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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 static inline struct address_space *faults_disabled_mapping(void)
  39 {
  40         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  41 }
  42
  43 static inline void set_fdm_dropped_locks(void)
  44 {
  45         current->faults_disabled_mapping =
  46                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  47 }
  48
  49 static inline bool fdm_dropped_locks(void)
  50 {
  51         return ((unsigned long) current->faults_disabled_mapping) & 1;
  52 }
  53
  54 struct quota_res {
  55         u64                             sectors;
  56 };
  57
  58 struct bch_writepage_io {
  59         struct closure                  cl;
  60         struct bch_inode_info           *inode;
  61
  62         /* must be last: */
  63         struct bch_write_op             op;
  64 };
  65
  66 struct dio_write {
  67         struct completion               done;
  68         struct kiocb                    *req;
  69         struct mm_struct                *mm;
  70         unsigned                        loop:1,
  71                                         sync:1,
  72                                         free_iov:1;
  73         struct quota_res                quota_res;
  74         u64                             written;
  75
  76         struct iov_iter                 iter;
  77         struct iovec                    inline_vecs[2];
  78
  79         /* must be last: */
  80         struct bch_write_op             op;
  81 };
  82
  83 struct dio_read {
  84         struct closure                  cl;
  85         struct kiocb                    *req;
  86         long                            ret;
  87         bool                            should_dirty;
  88         struct bch_read_bio             rbio;
  89 };
  90
  91 /* pagecache_block must be held */
  92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  93                                               loff_t start, loff_t end)
  94 {
  95         int ret;
  96
  97         /*
  98          * XXX: the way this is currently implemented, we can spin if a process
  99          * is continually redirtying a specific page
 100          */
 101         do {
 102                 if (!mapping->nrpages)
 103                         return 0;
 104
 105                 ret = filemap_write_and_wait_range(mapping, start, end);
 106                 if (ret)
 107                         break;
 108
 109                 if (!mapping->nrpages)
 110                         return 0;
 111
 112                 ret = invalidate_inode_pages2_range(mapping,
 113                                 start >> PAGE_SHIFT,
 114                                 end >> PAGE_SHIFT);
 115         } while (ret == -EBUSY);
 116
 117         return ret;
 118 }
 119
 120 /* quotas */
 121
 122 #ifdef CONFIG_BCACHEFS_QUOTA
 123
 124 static void bch2_quota_reservation_put(struct bch_fs *c,
 125                                        struct bch_inode_info *inode,
 126                                        struct quota_res *res)
 127 {
 128         if (!res->sectors)
 129                 return;
 130
 131         mutex_lock(&inode->ei_quota_lock);
 132         BUG_ON(res->sectors > inode->ei_quota_reserved);
 133
 134         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 135                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 136         inode->ei_quota_reserved -= res->sectors;
 137         mutex_unlock(&inode->ei_quota_lock);
 138
 139         res->sectors = 0;
 140 }
 141
 142 static int bch2_quota_reservation_add(struct bch_fs *c,
 143                                       struct bch_inode_info *inode,
 144                                       struct quota_res *res,
 145                                       unsigned sectors,
 146                                       bool check_enospc)
 147 {
 148         int ret;
 149
 150         mutex_lock(&inode->ei_quota_lock);
 151         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 152                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 153         if (likely(!ret)) {
 154                 inode->ei_quota_reserved += sectors;
 155                 res->sectors += sectors;
 156         }
 157         mutex_unlock(&inode->ei_quota_lock);
 158
 159         return ret;
 160 }
 161
 162 #else
 163
 164 static void bch2_quota_reservation_put(struct bch_fs *c,
 165                                        struct bch_inode_info *inode,
 166                                        struct quota_res *res)
 167 {
 168 }
 169
 170 static int bch2_quota_reservation_add(struct bch_fs *c,
 171                                       struct bch_inode_info *inode,
 172                                       struct quota_res *res,
 173                                       unsigned sectors,
 174                                       bool check_enospc)
 175 {
 176         return 0;
 177 }
 178
 179 #endif
 180
 181 /* i_size updates: */
 182
 183 struct inode_new_size {
 184         loff_t          new_size;
 185         u64             now;
 186         unsigned        fields;
 187 };
 188
 189 static int inode_set_size(struct bch_inode_info *inode,
 190                           struct bch_inode_unpacked *bi,
 191                           void *p)
 192 {
 193         struct inode_new_size *s = p;
 194
 195         bi->bi_size = s->new_size;
 196         if (s->fields & ATTR_ATIME)
 197                 bi->bi_atime = s->now;
 198         if (s->fields & ATTR_MTIME)
 199                 bi->bi_mtime = s->now;
 200         if (s->fields & ATTR_CTIME)
 201                 bi->bi_ctime = s->now;
 202
 203         return 0;
 204 }
 205
 206 int __must_check bch2_write_inode_size(struct bch_fs *c,
 207                                        struct bch_inode_info *inode,
 208                                        loff_t new_size, unsigned fields)
 209 {
 210         struct inode_new_size s = {
 211                 .new_size       = new_size,
 212                 .now            = bch2_current_time(c),
 213                 .fields         = fields,
 214         };
 215
 216         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 217 }
 218
 219 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 220                            struct quota_res *quota_res, s64 sectors)
 221 {
 222         if (!sectors)
 223                 return;
 224
 225         mutex_lock(&inode->ei_quota_lock);
 226 #ifdef CONFIG_BCACHEFS_QUOTA
 227         if (quota_res && sectors > 0) {
 228                 BUG_ON(sectors > quota_res->sectors);
 229                 BUG_ON(sectors > inode->ei_quota_reserved);
 230
 231                 quota_res->sectors -= sectors;
 232                 inode->ei_quota_reserved -= sectors;
 233         } else {
 234                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 235         }
 236 #endif
 237         inode->v.i_blocks += sectors;
 238         mutex_unlock(&inode->ei_quota_lock);
 239 }
 240
 241 /* page state: */
 242
 243 /* stored in page->private: */
 244
 245 struct bch_page_sector {
 246         /* Uncompressed, fully allocated replicas: */
 247         unsigned                nr_replicas:3;
 248
 249         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 250         unsigned                replicas_reserved:3;
 251
 252         /* i_sectors: */
 253         enum {
 254                 SECTOR_UNALLOCATED,
 255                 SECTOR_RESERVED,
 256                 SECTOR_DIRTY,
 257                 SECTOR_ALLOCATED,
 258         }                       state:2;
 259 };
 260
 261 struct bch_page_state {
 262         spinlock_t              lock;
 263         atomic_t                write_count;
 264         struct bch_page_sector  s[PAGE_SECTORS];
 265 };
 266
 267 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 268 {
 269         return page_has_private(page)
 270                 ? (struct bch_page_state *) page_private(page)
 271                 : NULL;
 272 }
 273
 274 static inline struct bch_page_state *bch2_page_state(struct page *page)
 275 {
 276         EBUG_ON(!PageLocked(page));
 277
 278         return __bch2_page_state(page);
 279 }
 280
 281 /* for newly allocated pages: */
 282 static void __bch2_page_state_release(struct page *page)
 283 {
 284         kfree(detach_page_private(page));
 285 }
 286
 287 static void bch2_page_state_release(struct page *page)
 288 {
 289         EBUG_ON(!PageLocked(page));
 290         __bch2_page_state_release(page);
 291 }
 292
 293 /* for newly allocated pages: */
 294 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 295                                                        gfp_t gfp)
 296 {
 297         struct bch_page_state *s;
 298
 299         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 300         if (!s)
 301                 return NULL;
 302
 303         spin_lock_init(&s->lock);
 304         attach_page_private(page, s);
 305         return s;
 306 }
 307
 308 static struct bch_page_state *bch2_page_state_create(struct page *page,
 309                                                      gfp_t gfp)
 310 {
 311         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 312 }
 313
 314 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 315 {
 316         /* XXX: this should not be open coded */
 317         return inode->ei_inode.bi_data_replicas
 318                 ? inode->ei_inode.bi_data_replicas - 1
 319                 : c->opts.data_replicas;
 320 }
 321
 322 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 323                                                   unsigned nr_replicas)
 324 {
 325         return max(0, (int) nr_replicas -
 326                    s->nr_replicas -
 327                    s->replicas_reserved);
 328 }
 329
 330 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 331                                 struct bch_inode_info *inode,
 332                                 struct page *page, bool check_enospc)
 333 {
 334         struct bch_page_state *s = bch2_page_state_create(page, 0);
 335         unsigned nr_replicas = inode_nr_replicas(c, inode);
 336         struct disk_reservation disk_res = { 0 };
 337         unsigned i, disk_res_sectors = 0;
 338         int ret;
 339
 340         if (!s)
 341                 return -ENOMEM;
 342
 343         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 344                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 345
 346         if (!disk_res_sectors)
 347                 return 0;
 348
 349         ret = bch2_disk_reservation_get(c, &disk_res,
 350                                         disk_res_sectors, 1,
 351                                         !check_enospc
 352                                         ? BCH_DISK_RESERVATION_NOFAIL
 353                                         : 0);
 354         if (unlikely(ret))
 355                 return ret;
 356
 357         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 358                 s->s[i].replicas_reserved +=
 359                         sectors_to_reserve(&s->s[i], nr_replicas);
 360
 361         return 0;
 362 }
 363
 364 struct bch2_page_reservation {
 365         struct disk_reservation disk;
 366         struct quota_res        quota;
 367 };
 368
 369 static void bch2_page_reservation_init(struct bch_fs *c,
 370                         struct bch_inode_info *inode,
 371                         struct bch2_page_reservation *res)
 372 {
 373         memset(res, 0, sizeof(*res));
 374
 375         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 376 }
 377
 378 static void bch2_page_reservation_put(struct bch_fs *c,
 379                         struct bch_inode_info *inode,
 380                         struct bch2_page_reservation *res)
 381 {
 382         bch2_disk_reservation_put(c, &res->disk);
 383         bch2_quota_reservation_put(c, inode, &res->quota);
 384 }
 385
 386 static int bch2_page_reservation_get(struct bch_fs *c,
 387                         struct bch_inode_info *inode, struct page *page,
 388                         struct bch2_page_reservation *res,
 389                         unsigned offset, unsigned len, bool check_enospc)
 390 {
 391         struct bch_page_state *s = bch2_page_state_create(page, 0);
 392         unsigned i, disk_sectors = 0, quota_sectors = 0;
 393         int ret;
 394
 395         if (!s)
 396                 return -ENOMEM;
 397
 398         for (i = round_down(offset, block_bytes(c)) >> 9;
 399              i < round_up(offset + len, block_bytes(c)) >> 9;
 400              i++) {
 401                 disk_sectors += sectors_to_reserve(&s->s[i],
 402                                                 res->disk.nr_replicas);
 403                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 404         }
 405
 406         if (disk_sectors) {
 407                 ret = bch2_disk_reservation_add(c, &res->disk,
 408                                                 disk_sectors,
 409                                                 !check_enospc
 410                                                 ? BCH_DISK_RESERVATION_NOFAIL
 411                                                 : 0);
 412                 if (unlikely(ret))
 413                         return ret;
 414         }
 415
 416         if (quota_sectors) {
 417                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 418                                                  quota_sectors,
 419                                                  check_enospc);
 420                 if (unlikely(ret)) {
 421                         struct disk_reservation tmp = {
