) &&
1890                     !bch2_check_range_allocated(c, dio->op.pos,
1891                                 bio_sectors(bio),
1892                                 dio->op.opts.data_replicas,
1893                                 dio->op.opts.compression != 0))
1894                         goto err;
1895
1896                 task_io_account_write(bio->bi_iter.bi_size);
1897
1898                 if (!dio->sync && !dio->loop && dio->iter.count) {
1899                         struct iovec *iov = dio->inline_vecs;
1900
1901                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1902                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1903                                               GFP_KERNEL);
1904                                 if (unlikely(!iov)) {
1905                                         dio->sync = sync = true;
1906                                         goto do_io;
1907                                 }
1908
1909                                 dio->free_iov = true;
1910                         }
1911
1912                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1913                         dio->iter.iov = iov;
1914                 }
1915 do_io:
1916                 dio->loop = true;
1917                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1918
1919                 if (sync)
1920                         wait_for_completion(&dio->done);
1921                 else
1922                         return -EIOCBQUEUED;
1923 loop:
1924                 i_sectors_acct(c, inode, &dio->quota_res,
1925                                dio->op.i_sectors_delta);
1926                 req->ki_pos += (u64) dio->op.written << 9;
1927                 dio->written += dio->op.written;
1928
1929                 spin_lock(&inode->v.i_lock);
1930                 if (req->ki_pos > inode->v.i_size)
1931                         i_size_write(&inode->v, req->ki_pos);
1932                 spin_unlock(&inode->v.i_lock);
1933
1934                 bio_for_each_segment_all(bv, bio, iter)
1935                         put_page(bv->bv_page);
1936
1937                 if (dio->op.error) {
1938                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
1939                         break;
1940                 }
1941
1942                 if (!dio->iter.count)
1943                         break;
1944
1945                 bio_reset(bio);
1946                 reinit_completion(&dio->done);
1947         }
1948
1949         ret = dio->op.error ?: ((long) dio->written << 9);
1950 err:
1951         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1952         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1953
1954         if (dio->free_iov)
1955                 kfree(dio->iter.iov);
1956
1957         bio_put(bio);
1958
1959         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1960         inode_dio_end(&inode->v);
1961
1962         if (!sync) {
1963                 req->ki_complete(req, ret, 0);
1964                 ret = -EIOCBQUEUED;
1965         }
1966         return ret;
1967 }
1968
1969 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1970 {
1971         struct dio_write *dio = container_of(op, struct dio_write, op);
1972
1973         if (dio->sync)
1974                 complete(&dio->done);
1975         else
1976                 bch2_dio_write_loop(dio);
1977 }
1978
1979 static noinline
1980 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1981 {
1982         struct file *file = req->ki_filp;
1983         struct address_space *mapping = file->f_mapping;
1984         struct bch_inode_info *inode = file_bch_inode(file);
1985         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1986         struct dio_write *dio;
1987         struct bio *bio;
1988         bool locked = true, extending;
1989         ssize_t ret;
1990
1991         prefetch(&c->opts);
1992         prefetch((void *) &c->opts + 64);
1993         prefetch(&inode->ei_inode);
1994         prefetch((void *) &inode->ei_inode + 64);
1995
1996         inode_lock(&inode->v);
1997
1998         ret = generic_write_checks(req, iter);
1999         if (unlikely(ret <= 0))
2000                 goto err;
2001
2002         ret = file_remove_privs(file);
2003         if (unlikely(ret))
2004                 goto err;
2005
2006         ret = file_update_time(file);
2007         if (unlikely(ret))
2008                 goto err;
2009
2010         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2011                 goto err;
2012
2013         inode_dio_begin(&inode->v);
2014         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2015
2016         extending = req->ki_pos + iter->count > inode->v.i_size;
2017         if (!extending) {
2018                 inode_unlock(&inode->v);
2019                 locked = false;
2020         }
2021
2022         bio = bio_alloc_bioset(GFP_KERNEL,
2023                                iov_iter_npages(iter, BIO_MAX_PAGES),
2024                                &c->dio_write_bioset);
2025         dio = container_of(bio, struct dio_write, op.wbio.bio);
2026         init_completion(&dio->done);
2027         dio->req                = req;
2028         dio->mm                 = current->mm;
2029         dio->loop               = false;
2030         dio->sync               = is_sync_kiocb(req) || extending;
2031         dio->free_iov           = false;
2032         dio->quota_res.sectors  = 0;
2033         dio->written            = 0;
2034         dio->iter               = *iter;
2035
2036         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2037                                          iter->count >> 9, true);
