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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_buf.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 static inline struct address_space *faults_disabled_mapping(void)
  39 {
  40         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  41 }
  42
  43 static inline void set_fdm_dropped_locks(void)
  44 {
  45         current->faults_disabled_mapping =
  46                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  47 }
  48
  49 static inline bool fdm_dropped_locks(void)
  50 {
  51         return ((unsigned long) current->faults_disabled_mapping) & 1;
  52 }
  53
  54 struct quota_res {
  55         u64                             sectors;
  56 };
  57
  58 struct bch_writepage_io {
  59         struct closure                  cl;
  60         struct bch_inode_info           *inode;
  61
  62         /* must be last: */
  63         struct bch_write_op             op;
  64 };
  65
  66 struct dio_write {
  67         struct completion               done;
  68         struct kiocb                    *req;
  69         struct mm_struct                *mm;
  70         unsigned                        loop:1,
  71                                         sync:1,
  72                                         free_iov:1;
  73         struct quota_res                quota_res;
  74         u64                             written;
  75
  76         struct iov_iter                 iter;
  77         struct iovec                    inline_vecs[2];
  78
  79         /* must be last: */
  80         struct bch_write_op             op;
  81 };
  82
  83 struct dio_read {
  84         struct closure                  cl;
  85         struct kiocb                    *req;
  86         long                            ret;
  87         bool                            should_dirty;
  88         struct bch_read_bio             rbio;
  89 };
  90
  91 /* pagecache_block must be held */
  92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  93                                               loff_t start, loff_t end)
  94 {
  95         int ret;
  96
  97         /*
  98          * XXX: the way this is currently implemented, we can spin if a process
  99          * is continually redirtying a specific page
 100          */
 101         do {
 102                 if (!mapping->nrpages &&
 103                     !mapping->nrexceptional)
 104                         return 0;
 105
 106                 ret = filemap_write_and_wait_range(mapping, start, end);
 107                 if (ret)
 108                         break;
 109
 110                 if (!mapping->nrpages)
 111                         return 0;
 112
 113                 ret = invalidate_inode_pages2_range(mapping,
 114                                 start >> PAGE_SHIFT,
 115                                 end >> PAGE_SHIFT);
 116         } while (ret == -EBUSY);
 117
 118         return ret;
 119 }
 120
 121 /* quotas */
 122
 123 #ifdef CONFIG_BCACHEFS_QUOTA
 124
 125 static void bch2_quota_reservation_put(struct bch_fs *c,
 126                                        struct bch_inode_info *inode,
 127                                        struct quota_res *res)
 128 {
 129         if (!res->sectors)
 130                 return;
 131
 132         mutex_lock(&inode->ei_quota_lock);
 133         BUG_ON(res->sectors > inode->ei_quota_reserved);
 134
 135         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 136                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 137         inode->ei_quota_reserved -= res->sectors;
 138         mutex_unlock(&inode->ei_quota_lock);
 139
 140         res->sectors = 0;
 141 }
 142
 143 static int bch2_quota_reservation_add(struct bch_fs *c,
 144                                       struct bch_inode_info *inode,
 145                                       struct quota_res *res,
 146                                       unsigned sectors,
 147                                       bool check_enospc)
 148 {
 149         int ret;
 150
 151         mutex_lock(&inode->ei_quota_lock);
 152         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 153                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 154         if (likely(!ret)) {
 155                 inode->ei_quota_reserved += sectors;
 156                 res->sectors += sectors;
 157         }
 158         mutex_unlock(&inode->ei_quota_lock);
 159
 160         return ret;
 161 }
 162
 163 #else
 164
 165 static void bch2_quota_reservation_put(struct bch_fs *c,
 166                                        struct bch_inode_info *inode,
 167                                        struct quota_res *res)
 168 {
 169 }
 170
 171 static int bch2_quota_reservation_add(struct bch_fs *c,
 172                                       struct bch_inode_info *inode,
 173                                       struct quota_res *res,
 174                                       unsigned sectors,
 175                                       bool check_enospc)
 176 {
 177         return 0;
 178 }
 179
 180 #endif
 181
 182 /* i_size updates: */
 183
 184 struct inode_new_size {
 185         loff_t          new_size;
 186         u64             now;
 187         unsigned        fields;
 188 };
 189
 190 static int inode_set_size(struct bch_inode_info *inode,
 191                           struct bch_inode_unpacked *bi,
 192                           void *p)
 193 {
 194         struct inode_new_size *s = p;
 195
 196         bi->bi_size = s->new_size;
 197         if (s->fields & ATTR_ATIME)
 198                 bi->bi_atime = s->now;
 199         if (s->fields & ATTR_MTIME)
 200                 bi->bi_mtime = s->now;
 201         if (s->fields & ATTR_CTIME)
 202                 bi->bi_ctime = s->now;
 203
 204         return 0;
 205 }
 206
 207 int __must_check bch2_write_inode_size(struct bch_fs *c,
 208                                        struct bch_inode_info *inode,
 209                                        loff_t new_size, unsigned fields)
 210 {
 211         struct inode_new_size s = {
 212                 .new_size       = new_size,
 213                 .now            = bch2_current_time(c),
 214                 .fields         = fields,
 215         };
 216
 217         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 218 }
 219
 220 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 221                            struct quota_res *quota_res, s64 sectors)
 222 {
 223         if (!sectors)
 224                 return;
 225
 226         mutex_lock(&inode->ei_quota_lock);
 227 #ifdef CONFIG_BCACHEFS_QUOTA
 228         if (quota_res && sectors > 0) {
 229                 BUG_ON(sectors > quota_res->sectors);
 230                 BUG_ON(sectors > inode->ei_quota_reserved);
 231
 232                 quota_res->sectors -= sectors;
 233                 inode->ei_quota_reserved -= sectors;
 234         } else {
 235                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 236         }
 237 #endif
 238         inode->v.i_blocks += sectors;
 239         mutex_unlock(&inode->ei_quota_lock);
 240 }
 241
 242 /* page state: */
 243
 244 /* stored in page->private: */
 245
 246 struct bch_page_sector {
 247         /* Uncompressed, fully allocated replicas: */
 248         unsigned                nr_replicas:3;
 249
 250         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 251         unsigned                replicas_reserved:3;
 252
 253         /* i_sectors: */
 254         enum {
 255                 SECTOR_UNALLOCATED,
 256                 SECTOR_RESERVED,
 257                 SECTOR_DIRTY,
 258                 SECTOR_ALLOCATED,
 259         }                       state:2;
 260 };
 261
 262 struct bch_page_state {
 263         spinlock_t              lock;
 264         atomic_t                write_count;
 265         struct bch_page_sector  s[PAGE_SECTORS];
 266 };
 267
 268 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 269 {
 270         return page_has_private(page)
 271                 ? (struct bch_page_state *) page_private(page)
 272                 : NULL;
 273 }
 274
 275 static inline struct bch_page_state *bch2_page_state(struct page *page)
 276 {
 277         EBUG_ON(!PageLocked(page));
 278
 279         return __bch2_page_state(page);
 280 }
 281
 282 /* for newly allocated pages: */
 283 static void __bch2_page_state_release(struct page *page)
 284 {
 285         kfree(detach_page_private(page));
 286 }
 287
 288 static void bch2_page_state_release(struct page *page)
 289 {
 290         EBUG_ON(!PageLocked(page));
 291         __bch2_page_state_release(page);
 292 }
 293
 294 /* for newly allocated pages: */
 295 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 296                                                        gfp_t gfp)
 297 {
 298         struct bch_page_state *s;
 299
 300         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 301         if (!s)
 302                 return NULL;
 303
 304         spin_lock_init(&s->lock);
 305         attach_page_private(page, s);
 306         return s;
 307 }
 308
 309 static struct bch_page_state *bch2_page_state_create(struct page *page,
 310                                                      gfp_t gfp)
 311 {
 312         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 313 }
 314
 315 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 316 {
 317         /* XXX: this should not be open coded */
 318         return inode->ei_inode.bi_data_replicas
 319                 ? inode->ei_inode.bi_data_replicas - 1
 320                 : c->opts.data_replicas;
 321 }
 322
 323 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 324                                                   unsigned nr_replicas)
 325 {
 326         return max(0, (int) nr_replicas -
 327                    s->nr_replicas -
 328                    s->replicas_reserved);
 329 }
 330
 331 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 332                                 struct bch_inode_info *inode,
 333                                 struct page *page, bool check_enospc)
 334 {
 335         struct bch_page_state *s = bch2_page_state_create(page, 0);
 336         unsigned nr_replicas = inode_nr_replicas(c, inode);
 337         struct disk_reservation disk_res = { 0 };
 338         unsigned i, disk_res_sectors = 0;
 339         int ret;
 340
 341         if (!s)
 342                 return -ENOMEM;
 343
 344         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 345                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 346
 347         if (!disk_res_sectors)
 348                 return 0;
 349
 350         ret = bch2_disk_reservation_get(c, &disk_res,
 351                                         disk_res_sectors, 1,
 352                                         !check_enospc
 353                                         ? BCH_DISK_RESERVATION_NOFAIL
 354                                         : 0);
 355         if (unlikely(ret))
 356                 return ret;
 357
 358         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 359                 s->s[i].replicas_reserved +=
 360                         sectors_to_reserve(&s->s[i], nr_replicas);
 361
 362         return 0;
 363 }
 364
 365 struct bch2_page_reservation {
 366         struct disk_reservation disk;
 367         struct quota_res        quota;
 368 };
 369
 370 static void bch2_page_reservation_init(struct bch_fs *c,
 371                         struct bch_inode_info *inode,
 372                         struct bch2_page_reservation *res)
 373 {
 374         memset(res, 0, sizeof(*res));
 375
 376         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 377 }
 378
 379 static void bch2_page_reservation_put(struct bch_fs *c,
 380                         struct bch_inode_info *inode,
 381                         struct bch2_page_reservation *res)
 382 {
 383         bch2_disk_reservation_put(c, &res->disk);
 384         bch2_quota_reservation_put(c, inode, &res->quota);
 385 }
 386
 387 static int bch2_page_reservation_get(struct bch_fs *c,
 388                         struct bch_inode_info *inode, struct page *page,
 389                         struct bch2_page_reservation *res,
 390                         unsigned offset, unsigned len, bool check_enospc)
 391 {
 392         struct bch_page_state *s = bch2_page_state_create(page, 0);
 393         unsigned i, disk_sectors = 0, quota_sectors = 0;
 394         int ret;
 395
 396         if (!s)
 397                 return -ENOMEM;
 398
 399         for (i = round_down(offset, block_bytes(c)) >> 9;
 400              i < round_up(offset + len, block_bytes(c)) >> 9;
 401              i++) {
 402                 disk_sectors += sectors_to_reserve(&s->s[i],
 403                                                 res->disk.nr_replicas);
 404                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 405         }
 406
 407         if (disk_sectors) {
 408                 ret = bch2_disk_reservation_add(c, &res->disk,
 409                                                 disk_sectors,
 410                                                 !check_enospc
 411                                                 ? BCH_DISK_RESERVATION_NOFAIL
 412                                                 : 0);
 413                 if (unlikely(ret))
 414                         return ret;
 415         }
 416
 417         if (quota_sectors) {
 418                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 419                                                  quota_sectors,
 420                                                  check_enospc);
