fs/bcachefs/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 #include "trace.h"
  23
  24 #include <linux/aio.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/falloc.h>
  27 #include <linux/migrate.h>
  28 #include <linux/mmu_context.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/rmap.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/task_io_accounting_ops.h>
  33 #include <linux/uio.h>
  34 #include <linux/writeback.h>
  35
  36 #include <trace/events/writeback.h>
  37
  38 static inline bool bio_full(struct bio *bio, unsigned len)
  39 {
  40         if (bio->bi_vcnt >= bio->bi_max_vecs)
  41                 return true;
  42         if (bio->bi_iter.bi_size > UINT_MAX - len)
  43                 return true;
  44         return false;
  45 }
  46
  47 static inline struct address_space *faults_disabled_mapping(void)
  48 {
  49         return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
  50 }
  51
  52 static inline void set_fdm_dropped_locks(void)
  53 {
  54         current->faults_disabled_mapping =
  55                 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
  56 }
  57
  58 static inline bool fdm_dropped_locks(void)
  59 {
  60         return ((unsigned long) current->faults_disabled_mapping) & 1;
  61 }
  62
  63 struct quota_res {
  64         u64                             sectors;
  65 };
  66
  67 struct bch_writepage_io {
  68         struct closure                  cl;
  69         struct bch_inode_info           *inode;
  70
  71         /* must be last: */
  72         struct bch_write_op             op;
  73 };
  74
  75 struct dio_write {
  76         struct completion               done;
  77         struct kiocb                    *req;
  78         struct mm_struct                *mm;
  79         unsigned                        loop:1,
  80                                         sync:1,
  81                                         free_iov:1;
  82         struct quota_res                quota_res;
  83         u64                             written;
  84
  85         struct iov_iter                 iter;
  86         struct iovec                    inline_vecs[2];
  87
  88         /* must be last: */
  89         struct bch_write_op             op;
  90 };
  91
  92 struct dio_read {
  93         struct closure                  cl;
  94         struct kiocb                    *req;
  95         long                            ret;
  96         struct bch_read_bio             rbio;
  97 };
  98
  99 /* pagecache_block must be held */
 100 static noinline int write_invalidate_inode_pages_range(struct address_space *mapping,
 101                                               loff_t start, loff_t end)
 102 {
 103         int ret;
 104
 105         /*
 106          * XXX: the way this is currently implemented, we can spin if a process
 107          * is continually redirtying a specific page
 108          */
 109         do {
 110                 if (!mapping->nrpages)
 111                         return 0;
 112
 113                 ret = filemap_write_and_wait_range(mapping, start, end);
 114                 if (ret)
 115                         break;
 116
 117                 if (!mapping->nrpages)
 118                         return 0;
 119
 120                 ret = invalidate_inode_pages2_range(mapping,
 121                                 start >> PAGE_SHIFT,
 122                                 end >> PAGE_SHIFT);
 123         } while (ret == -EBUSY);
 124
 125         return ret;
 126 }
 127
 128 /* quotas */
 129
 130 #ifdef CONFIG_BCACHEFS_QUOTA
 131
 132 static void bch2_quota_reservation_put(struct bch_fs *c,
 133                                        struct bch_inode_info *inode,
 134                                        struct quota_res *res)
 135 {
 136         if (!res->sectors)
 137                 return;
 138
 139         mutex_lock(&inode->ei_quota_lock);
 140         BUG_ON(res->sectors > inode->ei_quota_reserved);
 141
 142         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 143                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 144         inode->ei_quota_reserved -= res->sectors;
 145         mutex_unlock(&inode->ei_quota_lock);
 146
 147         res->sectors = 0;
 148 }
 149
 150 static int bch2_quota_reservation_add(struct bch_fs *c,
 151                                       struct bch_inode_info *inode,
 152                                       struct quota_res *res,
 153                                       unsigned sectors,
 154                                       bool check_enospc)
 155 {
 156         int ret;
 157
 158         mutex_lock(&inode->ei_quota_lock);
 159         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 160                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 161         if (likely(!ret)) {
 162                 inode->ei_quota_reserved += sectors;
 163                 res->sectors += sectors;
 164         }
 165         mutex_unlock(&inode->ei_quota_lock);
 166
 167         return ret;
 168 }
 169
 170 #else
 171
 172 static void bch2_quota_reservation_put(struct bch_fs *c,
 173                                        struct bch_inode_info *inode,
 174                                        struct quota_res *res)
 175 {
 176 }
 177
 178 static int bch2_quota_reservation_add(struct bch_fs *c,
 179                                       struct bch_inode_info *inode,
 180                                       struct quota_res *res,
 181                                       unsigned sectors,
 182                                       bool check_enospc)
 183 {
 184         return 0;
 185 }
 186
 187 #endif
 188
 189 /* i_size updates: */
 190
 191 struct inode_new_size {
 192         loff_t          new_size;
 193         u64             now;
 194         unsigned        fields;
 195 };
 196
 197 static int inode_set_size(struct bch_inode_info *inode,
 198                           struct bch_inode_unpacked *bi,
 199                           void *p)
 200 {
 201         struct inode_new_size *s = p;
 202
 203         bi->bi_size = s->new_size;
 204         if (s->fields & ATTR_ATIME)
 205                 bi->bi_atime = s->now;
 206         if (s->fields & ATTR_MTIME)
 207                 bi->bi_mtime = s->now;
 208         if (s->fields & ATTR_CTIME)
 209                 bi->bi_ctime = s->now;
 210
 211         return 0;
 212 }
 213
 214 int __must_check bch2_write_inode_size(struct bch_fs *c,
 215                                        struct bch_inode_info *inode,
 216                                        loff_t new_size, unsigned fields)
 217 {
 218         struct inode_new_size s = {
 219                 .new_size       = new_size,
 220                 .now            = bch2_current_time(c),
 221                 .fields         = fields,
 222         };
 223
 224         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 225 }
 226
 227 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 228                            struct quota_res *quota_res, s64 sectors)
 229 {
 230         if (!sectors)
 231                 return;
 232
 233         mutex_lock(&inode->ei_quota_lock);
 234 #ifdef CONFIG_BCACHEFS_QUOTA
 235         if (quota_res && sectors > 0) {
 236                 BUG_ON(sectors > quota_res->sectors);
 237                 BUG_ON(sectors > inode->ei_quota_reserved);
 238
 239                 quota_res->sectors -= sectors;
 240                 inode->ei_quota_reserved -= sectors;
 241         } else {
 242                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 243         }
 244 #endif
 245         inode->v.i_blocks += sectors;
 246         mutex_unlock(&inode->ei_quota_lock);
 247 }
 248
 249 /* page state: */
 250
 251 /* stored in page->private: */
 252
 253 struct bch_page_sector {
 254         /* Uncompressed, fully allocated replicas: */
 255         unsigned                nr_replicas:3;
 256
 257         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 258         unsigned                replicas_reserved:3;
 259
 260         /* i_sectors: */
 261         enum {
 262                 SECTOR_UNALLOCATED,
 263                 SECTOR_RESERVED,
 264                 SECTOR_DIRTY,
 265                 SECTOR_ALLOCATED,
 266         }                       state:2;
 267 };
 268
 269 struct bch_page_state {
 270         spinlock_t              lock;
 271         atomic_t                write_count;
 272         struct bch_page_sector  s[PAGE_SECTORS];
 273 };
 274
 275 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 276 {
 277         return page_has_private(page)
 278                 ? (struct bch_page_state *) page_private(page)
 279                 : NULL;
 280 }
 281
 282 static inline struct bch_page_state *bch2_page_state(struct page *page)
 283 {
 284         EBUG_ON(!PageLocked(page));
 285
 286         return __bch2_page_state(page);
 287 }
 288
 289 /* for newly allocated pages: */
 290 static void __bch2_page_state_release(struct page *page)
 291 {
 292         kfree(detach_page_private(page));
 293 }
 294
 295 static void bch2_page_state_release(struct page *page)
 296 {
 297         EBUG_ON(!PageLocked(page));
 298         __bch2_page_state_release(page);
 299 }
 300
 301 /* for newly allocated pages: */
 302 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 303                                                        gfp_t gfp)
 304 {
 305         struct bch_page_state *s;
 306
 307         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 308         if (!s)
 309                 return NULL;
 310
 311         spin_lock_init(&s->lock);
 312         attach_page_private(page, s);
 313         return s;
 314 }
 315
 316 static struct bch_page_state *bch2_page_state_create(struct page *page,
 317                                                      gfp_t gfp)
 318 {
 319         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 320 }
 321
 322 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 323 {
 324         /* XXX: this should not be open coded */
 325         return inode->ei_inode.bi_data_replicas
 326                 ? inode->ei_inode.bi_data_replicas - 1
 327                 : c->opts.data_replicas;
 328 }
 329
 330 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 331                                                   unsigned nr_replicas)
 332 {
 333         return max(0, (int) nr_replicas -
 334                    s->nr_replicas -
 335                    s->replicas_reserved);
 336 }
 337
 338 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 339                                 struct bch_inode_info *inode,
 340                                 struct page *page, bool check_enospc)
 341 {
 342         struct bch_page_state *s = bch2_page_state_create(page, 0);
 343         unsigned nr_replicas = inode_nr_replicas(c, inode);
 344         struct disk_reservation disk_res = { 0 };
 345         unsigned i, disk_res_sectors = 0;
 346         int ret;
 347
 348         if (!s)
 349                 return -ENOMEM;
 350
 351         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 352                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 353
 354         if (!disk_res_sectors)
 355                 return 0;
 356
 357         ret = bch2_disk_reservation_get(c, &disk_res,
 358                                         disk_res_sectors, 1,
 359                                         !check_enospc
 360                                         ? BCH_DISK_RESERVATION_NOFAIL
 361                                         : 0);
 362         if (unlikely(ret))
 363                 return ret;
 364
 365         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 366                 s->s[i].replicas_reserved +=
 367                         sectors_to_reserve(&s->s[i], nr_replicas);
 368
 369         return 0;
 370 }
 371
 372 struct bch2_page_reservation {
 373         struct disk_reservation disk;
 374         struct quota_res        quota;
 375 };
 376
 377 static void bch2_page_reservation_init(struct bch_fs *c,
 378                         struct bch_inode_info *inode,
 379                         struct bch2_page_reservation *res)
 380 {
 381         memset(res, 0, sizeof(*res));
 382
 383         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 384 }
 385
 386 static void bch2_page_reservation_put(struct bch_fs *c,
 387                         struct bch_inode_info *inode,
 388                         struct bch2_page_reservation *res)
 389 {
 390         bch2_disk_reservation_put(c, &res->disk);
 391         bch2_quota_reservation_put(c, inode, &res->quota);
 392 }
 393
 394 static int bch2_page_reservation_get(struct bch_fs *c,
 395                         struct bch_inode_info *inode, struct page *page,
 396                         struct bch2_page_reservation *res,
 397                         unsigned offset, unsigned len, bool check_enospc)
 398 {
 399         struct bch_page_state *s = bch2_page_state_create(page, 0);
 400         unsigned i, disk_sectors = 0, quota_sectors = 0;
 401         int ret;
 402
 403         if (!s)
 404                 return -ENOMEM;
 405
 406         for (i = round_down(offset, block_bytes(c)) >> 9;
 407              i < round_up(offset + len, block_bytes(c)) >> 9;
 408              i++) {
 409                 disk_sectors += sectors_to_reserve(&s->s[i],
 410                                                 res->disk.nr_replicas);
 411                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 412         }
 413