 422                                 .sectors = disk_sectors
 423                         };
 424
 425                         bch2_disk_reservation_put(c, &tmp);
 426                         res->disk.sectors -= disk_sectors;
 427                         return ret;
 428                 }
 429         }
 430
 431         return 0;
 432 }
 433
 434 static void bch2_clear_page_bits(struct page *page)
 435 {
 436         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 437         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 438         struct bch_page_state *s = bch2_page_state(page);
 439         struct disk_reservation disk_res = { 0 };
 440         int i, dirty_sectors = 0;
 441
 442         if (!s)
 443                 return;
 444
 445         EBUG_ON(!PageLocked(page));
 446         EBUG_ON(PageWriteback(page));
 447
 448         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 449                 disk_res.sectors += s->s[i].replicas_reserved;
 450                 s->s[i].replicas_reserved = 0;
 451
 452                 if (s->s[i].state == SECTOR_DIRTY) {
 453                         dirty_sectors++;
 454                         s->s[i].state = SECTOR_UNALLOCATED;
 455                 }
 456         }
 457
 458         bch2_disk_reservation_put(c, &disk_res);
 459
 460         if (dirty_sectors)
 461                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 462
 463         bch2_page_state_release(page);
 464 }
 465
 466 static void bch2_set_page_dirty(struct bch_fs *c,
 467                         struct bch_inode_info *inode, struct page *page,
 468                         struct bch2_page_reservation *res,
 469                         unsigned offset, unsigned len)
 470 {
 471         struct bch_page_state *s = bch2_page_state(page);
 472         unsigned i, dirty_sectors = 0;
 473
 474         WARN_ON((u64) page_offset(page) + offset + len >
 475                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 476
 477         spin_lock(&s->lock);
 478
 479         for (i = round_down(offset, block_bytes(c)) >> 9;
 480              i < round_up(offset + len, block_bytes(c)) >> 9;
 481              i++) {
 482                 unsigned sectors = sectors_to_reserve(&s->s[i],
 483                                                 res->disk.nr_replicas);
 484
 485                 /*
 486                  * This can happen if we race with the error path in
 487                  * bch2_writepage_io_done():
 488                  */
 489                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 490
 491                 s->s[i].replicas_reserved += sectors;
 492                 res->disk.sectors -= sectors;
 493
 494                 if (s->s[i].state == SECTOR_UNALLOCATED)
 495                         dirty_sectors++;
 496
 497                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 498         }
 499
 500         spin_unlock(&s->lock);
 501
 502         if (dirty_sectors)
 503                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 504
 505         if (!PageDirty(page))
 506                 __set_page_dirty_nobuffers(page);
 507 }
 508
 509 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 510 {
 511         struct file *file = vmf->vma->vm_file;
 512         struct address_space *mapping = file->f_mapping;
 513         struct address_space *fdm = faults_disabled_mapping();
 514         struct bch_inode_info *inode = file_bch_inode(file);
 515         int ret;
 516
 517         if (fdm == mapping)
 518                 return VM_FAULT_SIGBUS;
 519
 520         /* Lock ordering: */
 521         if (fdm > mapping) {
 522                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 523
 524                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 525                         goto got_lock;
 526
 527                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 528
 529                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 530                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 531
 532                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 533
 534                 /* Signal that lock has been dropped: */
 535                 set_fdm_dropped_locks();
 536                 return VM_FAULT_SIGBUS;
 537         }
 538
 539         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 540 got_lock:
 541         ret = filemap_fault(vmf);
 542         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 543
 544         return ret;
 545 }
 546
 547 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 548 {
 549         struct page *page = vmf->page;
 550         struct file *file = vmf->vma->vm_file;
 551         struct bch_inode_info *inode = file_bch_inode(file);
 552         struct address_space *mapping = file->f_mapping;
 553         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 554         struct bch2_page_reservation res;
 555         unsigned len;
 556         loff_t isize;
 557         int ret = VM_FAULT_LOCKED;
 558
 559         bch2_page_reservation_init(c, inode, &res);
 560
 561         sb_start_pagefault(inode->v.i_sb);
 562         file_update_time(file);
 563
 564         /*
 565          * Not strictly necessary, but helps avoid dio writes livelocking in
 566          * write_invalidate_inode_pages_range() - can drop this if/when we get
 567          * a write_invalidate_inode_pages_range() that works without dropping
 568          * page lock before invalidating page
 569          */
 570         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 571
 572         lock_page(page);
 573         isize = i_size_read(&inode->v);
 574
 575         if (page->mapping != mapping || page_offset(page) >= isize) {
 576                 unlock_page(page);
 577                 ret = VM_FAULT_NOPAGE;
 578                 goto out;
 579         }
 580
 581         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 582
 583         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 584                 unlock_page(page);
 585                 ret = VM_FAULT_SIGBUS;
 586                 goto out;
 587         }
 588
 589         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 590         bch2_page_reservation_put(c, inode, &res);
 591
 592         wait_for_stable_page(page);
 593 out:
 594         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 595         sb_end_pagefault(inode->v.i_sb);
 596
 597         return ret;
 598 }
 599
 600 void bch2_invalidatepage(struct page *page, unsigned int offset,
 601                          unsigned int length)
 602 {
 603         if (offset || length < PAGE_SIZE)
 604                 return;
 605
 606         bch2_clear_page_bits(page);
 607 }
 608
 609 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 610 {
 611         if (PageDirty(page))
 612                 return 0;
 613
 614         bch2_clear_page_bits(page);
 615         return 1;
 616 }
 617
 618 #ifdef CONFIG_MIGRATION
 619 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 620                       struct page *page, enum migrate_mode mode)
 621 {
 622         int ret;
 623
 624         EBUG_ON(!PageLocked(page));
 625         EBUG_ON(!PageLocked(newpage));
 626
 627         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 628         if (ret != MIGRATEPAGE_SUCCESS)
 629                 return ret;
 630
 631         if (PagePrivate(page))
 632                 attach_page_private(newpage, detach_page_private(page));
 633
 634         if (mode != MIGRATE_SYNC_NO_COPY)
 635                 migrate_page_copy(newpage, page);
 636         else
 637                 migrate_page_states(newpage, page);
 638         return MIGRATEPAGE_SUCCESS;
 639 }
 640 #endif
 641
 642 /* readpage(s): */
 643
 644 static void bch2_readpages_end_io(struct bio *bio)
 645 {
 646         struct bvec_iter_all iter;
 647         struct bio_vec *bv;
 648
 649         bio_for_each_segment_all(bv, bio, iter) {
 650                 struct page *page = bv->bv_page;
 651
 652                 if (!bio->bi_status) {
 653                         SetPageUptodate(page);
 654                 } else {
 655                         ClearPageUptodate(page);
 656                         SetPageError(page);
 657                 }
 658                 unlock_page(page);
 659         }
 660
 661         bio_put(bio);
 662 }
 663
 664 struct readpages_iter {
 665         struct address_space    *mapping;
 666         struct page             **pages;
 667         unsigned                nr_pages;
 668         unsigned                idx;
 669         pgoff_t                 offset;
 670 };
 671
 672 static int readpages_iter_init(struct readpages_iter *iter,
 673                                struct readahead_control *ractl)
 674 {
 675         unsigned i, nr_pages = readahead_count(ractl);
 676
 677         memset(iter, 0, sizeof(*iter));
 678
 679         iter->mapping   = ractl->mapping;
 680         iter->offset    = readahead_index(ractl);
 681         iter->nr_pages  = nr_pages;
 682
 683         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 684         if (!iter->pages)
 685                 return -ENOMEM;
 686
 687         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 688         for (i = 0; i < nr_pages; i++) {
 689                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 690                 put_page(iter->pages[i]);
 691         }
 692
 693         return 0;
 694 }
 695
 696 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 697 {
 698         if (iter->idx >= iter->nr_pages)
 699                 return NULL;
 700
 701         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 702
 703         return iter->pages[iter->idx];
 704 }
 705
 706 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 707 {
 708         struct bvec_iter iter;
 709         struct bio_vec bv;
 710         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 711                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 712         unsigned state = k.k->type == KEY_TYPE_reservation
 713                 ? SECTOR_RESERVED
 714                 : SECTOR_ALLOCATED;
 715
 716         bio_for_each_segment(bv, bio, iter) {
 717                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 718                 unsigned i;
 719
 720                 for (i = bv.bv_offset >> 9;
 721                      i < (bv.bv_offset + bv.bv_len) >> 9;
 722                      i++) {
 723                         s->s[i].nr_replicas = nr_ptrs;
 724                         s->s[i].state = state;
 725                 }
 726         }
 727 }
 728
 729 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 730 {
 731         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 732         struct bch_extent_crc_unpacked crc;
 733         const union bch_extent_entry *i;
 734
 735         bkey_for_each_crc(k.k, ptrs, crc, i)
 736                 if (crc.csum_type || crc.compression_type)
 737                         return true;
 738         return false;
 739 }
 740
 741 static void readpage_bio_extend(struct readpages_iter *iter,
 742                                 struct bio *bio,
 743                                 unsigned sectors_this_extent,
 744                                 bool get_more)
 745 {
 746         while (bio_sectors(bio) < sectors_this_extent &&
 747                bio->bi_vcnt < bio->bi_max_vecs) {
 748                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 749                 struct page *page = readpage_iter_next(iter);
 750                 int ret;
 751
 752                 if (page) {
 753                         if (iter->offset + iter->idx != page_offset)
 754                                 break;
 755
 756                         iter->idx++;
 757                 } else {
 758                         if (!get_more)
 759                                 break;
 760
 761                         page = xa_load(&iter->mapping->i_pages, page_offset);
 762                         if (page && !xa_is_value(page))
 763                                 break;
 764
 765                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 766                         if (!page)
 767                                 break;
 768
 769                         if (!__bch2_page_state_create(page, 0)) {
 770                                 put_page(page);
 771                                 break;
 772                         }
 773
 774                         ret = add_to_page_cache_lru(page, iter->mapping,
 775                                                     page_offset, GFP_NOFS);
 776                         if (ret) {
 777                                 __bch2_page_state_release(page);
 778                                 put_page(page);
 779                                 break;
 780                         }
 781
 782                         put_page(page);
 783                 }
 784
 785                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 786         }
 787 }
 788
 789 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 790                        struct bch_read_bio *rbio, u64 inum,
 791                        struct readpages_iter *readpages_iter)
 792 {
 793         struct bch_fs *c = trans->c;
 794         struct bkey_buf sk;
 795         int flags = BCH_READ_RETRY_IF_STALE|
 796                 BCH_READ_MAY_PROMOTE;
 797         int ret = 0;
 798
 799         rbio->c = c;
 800         rbio->start_time = local_clock();
 801
 802         bch2_bkey_buf_init(&sk);
 803 retry:
 804         bch2_trans_begin(trans);
 805
 806         while (1) {
 807                 struct bkey_s_c k;