2038         if (unlikely(ret))
2039                 goto err_put_bio;
2040
2041         ret = write_invalidate_inode_pages_range(mapping,
2042                                         req->ki_pos,
2043                                         req->ki_pos + iter->count - 1);
2044         if (unlikely(ret))
2045                 goto err_put_bio;
2046
2047         ret = bch2_dio_write_loop(dio);
2048 err:
2049         if (locked)
2050                 inode_unlock(&inode->v);
2051         return ret;
2052 err_put_bio:
2053         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2054         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2055         bio_put(bio);
2056         inode_dio_end(&inode->v);
2057         goto err;
2058 }
2059
2060 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2061 {
2062         struct file *file = iocb->ki_filp;
2063         struct bch_inode_info *inode = file_bch_inode(file);
2064         ssize_t ret;
2065
2066         if (iocb->ki_flags & IOCB_DIRECT)
2067                 return bch2_direct_write(iocb, from);
2068
2069         /* We can write back this queue in page reclaim */
2070         current->backing_dev_info = inode_to_bdi(&inode->v);
2071         inode_lock(&inode->v);
2072
2073         ret = generic_write_checks(iocb, from);
2074         if (ret <= 0)
2075                 goto unlock;
2076
2077         ret = file_remove_privs(file);
2078         if (ret)
2079                 goto unlock;
2080
2081         ret = file_update_time(file);
2082         if (ret)
2083                 goto unlock;
2084
2085         ret = bch2_buffered_write(iocb, from);
2086         if (likely(ret > 0))
2087                 iocb->ki_pos += ret;
2088 unlock:
2089         inode_unlock(&inode->v);
2090         current->backing_dev_info = NULL;
2091
2092         if (ret > 0)
2093                 ret = generic_write_sync(iocb, ret);
2094
2095         return ret;
2096 }
2097
2098 /* fsync: */
2099
2100 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2101 {
2102         struct bch_inode_info *inode = file_bch_inode(file);
2103         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2104         int ret, ret2;
2105
2106         ret = file_write_and_wait_range(file, start, end);
2107         if (ret)
2108                 return ret;
2109
2110         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2111                 goto out;
2112
2113         ret = sync_inode_metadata(&inode->v, 1);
2114         if (ret)
2115                 return ret;
2116 out:
2117         if (!c->opts.journal_flush_disabled)
2118                 ret = bch2_journal_flush_seq(&c->journal,
2119                                              inode->ei_journal_seq);
2120         ret2 = file_check_and_advance_wb_err(file);
2121
2122         return ret ?: ret2;
2123 }
2124
2125 /* truncate: */
2126
2127 static inline int range_has_data(struct bch_fs *c,
2128                                   struct bpos start,
2129                                   struct bpos end)
2130 {
2131         struct btree_trans trans;
2132         struct btree_iter *iter;
2133         struct bkey_s_c k;
2134         int ret = 0;
2135
2136         bch2_trans_init(&trans, c, 0, 0);
2137
2138         for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2139                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2140                         break;
2141
2142                 if (bkey_extent_is_data(k.k)) {
2143                         ret = 1;
2144                         break;
2145                 }
2146         }
2147         bch2_trans_iter_put(&trans, iter);
2148
2149         return bch2_trans_exit(&trans) ?: ret;
2150 }
2151
2152 static int __bch2_truncate_page(struct bch_inode_info *inode,
2153                                 pgoff_t index, loff_t start, loff_t end)
2154 {
2155         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2156         struct address_space *mapping = inode->v.i_mapping;
2157         struct bch_page_state *s;
2158         unsigned start_offset = start & (PAGE_SIZE - 1);
2159         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2160         unsigned i;
2161         struct page *page;
2162         int ret = 0;
2163
2164         /* Page boundary? Nothing to do */
2165         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2166               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2167                 return 0;
2168
2169         /* Above i_size? */
2170         if (index << PAGE_SHIFT >= inode->v.i_size)
2171                 return 0;
2172
2173         page = find_lock_page(mapping, index);
2174         if (!page) {
2175                 /*
2176                  * XXX: we're doing two index lookups when we end up reading the
2177                  * page
2178                  */
2179                 ret = range_has_data(c,
2180                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2181                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2182                 if (ret <= 0)
2183                         return ret;
2184
2185                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2186                 if (unlikely(!page)) {
2187                         ret = -ENOMEM;
2188                         goto out;
2189                 }
2190         }
2191
2192         s = bch2_page_state_create(page, 0);
2193         if (!s) {
2194                 ret = -ENOMEM;
2195                 goto unlock;
2196         }
2197
2198         if (!PageUptodate(page)) {
2199                 ret = bch2_read_single_page(page, mapping);
2200                 if (ret)
