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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "btree_update.h"
   7 #include "buckets.h"
   8 #include "clock.h"
   9 #include "error.h"
  10 #include "extents.h"
  11 #include "fs.h"
  12 #include "fs-io.h"
  13 #include "fsck.h"
  14 #include "inode.h"
  15 #include "journal.h"
  16 #include "io.h"
  17 #include "keylist.h"
  18 #include "quota.h"
  19 #include "reflink.h"
  20
  21 #include <linux/aio.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/falloc.h>
  24 #include <linux/migrate.h>
  25 #include <linux/mmu_context.h>
  26 #include <linux/pagevec.h>
  27 #include <linux/sched/signal.h>
  28 #include <linux/task_io_accounting_ops.h>
  29 #include <linux/uio.h>
  30 #include <linux/writeback.h>
  31
  32 #include <trace/events/bcachefs.h>
  33 #include <trace/events/writeback.h>
  34
  35 struct quota_res {
  36         u64                             sectors;
  37 };
  38
  39 struct bch_writepage_io {
  40         struct closure                  cl;
  41         struct bch_inode_info           *inode;
  42
  43         /* must be last: */
  44         struct bch_write_op             op;
  45 };
  46
  47 struct dio_write {
  48         struct closure                  cl;
  49         struct kiocb                    *req;
  50         struct mm_struct                *mm;
  51         unsigned                        loop:1,
  52                                         sync:1,
  53                                         free_iov:1;
  54         struct quota_res                quota_res;
  55
  56         struct iov_iter                 iter;
  57         struct iovec                    inline_vecs[2];
  58
  59         /* must be last: */
  60         struct bch_write_op             op;
  61 };
  62
  63 struct dio_read {
  64         struct closure                  cl;
  65         struct kiocb                    *req;
  66         long                            ret;
  67         struct bch_read_bio             rbio;
  68 };
  69
  70 /* pagecache_block must be held */
  71 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  72                                               loff_t start, loff_t end)
  73 {
  74         int ret;
  75
  76         /*
  77          * XXX: the way this is currently implemented, we can spin if a process
  78          * is continually redirtying a specific page
  79          */
  80         do {
  81                 if (!mapping->nrpages &&
  82                     !mapping->nrexceptional)
  83                         return 0;
  84
  85                 ret = filemap_write_and_wait_range(mapping, start, end);
  86                 if (ret)
  87                         break;
  88
  89                 if (!mapping->nrpages)
  90                         return 0;
  91
  92                 ret = invalidate_inode_pages2_range(mapping,
  93                                 start >> PAGE_SHIFT,
  94                                 end >> PAGE_SHIFT);
  95         } while (ret == -EBUSY);
  96
  97         return ret;
  98 }
  99
 100 /* quotas */
 101
 102 #ifdef CONFIG_BCACHEFS_QUOTA
 103
 104 static void bch2_quota_reservation_put(struct bch_fs *c,
 105                                        struct bch_inode_info *inode,
 106                                        struct quota_res *res)
 107 {
 108         if (!res->sectors)
 109                 return;
 110
 111         mutex_lock(&inode->ei_quota_lock);
 112         BUG_ON(res->sectors > inode->ei_quota_reserved);
 113
 114         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 115                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 116         inode->ei_quota_reserved -= res->sectors;
 117         mutex_unlock(&inode->ei_quota_lock);
 118
 119         res->sectors = 0;
 120 }
 121
 122 static int bch2_quota_reservation_add(struct bch_fs *c,
 123                                       struct bch_inode_info *inode,
 124                                       struct quota_res *res,
 125                                       unsigned sectors,
 126                                       bool check_enospc)
 127 {
 128         int ret;
 129
 130         mutex_lock(&inode->ei_quota_lock);
 131         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 132                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 133         if (likely(!ret)) {
 134                 inode->ei_quota_reserved += sectors;
 135                 res->sectors += sectors;
 136         }
 137         mutex_unlock(&inode->ei_quota_lock);
 138
 139         return ret;
 140 }
 141
 142 #else
 143
 144 static void bch2_quota_reservation_put(struct bch_fs *c,