 808                 unsigned bytes, sectors, offset_into_extent;
 809                 enum btree_id data_btree = BTREE_ID_extents;
 810
 811                 bch2_btree_iter_set_pos(iter,
 812                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 813
 814                 k = bch2_btree_iter_peek_slot(iter);
 815                 ret = bkey_err(k);
 816                 if (ret)
 817                         break;
 818
 819                 offset_into_extent = iter->pos.offset -
 820                         bkey_start_offset(k.k);
 821                 sectors = k.k->size - offset_into_extent;
 822
 823                 bch2_bkey_buf_reassemble(&sk, c, k);
 824
 825                 ret = bch2_read_indirect_extent(trans, &data_btree,
 826                                         &offset_into_extent, &sk);
 827                 if (ret)
 828                         break;
 829
 830                 k = bkey_i_to_s_c(sk.k);
 831
 832                 sectors = min(sectors, k.k->size - offset_into_extent);
 833
 834                 bch2_trans_unlock(trans);
 835
 836                 if (readpages_iter)
 837                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 838                                             extent_partial_reads_expensive(k));
 839
 840                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 841                 swap(rbio->bio.bi_iter.bi_size, bytes);
 842
 843                 if (rbio->bio.bi_iter.bi_size == bytes)
 844                         flags |= BCH_READ_LAST_FRAGMENT;
 845
 846                 if (bkey_extent_is_allocation(k.k))
 847                         bch2_add_page_sectors(&rbio->bio, k);
 848
 849                 bch2_read_extent(trans, rbio, iter->pos,
 850                                  data_btree, k, offset_into_extent, flags);
 851
 852                 if (flags & BCH_READ_LAST_FRAGMENT)
 853                         break;
 854
 855                 swap(rbio->bio.bi_iter.bi_size, bytes);
 856                 bio_advance(&rbio->bio, bytes);
 857         }
 858
 859         if (ret == -EINTR)
 860                 goto retry;
 861
 862         if (ret) {
 863                 bch_err_inum_ratelimited(c, inum,
 864                                 "read error %i from btree lookup", ret);
 865                 rbio->bio.bi_status = BLK_STS_IOERR;
 866                 bio_endio(&rbio->bio);
 867         }
 868
 869         bch2_bkey_buf_exit(&sk, c);
 870 }
 871
 872 void bch2_readahead(struct readahead_control *ractl)
 873 {
 874         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 875         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 876         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 877         struct btree_trans trans;
 878         struct btree_iter *iter;
 879         struct page *page;
 880         struct readpages_iter readpages_iter;
 881         int ret;
 882
 883         ret = readpages_iter_init(&readpages_iter, ractl);
 884         BUG_ON(ret);
 885
 886         bch2_trans_init(&trans, c, 0, 0);
 887         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 888                                    BTREE_ITER_SLOTS);
 889
 890         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 891
 892         while ((page = readpage_iter_next(&readpages_iter))) {
 893                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 894                 unsigned n = min_t(unsigned,
 895                                    readpages_iter.nr_pages -
 896                                    readpages_iter.idx,
 897                                    BIO_MAX_VECS);
 898                 struct bch_read_bio *rbio =
 899                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 900                                   opts);
 901
 902                 readpages_iter.idx++;
 903
 904                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 905                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 906                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 907                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 908
 909                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 910                            &readpages_iter);
 911         }
 912
 913         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 914
 915         bch2_trans_iter_put(&trans, iter);
 916         bch2_trans_exit(&trans);
 917         kfree(readpages_iter.pages);
 918 }
 919
 920 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 921                              u64 inum, struct page *page)
 922 {
 923         struct btree_trans trans;
 924         struct btree_iter *iter;
 925
 926         bch2_page_state_create(page, __GFP_NOFAIL);
 927
 928         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 929         rbio->bio.bi_iter.bi_sector =
 930                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 931         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 932
 933         bch2_trans_init(&trans, c, 0, 0);
 934         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 935                                    BTREE_ITER_SLOTS);
 936
 937         bchfs_read(&trans, iter, rbio, inum, NULL);
 938
 939         bch2_trans_iter_put(&trans, iter);
 940         bch2_trans_exit(&trans);
 941 }
 942
 943 int bch2_readpage(struct file *file, struct page *page)
 944 {
 945         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 946         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 947         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 948         struct bch_read_bio *rbio;
 949
 950         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 951         rbio->bio.bi_end_io = bch2_readpages_end_io;
 952
 953         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 954         return 0;
 955 }
 956
 957 static void bch2_read_single_page_end_io(struct bio *bio)
 958 {
 959         complete(bio->bi_private);
 960 }
 961
 962 static int bch2_read_single_page(struct page *page,
 963                                  struct address_space *mapping)
 964 {
 965         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 966         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 967         struct bch_read_bio *rbio;
 968         int ret;
 969         DECLARE_COMPLETION_ONSTACK(done);
 970
 971         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 972                          io_opts(c, &inode->ei_inode));
 973         rbio->bio.bi_private = &done;
 974         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 975
 976         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 977         wait_for_completion(&done);
 978
 979         ret = blk_status_to_errno(rbio->bio.bi_status);
 980         bio_put(&rbio->bio);
 981
 982         if (ret < 0)
 983                 return ret;
 984
 985         SetPageUptodate(page);
 986         return 0;
 987 }
 988
 989 /* writepages: */
 990
 991 struct bch_writepage_state {
 992         struct bch_writepage_io *io;
 993         struct bch_io_opts      opts;
 994 };
 995
 996 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 997                                                                   struct bch_inode_info *inode)
 998 {
 999         return (struct bch_writepage_state) {
1000                 .opts = io_opts(c, &inode->ei_inode)
1001         };
1002 }
1003
1004 static void bch2_writepage_io_free(struct closure *cl)
1005 {
1006         struct bch_writepage_io *io = container_of(cl,
1007                                         struct bch_writepage_io, cl);
1008
1009         bio_put(&io->op.wbio.bio);
1010 }
1011
1012 static void bch2_writepage_io_done(struct closure *cl)
1013 {
1014         struct bch_writepage_io *io = container_of(cl,
1015                                         struct bch_writepage_io, cl);
1016         struct bch_fs *c = io->op.c;
1017         struct bio *bio = &io->op.wbio.bio;
1018         struct bvec_iter_all iter;
1019         struct bio_vec *bvec;
1020         unsigned i;
1021
1022         up(&io->op.c->io_in_flight);
1023
1024         if (io->op.error) {
1025                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1026
1027                 bio_for_each_segment_all(bvec, bio, iter) {
1028                         struct bch_page_state *s;
1029
1030                         SetPageError(bvec->bv_page);
1031                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1032
1033                         s = __bch2_page_state(bvec->bv_page);
1034                         spin_lock(&s->lock);
1035                         for (i = 0; i < PAGE_SECTORS; i++)
1036                                 s->s[i].nr_replicas = 0;
1037                         spin_unlock(&s->lock);
1038                 }
1039         }
1040
1041         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1042                 bio_for_each_segment_all(bvec, bio, iter) {
1043                         struct bch_page_state *s;
1044
1045                         s = __bch2_page_state(bvec->bv_page);
1046                         spin_lock(&s->lock);
1047                         for (i = 0; i < PAGE_SECTORS; i++)
1048                                 s->s[i].nr_replicas = 0;
1049                         spin_unlock(&s->lock);
1050                 }
1051         }
1052
1053         /*
1054          * racing with fallocate can cause us to add fewer sectors than
1055          * expected - but we shouldn't add more sectors than expected:
1056          */
1057         BUG_ON(io->op.i_sectors_delta > 0);
1058
1059         /*
1060          * (error (due to going RO) halfway through a page can screw that up
1061          * slightly)
1062          * XXX wtf?
1063            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1064          */
1065
1066         /*
1067          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1068          * before calling end_page_writeback:
1069          */
1070         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1071
1072         bio_for_each_segment_all(bvec, bio, iter) {
1073                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1074
1075                 if (atomic_dec_and_test(&s->write_count))
1076                         end_page_writeback(bvec->bv_page);
1077         }
1078
1079         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1080 }
1081
1082 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1083 {
1084         struct bch_writepage_io *io = w->io;
1085
1086         down(&io->op.c->io_in_flight);
1087
1088         w->io = NULL;
1089         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1090         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1091 }
1092
1093 /*
1094  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1095  * possible, else allocating a new one:
1096  */
1097 static void bch2_writepage_io_alloc(struct bch_fs *c,
1098                                     struct writeback_control *wbc,
1099                                     struct bch_writepage_state *w,
1100                                     struct bch_inode_info *inode,
1101                                     u64 sector,
1102                                     unsigned nr_replicas)
1103 {
1104         struct bch_write_op *op;
1105
1106         w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1107                                               &c->writepage_bioset),
1108                              struct bch_writepage_io, op.wbio.bio);
1109
1110         closure_init(&w->io->cl, NULL);
1111         w->io->inode            = inode;
1112
1113         op                      = &w->io->op;
1114         bch2_write_op_init(op, c, w->opts);
1115         op->target              = w->opts.foreground_target;
1116         op_journal_seq_set(op, &inode->ei_journal_seq);
1117         op->nr_replicas         = nr_replicas;
1118         op->res.nr_replicas     = nr_replicas;
1119         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1120         op->pos                 = POS(inode->v.i_ino, sector);
1121         op->wbio.bio.bi_iter.bi_sector = sector;
1122         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1123 }
1124
1125 static int __bch2_writepage(struct page *page,
1126                             struct writeback_control *wbc,
1127                             void *data)
1128 {
1129         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1130         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1131         struct bch_writepage_state *w = data;
1132         struct bch_page_state *s, orig;
1133         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1134         loff_t i_size = i_size_read(&inode->v);
1135         pgoff_t end_index = i_size >> PAGE_SHIFT;
1136         int ret;
1137
1138         EBUG_ON(!PageUptodate(page));
1139
1140         /* Is the page fully inside i_size? */
1141         if (page->index < end_index)
1142                 goto do_io;
1143
1144         /* Is the page fully outside i_size? (truncate in progress) */
1145         offset = i_size & (PAGE_SIZE - 1);
1146         if (page->index > end_index || !offset) {
1147                 unlock_page(page);
1148                 return 0;
1149         }
1150
1151         /*
1152          * The page straddles i_size.  It must be zeroed out on each and every
1153          * writepage invocation because it may be mmapped.  "A file is mapped
1154          * in multiples of the page size.  For a file that is not a multiple of
1155          * the  page size, the remaining memory is zeroed when mapped, and
1156          * writes to that region are not written out to the file."