 414         if (disk_sectors) {
 415                 ret = bch2_disk_reservation_add(c, &res->disk,
 416                                                 disk_sectors,
 417                                                 !check_enospc
 418                                                 ? BCH_DISK_RESERVATION_NOFAIL
 419                                                 : 0);
 420                 if (unlikely(ret))
 421                         return ret;
 422         }
 423
 424         if (quota_sectors) {
 425                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 426                                                  quota_sectors,
 427                                                  check_enospc);
 428                 if (unlikely(ret)) {
 429                         struct disk_reservation tmp = {
 430                                 .sectors = disk_sectors
 431                         };
 432
 433                         bch2_disk_reservation_put(c, &tmp);
 434                         res->disk.sectors -= disk_sectors;
 435                         return ret;
 436                 }
 437         }
 438
 439         return 0;
 440 }
 441
 442 static void bch2_clear_page_bits(struct page *page)
 443 {
 444         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 445         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 446         struct bch_page_state *s = bch2_page_state(page);
 447         struct disk_reservation disk_res = { 0 };
 448         int i, dirty_sectors = 0;
 449
 450         if (!s)
 451                 return;
 452
 453         EBUG_ON(!PageLocked(page));
 454         EBUG_ON(PageWriteback(page));
 455
 456         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 457                 disk_res.sectors += s->s[i].replicas_reserved;
 458                 s->s[i].replicas_reserved = 0;
 459
 460                 if (s->s[i].state == SECTOR_DIRTY) {
 461                         dirty_sectors++;
 462                         s->s[i].state = SECTOR_UNALLOCATED;
 463                 }
 464         }
 465
 466         bch2_disk_reservation_put(c, &disk_res);
 467
 468         if (dirty_sectors)
 469                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 470
 471         bch2_page_state_release(page);
 472 }
 473
 474 static void bch2_set_page_dirty(struct bch_fs *c,
 475                         struct bch_inode_info *inode, struct page *page,
 476                         struct bch2_page_reservation *res,
 477                         unsigned offset, unsigned len)
 478 {
 479         struct bch_page_state *s = bch2_page_state(page);
 480         unsigned i, dirty_sectors = 0;
 481
 482         WARN_ON((u64) page_offset(page) + offset + len >
 483                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 484
 485         spin_lock(&s->lock);
 486
 487         for (i = round_down(offset, block_bytes(c)) >> 9;
 488              i < round_up(offset + len, block_bytes(c)) >> 9;
 489              i++) {
 490                 unsigned sectors = sectors_to_reserve(&s->s[i],
 491                                                 res->disk.nr_replicas);
 492
 493                 /*
 494                  * This can happen if we race with the error path in
 495                  * bch2_writepage_io_done():
 496                  */
 497                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 498
 499                 s->s[i].replicas_reserved += sectors;
 500                 res->disk.sectors -= sectors;
 501
 502                 if (s->s[i].state == SECTOR_UNALLOCATED)
 503                         dirty_sectors++;
 504
 505                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 506         }
 507
 508         spin_unlock(&s->lock);
 509
 510         if (dirty_sectors)
 511                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 512
 513         if (!PageDirty(page))
 514                 filemap_dirty_folio(inode->v.i_mapping, page_folio(page));
 515 }
 516
 517 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 518 {
 519         struct file *file = vmf->vma->vm_file;
 520         struct address_space *mapping = file->f_mapping;
 521         struct address_space *fdm = faults_disabled_mapping();
 522         struct bch_inode_info *inode = file_bch_inode(file);
 523         int ret;
 524
 525         if (fdm == mapping)
 526                 return VM_FAULT_SIGBUS;
 527
 528         /* Lock ordering: */
 529         if (fdm > mapping) {
 530                 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
 531
 532                 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
 533                         goto got_lock;
 534
 535                 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
 536
 537                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 538                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 539
 540                 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
 541
 542                 /* Signal that lock has been dropped: */
 543                 set_fdm_dropped_locks();
 544                 return VM_FAULT_SIGBUS;
 545         }
 546
 547         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 548 got_lock:
 549         ret = filemap_fault(vmf);
 550         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 551
 552         return ret;
 553 }
 554
 555 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 556 {
 557         struct page *page = vmf->page;
 558         struct file *file = vmf->vma->vm_file;
 559         struct bch_inode_info *inode = file_bch_inode(file);
 560         struct address_space *mapping = file->f_mapping;
 561         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 562         struct bch2_page_reservation res;
 563         unsigned len;
 564         loff_t isize;
 565         int ret = VM_FAULT_LOCKED;
 566
 567         bch2_page_reservation_init(c, inode, &res);
 568
 569         sb_start_pagefault(inode->v.i_sb);
 570         file_update_time(file);
 571
 572         /*
 573          * Not strictly necessary, but helps avoid dio writes livelocking in
 574          * write_invalidate_inode_pages_range() - can drop this if/when we get
 575          * a write_invalidate_inode_pages_range() that works without dropping
 576          * page lock before invalidating page
 577          */
 578         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 579
 580         lock_page(page);
 581         isize = i_size_read(&inode->v);
 582
 583         if (page->mapping != mapping || page_offset(page) >= isize) {
 584                 unlock_page(page);
 585                 ret = VM_FAULT_NOPAGE;
 586                 goto out;
 587         }
 588
 589         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 590
 591         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 592                 unlock_page(page);
 593                 ret = VM_FAULT_SIGBUS;
 594                 goto out;
 595         }
 596
 597         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 598         bch2_page_reservation_put(c, inode, &res);
 599
 600         wait_for_stable_page(page);
 601 out:
 602         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 603         sb_end_pagefault(inode->v.i_sb);
 604
 605         return ret;
 606 }
 607
 608 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
 609 {
 610         if (offset || length < folio_size(folio))
 611                 return;
 612
 613         bch2_clear_page_bits(&folio->page);
 614 }
 615
 616 bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
 617 {
 618         if (folio_test_dirty(folio) || folio_test_writeback(folio))
 619                 return false;
 620
 621         bch2_clear_page_bits(&folio->page);
 622         return true;
 623 }
 624
 625 /* readpage(s): */
 626
 627 static void bch2_readpages_end_io(struct bio *bio)
 628 {
 629         struct bvec_iter_all iter;
 630         struct bio_vec *bv;
 631
 632         bio_for_each_segment_all(bv, bio, iter) {
 633                 struct page *page = bv->bv_page;
 634
 635                 if (!bio->bi_status) {
 636                         SetPageUptodate(page);
 637                 } else {
 638                         ClearPageUptodate(page);
 639                         SetPageError(page);
 640                 }
 641                 unlock_page(page);
 642         }
 643
 644         bio_put(bio);
 645 }
 646
 647 struct readpages_iter {
 648         struct address_space    *mapping;
 649         struct page             **pages;
 650         unsigned                nr_pages;
 651         unsigned                idx;
 652         pgoff_t                 offset;
 653 };
 654
 655 static int readpages_iter_init(struct readpages_iter *iter,
 656                                struct readahead_control *ractl)
 657 {
 658         unsigned i, nr_pages = readahead_count(ractl);
 659
 660         memset(iter, 0, sizeof(*iter));
 661
 662         iter->mapping   = ractl->mapping;
 663         iter->offset    = readahead_index(ractl);
 664         iter->nr_pages  = nr_pages;
 665
 666         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 667         if (!iter->pages)
 668                 return -ENOMEM;
 669
 670         nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
 671         for (i = 0; i < nr_pages; i++) {
 672                 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
 673                 put_page(iter->pages[i]);
 674         }
 675
 676         return 0;
 677 }
 678
 679 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 680 {
 681         if (iter->idx >= iter->nr_pages)
 682                 return NULL;
 683
 684         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 685
 686         return iter->pages[iter->idx];
 687 }
 688
 689 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 690 {
 691         struct bvec_iter iter;
 692         struct bio_vec bv;
 693         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 694                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 695         unsigned state = k.k->type == KEY_TYPE_reservation
 696                 ? SECTOR_RESERVED
 697                 : SECTOR_ALLOCATED;
 698
 699         bio_for_each_segment(bv, bio, iter) {
 700                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 701                 unsigned i;
 702
 703                 for (i = bv.bv_offset >> 9;
 704                      i < (bv.bv_offset + bv.bv_len) >> 9;
 705                      i++) {
 706                         s->s[i].nr_replicas = nr_ptrs;
 707                         s->s[i].state = state;
 708                 }
 709         }
 710 }
 711
 712 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 713 {
 714         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 715         struct bch_extent_crc_unpacked crc;
 716         const union bch_extent_entry *i;
 717
 718         bkey_for_each_crc(k.k, ptrs, crc, i)
 719                 if (crc.csum_type || crc.compression_type)
 720                         return true;
 721         return false;
 722 }
 723
 724 static void readpage_bio_extend(struct readpages_iter *iter,
 725                                 struct bio *bio,
 726                                 unsigned sectors_this_extent,
 727                                 bool get_more)
 728 {
 729         while (bio_sectors(bio) < sectors_this_extent &&
 730                bio->bi_vcnt < bio->bi_max_vecs) {
 731                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 732                 struct page *page = readpage_iter_next(iter);
 733                 int ret;
 734
 735                 if (page) {
 736                         if (iter->offset + iter->idx != page_offset)
 737                                 break;
 738
 739                         iter->idx++;
 740                 } else {
 741                         if (!get_more)
 742                                 break;
 743
 744                         page = xa_load(&iter->mapping->i_pages, page_offset);
 745                         if (page && !xa_is_value(page))
 746                                 break;
 747
 748                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 749                         if (!page)
 750                                 break;
 751
 752                         if (!__bch2_page_state_create(page, 0)) {
 753                                 put_page(page);
 754                                 break;
 755                         }
 756
 757                         ret = add_to_page_cache_lru(page, iter->mapping,
 758                                                     page_offset, GFP_NOFS);
 759                         if (ret) {
 760                                 __bch2_page_state_release(page);
 761                                 put_page(page);
 762                                 break;
 763                         }
 764
 765                         put_page(page);
 766                 }
 767
 768                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 769         }
 770 }
 771
 772 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 773                        struct bch_read_bio *rbio, u64 inum,
 774                        struct readpages_iter *readpages_iter)
 775 {
 776         struct bch_fs *c = trans->c;
 777         struct bkey_buf sk;
 778         int flags = BCH_READ_RETRY_IF_STALE|
 779                 BCH_READ_MAY_PROMOTE;
 780         int ret = 0;