2201                         goto unlock;
2202         }
2203
2204         if (index != start >> PAGE_SHIFT)
2205                 start_offset = 0;
2206         if (index != end >> PAGE_SHIFT)
2207                 end_offset = PAGE_SIZE;
2208
2209         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2210              i < round_down(end_offset, block_bytes(c)) >> 9;
2211              i++) {
2212                 s->s[i].nr_replicas     = 0;
2213                 s->s[i].state           = SECTOR_UNALLOCATED;
2214         }
2215
2216         zero_user_segment(page, start_offset, end_offset);
2217
2218         /*
2219          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2220          *
2221          * XXX: because we aren't currently tracking whether the page has actual
2222          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2223          */
2224         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2225         BUG_ON(ret);
2226
2227         /*
2228          * This removes any writeable userspace mappings; we need to force
2229          * .page_mkwrite to be called again before any mmapped writes, to
2230          * redirty the full page:
2231          */
2232         page_mkclean(page);
2233         __set_page_dirty_nobuffers(page);
2234 unlock:
2235         unlock_page(page);
2236         put_page(page);
2237 out:
2238         return ret;
2239 }
2240
2241 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2242 {
2243         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2244                                     from, round_up(from, PAGE_SIZE));
2245 }
2246
2247 static int bch2_extend(struct bch_inode_info *inode,
2248                        struct bch_inode_unpacked *inode_u,
2249                        struct iattr *iattr)
2250 {
2251         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2252         struct address_space *mapping = inode->v.i_mapping;
2253         int ret;
2254
2255         /*
2256          * sync appends:
2257          *
2258          * this has to be done _before_ extending i_size:
2259          */
2260         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2261         if (ret)
2262                 return ret;
2263
2264         truncate_setsize(&inode->v, iattr->ia_size);
2265         setattr_copy(&inode->v, iattr);
2266
2267         mutex_lock(&inode->ei_update_lock);
2268         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2269                                     ATTR_MTIME|ATTR_CTIME);
2270         mutex_unlock(&inode->ei_update_lock);
2271
2272         return ret;
2273 }
2274
2275 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2276                                    struct bch_inode_unpacked *bi,
2277                                    void *p)
2278 {
2279         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2280
2281         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2282         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2283         return 0;
2284 }
2285
2286 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2287                                   struct bch_inode_unpacked *bi, void *p)
2288 {
2289         u64 *new_i_size = p;
2290
2291         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2292         bi->bi_size = *new_i_size;
2293         return 0;
2294 }
2295
2296 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2297 {
2298         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2299         struct address_space *mapping = inode->v.i_mapping;
2300         struct bch_inode_unpacked inode_u;
2301         struct btree_trans trans;
2302         struct btree_iter *iter;
2303         u64 new_i_size = iattr->ia_size;
2304         s64 i_sectors_delta = 0;
2305         int ret = 0;
2306
2307         inode_dio_wait(&inode->v);
2308         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2309
2310         /*
2311          * fetch current on disk i_size: inode is locked, i_size can only
2312          * increase underneath us:
2313          */
2314         bch2_trans_init(&trans, c, 0, 0);
2315         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2316         ret = PTR_ERR_OR_ZERO(iter);
2317         bch2_trans_iter_put(&trans, iter);
2318         bch2_trans_exit(&trans);
2319
2320         if (ret)
2321                 goto err;
2322
2323         /*
2324          * check this before next assertion; on filesystem error our normal
2325          * invariants are a bit broken (truncate has to truncate the page cache
2326          * before the inode).
2327          */
2328         ret = bch2_journal_error(&c->journal);
2329         if (ret)
2330                 goto err;
2331
2332         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2333                 inode->v.i_size < inode_u.bi_size);
2334
2335         if (iattr->ia_size > inode->v.i_size) {
2336                 ret = bch2_extend(inode, &inode_u, iattr);
2337                 goto err;
2338         }
2339
2340         ret = bch2_truncate_page(inode, iattr->ia_size);
2341         if (unlikely(ret))
2342                 goto err;
2343
2344         /*
2345          * When extending, we're going to write the new i_size to disk
2346          * immediately so we need to flush anything above the current on disk
2347          * i_size first:
2348          *
2349          * Also, when extending we need to flush the page that i_size currently
2350          * straddles - if it's mapped to userspace, we need to ensure that
2351          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2352          * again to allocate the part of the page that was extended.