1157          */
1158         zero_user_segment(page, offset, PAGE_SIZE);
1159 do_io:
1160         s = bch2_page_state_create(page, __GFP_NOFAIL);
1161
1162         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1163         if (ret) {
1164                 SetPageError(page);
1165                 mapping_set_error(page->mapping, ret);
1166                 unlock_page(page);
1167                 return 0;
1168         }
1169
1170         /* Before unlocking the page, get copy of reservations: */
1171         orig = *s;
1172
1173         for (i = 0; i < PAGE_SECTORS; i++) {
1174                 if (s->s[i].state < SECTOR_DIRTY)
1175                         continue;
1176
1177                 nr_replicas_this_write =
1178                         min_t(unsigned, nr_replicas_this_write,
1179                               s->s[i].nr_replicas +
1180                               s->s[i].replicas_reserved);
1181         }
1182
1183         for (i = 0; i < PAGE_SECTORS; i++) {
1184                 if (s->s[i].state < SECTOR_DIRTY)
1185                         continue;
1186
1187                 s->s[i].nr_replicas = w->opts.compression
1188                         ? 0 : nr_replicas_this_write;
1189
1190                 s->s[i].replicas_reserved = 0;
1191                 s->s[i].state = SECTOR_ALLOCATED;
1192         }
1193
1194         BUG_ON(atomic_read(&s->write_count));
1195         atomic_set(&s->write_count, 1);
1196
1197         BUG_ON(PageWriteback(page));
1198         set_page_writeback(page);
1199
1200         unlock_page(page);
1201
1202         offset = 0;
1203         while (1) {
1204                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1205                 u64 sector;
1206
1207                 while (offset < PAGE_SECTORS &&
1208                        orig.s[offset].state < SECTOR_DIRTY)
1209                         offset++;
1210
1211                 if (offset == PAGE_SECTORS)
1212                         break;
1213
1214                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1215
1216                 while (offset + sectors < PAGE_SECTORS &&
1217                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1218                         sectors++;
1219
1220                 for (i = offset; i < offset + sectors; i++) {
1221                         reserved_sectors += orig.s[i].replicas_reserved;
1222                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1223                 }
1224
1225                 if (w->io &&
1226                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1227                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1228                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1229                      (BIO_MAX_VECS * PAGE_SIZE) ||
1230                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1231                         bch2_writepage_do_io(w);
1232
1233                 if (!w->io)
1234                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1235                                                 nr_replicas_this_write);
1236
1237                 atomic_inc(&s->write_count);
1238
1239                 BUG_ON(inode != w->io->inode);
1240                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1241                                      sectors << 9, offset << 9));
1242
1243                 /* Check for writing past i_size: */
1244                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1245                         round_up(i_size, block_bytes(c)));
1246
1247                 w->io->op.res.sectors += reserved_sectors;
1248                 w->io->op.i_sectors_delta -= dirty_sectors;
1249                 w->io->op.new_i_size = i_size;
1250
1251                 offset += sectors;
1252         }
1253
1254         if (atomic_dec_and_test(&s->write_count))
1255                 end_page_writeback(page);
1256
1257         return 0;
1258 }
1259
1260 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1261 {
1262         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1263         struct bch_writepage_state w =
1264                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1265         struct blk_plug plug;
1266         int ret;
1267
1268         blk_start_plug(&plug);
1269         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1270         if (w.io)
1271                 bch2_writepage_do_io(&w);
1272         blk_finish_plug(&plug);
1273         return ret;
1274 }
1275
1276 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1277 {
1278         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1279         struct bch_writepage_state w =
1280                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1281         int ret;
1282
1283         ret = __bch2_writepage(page, wbc, &w);
1284         if (w.io)
1285                 bch2_writepage_do_io(&w);
1286
1287         return ret;
1288 }
1289
1290 /* buffered writes: */
1291
1292 int bch2_write_begin(struct file *file, struct address_space *mapping,
1293                      loff_t pos, unsigned len, unsigned flags,
1294                      struct page **pagep, void **fsdata)
1295 {
1296         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1297         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1298         struct bch2_page_reservation *res;
1299         pgoff_t index = pos >> PAGE_SHIFT;
1300         unsigned offset = pos & (PAGE_SIZE - 1);
1301         struct page *page;
1302         int ret = -ENOMEM;
1303
1304         res = kmalloc(sizeof(*res), GFP_KERNEL);
1305         if (!res)
1306                 return -ENOMEM;
1307
1308         bch2_page_reservation_init(c, inode, res);
1309         *fsdata = res;
1310
1311         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1312
1313         page = grab_cache_page_write_begin(mapping, index, flags);
1314         if (!page)
1315                 goto err_unlock;
1316
1317         if (PageUptodate(page))
1318                 goto out;
1319
1320         /* If we're writing entire page, don't need to read it in first: */
1321         if (len == PAGE_SIZE)
1322                 goto out;
1323
1324         if (!offset && pos + len >= inode->v.i_size) {
1325                 zero_user_segment(page, len, PAGE_SIZE);
1326                 flush_dcache_page(page);
1327                 goto out;
1328         }
1329
1330         if (index > inode->v.i_size >> PAGE_SHIFT) {
1331                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1332                 flush_dcache_page(page);
1333                 goto out;
1334         }
1335 readpage:
1336         ret = bch2_read_single_page(page, mapping);
1337         if (ret)
1338                 goto err;
1339 out:
1340         ret = bch2_page_reservation_get(c, inode, page, res,
1341                                         offset, len, true);
1342         if (ret) {
1343                 if (!PageUptodate(page)) {
1344                         /*
1345                          * If the page hasn't been read in, we won't know if we
1346                          * actually need a reservation - we don't actually need
1347                          * to read here, we just need to check if the page is
1348                          * fully backed by uncompressed data:
1349                          */
1350                         goto readpage;
1351                 }
1352
1353                 goto err;
1354         }
1355
1356         *pagep = page;
1357         return 0;
1358 err:
1359         unlock_page(page);
1360         put_page(page);
1361         *pagep = NULL;
1362 err_unlock:
1363         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1364         kfree(res);
1365         *fsdata = NULL;
1366         return ret;
1367 }
1368
1369 int bch2_write_end(struct file *file, struct address_space *mapping,
1370                    loff_t pos, unsigned len, unsigned copied,
1371                    struct page *page, void *fsdata)
1372 {
1373         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1374         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1375         struct bch2_page_reservation *res = fsdata;
1376         unsigned offset = pos & (PAGE_SIZE - 1);
1377
1378         lockdep_assert_held(&inode->v.i_rwsem);
1379
1380         if (unlikely(copied < len && !PageUptodate(page))) {
1381                 /*
1382                  * The page needs to be read in, but that would destroy
1383                  * our partial write - simplest thing is to just force
1384                  * userspace to redo the write:
1385                  */
1386                 zero_user(page, 0, PAGE_SIZE);
1387                 flush_dcache_page(page);
1388                 copied = 0;
1389         }
1390
1391         spin_lock(&inode->v.i_lock);
1392         if (pos + copied > inode->v.i_size)
1393                 i_size_write(&inode->v, pos + copied);
1394         spin_unlock(&inode->v.i_lock);
1395
1396         if (copied) {
1397                 if (!PageUptodate(page))
1398                         SetPageUptodate(page);
1399
1400                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1401
1402                 inode->ei_last_dirtied = (unsigned long) current;
1403         }
1404
1405         unlock_page(page);
1406         put_page(page);
1407         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1408
1409         bch2_page_reservation_put(c, inode, res);
1410         kfree(res);
1411
1412         return copied;
1413 }
1414
1415 #define WRITE_BATCH_PAGES       32
1416
1417 static int __bch2_buffered_write(struct bch_inode_info *inode,
1418                                  struct address_space *mapping,
1419                                  struct iov_iter *iter,
1420                                  loff_t pos, unsigned len)
1421 {
1422         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1423         struct page *pages[WRITE_BATCH_PAGES];
1424         struct bch2_page_reservation res;
1425         unsigned long index = pos >> PAGE_SHIFT;
1426         unsigned offset = pos & (PAGE_SIZE - 1);
1427         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1428         unsigned i, reserved = 0, set_dirty = 0;
1429         unsigned copied = 0, nr_pages_copied = 0;
1430         int ret = 0;
1431
1432         BUG_ON(!len);
1433         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1434
1435         bch2_page_reservation_init(c, inode, &res);
1436
1437         for (i = 0; i < nr_pages; i++) {
1438                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1439                 if (!pages[i]) {
1440                         nr_pages = i;
1441                         if (!i) {
1442                                 ret = -ENOMEM;
1443                                 goto out;
1444                         }
1445                         len = min_t(unsigned, len,
1446                                     nr_pages * PAGE_SIZE - offset);
1447                         break;
1448                 }
1449         }
1450
1451         if (offset && !PageUptodate(pages[0])) {
1452                 ret = bch2_read_single_page(pages[0], mapping);
1453                 if (ret)
1454                         goto out;
1455         }
1456
1457         if ((pos + len) & (PAGE_SIZE - 1) &&
1458             !PageUptodate(pages[nr_pages - 1])) {
1459                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1460                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1461                 } else {
1462                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1463                         if (ret)
1464                                 goto out;
1465                 }
1466         }
1467
1468         while (reserved < len) {
1469                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1470                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1471                 unsigned pg_len = min_t(unsigned, len - reserved,
1472                                         PAGE_SIZE - pg_offset);
1473 retry_reservation:
1474                 ret = bch2_page_reservation_get(c, inode, page, &res,
1475                                                 pg_offset, pg_len, true);
1476
1477                 if (ret && !PageUptodate(page)) {
1478                         ret = bch2_read_single_page(page, mapping);
1479                         if (!ret)
1480                                 goto retry_reservation;
1481                 }
1482
1483                 if (ret)
1484                         goto out;
1485
1486                 reserved += pg_len;
1487         }
1488
1489         if (mapping_writably_mapped(mapping))
1490                 for (i = 0; i < nr_pages; i++)
1491                         flush_dcache_page(pages[i]);
1492
1493         while (copied < len) {
1494                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1495                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1496                 unsigned pg_len = min_t(unsigned, len - copied,
1497                                         PAGE_SIZE - pg_offset);
1498                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1499                                                 iter, pg_offset, pg_len);
1500
1501                 if (!pg_copied)
1502                         break;
1503
1504                 if (!PageUptodate(page) &&
1505                     pg_copied != PAGE_SIZE &&
1506                     pos + copied + pg_copied < inode->v.i_size) {
1507                         zero_user(page, 0, PAGE_SIZE);
1508                         break;
1509                 }
1510
1511                 flush_dcache_page(page);
1512                 iov_iter_advance(iter, pg_copied);
1513                 copied += pg_copied;
1514
1515                 if (pg_copied != pg_len)
1516                         break;
1517         }
1518
1519         if (!copied)
1520                 goto out;
1521
1522         spin_lock(&inode->v.i_lock);
1523         if (pos + copied > inode->v.i_size)
1524                 i_size_write(&inode->v, pos + copied);
1525         spin_unlock(&inode->v.i_lock);
1526
1527         while (set_dirty < copied) {
1528                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1529                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1530                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1531                                         PAGE_SIZE - pg_offset);
1532
1533                 if (!PageUptodate(page))
1534                         SetPageUptodate(page);
1535
1536                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1537                 unlock_page(page);
1538                 put_page(page);
1539
1540                 set_dirty += pg_len;
1541         }
1542
1543         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1544         inode->ei_last_dirtied = (unsigned long) current;
1545 out:
1546         for (i = nr_pages_copied; i < nr_pages; i++) {
1547                 unlock_page(pages[i]);
1548                 put_page(pages[i]);
1549         }
1550
1551         bch2_page_reservation_put(c, inode, &res);
1552
1553         return copied ?: ret;
1554 }
1555
1556 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1557 {
1558         struct file *file = iocb->ki_filp;
1559         struct address_space *mapping = file->f_mapping;
1560         struct bch_inode_info *inode = file_bch_inode(file);
1561         loff_t pos = iocb->ki_pos;
1562         ssize_t written = 0;
1563         int ret = 0;
1564
1565         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1566
1567         do {
1568                 unsigned offset = pos & (PAGE_SIZE - 1);
1569                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1570                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1571 again:
1572                 /*
1573                  * Bring in the user page that we will copy from _first_.
1574                  * Otherwise there's a nasty deadlock on copying from the
1575                  * same page as we're writing to, without it being marked
1576                  * up-to-date.
1577                  *
1578                  * Not only is this an optimisation, but it is also required
1579                  * to check that the address is actually valid, when atomic
1580                  * usercopies are used, below.
1581                  */
1582                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1583                         bytes = min_t(unsigned long, iov_iter_count(iter),
1584                                       PAGE_SIZE - offset);
1585
1586                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1587                                 ret = -EFAULT;
1588                                 break;
1589                         }
1590                 }
1591
1592                 if (unlikely(fatal_signal_pending(current))) {
1593                         ret = -EINTR;
1594                         break;
1595                 }
1596
1597                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1598                 if (unlikely(ret < 0))
1599                         break;
1600
1601                 cond_resched();
1602
1603                 if (unlikely(ret == 0)) {
1604                         /*
1605                          * If we were unable to copy any data at all, we must
1606                          * fall back to a single segment length write.
1607                          *
1608                          * If we didn't fallback here, we could livelock
1609                          * because not all segments in the iov can be copied at
1610                          * once without a pagefault.