 145                                        struct bch_inode_info *inode,
 146                                        struct quota_res *res)
 147 {
 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         return 0;
 157 }
 158
 159 #endif
 160
 161 /* i_size updates: */
 162
 163 struct inode_new_size {
 164         loff_t          new_size;
 165         u64             now;
 166         unsigned        fields;
 167 };
 168
 169 static int inode_set_size(struct bch_inode_info *inode,
 170                           struct bch_inode_unpacked *bi,
 171                           void *p)
 172 {
 173         struct inode_new_size *s = p;
 174
 175         bi->bi_size = s->new_size;
 176         if (s->fields & ATTR_ATIME)
 177                 bi->bi_atime = s->now;
 178         if (s->fields & ATTR_MTIME)
 179                 bi->bi_mtime = s->now;
 180         if (s->fields & ATTR_CTIME)
 181                 bi->bi_ctime = s->now;
 182
 183         return 0;
 184 }
 185
 186 int __must_check bch2_write_inode_size(struct bch_fs *c,
 187                                        struct bch_inode_info *inode,
 188                                        loff_t new_size, unsigned fields)
 189 {
 190         struct inode_new_size s = {
 191                 .new_size       = new_size,
 192                 .now            = bch2_current_time(c),
 193                 .fields         = fields,
 194         };
 195
 196         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 197 }
 198
 199 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 200                            struct quota_res *quota_res, s64 sectors)
 201 {
 202         if (!sectors)
 203                 return;
 204
 205         mutex_lock(&inode->ei_quota_lock);
 206 #ifdef CONFIG_BCACHEFS_QUOTA
 207         if (quota_res && sectors > 0) {
 208                 BUG_ON(sectors > quota_res->sectors);
 209                 BUG_ON(sectors > inode->ei_quota_reserved);
 210
 211                 quota_res->sectors -= sectors;
 212                 inode->ei_quota_reserved -= sectors;
 213         } else {
 214                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 215         }
 216 #endif
 217         inode->v.i_blocks += sectors;
 218         mutex_unlock(&inode->ei_quota_lock);
 219 }
 220
 221 /* page state: */
 222
 223 /* stored in page->private: */
 224
 225 struct bch_page_sector {
 226         /* Uncompressed, fully allocated replicas: */
 227         unsigned                nr_replicas:3;
 228
 229         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 230         unsigned                replicas_reserved:3;
 231
 232         /* i_sectors: */
 233         enum {
 234                 SECTOR_UNALLOCATED,
 235                 SECTOR_RESERVED,
 236                 SECTOR_DIRTY,
 237                 SECTOR_ALLOCATED,
 238         }                       state:2;
 239 };
 240
 241 struct bch_page_state {
 242         spinlock_t              lock;
 243         atomic_t                write_count;
 244         struct bch_page_sector  s[PAGE_SECTORS];
 245 };
 246
 247 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 248 {
 249         return page_has_private(page)
 250                 ? (struct bch_page_state *) page_private(page)
 251                 : NULL;
 252 }
 253
 254 static inline struct bch_page_state *bch2_page_state(struct page *page)
 255 {
 256         EBUG_ON(!PageLocked(page));
 257
 258         return __bch2_page_state(page);
 259 }
 260
 261 /* for newly allocated pages: */
 262 static void __bch2_page_state_release(struct page *page)
 263 {
 264         struct bch_page_state *s = __bch2_page_state(page);
 265
 266         if (!s)
 267                 return;
 268
 269         ClearPagePrivate(page);
 270         set_page_private(page, 0);
 271         put_page(page);
 272         kfree(s);
 273 }
 274
 275 static void bch2_page_state_release(struct page *page)
 276 {
 277         struct bch_page_state *s = bch2_page_state(page);
 278
 279         if (!s)
 280                 return;
 281
 282         ClearPagePrivate(page);
 283         set_page_private(page, 0);
 284         put_page(page);
 285         kfree(s);
 286 }
 287
 288 /* for newly allocated pages: */
 289 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 290                                                        gfp_t gfp)
 291 {
 292         struct bch_page_state *s;
 293
 294         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 295         if (!s)
 296                 return NULL;
 297
 298         spin_lock_init(&s->lock);
 299         /*
 300          * migrate_page_move_mapping() assumes that pages with private data
 301          * have their count elevated by 1.