2353          */
2354         if (iattr->ia_size > inode_u.bi_size)
2355                 ret = filemap_write_and_wait_range(mapping,
2356                                 inode_u.bi_size,
2357                                 iattr->ia_size - 1);
2358         else if (iattr->ia_size & (PAGE_SIZE - 1))
2359                 ret = filemap_write_and_wait_range(mapping,
2360                                 round_down(iattr->ia_size, PAGE_SIZE),
2361                                 iattr->ia_size - 1);
2362         if (ret)
2363                 goto err;
2364
2365         mutex_lock(&inode->ei_update_lock);
2366         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2367                                &new_i_size, 0);
2368         mutex_unlock(&inode->ei_update_lock);
2369
2370         if (unlikely(ret))
2371                 goto err;
2372
2373         truncate_setsize(&inode->v, iattr->ia_size);
2374
2375         ret = bch2_fpunch(c, inode->v.i_ino,
2376                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2377                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2378         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2379
2380         if (unlikely(ret))
2381                 goto err;
2382
2383         setattr_copy(&inode->v, iattr);
2384
2385         mutex_lock(&inode->ei_update_lock);
2386         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2387                                ATTR_MTIME|ATTR_CTIME);
2388         mutex_unlock(&inode->ei_update_lock);
2389 err:
2390         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2391         return ret;
2392 }
2393
2394 /* fallocate: */
2395
2396 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2397 {
2398         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2399         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2400         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2401         int ret = 0;
2402
2403         inode_lock(&inode->v);
2404         inode_dio_wait(&inode->v);
2405         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2406
2407         ret = __bch2_truncate_page(inode,
2408                                    offset >> PAGE_SHIFT,
2409                                    offset, offset + len);
2410         if (unlikely(ret))
2411                 goto err;
2412
2413         if (offset >> PAGE_SHIFT !=
2414             (offset + len) >> PAGE_SHIFT) {
2415                 ret = __bch2_truncate_page(inode,
2416                                            (offset + len) >> PAGE_SHIFT,
2417                                            offset, offset + len);
2418                 if (unlikely(ret))
2419                         goto err;
2420         }
2421
2422         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2423
2424         if (discard_start < discard_end) {
2425                 s64 i_sectors_delta = 0;
2426
2427                 ret = bch2_fpunch(c, inode->v.i_ino,
2428                                   discard_start, discard_end,
2429                                   &inode->ei_journal_seq,
2430                                   &i_sectors_delta);
2431                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2432         }
2433 err:
2434         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2435         inode_unlock(&inode->v);
2436
2437         return ret;
2438 }
2439
2440 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2441                                    loff_t offset, loff_t len,
2442                                    bool insert)
2443 {
2444         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2445         struct address_space *mapping = inode->v.i_mapping;
2446         struct bkey_buf copy;
2447         struct btree_trans trans;
2448         struct btree_iter *src, *dst, *del;
2449         loff_t shift, new_size;
2450         u64 src_start;
2451         int ret = 0;
2452
2453         if ((offset | len) & (block_bytes(c) - 1))
2454                 return -EINVAL;
2455
2456         /*
2457          * We need i_mutex to keep the page cache consistent with the extents
2458          * btree, and the btree consistent with i_size - we don't need outside
2459          * locking for the extents btree itself, because we're using linked
2460          * iterators
2461          */
2462         inode_lock(&inode->v);
2463         inode_dio_wait(&inode->v);
2464         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2465
2466         if (insert) {
2467                 ret = -EFBIG;
2468                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2469                         goto err;
2470
2471                 ret = -EINVAL;
2472                 if (offset >= inode->v.i_size)
2473                         goto err;
2474
2475                 src_start       = U64_MAX;
2476                 shift           = len;
2477         } else {
2478                 ret = -EINVAL;
2479                 if (offset + len >= inode->v.i_size)
2480                         goto err;
2481
2482                 src_start       = offset + len;
2483                 shift           = -len;
2484         }
2485
2486         new_size = inode->v.i_size + shift;
2487
2488         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2489         if (ret)
2490                 goto err;
2491
2492         if (insert) {
2493                 i_size_write(&inode->v, new_size);
2494                 mutex_lock(&inode->ei_update_lock);
2495                 ret = bch2_write_inode_size(c, inode, new_size,
2496                                             ATTR_MTIME|ATTR_CTIME);
2497                 mutex_unlock(&inode->ei_update_lock);
2498         } else {
2499                 s64 i_sectors_delta = 0;
2500
2501                 ret = bch2_fpunch(c, inode->v.i_ino,
2502                                   offset >> 9, (offset + len) >> 9,
2503                                   &inode->ei_journal_seq,
2504                                   &i_sectors_delta);
2505                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2506
2507                 if (ret)
2508                         goto err;
2509         }
2510
2511         bch2_bkey_buf_init(&copy);
2512         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2513         src = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2514                         POS(inode->v.i_ino, src_start >> 9),
2515                         BTREE_ITER_INTENT);