 421                 if (unlikely(ret)) {
 422                         struct disk_reservation tmp = {
 423                                 .sectors = disk_sectors
 424                         };
 425
 426                         bch2_disk_reservation_put(c, &tmp);
 427                         res->disk.sectors -= disk_sectors;
 428                         return ret;
 429                 }
 430         }
 431
 432         return 0;
 433 }
 434
 435 static void bch2_clear_page_bits(struct page *page)
 436 {
 437         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 438         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 439         struct bch_page_state *s = bch2_page_state(page);
 440         struct disk_reservation disk_res = { 0 };
 441         int i, dirty_sectors = 0;
 442
 443         if (!s)
 444                 return;
 445
 446         EBUG_ON(!PageLocked(page));
 447         EBUG_ON(PageWriteback(page));
 448
 449         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 450                 disk_res.sectors += s->s[i].replicas_reserved;
 451                 s->s[i].replicas_reserved = 0;
 452
 453                 if (s->s[i].state == SECTOR_DIRTY) {
 454                         dirty_sectors++;
 455                         s->s[i].state = SECTOR_UNALLOCATED;
 456                 }
 457         }
 458
 459         bch2_disk_reservation_put(c, &disk_res);
 460
 461         if (dirty_sectors)
 462                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 463
 464         bch2_page_state_release(page);
 465 }
 466
 467 static void bch2_set_page_dirty(struct bch_fs *c,
 468                         struct bch_inode_info *inode, struct page *page,
 469                         struct bch2_page_reservation *res,
 470                         unsigned offset, unsigned len)
 471 {
 472         struct bch_page_state *s = bch2_page_state(page);
 473         unsigned i, dirty_sectors = 0;
 474
 475         WARN_ON((u64) page_offset(page) + offset + len >
 476                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 477
 478         spin_lock(&s->lock);
 479
 480         for (i = round_down(offset, block_bytes(c)) >> 9;
 481              i < round_up(offset + len, block_bytes(c)) >> 9;
 482              i++) {
 483                 unsigned sectors = sectors_to_reserve(&s->s[i],
 484                                                 res->disk.nr_replicas);
 485
 486                 /*
 487                  * This can happen if we race with the error path in
 488                  * bch2_writepage_io_done():
 489                  */
 490                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 491
 492                 s->s[i].replicas_reserved += sectors;
 493                 res->disk.sectors -= sectors;
 494
 495                 if (s->s[i].state == SECTOR_UNALLOCATED)
 496                         dirty_sectors++;
 497
 498                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 499         }
 500
 501         spin_unlock(&s->lock);
 502
 503         if (dirty_sectors)
 504                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 505
 506         if (!PageDirty(page))
 507                 __set_page_dirty_nobuffers(page);
 508 }
 509
 510 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 511 {
 512         struct file *file = vmf->vma->vm_file;
 513         struct address_space *mapping = file->f_mapping;
 514         struct address_space *fdm = faults_disabled_mapping();
 515         struct bch_inode_info *inode = file_bch_inode(file);
 516         int ret;
 517
 518         if (fdm == mapping)
 519                 return VM_FAULT_SIGBUS;
 520
 521         /* Lock ordering: */
 522         if (fdm > mapping) {
 523                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 524
 525                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 526                         goto got_lock;
 527
 528                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 529
 530                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 531                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 532
 533                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 534
 535                 /* Signal that lock has been dropped: */
 536                 set_fdm_dropped_locks();
 537                 return VM_FAULT_SIGBUS;
 538         }
 539
 540         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 541 got_lock:
 542         ret = filemap_fault(vmf);
 543         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 544
 545         return ret;
 546 }
 547
 548 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 549 {
 550         struct page *page = vmf->page;
 551         struct file *file = vmf->vma->vm_file;
 552         struct bch_inode_info *inode = file_bch_inode(file);
 553         struct address_space *mapping = file->f_mapping;
 554         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 555         struct bch2_page_reservation res;
 556         unsigned len;
 557         loff_t isize;
 558         int ret = VM_FAULT_LOCKED;
 559
 560         bch2_page_reservation_init(c, inode, &res);
 561
 562         sb_start_pagefault(inode->v.i_sb);
 563         file_update_time(file);
 564
 565         /*
 566          * Not strictly necessary, but helps avoid dio writes livelocking in
 567          * write_invalidate_inode_pages_range() - can drop this if/when we get
 568          * a write_invalidate_inode_pages_range() that works without dropping
 569          * page lock before invalidating page
 570          */
 571         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 572
 573         lock_page(page);
 574         isize = i_size_read(&inode->v);
 575
 576         if (page->mapping != mapping || page_offset(page) >= isize) {
 577                 unlock_page(page);
 578                 ret = VM_FAULT_NOPAGE;
 579                 goto out;
 580         }
 581
 582         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 583
 584         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 585                 unlock_page(page);
 586                 ret = VM_FAULT_SIGBUS;
 587                 goto out;
 588         }
 589
 590         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 591         bch2_page_reservation_put(c, inode, &res);
 592
 593         wait_for_stable_page(page);
 594 out:
 595         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 596         sb_end_pagefault(inode->v.i_sb);
 597
 598         return ret;
 599 }
 600
 601 void bch2_invalidatepage(struct page *page, unsigned int offset,
 602                          unsigned int length)
 603 {
 604         if (offset || length < PAGE_SIZE)
 605                 return;
 606
 607         bch2_clear_page_bits(page);
 608 }
 609
 610 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 611 {
 612         if (PageDirty(page))
 613                 return 0;
 614
 615         bch2_clear_page_bits(page);
 616         return 1;
 617 }
 618
 619 #ifdef CONFIG_MIGRATION
 620 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 621                       struct page *page, enum migrate_mode mode)
 622 {
 623         int ret;
 624
 625         EBUG_ON(!PageLocked(page));
 626         EBUG_ON(!PageLocked(newpage));
 627
 628         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 629         if (ret != MIGRATEPAGE_SUCCESS)
 630                 return ret;
 631
 632         if (PagePrivate(page))
 633                 attach_page_private(newpage, detach_page_private(page));
 634
 635         if (mode != MIGRATE_SYNC_NO_COPY)
 636                 migrate_page_copy(newpage, page);
 637         else
 638                 migrate_page_states(newpage, page);
 639         return MIGRATEPAGE_SUCCESS;
 640 }
 641 #endif
 642
 643 /* readpage(s): */
 644
 645 static void bch2_readpages_end_io(struct bio *bio)
 646 {
 647         struct bvec_iter_all iter;
 648         struct bio_vec *bv;
 649
 650         bio_for_each_segment_all(bv, bio, iter) {
 651                 struct page *page = bv->bv_page;
 652
 653                 if (!bio->bi_status) {
 654                         SetPageUptodate(page);
 655                 } else {
 656                         ClearPageUptodate(page);
 657                         SetPageError(page);
 658                 }
 659                 unlock_page(page);
 660         }
 661
 662         bio_put(bio);
 663 }
 664
 665 struct readpages_iter {
 666         struct address_space    *mapping;
 667         struct page             **pages;
 668         unsigned                nr_pages;
 669         unsigned                idx;
 670         pgoff_t                 offset;
 671 };
 672
 673 static int readpages_iter_init(struct readpages_iter *iter,
 674                                struct readahead_control *ractl)
 675 {
 676         unsigned i, nr_pages = readahead_count(ractl);
 677
 678         memset(iter, 0, sizeof(*iter));
 679
 680         iter->mapping   = ractl->mapping;
 681         iter->offset    = readahead_index(ractl);
 682         iter->nr_pages  = nr_pages;
 683
 684         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 685         if (!iter->pages)
 686                 return -ENOMEM;
 687
 688         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 689         for (i = 0; i < nr_pages; i++) {
 690                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 691                 put_page(iter->pages[i]);
 692         }
 693
 694         return 0;
 695 }
 696
 697 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 698 {
 699         if (iter->idx >= iter->nr_pages)
 700                 return NULL;
 701
 702         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 703
 704         return iter->pages[iter->idx];
 705 }
 706
 707 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 708 {
 709         struct bvec_iter iter;
 710         struct bio_vec bv;
 711         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 712                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 713         unsigned state = k.k->type == KEY_TYPE_reservation
 714                 ? SECTOR_RESERVED
 715                 : SECTOR_ALLOCATED;
 716
 717         bio_for_each_segment(bv, bio, iter) {
 718                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 719                 unsigned i;
 720
 721                 for (i = bv.bv_offset >> 9;
 722                      i < (bv.bv_offset + bv.bv_len) >> 9;
 723                      i++) {
 724                         s->s[i].nr_replicas = nr_ptrs;
 725                         s->s[i].state = state;
 726                 }
 727         }
 728 }
 729
 730 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 731 {
 732         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 733         struct bch_extent_crc_unpacked crc;
 734         const union bch_extent_entry *i;
 735
 736         bkey_for_each_crc(k.k, ptrs, crc, i)
 737                 if (crc.csum_type || crc.compression_type)
 738                         return true;
 739         return false;
 740 }
 741
 742 static void readpage_bio_extend(struct readpages_iter *iter,
 743                                 struct bio *bio,
 744                                 unsigned sectors_this_extent,
 745                                 bool get_more)
 746 {
 747         while (bio_sectors(bio) < sectors_this_extent &&
 748                bio->bi_vcnt < bio->bi_max_vecs) {
 749                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 750                 struct page *page = readpage_iter_next(iter);
 751                 int ret;
 752
 753                 if (page) {
 754                         if (iter->offset + iter->idx != page_offset)
 755                                 break;
 756
 757                         iter->idx++;
 758                 } else {
 759                         if (!get_more)
 760                                 break;
 761
 762                         page = xa_load(&iter->mapping->i_pages, page_offset);
 763                         if (page && !xa_is_value(page))
 764                                 break;
 765
 766                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 767                         if (!page)
 768                                 break;
 769
 770                         if (!__bch2_page_state_create(page, 0)) {
 771                                 put_page(page);
 772                                 break;
 773                         }
 774
 775                         ret = add_to_page_cache_lru(page, iter->mapping,
 776                                                     page_offset, GFP_NOFS);
 777                         if (ret) {
 778                                 __bch2_page_state_release(page);
 779                                 put_page(page);
 780                                 break;
 781                         }
 782
 783                         put_page(page);
 784                 }
 785
 786                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 787         }
 788 }
 789
 790 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 791                        struct bch_read_bio *rbio, u64 inum,
 792                        struct readpages_iter *readpages_iter)
 793 {
 794         struct bch_fs *c = trans->c;
 795         struct bkey_buf sk;
 796         int flags = BCH_READ_RETRY_IF_STALE|
 797                 BCH_READ_MAY_PROMOTE;
 798         int ret = 0;
 799
 800         rbio->c = c;
 801         rbio->start_time = local_clock();
 802
 803         bch2_bkey_buf_init(&sk);
 804 retry:
 805         while (1) {
 806                 struct bkey_s_c k;