1611                          */
1612                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1613                                       iov_iter_single_seg_count(iter));
1614                         goto again;
1615                 }
1616                 pos += ret;
1617                 written += ret;
1618                 ret = 0;
1619
1620                 balance_dirty_pages_ratelimited(mapping);
1621         } while (iov_iter_count(iter));
1622
1623         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1624
1625         return written ? written : ret;
1626 }
1627
1628 /* O_DIRECT reads */
1629
1630 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1631 {
1632         if (check_dirty) {
1633                 bio_check_pages_dirty(bio);
1634         } else {
1635                 bio_release_pages(bio, false);
1636                 bio_put(bio);
1637         }
1638 }
1639
1640 static void bch2_dio_read_complete(struct closure *cl)
1641 {
1642         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1643
1644         dio->req->ki_complete(dio->req, dio->ret, 0);
1645         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1646 }
1647
1648 static void bch2_direct_IO_read_endio(struct bio *bio)
1649 {
1650         struct dio_read *dio = bio->bi_private;
1651
1652         if (bio->bi_status)
1653                 dio->ret = blk_status_to_errno(bio->bi_status);
1654
1655         closure_put(&dio->cl);
1656 }
1657
1658 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1659 {
1660         struct dio_read *dio = bio->bi_private;
1661         bool should_dirty = dio->should_dirty;
1662
1663         bch2_direct_IO_read_endio(bio);
1664         bio_check_or_release(bio, should_dirty);
1665 }
1666
1667 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1668 {
1669         struct file *file = req->ki_filp;
1670         struct bch_inode_info *inode = file_bch_inode(file);
1671         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1672         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1673         struct dio_read *dio;
1674         struct bio *bio;
1675         loff_t offset = req->ki_pos;
1676         bool sync = is_sync_kiocb(req);
1677         size_t shorten;
1678         ssize_t ret;
1679
1680         if ((offset|iter->count) & (block_bytes(c) - 1))
1681                 return -EINVAL;
1682
1683         ret = min_t(loff_t, iter->count,
1684                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1685
1686         if (!ret)
1687                 return ret;
1688
1689         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1690         iter->count -= shorten;
1691
1692         bio = bio_alloc_bioset(GFP_KERNEL,
1693                                iov_iter_npages(iter, BIO_MAX_VECS),
1694                                &c->dio_read_bioset);
1695
1696         bio->bi_end_io = bch2_direct_IO_read_endio;
1697
1698         dio = container_of(bio, struct dio_read, rbio.bio);
1699         closure_init(&dio->cl, NULL);
1700
1701         /*
1702          * this is a _really_ horrible hack just to avoid an atomic sub at the
1703          * end:
1704          */
1705         if (!sync) {
1706                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1707                 atomic_set(&dio->cl.remaining,
1708                            CLOSURE_REMAINING_INITIALIZER -
1709                            CLOSURE_RUNNING +
1710                            CLOSURE_DESTRUCTOR);
1711         } else {
1712                 atomic_set(&dio->cl.remaining,
1713                            CLOSURE_REMAINING_INITIALIZER + 1);
1714         }
1715
1716         dio->req        = req;
1717         dio->ret        = ret;
1718         /*
1719          * This is one of the sketchier things I've encountered: we have to skip
1720          * the dirtying of requests that are internal from the kernel (i.e. from
1721          * loopback), because we'll deadlock on page_lock.
1722          */
1723         dio->should_dirty = iter_is_iovec(iter);
1724
1725         goto start;
1726         while (iter->count) {
1727                 bio = bio_alloc_bioset(GFP_KERNEL,
1728                                        iov_iter_npages(iter, BIO_MAX_VECS),
1729                                        &c->bio_read);
1730                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1731 start:
1732                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1733                 bio->bi_iter.bi_sector  = offset >> 9;
1734                 bio->bi_private         = dio;
1735
1736                 ret = bio_iov_iter_get_pages(bio, iter);
1737                 if (ret < 0) {
1738                         /* XXX: fault inject this path */
1739                         bio->bi_status = BLK_STS_RESOURCE;
1740                         bio_endio(bio);
1741                         break;
1742                 }
1743
1744                 offset += bio->bi_iter.bi_size;
1745
1746                 if (dio->should_dirty)
1747                         bio_set_pages_dirty(bio);
1748
1749                 if (iter->count)
1750                         closure_get(&dio->cl);
1751
1752                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1753         }
1754
1755         iter->count += shorten;
1756
1757         if (sync) {
1758                 closure_sync(&dio->cl);
1759                 closure_debug_destroy(&dio->cl);
1760                 ret = dio->ret;
1761                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1762                 return ret;
1763         } else {
1764                 return -EIOCBQUEUED;
1765         }
1766 }
1767
1768 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1769 {
1770         struct file *file = iocb->ki_filp;
1771         struct bch_inode_info *inode = file_bch_inode(file);
1772         struct address_space *mapping = file->f_mapping;
1773         size_t count = iov_iter_count(iter);
1774         ssize_t ret;
1775
1776         if (!count)
1777                 return 0; /* skip atime */
1778
1779         if (iocb->ki_flags & IOCB_DIRECT) {
1780                 struct blk_plug plug;
1781
1782                 ret = filemap_write_and_wait_range(mapping,
1783                                         iocb->ki_pos,
1784                                         iocb->ki_pos + count - 1);
1785                 if (ret < 0)
1786                         return ret;
1787
1788                 file_accessed(file);
1789
1790                 blk_start_plug(&plug);
1791                 ret = bch2_direct_IO_read(iocb, iter);
1792                 blk_finish_plug(&plug);
1793
1794                 if (ret >= 0)
1795                         iocb->ki_pos += ret;
1796         } else {
1797                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1798                 ret = generic_file_read_iter(iocb, iter);
1799                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1800         }
1801
1802         return ret;
1803 }
1804
1805 /* O_DIRECT writes */
1806
1807 static void bch2_dio_write_loop_async(struct bch_write_op *);
1808
1809 static long bch2_dio_write_loop(struct dio_write *dio)
1810 {
1811         bool kthread = (current->flags & PF_KTHREAD) != 0;
1812         struct kiocb *req = dio->req;
1813         struct address_space *mapping = req->ki_filp->f_mapping;
1814         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1815         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1816         struct bio *bio = &dio->op.wbio.bio;
1817         struct bvec_iter_all iter;
1818         struct bio_vec *bv;
1819         unsigned unaligned, iter_count;
1820         bool sync = dio->sync, dropped_locks;
1821         long ret;
1822
1823         if (dio->loop)
1824                 goto loop;
1825
1826         down(&c->io_in_flight);
1827
1828         while (1) {
1829                 iter_count = dio->iter.count;
1830
1831                 if (kthread)
1832                         kthread_use_mm(dio->mm);
1833                 BUG_ON(current->faults_disabled_mapping);
1834                 current->faults_disabled_mapping = mapping;
1835
1836                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1837
1838                 dropped_locks = fdm_dropped_locks();
1839
1840                 current->faults_disabled_mapping = NULL;
1841                 if (kthread)
1842                         kthread_unuse_mm(dio->mm);
1843
1844                 /*
1845                  * If the fault handler returned an error but also signalled
1846                  * that it dropped & retook ei_pagecache_lock, we just need to
1847                  * re-shoot down the page cache and retry:
1848                  */
1849                 if (dropped_locks && ret)
1850                         ret = 0;
1851
1852                 if (unlikely(ret < 0))
1853                         goto err;
1854
1855                 if (unlikely(dropped_locks)) {
1856                         ret = write_invalidate_inode_pages_range(mapping,
1857                                         req->ki_pos,
1858                                         req->ki_pos + iter_count - 1);
1859                         if (unlikely(ret))
1860                                 goto err;
1861
1862                         if (!bio->bi_iter.bi_size)
1863                                 continue;
1864                 }
1865
1866                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1867                 bio->bi_iter.bi_size -= unaligned;
1868                 iov_iter_revert(&dio->iter, unaligned);
1869
1870                 if (!bio->bi_iter.bi_size) {
1871                         /*
1872                          * bio_iov_iter_get_pages was only able to get <
1873                          * blocksize worth of pages:
1874                          */
1875                         ret = -EFAULT;
1876                         goto err;
1877                 }
1878
1879                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1880                 dio->op.end_io          = bch2_dio_write_loop_async;
1881                 dio->op.target          = dio->op.opts.foreground_target;
1882                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1883                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1884                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1885                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1886
1887                 if ((req->ki_flags & IOCB_DSYNC) &&
1888                     !c->opts.journal_flush_disabled)
1889                         dio->op.flags |= BCH_WRITE_FLUSH;
1890                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
1891
1892                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1893                                                 dio->op.opts.data_replicas, 0);
1894                 if (unlikely(ret) &&
1895                     !bch2_check_range_allocated(c, dio->op.pos,
1896                                 bio_sectors(bio),
1897                                 dio->op.opts.data_replicas,
1898                                 dio->op.opts.compression != 0))
1899                         goto err;
1900
1901                 task_io_account_write(bio->bi_iter.bi_size);
1902
1903                 if (!dio->sync && !dio->loop && dio->iter.count) {
1904                         struct iovec *iov = dio->inline_vecs;
1905
1906                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1907                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1908                                               GFP_KERNEL);
1909                                 if (unlikely(!iov)) {
1910                                         dio->sync = sync = true;
1911                                         goto do_io;
1912                                 }
1913
1914                                 dio->free_iov = true;
1915                         }
1916
1917                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1918                         dio->iter.iov = iov;
1919                 }
1920 do_io:
1921                 dio->loop = true;
1922                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1923
1924                 if (sync)
1925                         wait_for_completion(&dio->done);
1926                 else
1927                         return -EIOCBQUEUED;
1928 loop:
1929                 i_sectors_acct(c, inode, &dio->quota_res,
1930                                dio->op.i_sectors_delta);
1931                 req->ki_pos += (u64) dio->op.written << 9;
1932                 dio->written += dio->op.written;
1933
1934                 spin_lock(&inode->v.i_lock);
1935                 if (req->ki_pos > inode->v.i_size)
1936                         i_size_write(&inode->v, req->ki_pos);
1937                 spin_unlock(&inode->v.i_lock);
1938
1939                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1940                         bio_for_each_segment_all(bv, bio, iter)
1941                                 put_page(bv->bv_page);
1942                 bio->bi_vcnt = 0;
1943
1944                 if (dio->op.error) {
1945                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
1946                         break;
1947                 }
1948
1949                 if (!dio->iter.count)
1950                         break;
1951
1952                 bio_reset(bio);
1953                 reinit_completion(&dio->done);
1954         }
1955
1956         ret = dio->op.error ?: ((long) dio->written << 9);
1957 err:
1958         up(&c->io_in_flight);
1959         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1960         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1961
1962         if (dio->free_iov)
1963                 kfree(dio->iter.iov);
1964
1965         if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1966                 bio_for_each_segment_all(bv, bio, iter)
1967                         put_page(bv->bv_page);
1968         bio_put(bio);
1969
1970         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1971         inode_dio_end(&inode->v);