2516         dst = bch2_trans_copy_iter(&trans, src);
2517         del = bch2_trans_copy_iter(&trans, src);
2518
2519         while (ret == 0 || ret == -EINTR) {
2520                 struct disk_reservation disk_res =
2521                         bch2_disk_reservation_init(c, 0);
2522                 struct bkey_i delete;
2523                 struct bkey_s_c k;
2524                 struct bpos next_pos;
2525                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2526                 struct bpos atomic_end;
2527                 unsigned trigger_flags = 0;
2528
2529                 k = insert
2530                         ? bch2_btree_iter_peek_prev(src)
2531                         : bch2_btree_iter_peek(src);
2532                 if ((ret = bkey_err(k)))
2533                         continue;
2534
2535                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2536                         break;
2537
2538                 if (insert &&
2539                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2540                         break;
2541 reassemble:
2542                 bch2_bkey_buf_reassemble(&copy, c, k);
2543
2544                 if (insert &&
2545                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2546                         bch2_cut_front(move_pos, copy.k);
2547
2548                 copy.k->k.p.offset += shift >> 9;
2549                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2550
2551                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2552                 if (ret)
2553                         continue;
2554
2555                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2556                         if (insert) {
2557                                 move_pos = atomic_end;
2558                                 move_pos.offset -= shift >> 9;
2559                                 goto reassemble;
2560                         } else {
2561                                 bch2_cut_back(atomic_end, copy.k);
2562                         }
2563                 }
2564
2565                 bkey_init(&delete.k);
2566                 delete.k.p = copy.k->k.p;
2567                 delete.k.size = copy.k->k.size;
2568                 delete.k.p.offset -= shift >> 9;
2569                 bch2_btree_iter_set_pos(del, bkey_start_pos(&delete.k));
2570
2571                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2572
2573                 if (copy.k->k.size == k.k->size) {
2574                         /*
2575                          * If we're moving the entire extent, we can skip
2576                          * running triggers:
2577                          */
2578                         trigger_flags |= BTREE_TRIGGER_NORUN;
2579                 } else {
2580                         /* We might end up splitting compressed extents: */
2581                         unsigned nr_ptrs =
2582                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2583
2584                         ret = bch2_disk_reservation_get(c, &disk_res,
2585                                         copy.k->k.size, nr_ptrs,
2586                                         BCH_DISK_RESERVATION_NOFAIL);
2587                         BUG_ON(ret);
2588                 }
2589
2590                 ret =   bch2_trans_update(&trans, del, &delete, trigger_flags) ?:
2591                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2592                         bch2_trans_commit(&trans, &disk_res,
2593                                           &inode->ei_journal_seq,
2594                                           BTREE_INSERT_NOFAIL);
2595                 bch2_disk_reservation_put(c, &disk_res);
2596
2597                 if (!ret)
2598                         bch2_btree_iter_set_pos(src, next_pos);
2599         }
2600         bch2_trans_iter_put(&trans, del);
2601         bch2_trans_iter_put(&trans, dst);
2602         bch2_trans_iter_put(&trans, src);
2603         bch2_trans_exit(&trans);
2604         bch2_bkey_buf_exit(&copy, c);
2605
2606         if (ret)
2607                 goto err;
2608
2609         if (!insert) {
2610                 i_size_write(&inode->v, new_size);
2611                 mutex_lock(&inode->ei_update_lock);
2612                 ret = bch2_write_inode_size(c, inode, new_size,
2613                                             ATTR_MTIME|ATTR_CTIME);
2614                 mutex_unlock(&inode->ei_update_lock);
2615         }
2616 err:
2617         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2618         inode_unlock(&inode->v);
2619         return ret;
2620 }
2621
2622 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2623                             loff_t offset, loff_t len)
2624 {
2625         struct address_space *mapping = inode->v.i_mapping;
2626         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2627         struct btree_trans trans;
2628         struct btree_iter *iter;
2629         struct bpos end_pos;
2630         loff_t end              = offset + len;
2631         loff_t block_start      = round_down(offset,    block_bytes(c));
2632         loff_t block_end        = round_up(end,         block_bytes(c));
2633         unsigned sectors;
2634         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2635         int ret;
2636
2637         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2638
2639         inode_lock(&inode->v);
2640         inode_dio_wait(&inode->v);
2641         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2642
2643         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2644                 ret = inode_newsize_ok(&inode->v, end);
2645                 if (ret)
2646                         goto err;
2647         }
2648
2649         if (mode & FALLOC_FL_ZERO_RANGE) {
2650                 ret = __bch2_truncate_page(inode,
2651                                            offset >> PAGE_SHIFT,
2652                                            offset, end);
2653
2654                 if (!ret &&
2655                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2656                         ret = __bch2_truncate_page(inode,
2657                                                    end >> PAGE_SHIFT,
2658                                                    offset, end);
2659
2660                 if (unlikely(ret))
2661                         goto err;
2662
2663                 truncate_pagecache_range(&inode->v, offset, end - 1);
2664         }
2665