 302          */
 303         get_page(page);
 304         set_page_private(page, (unsigned long) s);
 305         SetPagePrivate(page);
 306         return s;
 307 }
 308
 309 static struct bch_page_state *bch2_page_state_create(struct page *page,
 310                                                      gfp_t gfp)
 311 {
 312         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 313 }
 314
 315 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 316 {
 317         /* XXX: this should not be open coded */
 318         return inode->ei_inode.bi_data_replicas
 319                 ? inode->ei_inode.bi_data_replicas - 1
 320                 : c->opts.data_replicas;
 321 }
 322
 323 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 324                                                   unsigned nr_replicas)
 325 {
 326         return max(0, (int) nr_replicas -
 327                    s->nr_replicas -
 328                    s->replicas_reserved);
 329 }
 330
 331 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 332                                 struct bch_inode_info *inode,
 333                                 struct page *page, bool check_enospc)
 334 {
 335         struct bch_page_state *s = bch2_page_state_create(page, 0);
 336         unsigned nr_replicas = inode_nr_replicas(c, inode);
 337         struct disk_reservation disk_res = { 0 };
 338         unsigned i, disk_res_sectors = 0;
 339         int ret;
 340
 341         if (!s)
 342                 return -ENOMEM;
 343
 344         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 345                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 346
 347         if (!disk_res_sectors)
 348                 return 0;
 349
 350         ret = bch2_disk_reservation_get(c, &disk_res,
 351                                         disk_res_sectors, 1,
 352                                         !check_enospc
 353                                         ? BCH_DISK_RESERVATION_NOFAIL
 354                                         : 0);
 355         if (unlikely(ret))
 356                 return ret;
 357
 358         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 359                 s->s[i].replicas_reserved +=
 360                         sectors_to_reserve(&s->s[i], nr_replicas);
 361
 362         return 0;
 363 }
 364
 365 struct bch2_page_reservation {
 366         struct disk_reservation disk;
 367         struct quota_res        quota;
 368 };
 369
 370 static void bch2_page_reservation_init(struct bch_fs *c,
 371                         struct bch_inode_info *inode,
 372                         struct bch2_page_reservation *res)
 373 {
 374         memset(res, 0, sizeof(*res));
 375
 376         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 377 }
 378
 379 static void bch2_page_reservation_put(struct bch_fs *c,
 380                         struct bch_inode_info *inode,
 381                         struct bch2_page_reservation *res)
 382 {
 383         bch2_disk_reservation_put(c, &res->disk);
 384         bch2_quota_reservation_put(c, inode, &res->quota);
 385 }
 386
 387 static int bch2_page_reservation_get(struct bch_fs *c,
 388                         struct bch_inode_info *inode, struct page *page,
 389                         struct bch2_page_reservation *res,
 390                         unsigned offset, unsigned len, bool check_enospc)
 391 {
 392         struct bch_page_state *s = bch2_page_state_create(page, 0);
 393         unsigned i, disk_sectors = 0, quota_sectors = 0;
 394         int ret;
 395
 396         if (!s)
 397                 return -ENOMEM;
 398
 399         for (i = round_down(offset, block_bytes(c)) >> 9;
 400              i < round_up(offset + len, block_bytes(c)) >> 9;
 401              i++) {
 402                 disk_sectors += sectors_to_reserve(&s->s[i],
 403                                                 res->disk.nr_replicas);
 404                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 405         }
 406
 407         if (disk_sectors) {
 408                 ret = bch2_disk_reservation_add(c, &res->disk,
 409                                                 disk_sectors,
 410                                                 !check_enospc
 411                                                 ? BCH_DISK_RESERVATION_NOFAIL