 807                 unsigned bytes, sectors, offset_into_extent;
 808                 enum btree_id data_btree = BTREE_ID_extents;
 809
 810                 bch2_btree_iter_set_pos(iter,
 811                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 812
 813                 k = bch2_btree_iter_peek_slot(iter);
 814                 ret = bkey_err(k);
 815                 if (ret)
 816                         break;
 817
 818                 offset_into_extent = iter->pos.offset -
 819                         bkey_start_offset(k.k);
 820                 sectors = k.k->size - offset_into_extent;
 821
 822                 bch2_bkey_buf_reassemble(&sk, c, k);
 823
 824                 ret = bch2_read_indirect_extent(trans, &data_btree,
 825                                         &offset_into_extent, &sk);
 826                 if (ret)
 827                         break;
 828
 829                 k = bkey_i_to_s_c(sk.k);
 830
 831                 sectors = min(sectors, k.k->size - offset_into_extent);
 832
 833                 bch2_trans_unlock(trans);
 834
 835                 if (readpages_iter)
 836                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 837                                             extent_partial_reads_expensive(k));
 838
 839                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 840                 swap(rbio->bio.bi_iter.bi_size, bytes);
 841
 842                 if (rbio->bio.bi_iter.bi_size == bytes)
 843                         flags |= BCH_READ_LAST_FRAGMENT;
 844
 845                 if (bkey_extent_is_allocation(k.k))
 846                         bch2_add_page_sectors(&rbio->bio, k);
 847
 848                 bch2_read_extent(trans, rbio, iter->pos,
 849                                  data_btree, k, offset_into_extent, flags);
 850
 851                 if (flags & BCH_READ_LAST_FRAGMENT)
 852                         break;
 853
 854                 swap(rbio->bio.bi_iter.bi_size, bytes);
 855                 bio_advance(&rbio->bio, bytes);
 856         }
 857
 858         if (ret == -EINTR)
 859                 goto retry;
 860
 861         if (ret) {
 862                 bch_err_inum_ratelimited(c, inum,
 863                                 "read error %i from btree lookup", ret);
 864                 rbio->bio.bi_status = BLK_STS_IOERR;
 865                 bio_endio(&rbio->bio);
 866         }
 867
 868         bch2_bkey_buf_exit(&sk, c);
 869 }
 870
 871 void bch2_readahead(struct readahead_control *ractl)
 872 {
 873         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 874         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 875         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 876         struct btree_trans trans;
 877         struct btree_iter *iter;
 878         struct page *page;
 879         struct readpages_iter readpages_iter;
 880         int ret;
 881
 882         ret = readpages_iter_init(&readpages_iter, ractl);
 883         BUG_ON(ret);
 884
 885         bch2_trans_init(&trans, c, 0, 0);
 886         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 887                                    BTREE_ITER_SLOTS);
 888
 889         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 890
 891         while ((page = readpage_iter_next(&readpages_iter))) {
 892                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 893                 unsigned n = min_t(unsigned,
 894                                    readpages_iter.nr_pages -
 895                                    readpages_iter.idx,
 896                                    BIO_MAX_VECS);
 897                 struct bch_read_bio *rbio =
 898                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 899                                   opts);
 900
 901                 readpages_iter.idx++;
 902
 903                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 904                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 905                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 906                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 907
 908                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 909                            &readpages_iter);
 910         }
 911
 912         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 913
 914         bch2_trans_iter_put(&trans, iter);
 915         bch2_trans_exit(&trans);
 916         kfree(readpages_iter.pages);
 917 }
 918
 919 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 920                              u64 inum, struct page *page)
 921 {
 922         struct btree_trans trans;
 923         struct btree_iter *iter;
 924
 925         bch2_page_state_create(page, __GFP_NOFAIL);
 926
 927         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 928         rbio->bio.bi_iter.bi_sector =
 929                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 930         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 931
 932         bch2_trans_init(&trans, c, 0, 0);
 933         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
 934                                    BTREE_ITER_SLOTS);
 935
 936         bchfs_read(&trans, iter, rbio, inum, NULL);
 937
 938         bch2_trans_iter_put(&trans, iter);
 939         bch2_trans_exit(&trans);
 940 }
 941
 942 int bch2_readpage(struct file *file, struct page *page)
 943 {
 944         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 945         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 946         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 947         struct bch_read_bio *rbio;
 948
 949         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 950         rbio->bio.bi_end_io = bch2_readpages_end_io;
 951
 952         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 953         return 0;
 954 }
 955
 956 static void bch2_read_single_page_end_io(struct bio *bio)
 957 {
 958         complete(bio->bi_private);
 959 }
 960
 961 static int bch2_read_single_page(struct page *page,
 962                                  struct address_space *mapping)
 963 {
 964         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 965         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 966         struct bch_read_bio *rbio;
 967         int ret;
 968         DECLARE_COMPLETION_ONSTACK(done);
 969
 970         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 971                          io_opts(c, &inode->ei_inode));
 972         rbio->bio.bi_private = &done;
 973         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 974
 975         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 976         wait_for_completion(&done);
 977
 978         ret = blk_status_to_errno(rbio->bio.bi_status);
 979         bio_put(&rbio->bio);
 980
 981         if (ret < 0)
 982                 return ret;
 983
 984         SetPageUptodate(page);
 985         return 0;
 986 }
 987
 988 /* writepages: */
 989
 990 struct bch_writepage_state {
 991         struct bch_writepage_io *io;
 992         struct bch_io_opts      opts;
 993 };
 994
 995 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 996                                                                   struct bch_inode_info *inode)
 997 {
 998         return (struct bch_writepage_state) {
 999                 .opts = io_opts(c, &inode->ei_inode)
1000         };
1001 }
1002
1003 static void bch2_writepage_io_free(struct closure *cl)
1004 {
1005         struct bch_writepage_io *io = container_of(cl,
1006                                         struct bch_writepage_io, cl);
1007
1008         bio_put(&io->op.wbio.bio);
1009 }
1010
1011 static void bch2_writepage_io_done(struct closure *cl)
1012 {
1013         struct bch_writepage_io *io = container_of(cl,
1014                                         struct bch_writepage_io, cl);
1015         struct bch_fs *c = io->op.c;
1016         struct bio *bio = &io->op.wbio.bio;
1017         struct bvec_iter_all iter;
1018         struct bio_vec *bvec;
1019         unsigned i;
1020
1021         up(&io->op.c->io_in_flight);
1022
1023         if (io->op.error) {
1024                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1025
1026                 bio_for_each_segment_all(bvec, bio, iter) {
1027                         struct bch_page_state *s;
1028
1029                         SetPageError(bvec->bv_page);
1030                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1031
1032                         s = __bch2_page_state(bvec->bv_page);
1033                         spin_lock(&s->lock);
1034                         for (i = 0; i < PAGE_SECTORS; i++)
1035                                 s->s[i].nr_replicas = 0;
1036                         spin_unlock(&s->lock);
1037                 }
1038         }
1039
1040         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1041                 bio_for_each_segment_all(bvec, bio, iter) {
1042                         struct bch_page_state *s;
1043
1044                         s = __bch2_page_state(bvec->bv_page);
1045                         spin_lock(&s->lock);
1046                         for (i = 0; i < PAGE_SECTORS; i++)
1047                                 s->s[i].nr_replicas = 0;
1048                         spin_unlock(&s->lock);
1049                 }
1050         }
1051
1052         /*
1053          * racing with fallocate can cause us to add fewer sectors than
1054          * expected - but we shouldn't add more sectors than expected:
1055          */
1056         BUG_ON(io->op.i_sectors_delta > 0);
1057
1058         /*
1059          * (error (due to going RO) halfway through a page can screw that up
1060          * slightly)
1061          * XXX wtf?
1062            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1063          */
1064
1065         /*
1066          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1067          * before calling end_page_writeback:
1068          */
1069         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1070
1071         bio_for_each_segment_all(bvec, bio, iter) {
1072                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1073
1074                 if (atomic_dec_and_test(&s->write_count))
1075                         end_page_writeback(bvec->bv_page);
1076         }
1077
1078         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1079 }
1080
1081 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1082 {
1083         struct bch_writepage_io *io = w->io;
1084
1085         down(&io->op.c->io_in_flight);
1086
1087         w->io = NULL;
1088         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1089         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1090 }
1091
1092 /*
1093  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1094  * possible, else allocating a new one:
1095  */
1096 static void bch2_writepage_io_alloc(struct bch_fs *c,
1097                                     struct writeback_control *wbc,
1098                                     struct bch_writepage_state *w,
1099                                     struct bch_inode_info *inode,
1100                                     u64 sector,
1101                                     unsigned nr_replicas)
1102 {
1103         struct bch_write_op *op;
1104
1105         w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1106                                               &c->writepage_bioset),
1107                              struct bch_writepage_io, op.wbio.bio);
1108
1109         closure_init(&w->io->cl, NULL);
1110         w->io->inode            = inode;
1111
1112         op                      = &w->io->op;
1113         bch2_write_op_init(op, c, w->opts);
1114         op->target              = w->opts.foreground_target;
1115         op_journal_seq_set(op, &inode->ei_journal_seq);
1116         op->nr_replicas         = nr_replicas;
1117         op->res.nr_replicas     = nr_replicas;
1118         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1119         op->pos                 = POS(inode->v.i_ino, sector);
1120         op->wbio.bio.bi_iter.bi_sector = sector;
1121         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1122 }
1123
1124 static int __bch2_writepage(struct page *page,
1125                             struct writeback_control *wbc,
1126                             void *data)
1127 {
1128         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1129         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1130         struct bch_writepage_state *w = data;
1131         struct bch_page_state *s, orig;
1132         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1133         loff_t i_size = i_size_read(&inode->v);
1134         pgoff_t end_index = i_size >> PAGE_SHIFT;
1135         int ret;
1136
1137         EBUG_ON(!PageUptodate(page));
1138
1139         /* Is the page fully inside i_size? */
1140         if (page->index < end_index)
1141                 goto do_io;
1142
1143         /* Is the page fully outside i_size? (truncate in progress) */
1144         offset = i_size & (PAGE_SIZE - 1);
1145         if (page->index > end_index || !offset) {
1146                 unlock_page(page);
1147                 return 0;
1148         }
1149
1150         /*
1151          * The page straddles i_size.  It must be zeroed out on each and every
1152          * writepage invocation because it may be mmapped.  "A file is mapped
1153          * in multiples of the page size.  For a file that is not a multiple of
1154          * the  page size, the remaining memory is zeroed when mapped, and
1155          * writes to that region are not written out to the file."