 781
 782         rbio->c = c;
 783         rbio->start_time = local_clock();
 784
 785         bch2_bkey_buf_init(&sk);
 786 retry:
 787         while (1) {
 788                 struct bkey_s_c k;
 789                 unsigned bytes, sectors, offset_into_extent;
 790
 791                 bch2_btree_iter_set_pos(iter,
 792                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 793
 794                 k = bch2_btree_iter_peek_slot(iter);
 795                 ret = bkey_err(k);
 796                 if (ret)
 797                         break;
 798
 799                 offset_into_extent = iter->pos.offset -
 800                         bkey_start_offset(k.k);
 801                 sectors = k.k->size - offset_into_extent;
 802
 803                 bch2_bkey_buf_reassemble(&sk, c, k);
 804
 805                 ret = bch2_read_indirect_extent(trans,
 806                                         &offset_into_extent, &sk);
 807                 if (ret)
 808                         break;
 809
 810                 k = bkey_i_to_s_c(sk.k);
 811
 812                 sectors = min(sectors, k.k->size - offset_into_extent);
 813
 814                 bch2_trans_unlock(trans);
 815
 816                 if (readpages_iter)
 817                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 818                                             extent_partial_reads_expensive(k));
 819
 820                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 821                 swap(rbio->bio.bi_iter.bi_size, bytes);
 822
 823                 if (rbio->bio.bi_iter.bi_size == bytes)
 824                         flags |= BCH_READ_LAST_FRAGMENT;
 825
 826                 if (bkey_extent_is_allocation(k.k))
 827                         bch2_add_page_sectors(&rbio->bio, k);
 828
 829                 bch2_read_extent(trans, rbio, k, offset_into_extent, flags);
 830
 831                 if (flags & BCH_READ_LAST_FRAGMENT)
 832                         break;
 833
 834                 swap(rbio->bio.bi_iter.bi_size, bytes);
 835                 bio_advance(&rbio->bio, bytes);
 836         }
 837
 838         if (ret == -EINTR)
 839                 goto retry;
 840
 841         if (ret) {
 842                 bch_err_inum_ratelimited(c, inum,
 843                                 "read error %i from btree lookup", ret);
 844                 rbio->bio.bi_status = BLK_STS_IOERR;
 845                 bio_endio(&rbio->bio);
 846         }
 847
 848         bch2_bkey_buf_exit(&sk, c);
 849 }
 850
 851 void bch2_readahead(struct readahead_control *ractl)
 852 {
 853         struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
 854         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 855         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 856         struct btree_trans trans;
 857         struct btree_iter *iter;
 858         struct page *page;
 859         struct readpages_iter readpages_iter;
 860         int ret;
 861
 862         ret = readpages_iter_init(&readpages_iter, ractl);
 863         BUG_ON(ret);
 864
 865         bch2_trans_init(&trans, c, 0, 0);
 866
 867         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 868                                    BTREE_ITER_SLOTS);
 869
 870         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 871
 872         while ((page = readpage_iter_next(&readpages_iter))) {
 873                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 874                 unsigned n = min_t(unsigned,
 875                                    readpages_iter.nr_pages -
 876                                    readpages_iter.idx,
 877                                    BIO_MAX_VECS);
 878                 struct bch_read_bio *rbio =
 879                         rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
 880                                                    GFP_NOFS, &c->bio_read),
 881                                   opts);
 882
 883                 readpages_iter.idx++;
 884
 885                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 886                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 887                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 888
 889                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 890                            &readpages_iter);
 891         }
 892
 893         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 894
 895         bch2_trans_exit(&trans);
 896         kfree(readpages_iter.pages);
 897 }
 898
 899 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 900                              u64 inum, struct page *page)
 901 {
 902         struct btree_trans trans;
 903         struct btree_iter *iter;
 904
 905         bch2_page_state_create(page, __GFP_NOFAIL);
 906
 907         rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
 908         rbio->bio.bi_iter.bi_sector =
 909                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 910         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 911
 912         bch2_trans_init(&trans, c, 0, 0);
 913         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 914                                    BTREE_ITER_SLOTS);
 915
 916         bchfs_read(&trans, iter, rbio, inum, NULL);
 917
 918         bch2_trans_exit(&trans);
 919 }
 920
 921 static void bch2_read_single_page_end_io(struct bio *bio)
 922 {
 923         complete(bio->bi_private);
 924 }
 925
 926 static int bch2_read_single_page(struct page *page,
 927                                  struct address_space *mapping)
 928 {
 929         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 930         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 931         struct bch_read_bio *rbio;
 932         int ret;
 933         DECLARE_COMPLETION_ONSTACK(done);
 934
 935         rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS, &c->bio_read),
 936                          io_opts(c, &inode->ei_inode));
 937         rbio->bio.bi_private = &done;
 938         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 939
 940         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 941         wait_for_completion(&done);
 942
 943         ret = blk_status_to_errno(rbio->bio.bi_status);
 944         bio_put(&rbio->bio);
 945
 946         if (ret < 0)
 947                 return ret;
 948
 949         SetPageUptodate(page);
 950         return 0;
 951 }
 952
 953 int bch2_read_folio(struct file *file, struct folio *folio)
 954 {
 955         struct page *page = &folio->page;
 956         int ret;
 957
 958         ret = bch2_read_single_page(page, page->mapping);
 959         folio_unlock(folio);
 960         return ret;
 961 }
 962
 963 /* writepages: */
 964
 965 struct bch_writepage_state {
 966         struct bch_writepage_io *io;
 967         struct bch_io_opts      opts;
 968 };
 969
 970 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 971                                                                   struct bch_inode_info *inode)
 972 {
 973         return (struct bch_writepage_state) {
 974                 .opts = io_opts(c, &inode->ei_inode)
 975         };
 976 }
 977
 978 static void bch2_writepage_io_free(struct closure *cl)
 979 {
 980         struct bch_writepage_io *io = container_of(cl,
 981                                         struct bch_writepage_io, cl);
 982
 983         bio_put(&io->op.wbio.bio);
 984 }
 985
 986 static void bch2_writepage_io_done(struct closure *cl)
 987 {
 988         struct bch_writepage_io *io = container_of(cl,
 989                                         struct bch_writepage_io, cl);
 990         struct bch_fs *c = io->op.c;
 991         struct bio *bio = &io->op.wbio.bio;
 992         struct bvec_iter_all iter;
 993         struct bio_vec *bvec;
 994         unsigned i;
 995
 996         if (io->op.error) {
 997                 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
 998
 999                 bio_for_each_segment_all(bvec, bio, iter) {
1000                         struct bch_page_state *s;
1001
1002                         SetPageError(bvec->bv_page);
1003                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1004
1005                         s = __bch2_page_state(bvec->bv_page);
1006                         spin_lock(&s->lock);
1007                         for (i = 0; i < PAGE_SECTORS; i++)
1008                                 s->s[i].nr_replicas = 0;
1009                         spin_unlock(&s->lock);
1010                 }
1011         }
1012
1013         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1014                 bio_for_each_segment_all(bvec, bio, iter) {
1015                         struct bch_page_state *s;
1016
1017                         s = __bch2_page_state(bvec->bv_page);
1018                         spin_lock(&s->lock);
1019                         for (i = 0; i < PAGE_SECTORS; i++)
1020                                 s->s[i].nr_replicas = 0;
1021                         spin_unlock(&s->lock);
1022                 }
1023         }
1024
1025         /*
1026          * racing with fallocate can cause us to add fewer sectors than
1027          * expected - but we shouldn't add more sectors than expected:
1028          */
1029         BUG_ON(io->op.i_sectors_delta > 0);
1030
1031         /*
1032          * (error (due to going RO) halfway through a page can screw that up
1033          * slightly)
1034          * XXX wtf?
1035            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1036          */
1037
1038         /*
1039          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1040          * before calling end_page_writeback:
1041          */
1042         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1043
1044         bio_for_each_segment_all(bvec, bio, iter) {
1045                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1046
1047                 if (atomic_dec_and_test(&s->write_count))
1048                         end_page_writeback(bvec->bv_page);
1049         }
1050
1051         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1052 }
1053
1054 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1055 {
1056         struct bch_writepage_io *io = w->io;
1057
1058         w->io = NULL;
1059         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1060         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1061 }
1062
1063 /*
1064  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1065  * possible, else allocating a new one:
1066  */
1067 static void bch2_writepage_io_alloc(struct bch_fs *c,
1068                                     struct writeback_control *wbc,
1069                                     struct bch_writepage_state *w,
1070                                     struct bch_inode_info *inode,
1071                                     u64 sector,
1072                                     unsigned nr_replicas)
1073 {
1074         struct bch_write_op *op;
1075
1076         w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1077                                               REQ_OP_WRITE,
1078                                               GFP_NOFS,
1079                                               &c->writepage_bioset),
1080                              struct bch_writepage_io, op.wbio.bio);
1081
1082         closure_init(&w->io->cl, NULL);
1083         w->io->inode            = inode;
1084
1085         op                      = &w->io->op;
1086         bch2_write_op_init(op, c, w->opts);
1087         op->target              = w->opts.foreground_target;
1088         op_journal_seq_set(op, &inode->ei_journal_seq);
1089         op->nr_replicas         = nr_replicas;
1090         op->res.nr_replicas     = nr_replicas;
1091         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1092         op->pos                 = POS(inode->v.i_ino, sector);
1093         op->wbio.bio.bi_iter.bi_sector = sector;
1094         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1095 }
1096
1097 static int __bch2_writepage(struct folio *folio,
1098                             struct writeback_control *wbc,
1099                             void *data)
1100 {
1101         struct page *page = &folio->page;
1102         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1103         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1104         struct bch_writepage_state *w = data;
1105         struct bch_page_state *s, orig;
1106         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1107         loff_t i_size = i_size_read(&inode->v);
1108         pgoff_t end_index = i_size >> PAGE_SHIFT;
1109         int ret;
1110
1111         EBUG_ON(!PageUptodate(page));
1112
1113         /* Is the page fully inside i_size? */
1114         if (page->index < end_index)
1115                 goto do_io;
1116
1117         /* Is the page fully outside i_size? (truncate in progress) */
1118         offset = i_size & (PAGE_SIZE - 1);
1119         if (page->index > end_index || !offset) {
1120                 unlock_page(page);
1121                 return 0;
1122         }
1123
1124         /*
1125          * The page straddles i_size.  It must be zeroed out on each and every
1126          * writepage invocation because it may be mmapped.  "A file is mapped
1127          * in multiples of the page size.  For a file that is not a multiple of
1128          * the  page size, the remaining memory is zeroed when mapped, and
1129          * writes to that region are not written out to the file."