1972
1973         if (!sync) {
1974                 req->ki_complete(req, ret, 0);
1975                 ret = -EIOCBQUEUED;
1976         }
1977         return ret;
1978 }
1979
1980 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1981 {
1982         struct dio_write *dio = container_of(op, struct dio_write, op);
1983
1984         if (dio->sync)
1985                 complete(&dio->done);
1986         else
1987                 bch2_dio_write_loop(dio);
1988 }
1989
1990 static noinline
1991 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1992 {
1993         struct file *file = req->ki_filp;
1994         struct address_space *mapping = file->f_mapping;
1995         struct bch_inode_info *inode = file_bch_inode(file);
1996         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1997         struct dio_write *dio;
1998         struct bio *bio;
1999         bool locked = true, extending;
2000         ssize_t ret;
2001
2002         prefetch(&c->opts);
2003         prefetch((void *) &c->opts + 64);
2004         prefetch(&inode->ei_inode);
2005         prefetch((void *) &inode->ei_inode + 64);
2006
2007         inode_lock(&inode->v);
2008
2009         ret = generic_write_checks(req, iter);
2010         if (unlikely(ret <= 0))
2011                 goto err;
2012
2013         ret = file_remove_privs(file);
2014         if (unlikely(ret))
2015                 goto err;
2016
2017         ret = file_update_time(file);
2018         if (unlikely(ret))
2019                 goto err;
2020
2021         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2022                 goto err;
2023
2024         inode_dio_begin(&inode->v);
2025         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2026
2027         extending = req->ki_pos + iter->count > inode->v.i_size;
2028         if (!extending) {
2029                 inode_unlock(&inode->v);
2030                 locked = false;
2031         }
2032
2033         bio = bio_alloc_bioset(GFP_KERNEL,
2034                                iov_iter_is_bvec(iter)
2035                                ? 0
2036                                : iov_iter_npages(iter, BIO_MAX_VECS),
2037                                &c->dio_write_bioset);
2038         dio = container_of(bio, struct dio_write, op.wbio.bio);
2039         init_completion(&dio->done);
2040         dio->req                = req;
2041         dio->mm                 = current->mm;
2042         dio->loop               = false;
2043         dio->sync               = is_sync_kiocb(req) || extending;
2044         dio->free_iov           = false;
2045         dio->quota_res.sectors  = 0;
2046         dio->written            = 0;
2047         dio->iter               = *iter;
2048
2049         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2050                                          iter->count >> 9, true);
2051         if (unlikely(ret))
2052                 goto err_put_bio;
2053
2054         ret = write_invalidate_inode_pages_range(mapping,
2055                                         req->ki_pos,
2056                                         req->ki_pos + iter->count - 1);
2057         if (unlikely(ret))
2058                 goto err_put_bio;
2059
2060         ret = bch2_dio_write_loop(dio);
2061 err:
2062         if (locked)
2063                 inode_unlock(&inode->v);
2064         return ret;
2065 err_put_bio:
2066         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2067         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2068         bio_put(bio);
2069         inode_dio_end(&inode->v);
2070         goto err;
2071 }
2072
2073 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2074 {
2075         struct file *file = iocb->ki_filp;
2076         struct bch_inode_info *inode = file_bch_inode(file);
2077         ssize_t ret;
2078
2079         if (iocb->ki_flags & IOCB_DIRECT)
2080                 return bch2_direct_write(iocb, from);
2081
2082         /* We can write back this queue in page reclaim */
2083         current->backing_dev_info = inode_to_bdi(&inode->v);
2084         inode_lock(&inode->v);
2085
2086         ret = generic_write_checks(iocb, from);
2087         if (ret <= 0)
2088                 goto unlock;
2089
2090         ret = file_remove_privs(file);
2091         if (ret)
2092                 goto unlock;
2093
2094         ret = file_update_time(file);
2095         if (ret)
2096                 goto unlock;
2097
2098         ret = bch2_buffered_write(iocb, from);
2099         if (likely(ret > 0))
2100                 iocb->ki_pos += ret;
2101 unlock:
2102         inode_unlock(&inode->v);
2103         current->backing_dev_info = NULL;
2104
2105         if (ret > 0)
2106                 ret = generic_write_sync(iocb, ret);
2107
2108         return ret;
2109 }
2110
2111 /* fsync: */
2112
2113 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2114 {
2115         struct bch_inode_info *inode = file_bch_inode(file);
2116         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2117         int ret, ret2;
2118
2119         ret = file_write_and_wait_range(file, start, end);
2120         if (ret)
2121                 return ret;
2122
2123         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2124                 goto out;
2125
2126         ret = sync_inode_metadata(&inode->v, 1);
2127         if (ret)
2128                 return ret;
2129 out:
2130         if (!c->opts.journal_flush_disabled)
2131                 ret = bch2_journal_flush_seq(&c->journal,
2132                                              inode->ei_journal_seq);
2133         ret2 = file_check_and_advance_wb_err(file);
2134
2135         return ret ?: ret2;
2136 }
2137
2138 /* truncate: */
2139
2140 static inline int range_has_data(struct bch_fs *c,
2141                                   struct bpos start,
2142                                   struct bpos end)
2143 {
2144         struct btree_trans trans;
2145         struct btree_iter *iter;
2146         struct bkey_s_c k;
2147         int ret = 0;
2148
2149         bch2_trans_init(&trans, c, 0, 0);
2150
2151         for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2152                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2153                         break;
2154
2155                 if (bkey_extent_is_data(k.k)) {
2156                         ret = 1;
2157                         break;
2158                 }
2159         }
2160         bch2_trans_iter_put(&trans, iter);
2161
2162         return bch2_trans_exit(&trans) ?: ret;
2163 }
2164
2165 static int __bch2_truncate_page(struct bch_inode_info *inode,
2166                                 pgoff_t index, loff_t start, loff_t end)
2167 {
2168         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2169         struct address_space *mapping = inode->v.i_mapping;
2170         struct bch_page_state *s;
2171         unsigned start_offset = start & (PAGE_SIZE - 1);
2172         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2173         unsigned i;
2174         struct page *page;
2175         int ret = 0;
2176
2177         /* Page boundary? Nothing to do */
2178         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2179               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2180                 return 0;
2181
2182         /* Above i_size? */
2183         if (index << PAGE_SHIFT >= inode->v.i_size)
2184                 return 0;
2185
2186         page = find_lock_page(mapping, index);
2187         if (!page) {
2188                 /*
2189                  * XXX: we're doing two index lookups when we end up reading the
2190                  * page
2191                  */
2192                 ret = range_has_data(c,
2193                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2194                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2195                 if (ret <= 0)
2196                         return ret;
2197
2198                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2199                 if (unlikely(!page)) {
2200                         ret = -ENOMEM;
2201                         goto out;
2202                 }
2203         }
2204
2205         s = bch2_page_state_create(page, 0);
2206         if (!s) {
2207                 ret = -ENOMEM;
2208                 goto unlock;
2209         }
2210
2211         if (!PageUptodate(page)) {
2212                 ret = bch2_read_single_page(page, mapping);
2213                 if (ret)
2214                         goto unlock;
2215         }
2216
2217         if (index != start >> PAGE_SHIFT)
2218                 start_offset = 0;
2219         if (index != end >> PAGE_SHIFT)
2220                 end_offset = PAGE_SIZE;
2221
2222         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2223              i < round_down(end_offset, block_bytes(c)) >> 9;
2224              i++) {
2225                 s->s[i].nr_replicas     = 0;
2226                 s->s[i].state           = SECTOR_UNALLOCATED;
2227         }
2228
2229         zero_user_segment(page, start_offset, end_offset);
2230
2231         /*
2232          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2233          *
2234          * XXX: because we aren't currently tracking whether the page has actual
2235          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2236          */
2237         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2238         BUG_ON(ret);
2239
2240         /*
2241          * This removes any writeable userspace mappings; we need to force
2242          * .page_mkwrite to be called again before any mmapped writes, to
2243          * redirty the full page:
2244          */
2245         page_mkclean(page);
2246         __set_page_dirty_nobuffers(page);
2247 unlock:
2248         unlock_page(page);
2249         put_page(page);
2250 out:
2251         return ret;
2252 }
2253
2254 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2255 {
2256         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2257                                     from, round_up(from, PAGE_SIZE));
2258 }
2259
2260 static int bch2_extend(struct user_namespace *mnt_userns,
2261                        struct bch_inode_info *inode,
2262                        struct bch_inode_unpacked *inode_u,
2263                        struct iattr *iattr)
2264 {
2265         struct address_space *mapping = inode->v.i_mapping;
2266         int ret;
2267
2268         /*
2269          * sync appends:
2270          *
2271          * this has to be done _before_ extending i_size:
2272          */
2273         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2274         if (ret)
2275                 return ret;
2276
2277         truncate_setsize(&inode->v, iattr->ia_size);
2278
2279         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2280 }
2281
2282 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2283                                    struct bch_inode_unpacked *bi,
2284                                    void *p)
2285 {
2286         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2287         return 0;
2288 }
2289
2290 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2291                                   struct bch_inode_unpacked *bi, void *p)
2292 {
2293         u64 *new_i_size = p;
2294
2295         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2296         bi->bi_size = *new_i_size;
2297         return 0;
2298 }
2299
2300 int bch2_truncate(struct user_namespace *mnt_userns,
2301                   struct bch_inode_info *inode, struct iattr *iattr)
2302 {
2303         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2304         struct address_space *mapping = inode->v.i_mapping;
2305         struct bch_inode_unpacked inode_u;
2306         u64 new_i_size = iattr->ia_size;
2307         s64 i_sectors_delta = 0;
2308         int ret = 0;
2309
2310         /*
2311          * If the truncate call with change the size of the file, the
2312          * cmtimes should be updated. If the size will not change, we
2313          * do not need to update the cmtimes.
2314          */
2315         if (iattr->ia_size != inode->v.i_size) {
2316                 if (!(iattr->ia_valid & ATTR_MTIME))
2317                         ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2318                 if (!(iattr->ia_valid & ATTR_CTIME))
2319                         ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2320                 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2321         }
2322
2323         inode_dio_wait(&inode->v);
2324         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2325
2326         ret = bch2_inode_find_by_inum(c, inode->v.i_ino, &inode_u);
2327         if (ret)
2328                 goto err;
2329
2330         /*
2331          * check this before next assertion; on filesystem error our normal
2332          * invariants are a bit broken (truncate has to truncate the page cache
2333          * before the inode).
2334          */
2335         ret = bch2_journal_error(&c->journal);
2336         if (ret)
2337                 goto err;
2338
2339         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2340                 inode->v.i_size < inode_u.bi_size);
2341
2342         if (iattr->ia_size > inode->v.i_size) {
2343                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2344                 goto err;
2345         }
2346
2347         iattr->ia_valid &= ~ATTR_SIZE;
2348
2349         ret = bch2_truncate_page(inode, iattr->ia_size);
2350         if (unlikely(ret))
2351                 goto err;
2352
2353         /*
2354          * When extending, we're going to write the new i_size to disk
2355          * immediately so we need to flush anything above the current on disk
2356          * i_size first:
2357          *
2358          * Also, when extending we need to flush the page that i_size currently
2359          * straddles - if it's mapped to userspace, we need to ensure that
2360          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2361          * again to allocate the part of the page that was extended.