 412                                                 : 0);
 413                 if (unlikely(ret))
 414                         return ret;
 415         }
 416
 417         if (quota_sectors) {
 418                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 419                                                  quota_sectors,
 420                                                  check_enospc);
 421                 if (unlikely(ret)) {
 422                         struct disk_reservation tmp = {
 423                                 .sectors = disk_sectors
 424                         };
 425
 426                         bch2_disk_reservation_put(c, &tmp);
 427                         res->disk.sectors -= disk_sectors;
 428                         return ret;
 429                 }
 430         }
 431
 432         return 0;
 433 }
 434
 435 static void bch2_clear_page_bits(struct page *page)
 436 {
 437         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 438         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 439         struct bch_page_state *s = bch2_page_state(page);
 440         struct disk_reservation disk_res = { 0 };
 441         int i, dirty_sectors = 0;
 442
 443         if (!s)
 444                 return;
 445
 446         EBUG_ON(!PageLocked(page));
 447         EBUG_ON(PageWriteback(page));
 448
 449         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 450                 disk_res.sectors += s->s[i].replicas_reserved;
 451                 s->s[i].replicas_reserved = 0;
 452
 453                 if (s->s[i].state == SECTOR_DIRTY) {
 454                         dirty_sectors++;
 455                         s->s[i].state = SECTOR_UNALLOCATED;
 456                 }
 457         }
 458
 459         bch2_disk_reservation_put(c, &disk_res);
 460
 461         if (dirty_sectors)
 462                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 463
 464         bch2_page_state_release(page);
 465 }
 466
 467 static void bch2_set_page_dirty(struct bch_fs *c,
 468                         struct bch_inode_info *inode, struct page *page,
 469                         struct bch2_page_reservation *res,
 470                         unsigned offset, unsigned len)
 471 {
 472         struct bch_page_state *s = bch2_page_state(page);
 473         unsigned i, dirty_sectors = 0;
 474
 475         WARN_ON((u64) page_offset(page) + offset + len >
 476                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 477
 478         spin_lock(&s->lock);
 479
 480         for (i = round_down(offset, block_bytes(c)) >> 9;
 481              i < round_up(offset + len, block_bytes(c)) >> 9;
 482              i++) {
 483                 unsigned sectors = sectors_to_reserve(&s->s[i],
 484                                                 res->disk.nr_replicas);
 485
 486                 BUG_ON(sectors > res->disk.sectors);
 487                 s->s[i].replicas_reserved += sectors;
 488                 res->disk.sectors -= sectors;
 489
 490                 if (s->s[i].state == SECTOR_UNALLOCATED)
 491                         dirty_sectors++;
 492
 493                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 494         }
 495
 496         spin_unlock(&s->lock);
 497
 498         if (dirty_sectors)
 499                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 500
 501         if (!PageDirty(page))
 502                 __set_page_dirty_nobuffers(page);
 503 }
 504
 505 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 506 {
 507         struct page *page = vmf->page;
 508         struct file *file = vmf->vma->vm_file;
 509         struct bch_inode_info *inode = file_bch_inode(file);
 510         struct address_space *mapping = inode->v.i_mapping;
 511         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 512         struct bch2_page_reservation res;
 513         unsigned len;
 514         loff_t isize;
 515         int ret = VM_FAULT_LOCKED;
 516
 517         bch2_page_reservation_init(c, inode, &res);
 518
 519         sb_start_pagefault(inode->v.i_sb);
 520         file_update_time(file);
 521
 522         /*
 523          * Not strictly necessary, but helps avoid dio writes livelocking in
 524          * write_invalidate_inode_pages_range() - can drop this if/when we get
 525          * a write_invalidate_inode_pages_range() that works without dropping