1130          */
1131         zero_user_segment(page, offset, PAGE_SIZE);
1132 do_io:
1133         s = bch2_page_state_create(page, __GFP_NOFAIL);
1134
1135         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1136         if (ret) {
1137                 SetPageError(page);
1138                 mapping_set_error(page->mapping, ret);
1139                 unlock_page(page);
1140                 return 0;
1141         }
1142
1143         /* Before unlocking the page, get copy of reservations: */
1144         orig = *s;
1145
1146         for (i = 0; i < PAGE_SECTORS; i++) {
1147                 if (s->s[i].state < SECTOR_DIRTY)
1148                         continue;
1149
1150                 nr_replicas_this_write =
1151                         min_t(unsigned, nr_replicas_this_write,
1152                               s->s[i].nr_replicas +
1153                               s->s[i].replicas_reserved);
1154         }
1155
1156         for (i = 0; i < PAGE_SECTORS; i++) {
1157                 if (s->s[i].state < SECTOR_DIRTY)
1158                         continue;
1159
1160                 s->s[i].nr_replicas = w->opts.compression
1161                         ? 0 : nr_replicas_this_write;
1162
1163                 s->s[i].replicas_reserved = 0;
1164                 s->s[i].state = SECTOR_ALLOCATED;
1165         }
1166
1167         BUG_ON(atomic_read(&s->write_count));
1168         atomic_set(&s->write_count, 1);
1169
1170         BUG_ON(PageWriteback(page));
1171         set_page_writeback(page);
1172
1173         unlock_page(page);
1174
1175         offset = 0;
1176         while (1) {
1177                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1178                 u64 sector;
1179
1180                 while (offset < PAGE_SECTORS &&
1181                        orig.s[offset].state < SECTOR_DIRTY)
1182                         offset++;
1183
1184                 if (offset == PAGE_SECTORS)
1185                         break;
1186
1187                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1188
1189                 while (offset + sectors < PAGE_SECTORS &&
1190                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1191                         sectors++;
1192
1193                 for (i = offset; i < offset + sectors; i++) {
1194                         reserved_sectors += orig.s[i].replicas_reserved;
1195                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1196                 }
1197
1198                 if (w->io &&
1199                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1200                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1201                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1202                      (BIO_MAX_VECS * PAGE_SIZE) ||
1203                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1204                         bch2_writepage_do_io(w);
1205
1206                 if (!w->io)
1207                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1208                                                 nr_replicas_this_write);
1209
1210                 atomic_inc(&s->write_count);
1211
1212                 BUG_ON(inode != w->io->inode);
1213                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1214                                      sectors << 9, offset << 9));
1215
1216                 /* Check for writing past i_size: */
1217                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1218                         round_up(i_size, block_bytes(c)));
1219
1220                 w->io->op.res.sectors += reserved_sectors;
1221                 w->io->op.i_sectors_delta -= dirty_sectors;
1222                 w->io->op.new_i_size = i_size;
1223
1224                 offset += sectors;
1225         }
1226
1227         if (atomic_dec_and_test(&s->write_count))
1228                 end_page_writeback(page);
1229
1230         return 0;
1231 }
1232
1233 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1234 {
1235         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1236         struct bch_writepage_state w =
1237                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1238         struct blk_plug plug;
1239         int ret;
1240
1241         blk_start_plug(&plug);
1242         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1243         if (w.io)
1244                 bch2_writepage_do_io(&w);
1245         blk_finish_plug(&plug);
1246         return ret;
1247 }
1248
1249 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1250 {
1251         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1252         struct bch_writepage_state w =
1253                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1254         int ret;
1255
1256         ret = __bch2_writepage(page_folio(page), wbc, &w);
1257         if (w.io)
1258                 bch2_writepage_do_io(&w);
1259
1260         return ret;
1261 }
1262
1263 /* buffered writes: */
1264
1265 int bch2_write_begin(struct file *file, struct address_space *mapping,
1266                      loff_t pos, unsigned len,
1267                      struct page **pagep, void **fsdata)
1268 {
1269         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1270         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1271         struct bch2_page_reservation *res;
1272         pgoff_t index = pos >> PAGE_SHIFT;
1273         unsigned offset = pos & (PAGE_SIZE - 1);
1274         struct page *page;
1275         int ret = -ENOMEM;
1276
1277         res = kmalloc(sizeof(*res), GFP_KERNEL);
1278         if (!res)
1279                 return -ENOMEM;
1280
1281         bch2_page_reservation_init(c, inode, res);
1282         *fsdata = res;
1283
1284         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1285
1286         page = grab_cache_page_write_begin(mapping, index);
1287         if (!page)
1288                 goto err_unlock;
1289
1290         if (PageUptodate(page))
1291                 goto out;
1292
1293         /* If we're writing entire page, don't need to read it in first: */
1294         if (len == PAGE_SIZE)
1295                 goto out;
1296
1297         if (!offset && pos + len >= inode->v.i_size) {
1298                 zero_user_segment(page, len, PAGE_SIZE);
1299                 flush_dcache_page(page);
1300                 goto out;
1301         }
1302
1303         if (index > inode->v.i_size >> PAGE_SHIFT) {
1304                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1305                 flush_dcache_page(page);
1306                 goto out;
1307         }
1308 readpage:
1309         ret = bch2_read_single_page(page, mapping);
1310         if (ret)
1311                 goto err;
1312 out:
1313         ret = bch2_page_reservation_get(c, inode, page, res,
1314                                         offset, len, true);
1315         if (ret) {
1316                 if (!PageUptodate(page)) {
1317                         /*
1318                          * If the page hasn't been read in, we won't know if we
1319                          * actually need a reservation - we don't actually need
1320                          * to read here, we just need to check if the page is
1321                          * fully backed by uncompressed data:
1322                          */
1323                         goto readpage;
1324                 }
1325
1326                 goto err;
1327         }
1328
1329         *pagep = page;
1330         return 0;
1331 err:
1332         unlock_page(page);
1333         put_page(page);
1334         *pagep = NULL;
1335 err_unlock:
1336         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1337         kfree(res);
1338         *fsdata = NULL;
1339         return ret;
1340 }
1341
1342 int bch2_write_end(struct file *file, struct address_space *mapping,
1343                    loff_t pos, unsigned len, unsigned copied,
1344                    struct page *page, void *fsdata)
1345 {
1346         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1347         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1348         struct bch2_page_reservation *res = fsdata;
1349         unsigned offset = pos & (PAGE_SIZE - 1);
1350
1351         lockdep_assert_held(&inode->v.i_rwsem);
1352
1353         if (unlikely(copied < len && !PageUptodate(page))) {
1354                 /*
1355                  * The page needs to be read in, but that would destroy
1356                  * our partial write - simplest thing is to just force
1357                  * userspace to redo the write:
1358                  */
1359                 zero_user(page, 0, PAGE_SIZE);
1360                 flush_dcache_page(page);
1361                 copied = 0;
1362         }
1363
1364         spin_lock(&inode->v.i_lock);
1365         if (pos + copied > inode->v.i_size)
1366                 i_size_write(&inode->v, pos + copied);
1367         spin_unlock(&inode->v.i_lock);
1368
1369         if (copied) {
1370                 if (!PageUptodate(page))
1371                         SetPageUptodate(page);
1372
1373                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1374
1375                 inode->ei_last_dirtied = (unsigned long) current;
1376         }
1377
1378         unlock_page(page);
1379         put_page(page);
1380         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1381
1382         bch2_page_reservation_put(c, inode, res);
1383         kfree(res);
1384
1385         return copied;
1386 }
1387
1388 #define WRITE_BATCH_PAGES       32
1389
1390 static int __bch2_buffered_write(struct bch_inode_info *inode,
1391                                  struct address_space *mapping,
1392                                  struct iov_iter *iter,
1393                                  loff_t pos, unsigned len)
1394 {
1395         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1396         struct page *pages[WRITE_BATCH_PAGES];
1397         struct bch2_page_reservation res;
1398         unsigned long index = pos >> PAGE_SHIFT;
1399         unsigned offset = pos & (PAGE_SIZE - 1);
1400         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1401         unsigned i, reserved = 0, set_dirty = 0;
1402         unsigned copied = 0, nr_pages_copied = 0;
1403         int ret = 0;
1404
1405         BUG_ON(!len);
1406         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1407
1408         bch2_page_reservation_init(c, inode, &res);
1409
1410         for (i = 0; i < nr_pages; i++) {
1411                 pages[i] = grab_cache_page_write_begin(mapping, index + i);
1412                 if (!pages[i]) {
1413                         nr_pages = i;
1414                         if (!i) {
1415                                 ret = -ENOMEM;
1416                                 goto out;
1417                         }
1418                         len = min_t(unsigned, len,
1419                                     nr_pages * PAGE_SIZE - offset);
1420                         break;
1421                 }
1422         }
1423
1424         if (offset && !PageUptodate(pages[0])) {
1425                 ret = bch2_read_single_page(pages[0], mapping);
1426                 if (ret)
1427                         goto out;
1428         }
1429
1430         if ((pos + len) & (PAGE_SIZE - 1) &&
1431             !PageUptodate(pages[nr_pages - 1])) {
1432                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1433                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1434                 } else {
1435                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1436                         if (ret)
1437                                 goto out;
1438                 }
1439         }
1440
1441         while (reserved < len) {
1442                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1443                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1444                 unsigned pg_len = min_t(unsigned, len - reserved,
1445                                         PAGE_SIZE - pg_offset);
1446 retry_reservation:
1447                 ret = bch2_page_reservation_get(c, inode, page, &res,
1448                                                 pg_offset, pg_len, true);
1449
1450                 if (ret && !PageUptodate(page)) {
1451                         ret = bch2_read_single_page(page, mapping);
1452                         if (!ret)
1453                                 goto retry_reservation;
1454                 }
1455
1456                 if (ret)
1457                         goto out;
1458
1459                 reserved += pg_len;
1460         }
1461
1462         if (mapping_writably_mapped(mapping))
1463                 for (i = 0; i < nr_pages; i++)
1464                         flush_dcache_page(pages[i]);
1465
1466         while (copied < len) {
1467                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1468                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1469                 unsigned pg_len = min_t(unsigned, len - copied,
1470                                         PAGE_SIZE - pg_offset);
1471                 unsigned pg_copied = copy_page_from_iter_atomic(page,
1472                                                 pg_offset, pg_len, iter);
1473
1474                 if (!pg_copied)
1475                         break;
1476
1477                 if (!PageUptodate(page) &&
1478                     pg_copied != PAGE_SIZE &&
1479                     pos + copied + pg_copied < inode->v.i_size) {
1480                         zero_user(page, 0, PAGE_SIZE);
1481                         break;
1482                 }
1483
1484                 flush_dcache_page(page);
1485                 copied += pg_copied;
1486
1487                 if (pg_copied != pg_len)
1488                         break;
1489         }
1490
1491         if (!copied)
1492                 goto out;
1493
1494         spin_lock(&inode->v.i_lock);
1495         if (pos + copied > inode->v.i_size)
1496                 i_size_write(&inode->v, pos + copied);
1497         spin_unlock(&inode->v.i_lock);
1498
1499         while (set_dirty < copied) {
1500                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1501                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1502                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1503                                         PAGE_SIZE - pg_offset);
1504
1505                 if (!PageUptodate(page))
1506                         SetPageUptodate(page);
1507
1508                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1509                 unlock_page(page);
1510                 put_page(page);
1511
1512                 set_dirty += pg_len;
1513         }
1514
1515         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1516         inode->ei_last_dirtied = (unsigned long) current;
1517 out:
1518         for (i = nr_pages_copied; i < nr_pages; i++) {
1519                 unlock_page(pages[i]);
1520                 put_page(pages[i]);
1521         }
1522
1523         bch2_page_reservation_put(c, inode, &res);
1524
1525         return copied ?: ret;
1526 }
1527
1528 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1529 {
1530         struct file *file = iocb->ki_filp;
1531         struct address_space *mapping = file->f_mapping;
1532         struct bch_inode_info *inode = file_bch_inode(file);
1533         loff_t pos = iocb->ki_pos;
1534         ssize_t written = 0;
1535         int ret = 0;
1536
1537         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1538
1539         do {
1540                 unsigned offset = pos & (PAGE_SIZE - 1);
1541                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1542                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1543 again:
1544                 /*
1545                  * Bring in the user page that we will copy from _first_.
1546                  * Otherwise there's a nasty deadlock on copying from the
1547                  * same page as we're writing to, without it being marked
1548                  * up-to-date.
1549                  *
1550                  * Not only is this an optimisation, but it is also required
1551                  * to check that the address is actually valid, when atomic
1552                  * usercopies are used, below.
1553                  */
1554                 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1555                         bytes = min_t(unsigned long, iov_iter_count(iter),
1556                                       PAGE_SIZE - offset);
1557
1558                         if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1559                                 ret = -EFAULT;
1560                                 break;
1561                         }
1562                 }
1563
1564                 if (unlikely(fatal_signal_pending(current))) {
1565                         ret = -EINTR;
1566                         break;
1567                 }
1568
1569                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1570                 if (unlikely(ret < 0))
1571                         break;
1572
1573                 cond_resched();
1574
1575                 if (unlikely(ret == 0)) {
1576                         /*
1577                          * If we were unable to copy any data at all, we must
1578                          * fall back to a single segment length write.
1579                          *
1580                          * If we didn't fallback here, we could livelock
1581                          * because not all segments in the iov can be copied at
1582                          * once without a pagefault.