2362          */
2363         if (iattr->ia_size > inode_u.bi_size)
2364                 ret = filemap_write_and_wait_range(mapping,
2365                                 inode_u.bi_size,
2366                                 iattr->ia_size - 1);
2367         else if (iattr->ia_size & (PAGE_SIZE - 1))
2368                 ret = filemap_write_and_wait_range(mapping,
2369                                 round_down(iattr->ia_size, PAGE_SIZE),
2370                                 iattr->ia_size - 1);
2371         if (ret)
2372                 goto err;
2373
2374         mutex_lock(&inode->ei_update_lock);
2375         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2376                                &new_i_size, 0);
2377         mutex_unlock(&inode->ei_update_lock);
2378
2379         if (unlikely(ret))
2380                 goto err;
2381
2382         truncate_setsize(&inode->v, iattr->ia_size);
2383
2384         ret = bch2_fpunch(c, inode->v.i_ino,
2385                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2386                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2387         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2388
2389         if (unlikely(ret))
2390                 goto err;
2391
2392         mutex_lock(&inode->ei_update_lock);
2393         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2394         mutex_unlock(&inode->ei_update_lock);
2395
2396         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2397 err:
2398         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2399         return ret;
2400 }
2401
2402 /* fallocate: */
2403
2404 static int inode_update_times_fn(struct bch_inode_info *inode,
2405                                  struct bch_inode_unpacked *bi, void *p)
2406 {
2407         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2408
2409         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2410         return 0;
2411 }
2412
2413 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2414 {
2415         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2416         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2417         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2418         int ret = 0;
2419
2420         inode_lock(&inode->v);
2421         inode_dio_wait(&inode->v);
2422         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2423
2424         ret = __bch2_truncate_page(inode,
2425                                    offset >> PAGE_SHIFT,
2426                                    offset, offset + len);
2427         if (unlikely(ret))
2428                 goto err;
2429
2430         if (offset >> PAGE_SHIFT !=
2431             (offset + len) >> PAGE_SHIFT) {
2432                 ret = __bch2_truncate_page(inode,
2433                                            (offset + len) >> PAGE_SHIFT,
2434                                            offset, offset + len);
2435                 if (unlikely(ret))
2436                         goto err;
2437         }
2438
2439         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2440
2441         if (discard_start < discard_end) {
2442                 s64 i_sectors_delta = 0;
2443
2444                 ret = bch2_fpunch(c, inode->v.i_ino,
2445                                   discard_start, discard_end,
2446                                   &inode->ei_journal_seq,
2447                                   &i_sectors_delta);
2448                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2449         }
2450
2451         mutex_lock(&inode->ei_update_lock);
2452         ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2453                                ATTR_MTIME|ATTR_CTIME) ?: ret;
2454         mutex_unlock(&inode->ei_update_lock);
2455 err:
2456         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2457         inode_unlock(&inode->v);
2458
2459         return ret;
2460 }
2461
2462 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2463                                    loff_t offset, loff_t len,
2464                                    bool insert)
2465 {
2466         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2467         struct address_space *mapping = inode->v.i_mapping;
2468         struct bkey_buf copy;
2469         struct btree_trans trans;
2470         struct btree_iter *src, *dst, *del;
2471         loff_t shift, new_size;
2472         u64 src_start;
2473         int ret = 0;
2474
2475         if ((offset | len) & (block_bytes(c) - 1))
2476                 return -EINVAL;
2477
2478         /*
2479          * We need i_mutex to keep the page cache consistent with the extents
2480          * btree, and the btree consistent with i_size - we don't need outside
2481          * locking for the extents btree itself, because we're using linked
2482          * iterators
2483          */
2484         inode_lock(&inode->v);
2485         inode_dio_wait(&inode->v);
2486         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2487
2488         if (insert) {
2489                 ret = -EFBIG;
2490                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2491                         goto err;
2492
2493                 ret = -EINVAL;
2494                 if (offset >= inode->v.i_size)
2495                         goto err;
2496
2497                 src_start       = U64_MAX;
2498                 shift           = len;
2499         } else {
2500                 ret = -EINVAL;
2501                 if (offset + len >= inode->v.i_size)
2502                         goto err;
2503
2504                 src_start       = offset + len;
2505                 shift           = -len;
2506         }
2507
2508         new_size = inode->v.i_size + shift;
2509
2510         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2511         if (ret)
2512                 goto err;
2513
2514         if (insert) {
2515                 i_size_write(&inode->v, new_size);
2516                 mutex_lock(&inode->ei_update_lock);
2517                 ret = bch2_write_inode_size(c, inode, new_size,
2518                                             ATTR_MTIME|ATTR_CTIME);
2519                 mutex_unlock(&inode->ei_update_lock);
2520         } else {
2521                 s64 i_sectors_delta = 0;
2522
2523                 ret = bch2_fpunch(c, inode->v.i_ino,
2524                                   offset >> 9, (offset + len) >> 9,
2525                                   &inode->ei_journal_seq,
2526                                   &i_sectors_delta);
2527                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2528
2529                 if (ret)
2530                         goto err;
2531         }
2532
2533         bch2_bkey_buf_init(&copy);
2534         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2535         src = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2536                         POS(inode->v.i_ino, src_start >> 9),
2537                         BTREE_ITER_INTENT);
2538         dst = bch2_trans_copy_iter(&trans, src);
2539         del = bch2_trans_copy_iter(&trans, src);
2540
2541         while (ret == 0 || ret == -EINTR) {
2542                 struct disk_reservation disk_res =
2543                         bch2_disk_reservation_init(c, 0);
2544                 struct bkey_i delete;
2545                 struct bkey_s_c k;
2546                 struct bpos next_pos;
2547                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2548                 struct bpos atomic_end;
2549                 unsigned trigger_flags = 0;
2550
2551                 bch2_trans_begin(&trans);
2552
2553                 k = insert
2554                         ? bch2_btree_iter_peek_prev(src)
2555                         : bch2_btree_iter_peek(src);
2556                 if ((ret = bkey_err(k)))
2557                         continue;
2558
2559                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2560                         break;
2561
2562                 if (insert &&
2563                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2564                         break;
2565 reassemble:
2566                 bch2_bkey_buf_reassemble(&copy, c, k);
2567
2568                 if (insert &&
2569                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2570                         bch2_cut_front(move_pos, copy.k);
2571
2572                 copy.k->k.p.offset += shift >> 9;
2573                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2574
2575                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2576                 if (ret)
2577                         continue;
2578
2579                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2580                         if (insert) {
2581                                 move_pos = atomic_end;
2582                                 move_pos.offset -= shift >> 9;
2583                                 goto reassemble;
2584                         } else {
2585                                 bch2_cut_back(atomic_end, copy.k);
2586                         }
2587                 }
2588
2589                 bkey_init(&delete.k);
2590                 delete.k.p = copy.k->k.p;
2591                 delete.k.size = copy.k->k.size;
2592                 delete.k.p.offset -= shift >> 9;
2593                 bch2_btree_iter_set_pos(del, bkey_start_pos(&delete.k));
2594
2595                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2596
2597                 if (copy.k->k.size == k.k->size) {
2598                         /*
2599                          * If we're moving the entire extent, we can skip
2600                          * running triggers:
2601                          */
2602                         trigger_flags |= BTREE_TRIGGER_NORUN;
2603                 } else {
2604                         /* We might end up splitting compressed extents: */
2605                         unsigned nr_ptrs =
2606                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2607
2608                         ret = bch2_disk_reservation_get(c, &disk_res,
2609                                         copy.k->k.size, nr_ptrs,
2610                                         BCH_DISK_RESERVATION_NOFAIL);
2611                         BUG_ON(ret);
2612                 }
2613
2614                 ret =   bch2_btree_iter_traverse(del) ?:
2615                         bch2_trans_update(&trans, del, &delete, trigger_flags) ?:
2616                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2617                         bch2_trans_commit(&trans, &disk_res,
2618                                           &inode->ei_journal_seq,
2619                                           BTREE_INSERT_NOFAIL);
2620                 bch2_disk_reservation_put(c, &disk_res);
2621
2622                 if (!ret)
2623                         bch2_btree_iter_set_pos(src, next_pos);
2624         }
2625         bch2_trans_iter_put(&trans, del);
2626         bch2_trans_iter_put(&trans, dst);
2627         bch2_trans_iter_put(&trans, src);
2628         bch2_trans_exit(&trans);
2629         bch2_bkey_buf_exit(&copy, c);
2630
2631         if (ret)
2632                 goto err;
2633
2634         if (!insert) {
2635                 i_size_write(&inode->v, new_size);
2636                 mutex_lock(&inode->ei_update_lock);
2637                 ret = bch2_write_inode_size(c, inode, new_size,
2638                                             ATTR_MTIME|ATTR_CTIME);
2639                 mutex_unlock(&inode->ei_update_lock);
2640         }
2641 err:
2642         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2643         inode_unlock(&inode->v);
2644         return ret;
2645 }
2646
2647 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2648                              u64 start_sector, u64 end_sector)
2649 {
2650         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2651         struct btree_trans trans;
2652         struct btree_iter *iter;
2653         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2654         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2655         int ret = 0;
2656
2657         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2658
2659         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2660                         POS(inode->v.i_ino, start_sector),
2661                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2662
2663         while (!ret && bkey_cmp(iter->pos, end_pos) < 0) {
2664                 s64 i_sectors_delta = 0;
2665                 struct disk_reservation disk_res = { 0 };
2666                 struct quota_res quota_res = { 0 };
2667                 struct bkey_i_reservation reservation;
2668                 struct bkey_s_c k;
2669                 unsigned sectors;
2670
2671                 bch2_trans_begin(&trans);
2672
2673                 k = bch2_btree_iter_peek_slot(iter);
2674                 if ((ret = bkey_err(k)))
2675                         goto bkey_err;
2676
2677                 /* already reserved */
2678                 if (k.k->type == KEY_TYPE_reservation &&
2679                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2680                         bch2_btree_iter_advance(iter);
2681                         continue;
2682                 }
2683
2684                 if (bkey_extent_is_data(k.k) &&
2685                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2686                         bch2_btree_iter_advance(iter);
2687                         continue;
2688                 }
2689
2690                 bkey_reservation_init(&reservation.k_i);
2691                 reservation.k.type      = KEY_TYPE_reservation;
2692                 reservation.k.p         = k.k->p;
2693                 reservation.k.size      = k.k->size;
2694
2695                 bch2_cut_front(iter->pos,       &reservation.k_i);
2696                 bch2_cut_back(end_pos,          &reservation.k_i);
2697
2698                 sectors = reservation.k.size;
2699                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2700
2701                 if (!bkey_extent_is_allocation(k.k)) {
2702                         ret = bch2_quota_reservation_add(c, inode,
2703                                         &quota_res,
2704                                         sectors, true);
2705                         if (unlikely(ret))
2706                                 goto bkey_err;
2707                 }
2708
2709                 if (reservation.v.nr_replicas < replicas ||
2710                     bch2_bkey_sectors_compressed(k)) {
2711                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2712                                                         replicas, 0);
2713                         if (unlikely(ret))
2714                                 goto bkey_err;
2715
2716                         reservation.v.nr_replicas = disk_res.nr_replicas;
2717                 }
2718
2719                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2720                                 &disk_res, &inode->ei_journal_seq,
2721                                 0, &i_sectors_delta, true);
2722                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2723 bkey_err:
2724                 bch2_quota_reservation_put(c, inode, &quota_res);
2725                 bch2_disk_reservation_put(c, &disk_res);
2726                 if (ret == -EINTR)
2727                         ret = 0;
2728         }
2729         bch2_trans_iter_put(&trans, iter);
2730         bch2_trans_exit(&trans);
2731         return ret;
2732 }
2733
2734 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2735                             loff_t offset, loff_t len)
2736 {
2737         struct address_space *mapping = inode->v.i_mapping;
2738         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2739         loff_t end              = offset + len;
2740         loff_t block_start      = round_down(offset,    block_bytes(c));
2741         loff_t block_end        = round_up(end,         block_bytes(c));
2742         int ret;
2743
2744         inode_lock(&inode->v);
2745         inode_dio_wait(&inode->v);
2746         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2747
2748         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2749                 ret = inode_newsize_ok(&inode->v, end);
2750                 if (ret)
2751                         goto err;
2752         }
2753
2754         if (mode & FALLOC_FL_ZERO_RANGE) {
2755                 ret = __bch2_truncate_page(inode,
2756                                            offset >> PAGE_SHIFT,
2757                                            offset, end);
2758
2759                 if (!ret &&
2760                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2761                         ret = __bch2_truncate_page(inode,
2762                                                    end >> PAGE_SHIFT,
2763                                                    offset, end);
2764
2765                 if (unlikely(ret))
2766                         goto err;
2767
2768                 truncate_pagecache_range(&inode->v, offset, end - 1);
2769         }
2770
2771         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
2772         if (ret)
2773                 goto err;
2774
2775         /*
2776          * Do we need to extend the file?