1156          */
1157         zero_user_segment(page, offset, PAGE_SIZE);
1158 do_io:
1159         s = bch2_page_state_create(page, __GFP_NOFAIL);
1160
1161         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1162         if (ret) {
1163                 SetPageError(page);
1164                 mapping_set_error(page->mapping, ret);
1165                 unlock_page(page);
1166                 return 0;
1167         }
1168
1169         /* Before unlocking the page, get copy of reservations: */
1170         orig = *s;
1171
1172         for (i = 0; i < PAGE_SECTORS; i++) {
1173                 if (s->s[i].state < SECTOR_DIRTY)
1174                         continue;
1175
1176                 nr_replicas_this_write =
1177                         min_t(unsigned, nr_replicas_this_write,
1178                               s->s[i].nr_replicas +
1179                               s->s[i].replicas_reserved);
1180         }
1181
1182         for (i = 0; i < PAGE_SECTORS; i++) {
1183                 if (s->s[i].state < SECTOR_DIRTY)
1184                         continue;
1185
1186                 s->s[i].nr_replicas = w->opts.compression
1187                         ? 0 : nr_replicas_this_write;
1188
1189                 s->s[i].replicas_reserved = 0;
1190                 s->s[i].state = SECTOR_ALLOCATED;
1191         }
1192
1193         BUG_ON(atomic_read(&s->write_count));
1194         atomic_set(&s->write_count, 1);
1195
1196         BUG_ON(PageWriteback(page));
1197         set_page_writeback(page);
1198
1199         unlock_page(page);
1200
1201         offset = 0;
1202         while (1) {
1203                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1204                 u64 sector;
1205
1206                 while (offset < PAGE_SECTORS &&
1207                        orig.s[offset].state < SECTOR_DIRTY)
1208                         offset++;
1209
1210                 if (offset == PAGE_SECTORS)
1211                         break;
1212
1213                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1214
1215                 while (offset + sectors < PAGE_SECTORS &&
1216                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1217                         sectors++;
1218
1219                 for (i = offset; i < offset + sectors; i++) {
1220                         reserved_sectors += orig.s[i].replicas_reserved;
1221                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1222                 }
1223
1224                 if (w->io &&
1225                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1226                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1227                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1228                      (BIO_MAX_VECS * PAGE_SIZE) ||
1229                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1230                         bch2_writepage_do_io(w);
1231
1232                 if (!w->io)
1233                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1234                                                 nr_replicas_this_write);
1235
1236                 atomic_inc(&s->write_count);
1237
1238                 BUG_ON(inode != w->io->inode);
1239                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1240                                      sectors << 9, offset << 9));
1241
1242                 /* Check for writing past i_size: */
1243                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1244                         round_up(i_size, block_bytes(c)));
1245
1246                 w->io->op.res.sectors += reserved_sectors;
1247                 w->io->op.i_sectors_delta -= dirty_sectors;
1248                 w->io->op.new_i_size = i_size;
1249
1250                 offset += sectors;
1251         }
1252
1253         if (atomic_dec_and_test(&s->write_count))
1254                 end_page_writeback(page);
1255
1256         return 0;
1257 }
1258
1259 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1260 {
1261         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1262         struct bch_writepage_state w =
1263                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1264         struct blk_plug plug;
1265         int ret;
1266
1267         blk_start_plug(&plug);
1268         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1269         if (w.io)
1270                 bch2_writepage_do_io(&w);
1271         blk_finish_plug(&plug);
1272         return ret;
1273 }
1274
1275 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1276 {
1277         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1278         struct bch_writepage_state w =
1279                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1280         int ret;
1281
1282         ret = __bch2_writepage(page, wbc, &w);
1283         if (w.io)
1284                 bch2_writepage_do_io(&w);
1285
1286         return ret;
1287 }
1288
1289 /* buffered writes: */
1290
1291 int bch2_write_begin(struct file *file, struct address_space *mapping,
1292                      loff_t pos, unsigned len, unsigned flags,
1293                      struct page **pagep, void **fsdata)
1294 {
1295         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1296         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1297         struct bch2_page_reservation *res;
1298         pgoff_t index = pos >> PAGE_SHIFT;
1299         unsigned offset = pos & (PAGE_SIZE - 1);
1300         struct page *page;
1301         int ret = -ENOMEM;
1302
1303         res = kmalloc(sizeof(*res), GFP_KERNEL);
1304         if (!res)
1305                 return -ENOMEM;
1306
1307         bch2_page_reservation_init(c, inode, res);
1308         *fsdata = res;
1309
1310         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1311
1312         page = grab_cache_page_write_begin(mapping, index, flags);
1313         if (!page)
1314                 goto err_unlock;
1315
1316         if (PageUptodate(page))
1317                 goto out;
1318
1319         /* If we're writing entire page, don't need to read it in first: */
1320         if (len == PAGE_SIZE)
1321                 goto out;
1322
1323         if (!offset && pos + len >= inode->v.i_size) {
1324                 zero_user_segment(page, len, PAGE_SIZE);
1325                 flush_dcache_page(page);
1326                 goto out;
1327         }
1328
1329         if (index > inode->v.i_size >> PAGE_SHIFT) {
1330                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1331                 flush_dcache_page(page);
1332                 goto out;
1333         }
1334 readpage:
1335         ret = bch2_read_single_page(page, mapping);
1336         if (ret)
1337                 goto err;
1338 out:
1339         ret = bch2_page_reservation_get(c, inode, page, res,
1340                                         offset, len, true);
1341         if (ret) {
1342                 if (!PageUptodate(page)) {
1343                         /*
1344                          * If the page hasn't been read in, we won't know if we
1345                          * actually need a reservation - we don't actually need
1346                          * to read here, we just need to check if the page is
1347                          * fully backed by uncompressed data:
1348                          */
1349                         goto readpage;
1350                 }
1351
1352                 goto err;
1353         }
1354
1355         *pagep = page;
1356         return 0;
1357 err:
1358         unlock_page(page);
1359         put_page(page);
1360         *pagep = NULL;
1361 err_unlock:
1362         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1363         kfree(res);
1364         *fsdata = NULL;
1365         return ret;
1366 }
1367
1368 int bch2_write_end(struct file *file, struct address_space *mapping,
1369                    loff_t pos, unsigned len, unsigned copied,
1370                    struct page *page, void *fsdata)
1371 {
1372         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1373         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1374         struct bch2_page_reservation *res = fsdata;
1375         unsigned offset = pos & (PAGE_SIZE - 1);
1376
1377         lockdep_assert_held(&inode->v.i_rwsem);
1378
1379         if (unlikely(copied < len && !PageUptodate(page))) {
1380                 /*
1381                  * The page needs to be read in, but that would destroy
1382                  * our partial write - simplest thing is to just force
1383                  * userspace to redo the write:
1384                  */
1385                 zero_user(page, 0, PAGE_SIZE);
1386                 flush_dcache_page(page);
1387                 copied = 0;
1388         }
1389
1390         spin_lock(&inode->v.i_lock);
1391         if (pos + copied > inode->v.i_size)
1392                 i_size_write(&inode->v, pos + copied);
1393         spin_unlock(&inode->v.i_lock);
1394
1395         if (copied) {
1396                 if (!PageUptodate(page))
1397                         SetPageUptodate(page);
1398
1399                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1400
1401                 inode->ei_last_dirtied = (unsigned long) current;
1402         }
1403
1404         unlock_page(page);
1405         put_page(page);
1406         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1407
1408         bch2_page_reservation_put(c, inode, res);
1409         kfree(res);
1410
1411         return copied;
1412 }
1413
1414 #define WRITE_BATCH_PAGES       32
1415
1416 static int __bch2_buffered_write(struct bch_inode_info *inode,
1417                                  struct address_space *mapping,
1418                                  struct iov_iter *iter,
1419                                  loff_t pos, unsigned len)
1420 {
1421         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1422         struct page *pages[WRITE_BATCH_PAGES];
1423         struct bch2_page_reservation res;
1424         unsigned long index = pos >> PAGE_SHIFT;
1425         unsigned offset = pos & (PAGE_SIZE - 1);
1426         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1427         unsigned i, reserved = 0, set_dirty = 0;
1428         unsigned copied = 0, nr_pages_copied = 0;
1429         int ret = 0;
1430
1431         BUG_ON(!len);
1432         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1433
1434         bch2_page_reservation_init(c, inode, &res);
1435
1436         for (i = 0; i < nr_pages; i++) {
1437                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1438                 if (!pages[i]) {
1439                         nr_pages = i;
1440                         if (!i) {
1441                                 ret = -ENOMEM;
1442                                 goto out;
1443                         }
1444                         len = min_t(unsigned, len,
1445                                     nr_pages * PAGE_SIZE - offset);
1446                         break;
1447                 }
1448         }
1449
1450         if (offset && !PageUptodate(pages[0])) {
1451                 ret = bch2_read_single_page(pages[0], mapping);
1452                 if (ret)
1453                         goto out;
1454         }
1455
1456         if ((pos + len) & (PAGE_SIZE - 1) &&
1457             !PageUptodate(pages[nr_pages - 1])) {
1458                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1459                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1460                 } else {
1461                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1462                         if (ret)
1463                                 goto out;
1464                 }
1465         }
1466
1467         while (reserved < len) {
1468                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1469                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1470                 unsigned pg_len = min_t(unsigned, len - reserved,
1471                                         PAGE_SIZE - pg_offset);
1472 retry_reservation:
1473                 ret = bch2_page_reservation_get(c, inode, page, &res,
1474                                                 pg_offset, pg_len, true);
1475
1476                 if (ret && !PageUptodate(page)) {
1477                         ret = bch2_read_single_page(page, mapping);
1478                         if (!ret)
1479                                 goto retry_reservation;
1480                 }
1481
1482                 if (ret)
1483                         goto out;
1484
1485                 reserved += pg_len;
1486         }
1487
1488         if (mapping_writably_mapped(mapping))
1489                 for (i = 0; i < nr_pages; i++)
1490                         flush_dcache_page(pages[i]);
1491
1492         while (copied < len) {
1493                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1494                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1495                 unsigned pg_len = min_t(unsigned, len - copied,
1496                                         PAGE_SIZE - pg_offset);
1497                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1498                                                 iter, pg_offset, pg_len);
1499
1500                 if (!pg_copied)
1501                         break;
1502
1503                 if (!PageUptodate(page) &&
1504                     pg_copied != PAGE_SIZE &&
1505                     pos + copied + pg_copied < inode->v.i_size) {
1506                         zero_user(page, 0, PAGE_SIZE);
1507                         break;
1508                 }
1509
1510                 flush_dcache_page(page);
1511                 iov_iter_advance(iter, pg_copied);
1512                 copied += pg_copied;
1513
1514                 if (pg_copied != pg_len)
1515                         break;
1516         }
1517
1518         if (!copied)
1519                 goto out;
1520
1521         spin_lock(&inode->v.i_lock);
1522         if (pos + copied > inode->v.i_size)
1523                 i_size_write(&inode->v, pos + copied);
1524         spin_unlock(&inode->v.i_lock);
1525
1526         while (set_dirty < copied) {
1527                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1528                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1529                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1530                                         PAGE_SIZE - pg_offset);
1531
1532                 if (!PageUptodate(page))
1533                         SetPageUptodate(page);
1534
1535                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1536                 unlock_page(page);
1537                 put_page(page);
1538
1539                 set_dirty += pg_len;
1540         }
1541
1542         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1543         inode->ei_last_dirtied = (unsigned long) current;
1544 out:
1545         for (i = nr_pages_copied; i < nr_pages; i++) {
1546                 unlock_page(pages[i]);
1547                 put_page(pages[i]);
1548         }
1549
1550         bch2_page_reservation_put(c, inode, &res);
1551
1552         return copied ?: ret;
1553 }
1554
1555 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1556 {
1557         struct file *file = iocb->ki_filp;
1558         struct address_space *mapping = file->f_mapping;
1559         struct bch_inode_info *inode = file_bch_inode(file);
1560         loff_t pos = iocb->ki_pos;
1561         ssize_t written = 0;
1562         int ret = 0;
1563
1564         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1565
1566         do {
1567                 unsigned offset = pos & (PAGE_SIZE - 1);
1568                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1569                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1570 again:
1571                 /*
1572                  * Bring in the user page that we will copy from _first_.
1573                  * Otherwise there's a nasty deadlock on copying from the
1574                  * same page as we're writing to, without it being marked
1575                  * up-to-date.
1576                  *
1577                  * Not only is this an optimisation, but it is also required
1578                  * to check that the address is actually valid, when atomic
1579                  * usercopies are used, below.
1580                  */
1581                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1582                         bytes = min_t(unsigned long, iov_iter_count(iter),
1583                                       PAGE_SIZE - offset);
1584
1585                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1586                                 ret = -EFAULT;
1587                                 break;
1588                         }
1589                 }
1590
1591                 if (unlikely(fatal_signal_pending(current))) {
1592                         ret = -EINTR;
1593                         break;
1594                 }
1595
1596                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1597                 if (unlikely(ret < 0))
1598                         break;
1599
1600                 cond_resched();
1601
1602                 if (unlikely(ret == 0)) {
1603                         /*
1604                          * If we were unable to copy any data at all, we must
1605                          * fall back to a single segment length write.
1606                          *
1607                          * If we didn't fallback here, we could livelock
1608                          * because not all segments in the iov can be copied at
1609                          * once without a pagefault.