1583                          */
1584                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1585                                       iov_iter_single_seg_count(iter));
1586                         goto again;
1587                 }
1588                 pos += ret;
1589                 written += ret;
1590                 ret = 0;
1591
1592                 balance_dirty_pages_ratelimited(mapping);
1593         } while (iov_iter_count(iter));
1594
1595         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1596
1597         return written ? written : ret;
1598 }
1599
1600 /* O_DIRECT reads */
1601
1602 static void bch2_dio_read_complete(struct closure *cl)
1603 {
1604         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1605
1606         dio->req->ki_complete(dio->req, dio->ret);
1607         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1608 }
1609
1610 static void bch2_direct_IO_read_endio(struct bio *bio)
1611 {
1612         struct dio_read *dio = bio->bi_private;
1613
1614         if (bio->bi_status)
1615                 dio->ret = blk_status_to_errno(bio->bi_status);
1616
1617         closure_put(&dio->cl);
1618 }
1619
1620 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1621 {
1622         bch2_direct_IO_read_endio(bio);
1623         bio_check_pages_dirty(bio);     /* transfers ownership */
1624 }
1625
1626 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1627 {
1628         struct file *file = req->ki_filp;
1629         struct bch_inode_info *inode = file_bch_inode(file);
1630         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1631         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1632         struct dio_read *dio;
1633         struct bio *bio;
1634         loff_t offset = req->ki_pos;
1635         bool sync = is_sync_kiocb(req);
1636         size_t shorten;
1637         ssize_t ret;
1638
1639         if ((offset|iter->count) & (block_bytes(c) - 1))
1640                 return -EINVAL;
1641
1642         ret = min_t(loff_t, iter->count,
1643                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1644
1645         if (!ret)
1646                 return ret;
1647
1648         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1649         iter->count -= shorten;
1650
1651         bio = bio_alloc_bioset(NULL,
1652                                iov_iter_npages(iter, BIO_MAX_VECS),
1653                                REQ_OP_READ,
1654                                GFP_KERNEL,
1655                                &c->dio_read_bioset);
1656
1657         bio->bi_end_io = bch2_direct_IO_read_endio;
1658
1659         dio = container_of(bio, struct dio_read, rbio.bio);
1660         closure_init(&dio->cl, NULL);
1661
1662         /*
1663          * this is a _really_ horrible hack just to avoid an atomic sub at the
1664          * end:
1665          */
1666         if (!sync) {
1667                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1668                 atomic_set(&dio->cl.remaining,
1669                            CLOSURE_REMAINING_INITIALIZER -
1670                            CLOSURE_RUNNING +
1671                            CLOSURE_DESTRUCTOR);
1672         } else {
1673                 atomic_set(&dio->cl.remaining,
1674                            CLOSURE_REMAINING_INITIALIZER + 1);
1675         }
1676
1677         dio->req        = req;
1678         dio->ret        = ret;
1679
1680         goto start;
1681         while (iter->count) {
1682                 bio = bio_alloc_bioset(NULL,
1683                                        iov_iter_npages(iter, BIO_MAX_VECS),
1684                                        REQ_OP_READ,
1685                                        GFP_KERNEL,
1686                                        &c->bio_read);
1687                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1688 start:
1689                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC;
1690                 bio->bi_iter.bi_sector  = offset >> 9;
1691                 bio->bi_private         = dio;
1692
1693                 ret = bio_iov_iter_get_pages(bio, iter);
1694                 if (ret < 0) {
1695                         /* XXX: fault inject this path */
1696                         bio->bi_status = BLK_STS_RESOURCE;
1697                         bio_endio(bio);
1698                         break;
1699                 }
1700
1701                 offset += bio->bi_iter.bi_size;
1702                 bio_set_pages_dirty(bio);
1703
1704                 if (iter->count)
1705                         closure_get(&dio->cl);
1706
1707                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1708         }
1709
1710         iter->count += shorten;
1711
1712         if (sync) {
1713                 closure_sync(&dio->cl);
1714                 closure_debug_destroy(&dio->cl);
1715                 ret = dio->ret;
1716                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1717                 return ret;
1718         } else {
1719                 return -EIOCBQUEUED;
1720         }
1721 }
1722
1723 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1724 {
1725         struct file *file = iocb->ki_filp;
1726         struct bch_inode_info *inode = file_bch_inode(file);
1727         struct address_space *mapping = file->f_mapping;
1728         size_t count = iov_iter_count(iter);
1729         ssize_t ret;
1730
1731         if (!count)
1732                 return 0; /* skip atime */
1733
1734         if (iocb->ki_flags & IOCB_DIRECT) {
1735                 struct blk_plug plug;
1736
1737                 if (unlikely(mapping->nrpages)) {
1738                         ret = filemap_write_and_wait_range(mapping,
1739                                                 iocb->ki_pos,
1740                                                 iocb->ki_pos + count - 1);
1741                         if (ret < 0)
1742                                 return ret;
1743                 }
1744
1745                 file_accessed(file);
1746
1747                 blk_start_plug(&plug);
1748                 ret = bch2_direct_IO_read(iocb, iter);
1749                 blk_finish_plug(&plug);
1750
1751                 if (ret >= 0)
1752                         iocb->ki_pos += ret;
1753         } else {
1754                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1755                 ret = generic_file_read_iter(iocb, iter);
1756                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1757         }
1758
1759         return ret;
1760 }
1761
1762 /* O_DIRECT writes */
1763
1764 /*
1765  * We're going to return -EIOCBQUEUED, but we haven't finished consuming the
1766  * iov_iter yet, so we need to stash a copy of the iovec: it might be on the
1767  * caller's stack, we're not guaranteed that it will live for the duration of
1768  * the IO:
1769  */
1770 static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
1771 {
1772         struct iovec *iov = dio->inline_vecs;
1773
1774         /*
1775          * iov_iter has a single embedded iovec - nothing to do:
1776          */
1777         if (iter_is_ubuf(&dio->iter))
1778                 return 0;
1779
1780         /*
1781          * We don't currently handle non-iovec iov_iters here - return an error,
1782          * and we'll fall back to doing the IO synchronously:
1783          */
1784         if (!iter_is_iovec(&dio->iter))
1785                 return -1;
1786
1787         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1788                 iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
1789                                     GFP_KERNEL);
1790                 if (unlikely(!iov))
1791                         return -ENOMEM;
1792
1793                 dio->free_iov = true;
1794         }
1795
1796         memcpy(iov, dio->iter.__iov, dio->iter.nr_segs * sizeof(*iov));
1797         dio->iter.__iov = iov;
1798         return 0;
1799 }
1800
1801 static void bch2_dio_write_loop_async(struct bch_write_op *);
1802
1803 static long bch2_dio_write_loop(struct dio_write *dio)
1804 {
1805         bool kthread = (current->flags & PF_KTHREAD) != 0;
1806         struct kiocb *req = dio->req;
1807         struct address_space *mapping = req->ki_filp->f_mapping;
1808         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1809         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1810         struct bio *bio = &dio->op.wbio.bio;
1811         struct bvec_iter_all iter;
1812         struct bio_vec *bv;
1813         unsigned unaligned, iter_count;
1814         bool sync = dio->sync, dropped_locks;
1815         long ret;
1816
1817         if (dio->loop)
1818                 goto loop;
1819
1820         while (1) {
1821                 iter_count = dio->iter.count;
1822
1823                 if (kthread)
1824                         kthread_use_mm(dio->mm);
1825                 BUG_ON(current->faults_disabled_mapping);
1826                 current->faults_disabled_mapping = mapping;
1827
1828                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1829
1830                 dropped_locks = fdm_dropped_locks();
1831
1832                 current->faults_disabled_mapping = NULL;
1833                 if (kthread)
1834                         kthread_unuse_mm(dio->mm);
1835
1836                 /*
1837                  * If the fault handler returned an error but also signalled
1838                  * that it dropped & retook ei_pagecache_lock, we just need to
1839                  * re-shoot down the page cache and retry:
1840                  */
1841                 if (dropped_locks && ret)
1842                         ret = 0;
1843
1844                 if (unlikely(ret < 0))
1845                         goto err;
1846
1847                 if (unlikely(dropped_locks)) {
1848                         ret = write_invalidate_inode_pages_range(mapping,
1849                                         req->ki_pos,
1850                                         req->ki_pos + iter_count - 1);
1851                         if (unlikely(ret))
1852                                 goto err;
1853
1854                         if (!bio->bi_iter.bi_size)
1855                                 continue;
1856                 }
1857
1858                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1859                 bio->bi_iter.bi_size -= unaligned;
1860                 iov_iter_revert(&dio->iter, unaligned);
1861
1862                 if (!bio->bi_iter.bi_size) {
1863                         /*
1864                          * bio_iov_iter_get_pages was only able to get <
1865                          * blocksize worth of pages:
1866                          */
1867                         bio_for_each_segment_all(bv, bio, iter)
1868                                 put_page(bv->bv_page);
1869                         ret = -EFAULT;
1870                         goto err;
1871                 }
1872
1873                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1874                 dio->op.end_io          = bch2_dio_write_loop_async;
1875                 dio->op.target          = dio->op.opts.foreground_target;
1876                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1877                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1878                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1879                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1880
1881                 if ((req->ki_flags & IOCB_DSYNC) &&
1882                     !c->opts.journal_flush_disabled)
1883                         dio->op.flags |= BCH_WRITE_FLUSH;
1884
1885                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1886                                                 dio->op.opts.data_replicas, 0);
1887                 if (unlikely(ret) &&
1888                     !bch2_check_range_allocated(c, dio->op.pos,
1889                                 bio_sectors(bio),
1890                                 dio->op.opts.data_replicas,
1891                                 dio->op.opts.compression != 0))
1892                         goto err;
1893
1894                 task_io_account_write(bio->bi_iter.bi_size);
1895
1896                 if (!dio->sync && !dio->loop && dio->iter.count) {
1897                         if (bch2_dio_write_copy_iov(dio)) {
1898                                 dio->sync = sync = true;
1899                                 goto do_io;
1900                         }
1901                 }
1902 do_io:
1903                 dio->loop = true;
1904                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1905
1906                 if (sync)
1907                         wait_for_completion(&dio->done);
1908                 else
1909                         return -EIOCBQUEUED;
1910 loop:
1911                 i_sectors_acct(c, inode, &dio->quota_res,
1912                                dio->op.i_sectors_delta);
1913                 req->ki_pos += (u64) dio->op.written << 9;
1914                 dio->written += dio->op.written;
1915
1916                 spin_lock(&inode->v.i_lock);
1917                 if (req->ki_pos > inode->v.i_size)
1918                         i_size_write(&inode->v, req->ki_pos);
1919                 spin_unlock(&inode->v.i_lock);
1920
1921                 bio_for_each_segment_all(bv, bio, iter)
1922                         put_page(bv->bv_page);
1923
1924                 if (dio->op.error) {
1925                         set_bit(EI_INODE_ERROR, &inode->ei_flags);
1926                         break;
1927                 }
1928
1929                 if (!dio->iter.count)
1930                         break;
1931
1932                 bio_reset(bio, NULL, REQ_OP_WRITE);
1933                 reinit_completion(&dio->done);
1934         }
1935
1936         ret = dio->op.error ?: ((long) dio->written << 9);
1937 err:
1938         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1939         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1940
1941         if (dio->free_iov)
1942                 kfree(dio->iter.__iov);
1943
1944         bio_put(bio);