2777          *
2778          * If we zeroed up to the end of the file, we dropped whatever writes
2779          * were going to write out the current i_size, so we have to extend
2780          * manually even if FL_KEEP_SIZE was set:
2781          */
2782         if (end >= inode->v.i_size &&
2783             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2784              (mode & FALLOC_FL_ZERO_RANGE))) {
2785
2786                 /*
2787                  * Sync existing appends before extending i_size,
2788                  * as in bch2_extend():
2789                  */
2790                 ret = filemap_write_and_wait_range(mapping,
2791                                         inode->ei_inode.bi_size, S64_MAX);
2792                 if (ret)
2793                         goto err;
2794
2795                 if (mode & FALLOC_FL_KEEP_SIZE)
2796                         end = inode->v.i_size;
2797                 else
2798                         i_size_write(&inode->v, end);
2799
2800                 mutex_lock(&inode->ei_update_lock);
2801                 ret = bch2_write_inode_size(c, inode, end, 0);
2802                 mutex_unlock(&inode->ei_update_lock);
2803         }
2804 err:
2805         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2806         inode_unlock(&inode->v);
2807         return ret;
2808 }
2809
2810 long bch2_fallocate_dispatch(struct file *file, int mode,
2811                              loff_t offset, loff_t len)
2812 {
2813         struct bch_inode_info *inode = file_bch_inode(file);
2814         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2815         long ret;
2816
2817         if (!percpu_ref_tryget(&c->writes))
2818                 return -EROFS;
2819
2820         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2821                 ret = bchfs_fallocate(inode, mode, offset, len);
2822         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2823                 ret = bchfs_fpunch(inode, offset, len);
2824         else if (mode == FALLOC_FL_INSERT_RANGE)
2825                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2826         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2827                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2828         else
2829                 ret = -EOPNOTSUPP;
2830
2831         percpu_ref_put(&c->writes);
2832
2833         return ret;
2834 }
2835
2836 static void mark_range_unallocated(struct bch_inode_info *inode,
2837                                    loff_t start, loff_t end)
2838 {
2839         pgoff_t index = start >> PAGE_SHIFT;
2840         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2841         struct pagevec pvec;
2842
2843         pagevec_init(&pvec);
2844
2845         do {
2846                 unsigned nr_pages, i, j;
2847
2848                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2849                                                 &index, end_index);
2850                 if (nr_pages == 0)
2851                         break;
2852
2853                 for (i = 0; i < nr_pages; i++) {
2854                         struct page *page = pvec.pages[i];
2855                         struct bch_page_state *s;
2856
2857                         lock_page(page);
2858                         s = bch2_page_state(page);
2859
2860                         if (s) {
2861                                 spin_lock(&s->lock);
2862                                 for (j = 0; j < PAGE_SECTORS; j++)
2863                                         s->s[j].nr_replicas = 0;
2864                                 spin_unlock(&s->lock);
2865                         }
2866
2867                         unlock_page(page);
2868                 }
2869                 pagevec_release(&pvec);
2870         } while (index <= end_index);
2871 }
2872
2873 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2874                              struct file *file_dst, loff_t pos_dst,
2875                              loff_t len, unsigned remap_flags)
2876 {
2877         struct bch_inode_info *src = file_bch_inode(file_src);
2878         struct bch_inode_info *dst = file_bch_inode(file_dst);
2879         struct bch_fs *c = src->v.i_sb->s_fs_info;
2880         s64 i_sectors_delta = 0;
2881         u64 aligned_len;
2882         loff_t ret = 0;
2883
2884         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2885                 return -EINVAL;
2886
2887         if (remap_flags & REMAP_FILE_DEDUP)
2888                 return -EOPNOTSUPP;
2889
2890         if ((pos_src & (block_bytes(c) - 1)) ||
2891             (pos_dst & (block_bytes(c) - 1)))
2892                 return -EINVAL;
2893
2894         if (src == dst &&
2895             abs(pos_src - pos_dst) < len)
2896                 return -EINVAL;
2897
2898         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2899
2900         file_update_time(file_dst);
2901
2902         inode_dio_wait(&src->v);
2903         inode_dio_wait(&dst->v);
2904
2905         ret = generic_remap_file_range_prep(file_src, pos_src,
2906                                             file_dst, pos_dst,
2907                                             &len, remap_flags);
2908         if (ret < 0 || len == 0)
2909                 goto err;
2910
2911         aligned_len = round_up((u64) len, block_bytes(c));
2912
2913         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2914                                 pos_dst, pos_dst + len - 1);
2915         if (ret)
2916                 goto err;
2917
2918         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2919
2920         ret = bch2_remap_range(c,
2921                                POS(dst->v.i_ino, pos_dst >> 9),
2922                                POS(src->v.i_ino, pos_src >> 9),
2923                                aligned_len >> 9,
2924                                &dst->ei_journal_seq,
2925                                pos_dst + len, &i_sectors_delta);
2926         if (ret < 0)
2927                 goto err;
2928
2929         /*
2930          * due to alignment, we might have remapped slightly more than requsted
2931          */
2932         ret = min((u64) ret << 9, (u64) len);
2933
2934         /* XXX get a quota reservation */
2935         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2936
2937         spin_lock(&dst->v.i_lock);
2938         if (pos_dst + ret > dst->v.i_size)
2939                 i_size_write(&dst->v, pos_dst + ret);
2940         spin_unlock(&dst->v.i_lock);
2941
2942         if (((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
2943              IS_SYNC(file_inode(file_dst))) &&
2944             !c->opts.journal_flush_disabled)
2945                 ret = bch2_journal_flush_seq(&c->journal, dst->ei_journal_seq);
2946 err:
2947         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2948
2949         return ret;
2950 }
2951
2952 /* fseek: */
2953
2954 static int page_data_offset(struct page *page, unsigned offset)
2955 {
2956         struct bch_page_state *s = bch2_page_state(page);
2957         unsigned i;
2958
2959         if (s)
2960                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2961                         if (s->s[i].state >= SECTOR_DIRTY)
2962                                 return i << 9;
2963
2964         return -1;
2965 }
2966
2967 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2968                                        loff_t start_offset,
2969                                        loff_t end_offset)
2970 {
2971         struct address_space *mapping = vinode->i_mapping;
2972         struct page *page;
2973         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2974         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2975         pgoff_t index           = start_index;
2976         loff_t ret;
2977         int offset;
2978
2979         while (index <= end_index) {
2980                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2981                         lock_page(page);
2982
2983                         offset = page_data_offset(page,
2984                                         page->index == start_index
2985                                         ? start_offset & (PAGE_SIZE - 1)
2986                                         : 0);
2987                         if (offset >= 0) {
2988                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2989                                             offset,
2990                                             start_offset, end_offset);
2991                                 unlock_page(page);
2992                                 put_page(page);
2993                                 return ret;
2994                         }
2995
2996                         unlock_page(page);
2997                         put_page(page);
2998                 } else {
2999                         break;
3000                 }
3001         }
3002
3003         return end_offset;
3004 }
3005
3006 static loff_t bch2_seek_data(struct file *file, u64 offset)
3007 {
3008         struct bch_inode_info *inode = file_bch_inode(file);
3009         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3010         struct btree_trans trans;
3011         struct btree_iter *iter;
3012         struct bkey_s_c k;
3013         u64 isize, next_data = MAX_LFS_FILESIZE;
3014         int ret;
3015
3016         isize = i_size_read(&inode->v);
3017         if (offset >= isize)
3018                 return -ENXIO;
3019
3020         bch2_trans_init(&trans, c, 0, 0);
3021
3022         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3023                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3024                 if (k.k->p.inode != inode->v.i_ino) {
3025                         break;
3026                 } else if (bkey_extent_is_data(k.k)) {
3027                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3028                         break;
3029                 } else if (k.k->p.offset >> 9 > isize)
3030                         break;
3031         }
3032         bch2_trans_iter_put(&trans, iter);
3033
3034         ret = bch2_trans_exit(&trans) ?: ret;
3035         if (ret)
3036                 return ret;
3037
3038         if (next_data > offset)
3039                 next_data = bch2_seek_pagecache_data(&inode->v,
3040                                                      offset, next_data);
3041
3042         if (next_data >= isize)
3043                 return -ENXIO;
3044
3045         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3046 }
3047
3048 static int __page_hole_offset(struct page *page, unsigned offset)
3049 {
3050         struct bch_page_state *s = bch2_page_state(page);
3051         unsigned i;
3052
3053         if (!s)
3054                 return 0;
3055
3056         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3057                 if (s->s[i].state < SECTOR_DIRTY)
3058                         return i << 9;
3059
3060         return -1;
3061 }
3062
3063 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3064 {
3065         pgoff_t index = offset >> PAGE_SHIFT;
3066         struct page *page;
3067         int pg_offset;
3068         loff_t ret = -1;
3069
3070         page = find_lock_page(mapping, index);
3071         if (!page)
3072                 return offset;
3073
3074         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3075         if (pg_offset >= 0)
3076                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3077
3078         unlock_page(page);
3079
3080         return ret;
3081 }
3082
3083 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3084                                        loff_t start_offset,
3085                                        loff_t end_offset)
3086 {
3087         struct address_space *mapping = vinode->i_mapping;
3088         loff_t offset = start_offset, hole;
3089
3090         while (offset < end_offset) {
3091                 hole = page_hole_offset(mapping, offset);
3092                 if (hole >= 0 && hole <= end_offset)
3093                         return max(start_offset, hole);
3094
3095                 offset += PAGE_SIZE;
3096                 offset &= PAGE_MASK;
3097         }
3098
3099         return end_offset;
3100 }
3101
3102 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3103 {
3104         struct bch_inode_info *inode = file_bch_inode(file);
3105         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3106         struct btree_trans trans;
3107         struct btree_iter *iter;
3108         struct bkey_s_c k;
3109         u64 isize, next_hole = MAX_LFS_FILESIZE;
3110         int ret;
3111
3112         isize = i_size_read(&inode->v);
3113         if (offset >= isize)
3114                 return -ENXIO;
3115
3116         bch2_trans_init(&trans, c, 0, 0);
3117
3118         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3119                            POS(inode->v.i_ino, offset >> 9),
3120                            BTREE_ITER_SLOTS, k, ret) {
3121                 if (k.k->p.inode != inode->v.i_ino) {
3122                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3123                                         offset, MAX_LFS_FILESIZE);
3124                         break;
3125                 } else if (!bkey_extent_is_data(k.k)) {
3126                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3127                                         max(offset, bkey_start_offset(k.k) << 9),
3128                                         k.k->p.offset << 9);
3129
3130                         if (next_hole < k.k->p.offset << 9)
3131                                 break;
3132                 } else {
3133                         offset = max(offset, bkey_start_offset(k.k) << 9);
3134                 }
3135         }
3136         bch2_trans_iter_put(&trans, iter);
3137
3138         ret = bch2_trans_exit(&trans) ?: ret;
3139         if (ret)
3140                 return ret;
3141
3142         if (next_hole > isize)
3143                 next_hole = isize;
3144
3145         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3146 }
3147
3148 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3149 {
3150         switch (whence) {
3151         case SEEK_SET:
3152         case SEEK_CUR:
3153         case SEEK_END:
3154                 return generic_file_llseek(file, offset, whence);
3155         case SEEK_DATA:
3156                 return bch2_seek_data(file, offset);
3157         case SEEK_HOLE:
3158                 return bch2_seek_hole(file, offset);
3159         }
3160
3161         return -EINVAL;
3162 }
3163
3164 void bch2_fs_fsio_exit(struct bch_fs *c)
3165 {
3166         bioset_exit(&c->dio_write_bioset);
3167         bioset_exit(&c->dio_read_bioset);
3168         bioset_exit(&c->writepage_bioset);
3169 }
3170
3171 int bch2_fs_fsio_init(struct bch_fs *c)
3172 {
3173         int ret = 0;
3174
3175         pr_verbose_init(c->opts, "");
3176
3177         if (bioset_init(&c->writepage_bioset,
3178                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3179                         BIOSET_NEED_BVECS) ||
3180             bioset_init(&c->dio_read_bioset,
3181                         4, offsetof(struct dio_read, rbio.bio),
3182                         BIOSET_NEED_BVECS) ||
3183             bioset_init(&c->dio_write_bioset,
3184                         4, offsetof(struct dio_write, op.wbio.bio),
3185                         BIOSET_NEED_BVECS))
3186                 ret = -ENOMEM;
3187
3188         pr_verbose_init(c->opts, "ret %i", ret);
3189         return ret;
3190 }
3191
3192 #endif /* NO_BCACHEFS_FS */