2666         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2667                         POS(inode->v.i_ino, block_start >> 9),
2668                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2669         end_pos = POS(inode->v.i_ino, block_end >> 9);
2670
2671         while (!ret && bkey_cmp(iter->pos, end_pos) < 0) {
2672                 s64 i_sectors_delta = 0;
2673                 struct disk_reservation disk_res = { 0 };
2674                 struct quota_res quota_res = { 0 };
2675                 struct bkey_i_reservation reservation;
2676                 struct bkey_s_c k;
2677
2678                 bch2_trans_begin(&trans);
2679
2680                 k = bch2_btree_iter_peek_slot(iter);
2681                 if ((ret = bkey_err(k)))
2682                         goto bkey_err;
2683
2684                 /* already reserved */
2685                 if (k.k->type == KEY_TYPE_reservation &&
2686                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2687                         bch2_btree_iter_next_slot(iter);
2688                         continue;
2689                 }
2690
2691                 if (bkey_extent_is_data(k.k) &&
2692                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2693                         bch2_btree_iter_next_slot(iter);
2694                         continue;
2695                 }
2696
2697                 bkey_reservation_init(&reservation.k_i);
2698                 reservation.k.type      = KEY_TYPE_reservation;
2699                 reservation.k.p         = k.k->p;
2700                 reservation.k.size      = k.k->size;
2701
2702                 bch2_cut_front(iter->pos,       &reservation.k_i);
2703                 bch2_cut_back(end_pos,          &reservation.k_i);
2704
2705                 sectors = reservation.k.size;
2706                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2707
2708                 if (!bkey_extent_is_allocation(k.k)) {
2709                         ret = bch2_quota_reservation_add(c, inode,
2710                                         &quota_res,
2711                                         sectors, true);
2712                         if (unlikely(ret))
2713                                 goto bkey_err;
2714                 }
2715
2716                 if (reservation.v.nr_replicas < replicas ||
2717                     bch2_bkey_sectors_compressed(k)) {
2718                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2719                                                         replicas, 0);
2720                         if (unlikely(ret))
2721                                 goto bkey_err;
2722
2723                         reservation.v.nr_replicas = disk_res.nr_replicas;
2724                 }
2725
2726                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2727                                 &disk_res, &inode->ei_journal_seq,
2728                                 0, &i_sectors_delta);
2729                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2730 bkey_err:
2731                 bch2_quota_reservation_put(c, inode, &quota_res);
2732                 bch2_disk_reservation_put(c, &disk_res);
2733                 if (ret == -EINTR)
2734                         ret = 0;
2735         }
2736         bch2_trans_iter_put(&trans, iter);
2737
2738         if (ret)
2739                 goto err;
2740
2741         /*
2742          * Do we need to extend the file?
2743          *
2744          * If we zeroed up to the end of the file, we dropped whatever writes
2745          * were going to write out the current i_size, so we have to extend
2746          * manually even if FL_KEEP_SIZE was set:
2747          */
2748         if (end >= inode->v.i_size &&
2749             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2750              (mode & FALLOC_FL_ZERO_RANGE))) {
2751                 struct btree_iter *inode_iter;
2752                 struct bch_inode_unpacked inode_u;
2753
2754                 do {
2755                         bch2_trans_begin(&trans);
2756                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2757                                                      inode->v.i_ino, 0);
2758                         ret = PTR_ERR_OR_ZERO(inode_iter);
2759                 } while (ret == -EINTR);
2760
2761                 bch2_trans_iter_put(&trans, inode_iter);
2762                 bch2_trans_unlock(&trans);
2763
2764                 if (ret)
2765                         goto err;
2766
2767                 /*
2768                  * Sync existing appends before extending i_size,
2769                  * as in bch2_extend():
2770                  */
2771                 ret = filemap_write_and_wait_range(mapping,
2772                                         inode_u.bi_size, S64_MAX);
2773                 if (ret)
2774                         goto err;
2775
2776                 if (mode & FALLOC_FL_KEEP_SIZE)
2777                         end = inode->v.i_size;
2778                 else
2779                         i_size_write(&inode->v, end);
2780
2781                 mutex_lock(&inode->ei_update_lock);
2782                 ret = bch2_write_inode_size(c, inode, end, 0);
2783                 mutex_unlock(&inode->ei_update_lock);
2784         }
2785 err:
2786         bch2_trans_exit(&trans);
2787         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2788         inode_unlock(&inode->v);
2789         return ret;
2790 }
2791
2792 long bch2_fallocate_dispatch(struct file *file, int mode,
2793                              loff_t offset, loff_t len)
2794 {
2795         struct bch_inode_info *inode = file_bch_inode(file);
2796         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2797         long ret;
2798
2799         if (!percpu_ref_tryget(&c->writes))
2800                 return -EROFS;
2801
2802         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2803                 ret = bchfs_fallocate(inode, mode, offset, len);
2804         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2805                 ret = bchfs_fpunch(inode, offset, len);
2806         else if (mode == FALLOC_FL_INSERT_RANGE)
2807                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2808         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2809                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2810         else
2811                 ret = -EOPNOTSUPP;
2812
2813         percpu_ref_put(&c->writes);
2814
2815         return ret;