1945
1946         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1947         inode_dio_end(&inode->v);
1948
1949         if (!sync) {
1950                 req->ki_complete(req, ret);
1951                 ret = -EIOCBQUEUED;
1952         }
1953         return ret;
1954 }
1955
1956 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1957 {
1958         struct dio_write *dio = container_of(op, struct dio_write, op);
1959
1960         if (dio->sync)
1961                 complete(&dio->done);
1962         else
1963                 bch2_dio_write_loop(dio);
1964 }
1965
1966 static noinline
1967 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1968 {
1969         struct file *file = req->ki_filp;
1970         struct address_space *mapping = file->f_mapping;
1971         struct bch_inode_info *inode = file_bch_inode(file);
1972         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1973         struct dio_write *dio;
1974         struct bio *bio;
1975         bool locked = true, extending;
1976         ssize_t ret;
1977
1978         prefetch(&c->opts);
1979         prefetch((void *) &c->opts + 64);
1980         prefetch(&inode->ei_inode);
1981         prefetch((void *) &inode->ei_inode + 64);
1982
1983         inode_lock(&inode->v);
1984
1985         ret = generic_write_checks(req, iter);
1986         if (unlikely(ret <= 0))
1987                 goto err;
1988
1989         ret = file_remove_privs(file);
1990         if (unlikely(ret))
1991                 goto err;
1992
1993         ret = file_update_time(file);
1994         if (unlikely(ret))
1995                 goto err;
1996
1997         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1998                 goto err;
1999
2000         inode_dio_begin(&inode->v);
2001         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2002
2003         extending = req->ki_pos + iter->count > inode->v.i_size;
2004         if (!extending) {
2005                 inode_unlock(&inode->v);
2006                 locked = false;
2007         }
2008
2009         bio = bio_alloc_bioset(NULL,
2010                                iov_iter_npages(iter, BIO_MAX_VECS),
2011                                REQ_OP_WRITE,
2012                                GFP_KERNEL,
2013                                &c->dio_write_bioset);
2014         dio = container_of(bio, struct dio_write, op.wbio.bio);
2015         init_completion(&dio->done);
2016         dio->req                = req;
2017         dio->mm                 = current->mm;
2018         dio->loop               = false;
2019         dio->sync               = is_sync_kiocb(req) || extending;
2020         dio->free_iov           = false;
2021         dio->quota_res.sectors  = 0;
2022         dio->written            = 0;
2023         dio->iter               = *iter;
2024
2025         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2026                                          iter->count >> 9, true);
2027         if (unlikely(ret))
2028                 goto err_put_bio;
2029
2030         if (unlikely(mapping->nrpages)) {
2031                 ret = write_invalidate_inode_pages_range(mapping,
2032                                                 req->ki_pos,
2033                                                 req->ki_pos + iter->count - 1);
2034                 if (unlikely(ret))
2035                         goto err_put_bio;
2036         }
2037
2038         ret = bch2_dio_write_loop(dio);
2039 err:
2040         if (locked)
2041                 inode_unlock(&inode->v);
2042         return ret;
2043 err_put_bio:
2044         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2045         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2046         bio_put(bio);
2047         inode_dio_end(&inode->v);
2048         goto err;
2049 }
2050
2051 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2052 {
2053         struct file *file = iocb->ki_filp;
2054         struct bch_inode_info *inode = file_bch_inode(file);
2055         ssize_t ret;
2056
2057         if (iocb->ki_flags & IOCB_DIRECT)
2058                 return bch2_direct_write(iocb, from);
2059
2060         inode_lock(&inode->v);
2061
2062         ret = generic_write_checks(iocb, from);
2063         if (ret <= 0)
2064                 goto unlock;
2065
2066         ret = file_remove_privs(file);
2067         if (ret)
2068                 goto unlock;
2069
2070         ret = file_update_time(file);
2071         if (ret)
2072                 goto unlock;
2073
2074         ret = bch2_buffered_write(iocb, from);
2075         if (likely(ret > 0))
2076                 iocb->ki_pos += ret;
2077 unlock:
2078         inode_unlock(&inode->v);
2079
2080         if (ret > 0)
2081                 ret = generic_write_sync(iocb, ret);
2082
2083         return ret;
2084 }
2085
2086 /* fsync: */
2087
2088 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2089 {
2090         struct bch_inode_info *inode = file_bch_inode(file);
2091         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2092         int ret, ret2;
2093
2094         ret = file_write_and_wait_range(file, start, end);
2095         if (ret)
2096                 return ret;
2097
2098         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2099                 goto out;
2100
2101         ret = sync_inode_metadata(&inode->v, 1);
2102         if (ret)
2103                 return ret;
2104 out:
2105         if (!c->opts.journal_flush_disabled)
2106                 ret = bch2_journal_flush_seq(&c->journal,
2107                                              inode->ei_journal_seq);
2108         ret2 = file_check_and_advance_wb_err(file);
2109
2110         return ret ?: ret2;
2111 }
2112
2113 /* truncate: */
2114
2115 static inline int range_has_data(struct bch_fs *c,
2116                                   struct bpos start,
2117                                   struct bpos end)
2118 {
2119         struct btree_trans trans;
2120         struct btree_iter *iter;
2121         struct bkey_s_c k;
2122         int ret = 0;
2123
2124         bch2_trans_init(&trans, c, 0, 0);
2125
2126         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2127                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2128                         break;
2129
2130                 if (bkey_extent_is_data(k.k)) {
2131                         ret = 1;
2132                         break;
2133                 }
2134         }
2135
2136         return bch2_trans_exit(&trans) ?: ret;
2137 }
2138
2139 static int __bch2_truncate_page(struct bch_inode_info *inode,
2140                                 pgoff_t index, loff_t start, loff_t end)
2141 {
2142         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2143         struct address_space *mapping = inode->v.i_mapping;
2144         struct bch_page_state *s;
2145         unsigned start_offset = start & (PAGE_SIZE - 1);
2146         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2147         unsigned i;
2148         struct page *page;
2149         int ret = 0;
2150
2151         /* Page boundary? Nothing to do */
2152         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2153               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2154                 return 0;
2155
2156         /* Above i_size? */
2157         if (index << PAGE_SHIFT >= inode->v.i_size)
2158                 return 0;
2159
2160         page = find_lock_page(mapping, index);
2161         if (!page) {
2162                 /*
2163                  * XXX: we're doing two index lookups when we end up reading the
2164                  * page
2165                  */
2166                 ret = range_has_data(c,
2167                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2168                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2169                 if (ret <= 0)
2170                         return ret;
2171
2172                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2173                 if (unlikely(!page)) {
2174                         ret = -ENOMEM;
2175                         goto out;
2176                 }
2177         }
2178
2179         s = bch2_page_state_create(page, 0);
2180         if (!s) {
2181                 ret = -ENOMEM;
2182                 goto unlock;
2183         }
2184
2185         if (!PageUptodate(page)) {
2186                 ret = bch2_read_single_page(page, mapping);
2187                 if (ret)
2188                         goto unlock;
2189         }
2190
2191         if (index != start >> PAGE_SHIFT)
2192                 start_offset = 0;
2193         if (index != end >> PAGE_SHIFT)
2194                 end_offset = PAGE_SIZE;
2195
2196         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2197              i < round_down(end_offset, block_bytes(c)) >> 9;
2198              i++) {
2199                 s->s[i].nr_replicas     = 0;
2200                 s->s[i].state           = SECTOR_UNALLOCATED;
2201         }
2202
2203         zero_user_segment(page, start_offset, end_offset);
2204
2205         /*
2206          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2207          *
2208          * XXX: because we aren't currently tracking whether the page has actual
2209          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2210          */
2211         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2212         BUG_ON(ret);
2213
2214         /*
2215          * This removes any writeable userspace mappings; we need to force
2216          * .page_mkwrite to be called again before any mmapped writes, to
2217          * redirty the full page:
2218          */
2219         page_mkclean(page);
2220         filemap_dirty_folio(mapping, page_folio(page));
2221 unlock:
2222         unlock_page(page);
2223         put_page(page);
2224 out:
2225         return ret;
2226 }
2227
2228 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2229 {
2230         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2231                                     from, round_up(from, PAGE_SIZE));
2232 }
2233
2234 static int bch2_extend(struct bch_inode_info *inode,
2235                        struct bch_inode_unpacked *inode_u,
2236                        struct iattr *iattr)
2237 {
2238         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2239         struct address_space *mapping = inode->v.i_mapping;
2240         int ret;
2241
2242         /*
2243          * sync appends:
2244          *
2245          * this has to be done _before_ extending i_size:
2246          */
2247         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2248         if (ret)
2249                 return ret;
2250
2251         truncate_setsize(&inode->v, iattr->ia_size);
2252         /* ATTR_MODE will never be set here, ns argument isn't needed: */
2253         setattr_copy(NULL, &inode->v, iattr);
2254
2255         mutex_lock(&inode->ei_update_lock);
2256         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2257                                     ATTR_MTIME|ATTR_CTIME);
2258         mutex_unlock(&inode->ei_update_lock);
2259
2260         return ret;
2261 }
2262
2263 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2264                                    struct bch_inode_unpacked *bi,
2265                                    void *p)
2266 {
2267         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2268
2269         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2270         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2271         return 0;
2272 }
2273
2274 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2275                                   struct bch_inode_unpacked *bi, void *p)
2276 {
2277         u64 *new_i_size = p;
2278
2279         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2280         bi->bi_size = *new_i_size;
2281         return 0;
2282 }
2283
2284 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2285 {
2286         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2287         struct address_space *mapping = inode->v.i_mapping;
2288         struct bch_inode_unpacked inode_u;
2289         struct btree_trans trans;
2290         struct btree_iter *iter;
2291         u64 new_i_size = iattr->ia_size;
2292         s64 i_sectors_delta = 0;
2293         int ret = 0;
2294
2295         inode_dio_wait(&inode->v);
2296         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2297
2298         /*
2299          * fetch current on disk i_size: inode is locked, i_size can only
2300          * increase underneath us:
2301          */
2302         bch2_trans_init(&trans, c, 0, 0);
2303         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2304         ret = PTR_ERR_OR_ZERO(iter);
2305         bch2_trans_exit(&trans);
2306
2307         if (ret)
2308                 goto err;
2309
2310         /*
2311          * check this before next assertion; on filesystem error our normal
2312          * invariants are a bit broken (truncate has to truncate the page cache
2313          * before the inode).
2314          */
2315         ret = bch2_journal_error(&c->journal);
2316         if (ret)
2317                 goto err;
2318
2319         WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2320                 inode->v.i_size < inode_u.bi_size);
2321
2322         if (iattr->ia_size > inode->v.i_size) {
2323                 ret = bch2_extend(inode, &inode_u, iattr);
2324                 goto err;
2325         }
2326
2327         ret = bch2_truncate_page(inode, iattr->ia_size);
2328         if (unlikely(ret))
2329                 goto err;
2330
2331         /*
2332          * When extending, we're going to write the new i_size to disk
2333          * immediately so we need to flush anything above the current on disk
2334          * i_size first:
2335          *
2336          * Also, when extending we need to flush the page that i_size currently
2337          * straddles - if it's mapped to userspace, we need to ensure that
2338          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2339          * again to allocate the part of the page that was extended.