1610                          */
1611                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1612                                       iov_iter_single_seg_count(iter));
1613                         goto again;
1614                 }
1615                 pos += ret;
1616                 written += ret;
1617                 ret = 0;
1618
1619                 balance_dirty_pages_ratelimited(mapping);
1620         } while (iov_iter_count(iter));
1621
1622         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1623
1624         return written ? written : ret;
1625 }
1626
1627 /* O_DIRECT reads */
1628
1629 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1630 {
1631         if (check_dirty) {
1632                 bio_check_pages_dirty(bio);
1633         } else {
1634                 bio_release_pages(bio, false);
1635                 bio_put(bio);
1636         }
1637 }
1638
1639 static void bch2_dio_read_complete(struct closure *cl)
1640 {
1641         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1642
1643         dio->req->ki_complete(dio->req, dio->ret, 0);
1644         bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1645 }
1646
1647 static void bch2_direct_IO_read_endio(struct bio *bio)
1648 {
1649         struct dio_read *dio = bio->bi_private;
1650
1651         if (bio->bi_status)
1652                 dio->ret = blk_status_to_errno(bio->bi_status);
1653
1654         closure_put(&dio->cl);
1655 }
1656
1657 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1658 {
1659         struct dio_read *dio = bio->bi_private;
1660         bool should_dirty = dio->should_dirty;
1661
1662         bch2_direct_IO_read_endio(bio);
1663         bio_check_or_release(bio, should_dirty);
1664 }
1665
1666 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1667 {
1668         struct file *file = req->ki_filp;
1669         struct bch_inode_info *inode = file_bch_inode(file);
1670         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1671         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1672         struct dio_read *dio;
1673         struct bio *bio;
1674         loff_t offset = req->ki_pos;
1675         bool sync = is_sync_kiocb(req);
1676         size_t shorten;
1677         ssize_t ret;
1678
1679         if ((offset|iter->count) & (block_bytes(c) - 1))
1680                 return -EINVAL;
1681
1682         ret = min_t(loff_t, iter->count,
1683                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1684
1685         if (!ret)
1686                 return ret;
1687
1688         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1689         iter->count -= shorten;
1690
1691         bio = bio_alloc_bioset(GFP_KERNEL,
1692                                iov_iter_npages(iter, BIO_MAX_VECS),
1693                                &c->dio_read_bioset);
1694
1695         bio->bi_end_io = bch2_direct_IO_read_endio;
1696
1697         dio = container_of(bio, struct dio_read, rbio.bio);
1698         closure_init(&dio->cl, NULL);
1699
1700         /*
1701          * this is a _really_ horrible hack just to avoid an atomic sub at the
1702          * end:
1703          */
1704         if (!sync) {
1705                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1706                 atomic_set(&dio->cl.remaining,
1707                            CLOSURE_REMAINING_INITIALIZER -
1708                            CLOSURE_RUNNING +
1709                            CLOSURE_DESTRUCTOR);
1710         } else {
1711                 atomic_set(&dio->cl.remaining,
1712                            CLOSURE_REMAINING_INITIALIZER + 1);
1713         }
1714
1715         dio->req        = req;
1716         dio->ret        = ret;
1717         /*
1718          * This is one of the sketchier things I've encountered: we have to skip
1719          * the dirtying of requests that are internal from the kernel (i.e. from
1720          * loopback), because we'll deadlock on page_lock.
1721          */
1722         dio->should_dirty = iter_is_iovec(iter);
1723
1724         goto start;
1725         while (iter->count) {
1726                 bio = bio_alloc_bioset(GFP_KERNEL,
1727                                        iov_iter_npages(iter, BIO_MAX_VECS),
1728                                        &c->bio_read);
1729                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1730 start:
1731                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1732                 bio->bi_iter.bi_sector  = offset >> 9;
1733                 bio->bi_private         = dio;
1734
1735                 ret = bio_iov_iter_get_pages(bio, iter);
1736                 if (ret < 0) {
1737                         /* XXX: fault inject this path */
1738                         bio->bi_status = BLK_STS_RESOURCE;
1739                         bio_endio(bio);
1740                         break;
1741                 }
1742
1743                 offset += bio->bi_iter.bi_size;
1744
1745                 if (dio->should_dirty)
1746                         bio_set_pages_dirty(bio);
1747
1748                 if (iter->count)
1749                         closure_get(&dio->cl);
1750
1751                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1752         }
1753
1754         iter->count += shorten;
1755
1756         if (sync) {
1757                 closure_sync(&dio->cl);
1758                 closure_debug_destroy(&dio->cl);
1759                 ret = dio->ret;
1760                 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1761                 return ret;
1762         } else {
1763                 return -EIOCBQUEUED;
1764         }
1765 }
1766
1767 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1768 {
1769         struct file *file = iocb->ki_filp;
1770         struct bch_inode_info *inode = file_bch_inode(file);
1771         struct address_space *mapping = file->f_mapping;
1772         size_t count = iov_iter_count(iter);
1773         ssize_t ret;
1774
1775         if (!count)
1776                 return 0; /* skip atime */
1777
1778         if (iocb->ki_flags & IOCB_DIRECT) {
1779                 struct blk_plug plug;
1780
1781                 ret = filemap_write_and_wait_range(mapping,
1782                                         iocb->ki_pos,
1783                                         iocb->ki_pos + count - 1);
1784                 if (ret < 0)
1785                         return ret;
1786
1787                 file_accessed(file);
1788
1789                 blk_start_plug(&plug);
1790                 ret = bch2_direct_IO_read(iocb, iter);
1791                 blk_finish_plug(&plug);
1792
1793                 if (ret >= 0)
1794                         iocb->ki_pos += ret;
1795         } else {
1796                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1797                 ret = generic_file_read_iter(iocb, iter);
1798                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1799         }
1800
1801         return ret;
1802 }
1803
1804 /* O_DIRECT writes */
1805
1806 static void bch2_dio_write_loop_async(struct bch_write_op *);
1807
1808 static long bch2_dio_write_loop(struct dio_write *dio)
1809 {
1810         bool kthread = (current->flags & PF_KTHREAD) != 0;
1811         struct kiocb *req = dio->req;
1812         struct address_space *mapping = req->ki_filp->f_mapping;
1813         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1814         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1815         struct bio *bio = &dio->op.wbio.bio;
1816         struct bvec_iter_all iter;
1817         struct bio_vec *bv;
1818         unsigned unaligned, iter_count;
1819         bool sync = dio->sync, dropped_locks;
1820         long ret;
1821
1822         if (dio->loop)
1823                 goto loop;
1824
1825         down(&c->io_in_flight);
1826
1827         while (1) {
1828                 iter_count = dio->iter.count;
1829
1830                 if (kthread)
1831                         kthread_use_mm(dio->mm);
1832                 BUG_ON(current->faults_disabled_mapping);
1833                 current->faults_disabled_mapping = mapping;
1834
1835                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1836
1837                 dropped_locks = fdm_dropped_locks();
1838
1839                 current->faults_disabled_mapping = NULL;
1840                 if (kthread)
1841                         kthread_unuse_mm(dio->mm);
1842
1843                 /*
1844                  * If the fault handler returned an error but also signalled
1845                  * that it dropped & retook ei_pagecache_lock, we just need to
1846                  * re-shoot down the page cache and retry:
1847                  */
1848                 if (dropped_locks && ret)
1849                         ret = 0;
1850
1851                 if (unlikely(ret < 0))
1852                         goto err;
1853
1854                 if (unlikely(dropped_locks)) {
1855                         ret = write_invalidate_inode_pages_range(mapping,
1856                                         req->ki_pos,
1857                                         req->ki_pos + iter_count - 1);
1858                         if (unlikely(ret))
1859                                 goto err;
1860
1861                         if (!bio->bi_iter.bi_size)
1862                                 continue;
1863                 }
1864
1865                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1866                 bio->bi_iter.bi_size -= unaligned;
1867                 iov_iter_revert(&dio->iter, unaligned);
1868
1869                 if (!bio->bi_iter.bi_size) {
1870                         /*
1871                          * bio_iov_iter_get_pages was only able to get <
1872                          * blocksize worth of pages:
1873                          */
1874                         bio_for_each_segment_all(bv, bio, iter)
1875                                 put_page(bv->bv_page);
1876                         ret = -EFAULT;
1877                         goto err;
1878                 }
1879
1880                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1881                 dio->op.end_io          = bch2_dio_write_loop_async;
1882                 dio->op.target          = dio->op.opts.foreground_target;
1883                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1884                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1885                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1886                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1887
1888                 if ((req->ki_flags & IOCB_DSYNC) &&
1889                     !c->opts.journal_flush_disabled)
1890                         dio->op.flags |= BCH_WRITE_FLUSH;
1891                 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
1892
1893                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1894                                                 dio->op.opts.data_replicas, 0);
1895                 if (unlikely(ret) &&
1896                     !bch2_check_range_allocated(c, dio->op.pos,
1897                                 bio_sectors(bio),
1898                                 dio->op.opts.data_replicas,
1899                                 dio->op.opts.compression != 0))
1900                         goto err;
1901
1902                 task_io_account_write(bio->bi_iter.bi_size);
1903
1904                 if (!dio->sync && !dio->loop && dio->iter.count) {
1905                         struct iovec *iov = dio->inline_vecs;
1906
1907                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1908                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1909                                               GFP_KERNEL);
1910                                 if (unlikely(!iov)) {
1911                                         dio->sync = sync = true;
1912                                         goto do_io;
1913                                 }
1914
1915                                 dio->free_iov = true;
1916                         }
1917
1918                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1919                         dio->iter.iov = iov;
1920                 }
1921 do_io:
1922                 dio->loop = true;
1923                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1924
1925                 if (sync)
1926                         wait_for_completion(&dio->done);
1927                 else
1928                         return -EIOCBQUEUED;
1929 loop:
1930                 i_sectors_acct(c, inode, &dio->quota_res,
1931                                dio->op.i_sectors_delta);
1932                 req->ki_pos += (u64) dio->op.written << 9;
1933                 dio->written += dio->op.written;
1934
1935                 spin_lock(&inode->v.i_lock);
1936                 if (req->ki_pos > inode->v.i_size)
1937                         i_size_write(&inode->v, req->ki_pos);
1938                 spin_unlock(&inode->v.i_lock);
1939
1940                 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1941                         bio_for_each_segment_all(bv, bio, iter)
1942                                 put_page(bv->bv_page);
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         bio_put(bio);
1966
1967         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1968         inode_dio_end(&inode->v);
1969
1970         if (!sync) {
1971                 req->ki_complete(req, ret, 0);
1972                 ret = -EIOCBQUEUED;
1973         }
1974         return ret;
1975 }
1976
1977 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1978 {
1979         struct dio_write *dio = container_of(op, struct dio_write, op);
1980
1981         if (dio->sync)
1982                 complete(&dio->done);
1983         else
1984                 bch2_dio_write_loop(dio);
1985 }
1986