2816 }
2817
2818 static void mark_range_unallocated(struct bch_inode_info *inode,
2819                                    loff_t start, loff_t end)
2820 {
2821         pgoff_t index = start >> PAGE_SHIFT;
2822         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2823         struct pagevec pvec;
2824
2825         pagevec_init(&pvec);
2826
2827         do {
2828                 unsigned nr_pages, i, j;
2829
2830                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2831                                                 &index, end_index);
2832                 if (nr_pages == 0)
2833                         break;
2834
2835                 for (i = 0; i < nr_pages; i++) {
2836                         struct page *page = pvec.pages[i];
2837                         struct bch_page_state *s;
2838
2839                         lock_page(page);
2840                         s = bch2_page_state(page);
2841
2842                         if (s) {
2843                                 spin_lock(&s->lock);
2844                                 for (j = 0; j < PAGE_SECTORS; j++)
2845                                         s->s[j].nr_replicas = 0;
2846                                 spin_unlock(&s->lock);
2847                         }
2848
2849                         unlock_page(page);
2850                 }
2851                 pagevec_release(&pvec);
2852         } while (index <= end_index);
2853 }
2854
2855 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2856                              struct file *file_dst, loff_t pos_dst,
2857                              loff_t len, unsigned remap_flags)
2858 {
2859         struct bch_inode_info *src = file_bch_inode(file_src);
2860         struct bch_inode_info *dst = file_bch_inode(file_dst);
2861         struct bch_fs *c = src->v.i_sb->s_fs_info;
2862         s64 i_sectors_delta = 0;
2863         u64 aligned_len;
2864         loff_t ret = 0;
2865
2866         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2867                 return -EINVAL;
2868
2869         if (remap_flags & REMAP_FILE_DEDUP)
2870                 return -EOPNOTSUPP;
2871
2872         if ((pos_src & (block_bytes(c) - 1)) ||
2873             (pos_dst & (block_bytes(c) - 1)))
2874                 return -EINVAL;
2875
2876         if (src == dst &&
2877             abs(pos_src - pos_dst) < len)
2878                 return -EINVAL;
2879
2880         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2881
2882         file_update_time(file_dst);
2883
2884         inode_dio_wait(&src->v);
2885         inode_dio_wait(&dst->v);
2886
2887         ret = generic_remap_file_range_prep(file_src, pos_src,
2888                                             file_dst, pos_dst,
2889                                             &len, remap_flags);
2890         if (ret < 0 || len == 0)
2891                 goto err;
2892
2893         aligned_len = round_up((u64) len, block_bytes(c));
2894
2895         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2896                                 pos_dst, pos_dst + len - 1);
2897         if (ret)
2898                 goto err;
2899
2900         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2901
2902         ret = bch2_remap_range(c,
2903                                POS(dst->v.i_ino, pos_dst >> 9),
2904                                POS(src->v.i_ino, pos_src >> 9),
2905                                aligned_len >> 9,
2906                                &dst->ei_journal_seq,
2907                                pos_dst + len, &i_sectors_delta);
2908         if (ret < 0)
2909                 goto err;
2910
2911         /*
2912          * due to alignment, we might have remapped slightly more than requsted
2913          */
2914         ret = min((u64) ret << 9, (u64) len);
2915
2916         /* XXX get a quota reservation */
2917         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2918
2919         spin_lock(&dst->v.i_lock);
2920         if (pos_dst + ret > dst->v.i_size)
2921                 i_size_write(&dst->v, pos_dst + ret);
2922         spin_unlock(&dst->v.i_lock);
2923 err:
2924         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2925
2926         return ret;
2927 }
2928
2929 /* fseek: */
2930
2931 static int page_data_offset(struct page *page, unsigned offset)
2932 {
2933         struct bch_page_state *s = bch2_page_state(page);
2934         unsigned i;
2935
2936         if (s)
2937                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2938                         if (s->s[i].state >= SECTOR_DIRTY)
2939                                 return i << 9;
2940
2941         return -1;
2942 }
2943
2944 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2945                                        loff_t start_offset,
2946                                        loff_t end_offset)
2947 {
2948         struct address_space *mapping = vinode->i_mapping;
2949         struct page *page;
2950         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2951         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2952         pgoff_t index           = start_index;
2953         loff_t ret;
2954         int offset;
2955
2956         while (index <= end_index) {
2957                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2958                         lock_page(page);
2959
2960                         offset = page_data_offset(page,
2961                                         page->index == start_index
2962                                         ? start_offset & (PAGE_SIZE - 1)
2963                                         : 0);
2964                         if (offset >= 0) {
2965                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2966                                             offset,
2967                                             start_offset, end_offset);
2968                                 unlock_page(page);
2969                                 put_page(page);
2970                                 return ret;
2971                         }
2972
2973                         unlock_page(page);
2974                         put_page(page);
2975                 } else {
2976                         break;
2977                 }
2978         }
2979
2980         return end_offset;
2981 }
2982
2983 static loff_t bch2_seek_data(struct file *file, u64 offset)
2984 {
2985         struct bch_inode_info *inode = file_bch_inode(file);