2340          */
2341         if (iattr->ia_size > inode_u.bi_size)
2342                 ret = filemap_write_and_wait_range(mapping,
2343                                 inode_u.bi_size,
2344                                 iattr->ia_size - 1);
2345         else if (iattr->ia_size & (PAGE_SIZE - 1))
2346                 ret = filemap_write_and_wait_range(mapping,
2347                                 round_down(iattr->ia_size, PAGE_SIZE),
2348                                 iattr->ia_size - 1);
2349         if (ret)
2350                 goto err;
2351
2352         mutex_lock(&inode->ei_update_lock);
2353         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2354                                &new_i_size, 0);
2355         mutex_unlock(&inode->ei_update_lock);
2356
2357         if (unlikely(ret))
2358                 goto err;
2359
2360         truncate_setsize(&inode->v, iattr->ia_size);
2361
2362         ret = bch2_fpunch(c, inode->v.i_ino,
2363                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2364                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2365         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2366
2367         if (unlikely(ret))
2368                 goto err;
2369
2370         /* ATTR_MODE will never be set here, ns argument isn't needed: */
2371         setattr_copy(NULL, &inode->v, iattr);
2372
2373         mutex_lock(&inode->ei_update_lock);
2374         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2375                                ATTR_MTIME|ATTR_CTIME);
2376         mutex_unlock(&inode->ei_update_lock);
2377 err:
2378         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2379         return ret;
2380 }
2381
2382 /* fallocate: */
2383
2384 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2385 {
2386         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2387         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2388         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2389         int ret = 0;
2390
2391         inode_lock(&inode->v);
2392         inode_dio_wait(&inode->v);
2393         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2394
2395         ret = __bch2_truncate_page(inode,
2396                                    offset >> PAGE_SHIFT,
2397                                    offset, offset + len);
2398         if (unlikely(ret))
2399                 goto err;
2400
2401         if (offset >> PAGE_SHIFT !=
2402             (offset + len) >> PAGE_SHIFT) {
2403                 ret = __bch2_truncate_page(inode,
2404                                            (offset + len) >> PAGE_SHIFT,
2405                                            offset, offset + len);
2406                 if (unlikely(ret))
2407                         goto err;
2408         }
2409
2410         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2411
2412         if (discard_start < discard_end) {
2413                 s64 i_sectors_delta = 0;
2414
2415                 ret = bch2_fpunch(c, inode->v.i_ino,
2416                                   discard_start, discard_end,
2417                                   &inode->ei_journal_seq,
2418                                   &i_sectors_delta);
2419                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2420         }
2421 err:
2422         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2423         inode_unlock(&inode->v);
2424
2425         return ret;
2426 }
2427
2428 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2429                                    loff_t offset, loff_t len,
2430                                    bool insert)
2431 {
2432         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2433         struct address_space *mapping = inode->v.i_mapping;
2434         struct bkey_buf copy;
2435         struct btree_trans trans;
2436         struct btree_iter *src, *dst;
2437         loff_t shift, new_size;
2438         u64 src_start;
2439         int ret;
2440
2441         if ((offset | len) & (block_bytes(c) - 1))
2442                 return -EINVAL;
2443
2444         bch2_bkey_buf_init(&copy);
2445         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2446
2447         /*
2448          * We need i_mutex to keep the page cache consistent with the extents
2449          * btree, and the btree consistent with i_size - we don't need outside
2450          * locking for the extents btree itself, because we're using linked
2451          * iterators
2452          */
2453         inode_lock(&inode->v);
2454         inode_dio_wait(&inode->v);
2455         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2456
2457         if (insert) {
2458                 ret = -EFBIG;
2459                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2460                         goto err;
2461
2462                 ret = -EINVAL;
2463                 if (offset >= inode->v.i_size)
2464                         goto err;
2465
2466                 src_start       = U64_MAX;
2467                 shift           = len;
2468         } else {
2469                 ret = -EINVAL;
2470                 if (offset + len >= inode->v.i_size)
2471                         goto err;
2472
2473                 src_start       = offset + len;
2474                 shift           = -len;
2475         }
2476
2477         new_size = inode->v.i_size + shift;
2478
2479         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2480         if (ret)
2481                 goto err;
2482
2483         if (insert) {
2484                 i_size_write(&inode->v, new_size);
2485                 mutex_lock(&inode->ei_update_lock);
2486                 ret = bch2_write_inode_size(c, inode, new_size,
2487                                             ATTR_MTIME|ATTR_CTIME);
2488                 mutex_unlock(&inode->ei_update_lock);
2489         } else {
2490                 s64 i_sectors_delta = 0;
2491
2492                 ret = bch2_fpunch(c, inode->v.i_ino,
2493                                   offset >> 9, (offset + len) >> 9,
2494                                   &inode->ei_journal_seq,
2495                                   &i_sectors_delta);
2496                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2497
2498                 if (ret)
2499                         goto err;
2500         }
2501
2502         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2503                         POS(inode->v.i_ino, src_start >> 9),
2504                         BTREE_ITER_INTENT);
2505         dst = bch2_trans_copy_iter(&trans, src);
2506
2507         while (1) {
2508                 struct disk_reservation disk_res =
2509                         bch2_disk_reservation_init(c, 0);
2510                 struct bkey_i delete;
2511                 struct bkey_s_c k;
2512                 struct bpos next_pos;
2513                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2514                 struct bpos atomic_end;
2515                 unsigned trigger_flags = 0;
2516
2517                 k = insert
2518                         ? bch2_btree_iter_peek_prev(src)
2519                         : bch2_btree_iter_peek(src);
2520                 if ((ret = bkey_err(k)))
2521                         goto bkey_err;
2522
2523                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2524                         break;
2525
2526                 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2527
2528                 if (insert &&
2529                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2530                         break;
2531 reassemble:
2532                 bch2_bkey_buf_reassemble(&copy, c, k);
2533
2534                 if (insert &&
2535                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2536                         bch2_cut_front(move_pos, copy.k);
2537
2538                 copy.k->k.p.offset += shift >> 9;
2539                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2540
2541                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2542                 if (ret)
2543                         goto bkey_err;
2544
2545                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2546                         if (insert) {
2547                                 move_pos = atomic_end;
2548                                 move_pos.offset -= shift >> 9;
2549                                 goto reassemble;
2550                         } else {
2551                                 bch2_cut_back(atomic_end, copy.k);
2552                         }
2553                 }
2554
2555                 bkey_init(&delete.k);
2556                 delete.k.p = copy.k->k.p;
2557                 delete.k.size = copy.k->k.size;
2558                 delete.k.p.offset -= shift >> 9;
2559
2560                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2561
2562                 if (copy.k->k.size == k.k->size) {
2563                         /*
2564                          * If we're moving the entire extent, we can skip
2565                          * running triggers:
2566                          */
2567                         trigger_flags |= BTREE_TRIGGER_NORUN;
2568                 } else {
2569                         /* We might end up splitting compressed extents: */
2570                         unsigned nr_ptrs =
2571                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2572
2573                         ret = bch2_disk_reservation_get(c, &disk_res,
2574                                         copy.k->k.size, nr_ptrs,
2575                                         BCH_DISK_RESERVATION_NOFAIL);
2576                         BUG_ON(ret);
2577                 }
2578
2579                 bch2_btree_iter_set_pos(src, bkey_start_pos(&delete.k));
2580
2581                 ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
2582                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2583                         bch2_trans_commit(&trans, &disk_res,
2584                                           &inode->ei_journal_seq,
2585                                           BTREE_INSERT_NOFAIL);
2586                 bch2_disk_reservation_put(c, &disk_res);
2587 bkey_err:
2588                 if (!ret)
2589                         bch2_btree_iter_set_pos(src, next_pos);
2590
2591                 if (ret == -EINTR)
2592                         ret = 0;
2593                 if (ret)
2594                         goto err;
2595
2596                 bch2_trans_cond_resched(&trans);
2597         }
2598         bch2_trans_unlock(&trans);
2599
2600         if (!insert) {
2601                 i_size_write(&inode->v, new_size);
2602                 mutex_lock(&inode->ei_update_lock);
2603                 ret = bch2_write_inode_size(c, inode, new_size,
2604                                             ATTR_MTIME|ATTR_CTIME);
2605                 mutex_unlock(&inode->ei_update_lock);
2606         }
2607 err:
2608         bch2_trans_exit(&trans);
2609         bch2_bkey_buf_exit(&copy, c);
2610         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2611         inode_unlock(&inode->v);
2612         return ret;
2613 }
2614
2615 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2616                             loff_t offset, loff_t len)
2617 {
2618         struct address_space *mapping = inode->v.i_mapping;
2619         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2620         struct btree_trans trans;
2621         struct btree_iter *iter;
2622         struct bpos end_pos;
2623         loff_t end              = offset + len;
2624         loff_t block_start      = round_down(offset,    block_bytes(c));
2625         loff_t block_end        = round_up(end,         block_bytes(c));
2626         unsigned sectors;
2627         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2628         int ret;
2629
2630         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2631
2632         inode_lock(&inode->v);
2633         inode_dio_wait(&inode->v);
2634         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2635
2636         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2637                 ret = inode_newsize_ok(&inode->v, end);
2638                 if (ret)
2639                         goto err;
2640         }
2641
2642         if (mode & FALLOC_FL_ZERO_RANGE) {
2643                 ret = __bch2_truncate_page(inode,
2644                                            offset >> PAGE_SHIFT,
2645                                            offset, end);
2646
2647                 if (!ret &&
2648                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2649                         ret = __bch2_truncate_page(inode,
2650                                                    end >> PAGE_SHIFT,
2651                                                    offset, end);
2652
2653                 if (unlikely(ret))
2654                         goto err;
2655
2656                 truncate_pagecache_range(&inode->v, offset, end - 1);
2657         }
2658
2659         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2660                         POS(inode->v.i_ino, block_start >> 9),
2661                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2662         end_pos = POS(inode->v.i_ino, block_end >> 9);
2663
2664         while (bkey_cmp(iter->pos, end_pos) < 0) {
2665                 s64 i_sectors_delta = 0;
2666                 struct disk_reservation disk_res = { 0 };
2667                 struct quota_res quota_res = { 0 };
2668                 struct bkey_i_reservation reservation;
2669                 struct bkey_s_c k;
2670
2671                 bch2_trans_begin(&trans);
2672
2673                 k = bch2_btree_iter_peek_slot(iter);
2674                 if ((ret = bkey_err(k)))
2675                         goto bkey_err;
2676
2677                 /* already reserved */
2678                 if (k.k->type == KEY_TYPE_reservation &&
2679                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2680                         bch2_btree_iter_next_slot(iter);
2681                         continue;
2682                 }
2683
2684                 if (bkey_extent_is_data(k.k) &&
2685                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2686                         bch2_btree_iter_next_slot(iter);
2687                         continue;
2688                 }
2689
2690                 bkey_reservation_init(&reservation.k_i);
2691                 reservation.k.type      = KEY_TYPE_reservation;
2692                 reservation.k.p         = k.k->p;
2693                 reservation.k.size      = k.k->size;
2694
2695                 bch2_cut_front(iter->pos,       &reservation.k_i);
2696                 bch2_cut_back(end_pos,          &reservation.k_i);
2697
2698                 sectors = reservation.k.size;
2699                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2700
2701                 if (!bkey_extent_is_allocation(k.k)) {
2702                         ret = bch2_quota_reservation_add(c, inode,
2703                                         &quota_res,
2704                                         sectors, true);
2705                         if (unlikely(ret))
2706                                 goto bkey_err;
2707                 }
2708
2709                 if (reservation.v.nr_replicas < replicas ||
2710                     bch2_bkey_sectors_compressed(k)) {
2711                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2712                                                         replicas, 0);
2713                         if (unlikely(ret))
2714                                 goto bkey_err;
2715
2716                         reservation.v.nr_replicas = disk_res.nr_replicas;
2717                 }
2718
2719                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2720                                 &disk_res, &inode->ei_journal_seq,
2721                                 0, &i_sectors_delta);
2722                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2723 bkey_err:
2724                 bch2_quota_reservation_put(c, inode, &quota_res);
2725                 bch2_disk_reservation_put(c, &disk_res);
2726                 if (ret == -EINTR)
2727                         ret = 0;
2728                 if (ret)
2729                         goto err;
2730         }
2731
2732         /*
2733          * Do we need to extend the file?