1987 static noinline
1988 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1989 {
1990         struct file *file = req->ki_filp;
1991         struct address_space *mapping = file->f_mapping;
1992         struct bch_inode_info *inode = file_bch_inode(file);
1993         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1994         struct dio_write *dio;
1995         struct bio *bio;
1996         bool locked = true, extending;
1997         ssize_t ret;
1998
1999         prefetch(&c->opts);
2000         prefetch((void *) &c->opts + 64);
2001         prefetch(&inode->ei_inode);
2002         prefetch((void *) &inode->ei_inode + 64);
2003
2004         inode_lock(&inode->v);
2005
2006         ret = generic_write_checks(req, iter);
2007         if (unlikely(ret <= 0))
2008                 goto err;
2009
2010         ret = file_remove_privs(file);
2011         if (unlikely(ret))
2012                 goto err;
2013
2014         ret = file_update_time(file);
2015         if (unlikely(ret))
2016                 goto err;
2017
2018         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2019                 goto err;
2020
2021         inode_dio_begin(&inode->v);
2022         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2023
2024         extending = req->ki_pos + iter->count > inode->v.i_size;
2025         if (!extending) {
2026                 inode_unlock(&inode->v);
2027                 locked = false;
2028         }
2029
2030         bio = bio_alloc_bioset(GFP_KERNEL,
2031                                iov_iter_is_bvec(iter)
2032                                ? 0
2033                                : iov_iter_npages(iter, BIO_MAX_VECS),
2034                                &c->dio_write_bioset);
2035         dio = container_of(bio, struct dio_write, op.wbio.bio);
2036         init_completion(&dio->done);
2037         dio->req                = req;
2038         dio->mm                 = current->mm;
2039         dio->loop               = false;
2040         dio->sync               = is_sync_kiocb(req) || extending;
2041         dio->free_iov           = false;
2042         dio->quota_res.sectors  = 0;
2043         dio->written            = 0;
2044         dio->iter               = *iter;
2045
2046         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2047                                          iter->count >> 9, true);
2048         if (unlikely(ret))
2049                 goto err_put_bio;
2050
2051         ret = write_invalidate_inode_pages_range(mapping,
2052                                         req->ki_pos,
2053                                         req->ki_pos + iter->count - 1);
2054         if (unlikely(ret))
2055                 goto err_put_bio;
2056
2057         ret = bch2_dio_write_loop(dio);
2058 err:
2059         if (locked)
2060                 inode_unlock(&inode->v);
2061         return ret;
2062 err_put_bio:
2063         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2064         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2065         bio_put(bio);
2066         inode_dio_end(&inode->v);
2067         goto err;
2068 }
2069
2070 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2071 {
2072         struct file *file = iocb->ki_filp;
2073         struct bch_inode_info *inode = file_bch_inode(file);
2074         ssize_t ret;
2075
2076         if (iocb->ki_flags & IOCB_DIRECT)
2077                 return bch2_direct_write(iocb, from);
2078
2079         /* We can write back this queue in page reclaim */
2080         current->backing_dev_info = inode_to_bdi(&inode->v);
2081         inode_lock(&inode->v);
2082
2083         ret = generic_write_checks(iocb, from);
2084         if (ret <= 0)
2085                 goto unlock;
2086
2087         ret = file_remove_privs(file);
2088         if (ret)
2089                 goto unlock;
2090
2091         ret = file_update_time(file);
2092         if (ret)
2093                 goto unlock;
2094
2095         ret = bch2_buffered_write(iocb, from);
2096         if (likely(ret > 0))
2097                 iocb->ki_pos += ret;
2098 unlock:
2099         inode_unlock(&inode->v);
2100         current->backing_dev_info = NULL;
2101
2102         if (ret > 0)
2103                 ret = generic_write_sync(iocb, ret);
2104
2105         return ret;
2106 }
2107
2108 /* fsync: */
2109
2110 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2111 {
2112         struct bch_inode_info *inode = file_bch_inode(file);
2113         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2114         int ret, ret2;
2115
2116         ret = file_write_and_wait_range(file, start, end);
2117         if (ret)
2118                 return ret;
2119
2120         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2121                 goto out;
2122
2123         ret = sync_inode_metadata(&inode->v, 1);
2124         if (ret)
2125                 return ret;
2126 out:
2127         if (!c->opts.journal_flush_disabled)
2128                 ret = bch2_journal_flush_seq(&c->journal,
2129                                              inode->ei_journal_seq);
2130         ret2 = file_check_and_advance_wb_err(file);
2131
2132         return ret ?: ret2;
2133 }
2134
2135 /* truncate: */
2136
2137 static inline int range_has_data(struct bch_fs *c,
2138                                   struct bpos start,
2139                                   struct bpos end)
2140 {
2141         struct btree_trans trans;
2142         struct btree_iter *iter;
2143         struct bkey_s_c k;
2144         int ret = 0;
2145
2146         bch2_trans_init(&trans, c, 0, 0);
2147
2148         for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2149                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2150                         break;
2151
2152                 if (bkey_extent_is_data(k.k)) {
2153                         ret = 1;
2154                         break;
2155                 }
2156         }
2157         bch2_trans_iter_put(&trans, iter);
2158
2159         return bch2_trans_exit(&trans) ?: ret;
2160 }
2161
2162 static int __bch2_truncate_page(struct bch_inode_info *inode,
2163                                 pgoff_t index, loff_t start, loff_t end)
2164 {
2165         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2166         struct address_space *mapping = inode->v.i_mapping;
2167         struct bch_page_state *s;
2168         unsigned start_offset = start & (PAGE_SIZE - 1);
2169         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2170         unsigned i;
2171         struct page *page;
2172         int ret = 0;
2173
2174         /* Page boundary? Nothing to do */
2175         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2176               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2177                 return 0;
2178
2179         /* Above i_size? */
2180         if (index << PAGE_SHIFT >= inode->v.i_size)
2181                 return 0;
2182
2183         page = find_lock_page(mapping, index);
2184         if (!page) {
2185                 /*
2186                  * XXX: we're doing two index lookups when we end up reading the
2187                  * page
2188                  */
2189                 ret = range_has_data(c,
2190                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2191                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2192                 if (ret <= 0)
2193                         return ret;
2194
2195                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2196                 if (unlikely(!page)) {
2197                         ret = -ENOMEM;
2198                         goto out;
2199                 }
2200         }
2201
2202         s = bch2_page_state_create(page, 0);
2203         if (!s) {
2204                 ret = -ENOMEM;
2205                 goto unlock;
2206         }
2207
2208         if (!PageUptodate(page)) {
2209                 ret = bch2_read_single_page(page, mapping);
2210                 if (ret)
2211                         goto unlock;
2212         }
2213
2214         if (index != start >> PAGE_SHIFT)
2215                 start_offset = 0;
2216         if (index != end >> PAGE_SHIFT)
2217                 end_offset = PAGE_SIZE;
2218
2219         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2220              i < round_down(end_offset, block_bytes(c)) >> 9;
2221              i++) {
2222                 s->s[i].nr_replicas     = 0;
2223                 s->s[i].state           = SECTOR_UNALLOCATED;
2224         }
2225
2226         zero_user_segment(page, start_offset, end_offset);
2227
2228         /*
2229          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2230          *
2231          * XXX: because we aren't currently tracking whether the page has actual
2232          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2233          */
2234         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2235         BUG_ON(ret);
2236
2237         /*
2238          * This removes any writeable userspace mappings; we need to force
2239          * .page_mkwrite to be called again before any mmapped writes, to
2240          * redirty the full page:
2241          */
2242         page_mkclean(page);
2243         __set_page_dirty_nobuffers(page);
2244 unlock:
2245         unlock_page(page);
2246         put_page(page);
2247 out:
2248         return ret;
2249 }
2250
2251 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2252 {
2253         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2254                                     from, round_up(from, PAGE_SIZE));
2255 }
2256
2257 static int bch2_extend(struct user_namespace *mnt_userns,
2258                        struct bch_inode_info *inode,
2259                        struct bch_inode_unpacked *inode_u,
2260                        struct iattr *iattr)
2261 {
2262         struct address_space *mapping = inode->v.i_mapping;
2263         int ret;
2264
2265         /*
2266          * sync appends:
2267          *
2268          * this has to be done _before_ extending i_size:
2269          */
2270         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2271         if (ret)
2272                 return ret;
2273
2274         truncate_setsize(&inode->v, iattr->ia_size);
2275
2276         return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2277 }
2278
2279 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2280                                    struct bch_inode_unpacked *bi,
2281                                    void *p)
2282 {
2283         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2284         return 0;
2285 }
2286
2287 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2288                                   struct bch_inode_unpacked *bi, void *p)
2289 {
2290         u64 *new_i_size = p;
2291
2292         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2293         bi->bi_size = *new_i_size;
2294         return 0;
2295 }
2296
2297 int bch2_truncate(struct user_namespace *mnt_userns,
2298                   struct bch_inode_info *inode, struct iattr *iattr)
2299 {
2300         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2301         struct address_space *mapping = inode->v.i_mapping;
2302         struct bch_inode_unpacked inode_u;
2303         struct btree_trans trans;
2304         struct btree_iter *iter;
2305         u64 new_i_size = iattr->ia_size;
2306         s64 i_sectors_delta = 0;
2307         int ret = 0;
2308
2309         /*
2310          * Don't update timestamps if we're not doing anything:
2311          */
2312         if (iattr->ia_size == inode->v.i_size)
2313                 return 0;
2314
2315         if (!(iattr->ia_valid & ATTR_MTIME))
2316                 ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2317         if (!(iattr->ia_valid & ATTR_CTIME))
2318                 ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2319         iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2320
2321         inode_dio_wait(&inode->v);
2322         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2323
2324         /*
2325          * fetch current on disk i_size: inode is locked, i_size can only
2326          * increase underneath us:
2327          */
2328         bch2_trans_init(&trans, c, 0, 0);
2329         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2330         ret = PTR_ERR_OR_ZERO(iter);
2331         bch2_trans_iter_put(&trans, iter);
2332         bch2_trans_exit(&trans);
2333
2334         if (ret)
2335                 goto err;
2336
2337         /*
2338          * check this before next assertion; on filesystem error our normal
2339          * invariants are a bit broken (truncate has to truncate the page cache
2340          * before the inode).
2341          */
2342         ret = bch2_journal_error(&c->journal);
2343         if (ret)
2344                 goto err;
2345
2346         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2347                 inode->v.i_size < inode_u.bi_size);
2348
2349         if (iattr->ia_size > inode->v.i_size) {
2350                 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2351                 goto err;
2352         }
2353
2354         iattr->ia_valid &= ~ATTR_SIZE;
2355
2356         ret = bch2_truncate_page(inode, iattr->ia_size);
2357         if (unlikely(ret))
2358                 goto err;
2359
2360         /*
2361          * When extending, we're going to write the new i_size to disk
2362          * immediately so we need to flush anything above the current on disk
2363          * i_size first:
2364          *
2365          * Also, when extending we need to flush the page that i_size currently
2366          * straddles - if it's mapped to userspace, we need to ensure that
2367          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2368          * again to allocate the part of the page that was extended.