2986         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2987         struct btree_trans trans;
2988         struct btree_iter *iter;
2989         struct bkey_s_c k;
2990         u64 isize, next_data = MAX_LFS_FILESIZE;
2991         int ret;
2992
2993         isize = i_size_read(&inode->v);
2994         if (offset >= isize)
2995                 return -ENXIO;
2996
2997         bch2_trans_init(&trans, c, 0, 0);
2998
2999         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3000                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3001                 if (k.k->p.inode != inode->v.i_ino) {
3002                         break;
3003                 } else if (bkey_extent_is_data(k.k)) {
3004                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3005                         break;
3006                 } else if (k.k->p.offset >> 9 > isize)
3007                         break;
3008         }
3009         bch2_trans_iter_put(&trans, iter);
3010
3011         ret = bch2_trans_exit(&trans) ?: ret;
3012         if (ret)
3013                 return ret;
3014
3015         if (next_data > offset)
3016                 next_data = bch2_seek_pagecache_data(&inode->v,
3017                                                      offset, next_data);
3018
3019         if (next_data >= isize)
3020                 return -ENXIO;
3021
3022         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3023 }
3024
3025 static int __page_hole_offset(struct page *page, unsigned offset)
3026 {
3027         struct bch_page_state *s = bch2_page_state(page);
3028         unsigned i;
3029
3030         if (!s)
3031                 return 0;
3032
3033         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3034                 if (s->s[i].state < SECTOR_DIRTY)
3035                         return i << 9;
3036
3037         return -1;
3038 }
3039
3040 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3041 {
3042         pgoff_t index = offset >> PAGE_SHIFT;
3043         struct page *page;
3044         int pg_offset;
3045         loff_t ret = -1;
3046
3047         page = find_lock_page(mapping, index);
3048         if (!page)
3049                 return offset;
3050
3051         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3052         if (pg_offset >= 0)
3053                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3054
3055         unlock_page(page);
3056
3057         return ret;
3058 }
3059
3060 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3061                                        loff_t start_offset,
3062                                        loff_t end_offset)
3063 {
3064         struct address_space *mapping = vinode->i_mapping;
3065         loff_t offset = start_offset, hole;
3066
3067         while (offset < end_offset) {
3068                 hole = page_hole_offset(mapping, offset);
3069                 if (hole >= 0 && hole <= end_offset)
3070                         return max(start_offset, hole);
3071
3072                 offset += PAGE_SIZE;
3073                 offset &= PAGE_MASK;
3074         }
3075
3076         return end_offset;
3077 }
3078
3079 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3080 {
3081         struct bch_inode_info *inode = file_bch_inode(file);
3082         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3083         struct btree_trans trans;
3084         struct btree_iter *iter;
3085         struct bkey_s_c k;
3086         u64 isize, next_hole = MAX_LFS_FILESIZE;
3087         int ret;
3088
3089         isize = i_size_read(&inode->v);
3090         if (offset >= isize)
3091                 return -ENXIO;
3092
3093         bch2_trans_init(&trans, c, 0, 0);
3094
3095         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3096                            POS(inode->v.i_ino, offset >> 9),
3097                            BTREE_ITER_SLOTS, k, ret) {
3098                 if (k.k->p.inode != inode->v.i_ino) {
3099                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3100                                         offset, MAX_LFS_FILESIZE);
3101                         break;
3102                 } else if (!bkey_extent_is_data(k.k)) {
3103                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3104                                         max(offset, bkey_start_offset(k.k) << 9),
3105                                         k.k->p.offset << 9);
3106
3107                         if (next_hole < k.k->p.offset << 9)
3108                                 break;
3109                 } else {
3110                         offset = max(offset, bkey_start_offset(k.k) << 9);
3111                 }
3112         }
3113         bch2_trans_iter_put(&trans, iter);
3114
3115         ret = bch2_trans_exit(&trans) ?: ret;
3116         if (ret)
3117                 return ret;
3118
3119         if (next_hole > isize)
3120                 next_hole = isize;
3121
3122         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3123 }
3124
3125 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3126 {
3127         switch (whence) {
3128         case SEEK_SET:
3129         case SEEK_CUR:
3130         case SEEK_END:
3131                 return generic_file_llseek(file, offset, whence);
3132         case SEEK_DATA:
3133                 return bch2_seek_data(file, offset);
3134         case SEEK_HOLE:
3135                 return bch2_seek_hole(file, offset);
3136         }
3137
3138         return -EINVAL;
3139 }
3140
3141 void bch2_fs_fsio_exit(struct bch_fs *c)
3142 {
3143         bioset_exit(&c->dio_write_bioset);
3144         bioset_exit(&c->dio_read_bioset);
3145         bioset_exit(&c->writepage_bioset);
3146 }
3147
3148 int bch2_fs_fsio_init(struct bch_fs *c)
3149 {
3150         int ret = 0;
3151
3152         pr_verbose_init(c->opts, "");
3153
3154         if (bioset_init(&c->writepage_bioset,
3155                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3156                         BIOSET_NEED_BVECS) ||
3157             bioset_init(&c->dio_read_bioset,
3158                         4, offsetof(struct dio_read, rbio.bio),
3159                         BIOSET_NEED_BVECS) ||
3160             bioset_init(&c->dio_write_bioset,
3161                         4, offsetof(struct dio_write, op.wbio.bio),
3162                         BIOSET_NEED_BVECS))
3163                 ret = -ENOMEM;
3164
3165         pr_verbose_init(c->opts, "ret %i", ret);
3166         return ret;
3167 }
3168
3169 #endif /* NO_BCACHEFS_FS */