2734          *
2735          * If we zeroed up to the end of the file, we dropped whatever writes
2736          * were going to write out the current i_size, so we have to extend
2737          * manually even if FL_KEEP_SIZE was set:
2738          */
2739         if (end >= inode->v.i_size &&
2740             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2741              (mode & FALLOC_FL_ZERO_RANGE))) {
2742                 struct btree_iter *inode_iter;
2743                 struct bch_inode_unpacked inode_u;
2744
2745                 do {
2746                         bch2_trans_begin(&trans);
2747                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2748                                                      inode->v.i_ino, 0);
2749                         ret = PTR_ERR_OR_ZERO(inode_iter);
2750                 } while (ret == -EINTR);
2751
2752                 bch2_trans_unlock(&trans);
2753
2754                 if (ret)
2755                         goto err;
2756
2757                 /*
2758                  * Sync existing appends before extending i_size,
2759                  * as in bch2_extend():
2760                  */
2761                 ret = filemap_write_and_wait_range(mapping,
2762                                         inode_u.bi_size, S64_MAX);
2763                 if (ret)
2764                         goto err;
2765
2766                 if (mode & FALLOC_FL_KEEP_SIZE)
2767                         end = inode->v.i_size;
2768                 else
2769                         i_size_write(&inode->v, end);
2770
2771                 mutex_lock(&inode->ei_update_lock);
2772                 ret = bch2_write_inode_size(c, inode, end, 0);
2773                 mutex_unlock(&inode->ei_update_lock);
2774         }
2775 err:
2776         bch2_trans_exit(&trans);
2777         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2778         inode_unlock(&inode->v);
2779         return ret;
2780 }
2781
2782 long bch2_fallocate_dispatch(struct file *file, int mode,
2783                              loff_t offset, loff_t len)
2784 {
2785         struct bch_inode_info *inode = file_bch_inode(file);
2786         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2787         long ret;
2788
2789         if (!percpu_ref_tryget(&c->writes))
2790                 return -EROFS;
2791
2792         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2793                 ret = bchfs_fallocate(inode, mode, offset, len);
2794         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2795                 ret = bchfs_fpunch(inode, offset, len);
2796         else if (mode == FALLOC_FL_INSERT_RANGE)
2797                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2798         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2799                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2800         else
2801                 ret = -EOPNOTSUPP;
2802
2803         percpu_ref_put(&c->writes);
2804
2805         return ret;
2806 }
2807
2808 static void mark_range_unallocated(struct bch_inode_info *inode,
2809                                    loff_t start, loff_t end)
2810 {
2811         pgoff_t index = start >> PAGE_SHIFT;
2812         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2813         struct folio_batch fbatch;
2814         unsigned i, j;
2815
2816         folio_batch_init(&fbatch);
2817
2818         while (filemap_get_folios(inode->v.i_mapping,
2819                                   &index, end_index, &fbatch)) {
2820                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2821                         struct folio *folio = fbatch.folios[i];
2822                         struct bch_page_state *s;
2823
2824                         folio_lock(folio);
2825                         s = bch2_page_state(&folio->page);
2826
2827                         if (s) {
2828                                 spin_lock(&s->lock);
2829                                 for (j = 0; j < PAGE_SECTORS; j++)
2830                                         s->s[j].nr_replicas = 0;
2831                                 spin_unlock(&s->lock);
2832                         }
2833
2834                         folio_unlock(folio);
2835                 }
2836                 folio_batch_release(&fbatch);
2837                 cond_resched();
2838         }
2839 }
2840
2841 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2842                              struct file *file_dst, loff_t pos_dst,
2843                              loff_t len, unsigned remap_flags)
2844 {
2845         struct bch_inode_info *src = file_bch_inode(file_src);
2846         struct bch_inode_info *dst = file_bch_inode(file_dst);
2847         struct bch_fs *c = src->v.i_sb->s_fs_info;
2848         s64 i_sectors_delta = 0;
2849         u64 aligned_len;
2850         loff_t ret = 0;
2851
2852         if (!c->opts.reflink)
2853                 return -EOPNOTSUPP;
2854
2855         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2856                 return -EINVAL;
2857
2858         if (remap_flags & REMAP_FILE_DEDUP)
2859                 return -EOPNOTSUPP;
2860
2861         if ((pos_src & (block_bytes(c) - 1)) ||
2862             (pos_dst & (block_bytes(c) - 1)))
2863                 return -EINVAL;
2864
2865         if (src == dst &&
2866             abs(pos_src - pos_dst) < len)
2867                 return -EINVAL;
2868
2869         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2870
2871         file_update_time(file_dst);
2872
2873         inode_dio_wait(&src->v);
2874         inode_dio_wait(&dst->v);
2875
2876         ret = generic_remap_file_range_prep(file_src, pos_src,
2877                                             file_dst, pos_dst,
2878                                             &len, remap_flags);
2879         if (ret < 0 || len == 0)
2880                 goto err;
2881
2882         aligned_len = round_up((u64) len, block_bytes(c));
2883
2884         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2885                                 pos_dst, pos_dst + len - 1);
2886         if (ret)
2887                 goto err;
2888
2889         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2890
2891         ret = bch2_remap_range(c,
2892                                POS(dst->v.i_ino, pos_dst >> 9),
2893                                POS(src->v.i_ino, pos_src >> 9),
2894                                aligned_len >> 9,
2895                                &dst->ei_journal_seq,
2896                                pos_dst + len, &i_sectors_delta);
2897         if (ret < 0)
2898                 goto err;
2899
2900         /*
2901          * due to alignment, we might have remapped slightly more than requsted
2902          */
2903         ret = min((u64) ret << 9, (u64) len);
2904
2905         /* XXX get a quota reservation */
2906         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2907
2908         spin_lock(&dst->v.i_lock);
2909         if (pos_dst + ret > dst->v.i_size)
2910                 i_size_write(&dst->v, pos_dst + ret);
2911         spin_unlock(&dst->v.i_lock);
2912 err:
2913         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2914
2915         return ret;
2916 }
2917
2918 /* fseek: */
2919
2920 static int folio_data_offset(struct folio *folio, unsigned offset)
2921 {
2922         struct bch_page_state *s = bch2_page_state(&folio->page);
2923         unsigned i;
2924
2925         if (s)
2926                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2927                         if (s->s[i].state >= SECTOR_DIRTY)
2928                                 return i << 9;
2929
2930         return -1;
2931 }
2932
2933 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2934                                        loff_t start_offset,
2935                                        loff_t end_offset)
2936 {
2937         struct folio_batch fbatch;
2938         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2939         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2940         pgoff_t index           = start_index;
2941         unsigned i;
2942         loff_t ret;
2943         int offset;
2944
2945         folio_batch_init(&fbatch);
2946
2947         while (filemap_get_folios(vinode->i_mapping,
2948                                   &index, end_index, &fbatch)) {
2949                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
2950                         struct folio *folio = fbatch.folios[i];
2951
2952                         folio_lock(folio);
2953                         offset = folio_data_offset(folio,
2954                                         folio->index == start_index
2955                                         ? start_offset & (PAGE_SIZE - 1)
2956                                         : 0);
2957                         if (offset >= 0) {
2958                                 ret = clamp(((loff_t) folio->index << PAGE_SHIFT) +
2959                                             offset,
2960                                             start_offset, end_offset);
2961                                 folio_unlock(folio);
2962                                 folio_batch_release(&fbatch);
2963                                 return ret;
2964                         }
2965                         folio_unlock(folio);
2966                 }
2967                 folio_batch_release(&fbatch);
2968                 cond_resched();
2969         }
2970
2971         return end_offset;
2972 }
2973
2974 static loff_t bch2_seek_data(struct file *file, u64 offset)
2975 {
2976         struct bch_inode_info *inode = file_bch_inode(file);
2977         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2978         struct btree_trans trans;
2979         struct btree_iter *iter;
2980         struct bkey_s_c k;
2981         u64 isize, next_data = MAX_LFS_FILESIZE;
2982         int ret;
2983
2984         isize = i_size_read(&inode->v);
2985         if (offset >= isize)
2986                 return -ENXIO;
2987
2988         bch2_trans_init(&trans, c, 0, 0);
2989
2990         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2991                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2992                 if (k.k->p.inode != inode->v.i_ino) {
2993                         break;
2994                 } else if (bkey_extent_is_data(k.k)) {
2995                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2996                         break;
2997                 } else if (k.k->p.offset >> 9 > isize)
2998                         break;
2999         }
3000
3001         ret = bch2_trans_exit(&trans) ?: ret;
3002         if (ret)
3003                 return ret;
3004
3005         if (next_data > offset)
3006                 next_data = bch2_seek_pagecache_data(&inode->v,
3007                                                      offset, next_data);
3008
3009         if (next_data >= isize)
3010                 return -ENXIO;
3011
3012         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3013 }
3014
3015 static int __page_hole_offset(struct page *page, unsigned offset)
3016 {
3017         struct bch_page_state *s = bch2_page_state(page);
3018         unsigned i;
3019
3020         if (!s)
3021                 return 0;
3022
3023         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3024                 if (s->s[i].state < SECTOR_DIRTY)
3025                         return i << 9;
3026
3027         return -1;
3028 }
3029
3030 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3031 {
3032         pgoff_t index = offset >> PAGE_SHIFT;
3033         struct page *page;
3034         int pg_offset;
3035         loff_t ret = -1;
3036
3037         page = find_lock_page(mapping, index);
3038         if (!page)
3039                 return offset;
3040
3041         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3042         if (pg_offset >= 0)
3043                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3044
3045         unlock_page(page);
3046
3047         return ret;
3048 }
3049
3050 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3051                                        loff_t start_offset,
3052                                        loff_t end_offset)
3053 {
3054         struct address_space *mapping = vinode->i_mapping;
3055         loff_t offset = start_offset, hole;
3056
3057         while (offset < end_offset) {
3058                 hole = page_hole_offset(mapping, offset);
3059                 if (hole >= 0 && hole <= end_offset)
3060                         return max(start_offset, hole);
3061
3062                 offset += PAGE_SIZE;
3063                 offset &= PAGE_MASK;
3064         }
3065
3066         return end_offset;
3067 }
3068
3069 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3070 {
3071         struct bch_inode_info *inode = file_bch_inode(file);
3072         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3073         struct btree_trans trans;
3074         struct btree_iter *iter;
3075         struct bkey_s_c k;
3076         u64 isize, next_hole = MAX_LFS_FILESIZE;
3077         int ret;
3078
3079         isize = i_size_read(&inode->v);
3080         if (offset >= isize)
3081                 return -ENXIO;
3082
3083         bch2_trans_init(&trans, c, 0, 0);
3084
3085         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3086                            POS(inode->v.i_ino, offset >> 9),
3087                            BTREE_ITER_SLOTS, k, ret) {
3088                 if (k.k->p.inode != inode->v.i_ino) {
3089                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3090                                         offset, MAX_LFS_FILESIZE);
3091                         break;
3092                 } else if (!bkey_extent_is_data(k.k)) {
3093                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3094                                         max(offset, bkey_start_offset(k.k) << 9),
3095                                         k.k->p.offset << 9);
3096
3097                         if (next_hole < k.k->p.offset << 9)
3098                                 break;
3099                 } else {
3100                         offset = max(offset, bkey_start_offset(k.k) << 9);
3101                 }
3102         }
3103
3104         ret = bch2_trans_exit(&trans) ?: ret;
3105         if (ret)
3106                 return ret;
3107
3108         if (next_hole > isize)
3109                 next_hole = isize;
3110
3111         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3112 }
3113
3114 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3115 {
3116         switch (whence) {
3117         case SEEK_SET:
3118         case SEEK_CUR:
3119         case SEEK_END:
3120                 return generic_file_llseek(file, offset, whence);
3121         case SEEK_DATA:
3122                 return bch2_seek_data(file, offset);
3123         case SEEK_HOLE:
3124                 return bch2_seek_hole(file, offset);
3125         }
3126
3127         return -EINVAL;
3128 }
3129
3130 void bch2_fs_fsio_exit(struct bch_fs *c)
3131 {
3132         bioset_exit(&c->dio_write_bioset);
3133         bioset_exit(&c->dio_read_bioset);
3134         bioset_exit(&c->writepage_bioset);
3135 }
3136
3137 int bch2_fs_fsio_init(struct bch_fs *c)
3138 {
3139         int ret = 0;
3140
3141         pr_verbose_init(c->opts, "");
3142
3143         if (bioset_init(&c->writepage_bioset,
3144                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3145                         BIOSET_NEED_BVECS) ||
3146             bioset_init(&c->dio_read_bioset,
3147                         4, offsetof(struct dio_read, rbio.bio),
3148                         BIOSET_NEED_BVECS) ||
3149             bioset_init(&c->dio_write_bioset,
3150                         4, offsetof(struct dio_write, op.wbio.bio),
3151                         BIOSET_NEED_BVECS))
3152                 ret = -ENOMEM;
3153
3154         pr_verbose_init(c->opts, "ret %i", ret);
3155         return ret;
3156 }
3157
3158 #endif /* NO_BCACHEFS_FS */