2369          */
2370         if (iattr->ia_size > inode_u.bi_size)
2371                 ret = filemap_write_and_wait_range(mapping,
2372                                 inode_u.bi_size,
2373                                 iattr->ia_size - 1);
2374         else if (iattr->ia_size & (PAGE_SIZE - 1))
2375                 ret = filemap_write_and_wait_range(mapping,
2376                                 round_down(iattr->ia_size, PAGE_SIZE),
2377                                 iattr->ia_size - 1);
2378         if (ret)
2379                 goto err;
2380
2381         mutex_lock(&inode->ei_update_lock);
2382         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2383                                &new_i_size, 0);
2384         mutex_unlock(&inode->ei_update_lock);
2385
2386         if (unlikely(ret))
2387                 goto err;
2388
2389         truncate_setsize(&inode->v, iattr->ia_size);
2390
2391         ret = bch2_fpunch(c, inode->v.i_ino,
2392                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2393                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2394         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2395
2396         if (unlikely(ret))
2397                 goto err;
2398
2399         mutex_lock(&inode->ei_update_lock);
2400         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2401         mutex_unlock(&inode->ei_update_lock);
2402
2403         ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2404 err:
2405         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2406         return ret;
2407 }
2408
2409 /* fallocate: */
2410
2411 static int inode_update_times_fn(struct bch_inode_info *inode,
2412                                  struct bch_inode_unpacked *bi, void *p)
2413 {
2414         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2415
2416         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2417         return 0;
2418 }
2419
2420 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2421 {
2422         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2423         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2424         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2425         int ret = 0;
2426
2427         inode_lock(&inode->v);
2428         inode_dio_wait(&inode->v);
2429         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2430
2431         ret = __bch2_truncate_page(inode,
2432                                    offset >> PAGE_SHIFT,
2433                                    offset, offset + len);
2434         if (unlikely(ret))
2435                 goto err;
2436
2437         if (offset >> PAGE_SHIFT !=
2438             (offset + len) >> PAGE_SHIFT) {
2439                 ret = __bch2_truncate_page(inode,
2440                                            (offset + len) >> PAGE_SHIFT,
2441                                            offset, offset + len);
2442                 if (unlikely(ret))
2443                         goto err;
2444         }
2445
2446         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2447
2448         if (discard_start < discard_end) {
2449                 s64 i_sectors_delta = 0;
2450
2451                 ret = bch2_fpunch(c, inode->v.i_ino,
2452                                   discard_start, discard_end,
2453                                   &inode->ei_journal_seq,
2454                                   &i_sectors_delta);
2455                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2456         }
2457
2458         mutex_lock(&inode->ei_update_lock);
2459         ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2460                                ATTR_MTIME|ATTR_CTIME) ?: ret;
2461         mutex_unlock(&inode->ei_update_lock);
2462 err:
2463         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2464         inode_unlock(&inode->v);
2465
2466         return ret;
2467 }
2468
2469 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2470                                    loff_t offset, loff_t len,
2471                                    bool insert)
2472 {
2473         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2474         struct address_space *mapping = inode->v.i_mapping;
2475         struct bkey_buf copy;
2476         struct btree_trans trans;
2477         struct btree_iter *src, *dst, *del;
2478         loff_t shift, new_size;
2479         u64 src_start;
2480         int ret = 0;
2481
2482         if ((offset | len) & (block_bytes(c) - 1))
2483                 return -EINVAL;
2484
2485         /*
2486          * We need i_mutex to keep the page cache consistent with the extents
2487          * btree, and the btree consistent with i_size - we don't need outside
2488          * locking for the extents btree itself, because we're using linked
2489          * iterators
2490          */
2491         inode_lock(&inode->v);
2492         inode_dio_wait(&inode->v);
2493         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2494
2495         if (insert) {
2496                 ret = -EFBIG;
2497                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2498                         goto err;
2499
2500                 ret = -EINVAL;
2501                 if (offset >= inode->v.i_size)
2502                         goto err;
2503
2504                 src_start       = U64_MAX;
2505                 shift           = len;
2506         } else {
2507                 ret = -EINVAL;
2508                 if (offset + len >= inode->v.i_size)
2509                         goto err;
2510
2511                 src_start       = offset + len;
2512                 shift           = -len;
2513         }
2514
2515         new_size = inode->v.i_size + shift;
2516
2517         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2518         if (ret)
2519                 goto err;
2520
2521         if (insert) {
2522                 i_size_write(&inode->v, new_size);
2523                 mutex_lock(&inode->ei_update_lock);
2524                 ret = bch2_write_inode_size(c, inode, new_size,
2525                                             ATTR_MTIME|ATTR_CTIME);
2526                 mutex_unlock(&inode->ei_update_lock);
2527         } else {
2528                 s64 i_sectors_delta = 0;
2529
2530                 ret = bch2_fpunch(c, inode->v.i_ino,
2531                                   offset >> 9, (offset + len) >> 9,
2532                                   &inode->ei_journal_seq,
2533                                   &i_sectors_delta);
2534                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2535
2536                 if (ret)
2537                         goto err;
2538         }
2539
2540         bch2_bkey_buf_init(&copy);
2541         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2542         src = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2543                         POS(inode->v.i_ino, src_start >> 9),
2544                         BTREE_ITER_INTENT);
2545         dst = bch2_trans_copy_iter(&trans, src);
2546         del = bch2_trans_copy_iter(&trans, src);
2547
2548         while (ret == 0 || ret == -EINTR) {
2549                 struct disk_reservation disk_res =
2550                         bch2_disk_reservation_init(c, 0);
2551                 struct bkey_i delete;
2552                 struct bkey_s_c k;
2553                 struct bpos next_pos;
2554                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2555                 struct bpos atomic_end;
2556                 unsigned trigger_flags = 0;
2557
2558                 k = insert
2559                         ? bch2_btree_iter_peek_prev(src)
2560                         : bch2_btree_iter_peek(src);
2561                 if ((ret = bkey_err(k)))
2562                         continue;
2563
2564                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2565                         break;
2566
2567                 if (insert &&
2568                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2569                         break;
2570 reassemble:
2571                 bch2_bkey_buf_reassemble(&copy, c, k);
2572
2573                 if (insert &&
2574                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2575                         bch2_cut_front(move_pos, copy.k);
2576
2577                 copy.k->k.p.offset += shift >> 9;
2578                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2579
2580                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2581                 if (ret)
2582                         continue;
2583
2584                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2585                         if (insert) {
2586                                 move_pos = atomic_end;
2587                                 move_pos.offset -= shift >> 9;
2588                                 goto reassemble;
2589                         } else {
2590                                 bch2_cut_back(atomic_end, copy.k);
2591                         }
2592                 }
2593
2594                 bkey_init(&delete.k);
2595                 delete.k.p = copy.k->k.p;
2596                 delete.k.size = copy.k->k.size;
2597                 delete.k.p.offset -= shift >> 9;
2598                 bch2_btree_iter_set_pos(del, bkey_start_pos(&delete.k));
2599
2600                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2601
2602                 if (copy.k->k.size == k.k->size) {
2603                         /*
2604                          * If we're moving the entire extent, we can skip
2605                          * running triggers:
2606                          */
2607                         trigger_flags |= BTREE_TRIGGER_NORUN;
2608                 } else {
2609                         /* We might end up splitting compressed extents: */
2610                         unsigned nr_ptrs =
2611                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2612
2613                         ret = bch2_disk_reservation_get(c, &disk_res,
2614                                         copy.k->k.size, nr_ptrs,
2615                                         BCH_DISK_RESERVATION_NOFAIL);
2616                         BUG_ON(ret);
2617                 }
2618
2619                 ret =   bch2_btree_iter_traverse(del) ?:
2620                         bch2_trans_update(&trans, del, &delete, trigger_flags) ?:
2621                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2622                         bch2_trans_commit(&trans, &disk_res,
2623                                           &inode->ei_journal_seq,
2624                                           BTREE_INSERT_NOFAIL);
2625                 bch2_disk_reservation_put(c, &disk_res);
2626
2627                 if (!ret)
2628                         bch2_btree_iter_set_pos(src, next_pos);
2629         }
2630         bch2_trans_iter_put(&trans, del);
2631         bch2_trans_iter_put(&trans, dst);
2632         bch2_trans_iter_put(&trans, src);
2633         bch2_trans_exit(&trans);
2634         bch2_bkey_buf_exit(&copy, c);
2635
2636         if (ret)
2637                 goto err;
2638
2639         if (!insert) {
2640                 i_size_write(&inode->v, new_size);
2641                 mutex_lock(&inode->ei_update_lock);
2642                 ret = bch2_write_inode_size(c, inode, new_size,
2643                                             ATTR_MTIME|ATTR_CTIME);
2644                 mutex_unlock(&inode->ei_update_lock);
2645         }
2646 err:
2647         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2648         inode_unlock(&inode->v);
2649         return ret;
2650 }
2651
2652 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2653                              u64 start_sector, u64 end_sector)
2654 {
2655         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2656         struct btree_trans trans;
2657         struct btree_iter *iter;
2658         struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2659         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2660         int ret = 0;
2661
2662         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2663
2664         iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2665                         POS(inode->v.i_ino, start_sector),
2666                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2667
2668         while (!ret && bkey_cmp(iter->pos, end_pos) < 0) {
2669                 s64 i_sectors_delta = 0;
2670                 struct disk_reservation disk_res = { 0 };
2671                 struct quota_res quota_res = { 0 };
2672                 struct bkey_i_reservation reservation;
2673                 struct bkey_s_c k;
2674                 unsigned sectors;
2675
2676                 bch2_trans_begin(&trans);
2677
2678                 k = bch2_btree_iter_peek_slot(iter);
2679                 if ((ret = bkey_err(k)))
2680                         goto bkey_err;
2681
2682                 /* already reserved */
2683                 if (k.k->type == KEY_TYPE_reservation &&
2684                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2685                         bch2_btree_iter_next_slot(iter);
2686                         continue;
2687                 }
2688
2689                 if (bkey_extent_is_data(k.k) &&
2690                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2691                         bch2_btree_iter_next_slot(iter);
2692                         continue;
2693                 }
2694
2695                 bkey_reservation_init(&reservation.k_i);
2696                 reservation.k.type      = KEY_TYPE_reservation;
2697                 reservation.k.p         = k.k->p;
2698                 reservation.k.size      = k.k->size;
2699
2700                 bch2_cut_front(iter->pos,       &reservation.k_i);
2701                 bch2_cut_back(end_pos,          &reservation.k_i);
2702
2703                 sectors = reservation.k.size;
2704                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2705
2706                 if (!bkey_extent_is_allocation(k.k)) {
2707                         ret = bch2_quota_reservation_add(c, inode,
2708                                         &quota_res,
2709                                         sectors, true);
2710                         if (unlikely(ret))
2711                                 goto bkey_err;
2712                 }
2713
2714                 if (reservation.v.nr_replicas < replicas ||
2715                     bch2_bkey_sectors_compressed(k)) {
2716                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2717                                                         replicas, 0);
2718                         if (unlikely(ret))
2719                                 goto bkey_err;
2720
2721                         reservation.v.nr_replicas = disk_res.nr_replicas;
2722                 }
2723
2724                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2725                                 &disk_res, &inode->ei_journal_seq,
2726                                 0, &i_sectors_delta, true);
2727                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2728 bkey_err:
2729                 bch2_quota_reservation_put(c, inode, &quota_res);
2730                 bch2_disk_reservation_put(c, &disk_res);
2731                 if (ret == -EINTR)
2732                         ret = 0;
2733         }
2734         bch2_trans_iter_put(&trans, iter);
2735         bch2_trans_exit(&trans);
2736         return ret;
2737 }
2738
2739 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2740                             loff_t offset, loff_t len)
2741 {
2742         struct address_space *mapping = inode->v.i_mapping;
2743         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2744         loff_t end              = offset + len;
2745         loff_t block_start      = round_down(offset,    block_bytes(c));
2746         loff_t block_end        = round_up(end,         block_bytes(c));
2747         int ret;
2748
2749         inode_lock(&inode->v);
2750         inode_dio_wait(&inode->v);
2751         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2752
2753         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2754                 ret = inode_newsize_ok(&inode->v, end);
2755                 if (ret)
2756                         goto err;
2757         }
2758
2759         if (mode & FALLOC_FL_ZERO_RANGE) {
2760                 ret = __bch2_truncate_page(inode,
2761                                            offset >> PAGE_SHIFT,
2762                                            offset, end);
2763
2764                 if (!ret &&
2765                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2766                         ret = __bch2_truncate_page(inode,
2767                                                    end >> PAGE_SHIFT,
2768                                                    offset, end);
2769
2770                 if (unlikely(ret))
2771                         goto err;
2772
2773                 truncate_pagecache_range(&inode->v, offset, end - 1);
2774         }
2775
2776         ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
2777         if (ret)
2778                 goto err;
2779
2780         /*
2781          * Do we need to extend the file?
2782          *
2783          * If we zeroed up to the end of the file, we dropped whatever writes
2784          * were going to write out the current i_size, so we have to extend
2785          * manually even if FL_KEEP_SIZE was set:
2786          */
2787         if (end >= inode->v.i_size &&
2788             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2789              (mode & FALLOC_FL_ZERO_RANGE))) {
2790
2791                 /*
2792                  * Sync existing appends before extending i_size,
2793                  * as in bch2_extend():
2794                  */
2795                 ret = filemap_write_and_wait_range(mapping,
2796                                         inode->ei_inode.bi_size, S64_MAX);
2797                 if (ret)
2798                         goto err;
2799
2800                 if (mode & FALLOC_FL_KEEP_SIZE)
2801                         end = inode->v.i_size;
2802                 else
2803                         i_size_write(&inode->v, end);
2804
2805                 mutex_lock(&inode->ei_update_lock);
2806                 ret = bch2_write_inode_size(c, inode, end, 0);
2807                 mutex_unlock(&inode->ei_update_lock);
2808         }
2809 err:
2810         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2811         inode_unlock(&inode->v);
2812         return ret;
2813 }
2814
2815 long bch2_fallocate_dispatch(struct file *file, int mode,
2816                              loff_t offset, loff_t len)
2817 {
2818         struct bch_inode_info *inode = file_bch_inode(file);
2819         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2820         long ret;
2821
2822         if (!percpu_ref_tryget(&c->writes))
2823                 return -EROFS;
2824
2825         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2826                 ret = bchfs_fallocate(inode, mode, offset, len);
2827         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2828                 ret = bchfs_fpunch(inode, offset, len);
2829         else if (mode == FALLOC_FL_INSERT_RANGE)
2830                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2831         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2832                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2833         else
2834                 ret = -EOPNOTSUPP;
2835
2836         percpu_ref_put(&c->writes);
2837
2838         return ret;
2839 }
2840
2841 static void mark_range_unallocated(struct bch_inode_info *inode,
2842                                    loff_t start, loff_t end)
2843 {
2844         pgoff_t index = start >> PAGE_SHIFT;
2845         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2846         struct pagevec pvec;
2847
2848         pagevec_init(&pvec);
2849
2850         do {
2851                 unsigned nr_pages, i, j;
2852
2853                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2854                                                 &index, end_index);
2855                 if (nr_pages == 0)
2856                         break;
2857
2858                 for (i = 0; i < nr_pages; i++) {
2859                         struct page *page = pvec.pages[i];
2860                         struct bch_page_state *s;
2861
2862                         lock_page(page);
2863                         s = bch2_page_state(page);
2864
2865                         if (s) {
2866                                 spin_lock(&s->lock);
2867                                 for (j = 0; j < PAGE_SECTORS; j++)
2868                                         s->s[j].nr_replicas = 0;
2869                                 spin_unlock(&s->lock);
2870                         }
2871
2872                         unlock_page(page);
2873                 }
2874                 pagevec_release(&pvec);
2875         } while (index <= end_index);
2876 }
2877
2878 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2879                              struct file *file_dst, loff_t pos_dst,
2880                              loff_t len, unsigned remap_flags)
2881 {
2882         struct bch_inode_info *src = file_bch_inode(file_src);
2883         struct bch_inode_info *dst = file_bch_inode(file_dst);
2884         struct bch_fs *c = src->v.i_sb->s_fs_info;
2885         s64 i_sectors_delta = 0;
2886         u64 aligned_len;
2887         loff_t ret = 0;
2888
2889         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2890                 return -EINVAL;
2891
2892         if (remap_flags & REMAP_FILE_DEDUP)
2893                 return -EOPNOTSUPP;
2894
2895         if ((pos_src & (block_bytes(c) - 1)) ||
2896             (pos_dst & (block_bytes(c) - 1)))
2897                 return -EINVAL;
2898
2899         if (src == dst &&
2900             abs(pos_src - pos_dst) < len)
2901                 return -EINVAL;
2902
2903         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2904
2905         file_update_time(file_dst);
2906
2907         inode_dio_wait(&src->v);
2908         inode_dio_wait(&dst->v);
2909
2910         ret = generic_remap_file_range_prep(file_src, pos_src,
2911                                             file_dst, pos_dst,
2912                                             &len, remap_flags);
2913         if (ret < 0 || len == 0)
2914                 goto err;
2915
2916         aligned_len = round_up((u64) len, block_bytes(c));
2917
2918         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2919                                 pos_dst, pos_dst + len - 1);
2920         if (ret)
2921                 goto err;
2922
2923         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2924
2925         ret = bch2_remap_range(c,
2926                                POS(dst->v.i_ino, pos_dst >> 9),
2927                                POS(src->v.i_ino, pos_src >> 9),
2928                                aligned_len >> 9,
2929                                &dst->ei_journal_seq,
2930                                pos_dst + len, &i_sectors_delta);
2931         if (ret < 0)
2932                 goto err;
2933
2934         /*
2935          * due to alignment, we might have remapped slightly more than requsted
2936          */
2937         ret = min((u64) ret << 9, (u64) len);
2938
2939         /* XXX get a quota reservation */
2940         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2941
2942         spin_lock(&dst->v.i_lock);
2943         if (pos_dst + ret > dst->v.i_size)
2944                 i_size_write(&dst->v, pos_dst + ret);
2945         spin_unlock(&dst->v.i_lock);
2946
2947         if (((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
2948              IS_SYNC(file_inode(file_dst))) &&
2949             !c->opts.journal_flush_disabled)
2950                 ret = bch2_journal_flush_seq(&c->journal, dst->ei_journal_seq);
2951 err:
2952         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2953
2954         return ret;
2955 }
2956
2957 /* fseek: */
2958
2959 static int page_data_offset(struct page *page, unsigned offset)
2960 {
2961         struct bch_page_state *s = bch2_page_state(page);
2962         unsigned i;
2963
2964         if (s)
2965                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2966                         if (s->s[i].state >= SECTOR_DIRTY)
2967                                 return i << 9;
2968
2969         return -1;
2970 }
2971
2972 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2973                                        loff_t start_offset,
2974                                        loff_t end_offset)
2975 {
2976         struct address_space *mapping = vinode->i_mapping;
2977         struct page *page;
2978         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2979         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2980         pgoff_t index           = start_index;
2981         loff_t ret;
2982         int offset;
2983
2984         while (index <= end_index) {
2985                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2986                         lock_page(page);
2987
2988                         offset = page_data_offset(page,
2989                                         page->index == start_index
2990                                         ? start_offset & (PAGE_SIZE - 1)
2991                                         : 0);
2992                         if (offset >= 0) {
2993                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2994                                             offset,
2995                                             start_offset, end_offset);
2996                                 unlock_page(page);
2997                                 put_page(page);
2998                                 return ret;
2999                         }
3000
3001                         unlock_page(page);
3002                         put_page(page);
3003                 } else {
3004                         break;
3005                 }
3006         }
3007
3008         return end_offset;
3009 }
3010
3011 static loff_t bch2_seek_data(struct file *file, u64 offset)
3012 {
3013         struct bch_inode_info *inode = file_bch_inode(file);
3014         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3015         struct btree_trans trans;
3016         struct btree_iter *iter;
3017         struct bkey_s_c k;
3018         u64 isize, next_data = MAX_LFS_FILESIZE;
3019         int ret;
3020
3021         isize = i_size_read(&inode->v);
3022         if (offset >= isize)
3023                 return -ENXIO;
3024
3025         bch2_trans_init(&trans, c, 0, 0);
3026
3027         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3028                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3029                 if (k.k->p.inode != inode->v.i_ino) {
3030                         break;
3031                 } else if (bkey_extent_is_data(k.k)) {
3032                         next_data = max(offset, bkey_start_offset(k.k) << 9);
3033                         break;
3034                 } else if (k.k->p.offset >> 9 > isize)
3035                         break;
3036         }
3037         bch2_trans_iter_put(&trans, iter);
3038
3039         ret = bch2_trans_exit(&trans) ?: ret;
3040         if (ret)
3041                 return ret;
3042
3043         if (next_data > offset)
3044                 next_data = bch2_seek_pagecache_data(&inode->v,
3045                                                      offset, next_data);
3046
3047         if (next_data >= isize)
3048                 return -ENXIO;
3049
3050         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3051 }
3052
3053 static int __page_hole_offset(struct page *page, unsigned offset)
3054 {
3055         struct bch_page_state *s = bch2_page_state(page);
3056         unsigned i;
3057
3058         if (!s)
3059                 return 0;
3060
3061         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3062                 if (s->s[i].state < SECTOR_DIRTY)
3063                         return i << 9;
3064
3065         return -1;
3066 }
3067
3068 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3069 {
3070         pgoff_t index = offset >> PAGE_SHIFT;
3071         struct page *page;
3072         int pg_offset;
3073         loff_t ret = -1;
3074
3075         page = find_lock_page(mapping, index);
3076         if (!page)
3077                 return offset;
3078
3079         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3080         if (pg_offset >= 0)
3081                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3082
3083         unlock_page(page);
3084
3085         return ret;
3086 }
3087
3088 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3089                                        loff_t start_offset,
3090                                        loff_t end_offset)
3091 {
3092         struct address_space *mapping = vinode->i_mapping;
3093         loff_t offset = start_offset, hole;
3094
3095         while (offset < end_offset) {
3096                 hole = page_hole_offset(mapping, offset);
3097                 if (hole >= 0 && hole <= end_offset)
3098                         return max(start_offset, hole);
3099
3100                 offset += PAGE_SIZE;
3101                 offset &= PAGE_MASK;
3102         }
3103
3104         return end_offset;
3105 }
3106
3107 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3108 {
3109         struct bch_inode_info *inode = file_bch_inode(file);
3110         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3111         struct btree_trans trans;
3112         struct btree_iter *iter;
3113         struct bkey_s_c k;
3114         u64 isize, next_hole = MAX_LFS_FILESIZE;
3115         int ret;
3116
3117         isize = i_size_read(&inode->v);
3118         if (offset >= isize)
3119                 return -ENXIO;
3120
3121         bch2_trans_init(&trans, c, 0, 0);
3122
3123         for_each_btree_key(&trans, iter, BTREE_ID_extents,
3124                            POS(inode->v.i_ino, offset >> 9),
3125                            BTREE_ITER_SLOTS, k, ret) {
3126                 if (k.k->p.inode != inode->v.i_ino) {
3127                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3128                                         offset, MAX_LFS_FILESIZE);
3129                         break;
3130                 } else if (!bkey_extent_is_data(k.k)) {
3131                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3132                                         max(offset, bkey_start_offset(k.k) << 9),
3133                                         k.k->p.offset << 9);
3134
3135                         if (next_hole < k.k->p.offset << 9)
3136                                 break;
3137                 } else {
3138                         offset = max(offset, bkey_start_offset(k.k) << 9);
3139                 }
3140         }
3141         bch2_trans_iter_put(&trans, iter);
3142
3143         ret = bch2_trans_exit(&trans) ?: ret;
3144         if (ret)
3145                 return ret;
3146
3147         if (next_hole > isize)
3148                 next_hole = isize;
3149
3150         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3151 }
3152
3153 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3154 {
3155         switch (whence) {
3156         case SEEK_SET:
3157         case SEEK_CUR:
3158         case SEEK_END:
3159                 return generic_file_llseek(file, offset, whence);
3160         case SEEK_DATA:
3161                 return bch2_seek_data(file, offset);
3162         case SEEK_HOLE:
3163                 return bch2_seek_hole(file, offset);
3164         }
3165
3166         return -EINVAL;
3167 }
3168
3169 void bch2_fs_fsio_exit(struct bch_fs *c)
3170 {
3171         bioset_exit(&c->dio_write_bioset);
3172         bioset_exit(&c->dio_read_bioset);
3173         bioset_exit(&c->writepage_bioset);
3174 }
3175
3176 int bch2_fs_fsio_init(struct bch_fs *c)
3177 {
3178         int ret = 0;
3179
3180         pr_verbose_init(c->opts, "");
3181
3182         if (bioset_init(&c->writepage_bioset,
3183                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3184                         BIOSET_NEED_BVECS) ||
3185             bioset_init(&c->dio_read_bioset,
3186                         4, offsetof(struct dio_read, rbio.bio),
3187                         BIOSET_NEED_BVECS) ||
3188             bioset_init(&c->dio_write_bioset,
3189                         4, offsetof(struct dio_write, op.wbio.bio),
3190                         BIOSET_NEED_BVECS))
3191                 ret = -ENOMEM;
3192
3193         pr_verbose_init(c->opts, "ret %i", ret);
3194         return ret;
3195 }
3196
3197 #endif /* NO_BCACHEFS_FS */