 526          * page lock before invalidating page
 527          */
 528         if (current->pagecache_lock != &mapping->add_lock)
 529                 pagecache_add_get(&mapping->add_lock);
 530
 531         lock_page(page);
 532         isize = i_size_read(&inode->v);
 533
 534         if (page->mapping != mapping || page_offset(page) >= isize) {
 535                 unlock_page(page);
 536                 ret = VM_FAULT_NOPAGE;
 537                 goto out;
 538         }
 539
 540         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 541
 542         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 543                 unlock_page(page);
 544                 ret = VM_FAULT_SIGBUS;
 545                 goto out;
 546         }
 547
 548         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 549         wait_for_stable_page(page);
 550 out:
 551         if (current->pagecache_lock != &mapping->add_lock)
 552                 pagecache_add_put(&mapping->add_lock);
 553         sb_end_pagefault(inode->v.i_sb);
 554
 555         bch2_page_reservation_put(c, inode, &res);
 556
 557         return ret;
 558 }
 559
 560 void bch2_invalidatepage(struct page *page, unsigned int offset,
 561                          unsigned int length)
 562 {
 563         if (offset || length < PAGE_SIZE)
 564                 return;
 565
 566         bch2_clear_page_bits(page);
 567 }
 568
 569 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 570 {
 571         if (PageDirty(page))
 572                 return 0;
 573
 574         bch2_clear_page_bits(page);
 575         return 1;
 576 }
 577
 578 #ifdef CONFIG_MIGRATION
 579 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 580                       struct page *page, enum migrate_mode mode)
 581 {
 582         int ret;
 583
 584         EBUG_ON(!PageLocked(page));
 585         EBUG_ON(!PageLocked(newpage));
 586
 587         ret = migrate_page_move_mapping(mapping, newpage, page, mode, 0);
 588         if (ret != MIGRATEPAGE_SUCCESS)
 589                 return ret;
 590
 591         if (PagePrivate(page)) {
 592                 ClearPagePrivate(page);
 593                 get_page(newpage);
 594                 set_page_private(newpage, page_private(page));
 595                 set_page_private(page, 0);
 596                 put_page(page);
 597                 SetPagePrivate(newpage);
 598         }
 599
 600         if (mode != MIGRATE_SYNC_NO_COPY)
 601                 migrate_page_copy(newpage, page);
 602         else
 603                 migrate_page_states(newpage, page);
 604         return MIGRATEPAGE_SUCCESS;
 605 }
 606 #endif
 607
 608 /* readpage(s): */
 609
 610 static void bch2_readpages_end_io(struct bio *bio)
 611 {
 612         struct bvec_iter_all iter;
 613         struct bio_vec *bv;
 614
 615         bio_for_each_segment_all(bv, bio, iter) {
 616                 struct page *page = bv->bv_page;
 617
 618                 if (!bio->bi_status) {
 619                         SetPageUptodate(page);
 620                 } else {
 621                         ClearPageUptodate(page);
 622                         SetPageError(page);
 623                 }
 624                 unlock_page(page);
 625         }
 626
 627         bio_put(bio);
 628 }
 629
 630 static inline void page_state_init_for_read(struct page *page)
 631 {
 632         SetPagePrivate(page);
 633         page->private = 0;
 634 }
 635
 636 struct readpages_iter {
 637         struct address_space    *mapping;
 638         struct page             **pages;
 639         unsigned                nr_pages;
 640         unsigned                nr_added;
 641         unsigned                idx;
 642         pgoff_t                 offset;
 643 };
 644
 645 static int readpages_iter_init(struct readpages_iter *iter,
 646                                struct address_space *mapping,
 647                                struct list_head *pages, unsigned nr_pages)
 648 {
 649         memset(iter, 0, sizeof(*iter));
 650
 651         iter->mapping   = mapping;
 652         iter->offset    = list_last_entry(pages, struct page, lru)->index;
 653
 654         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 655         if (!iter->pages)
 656                 return -ENOMEM;
 657
 658         while (!list_empty(pages)) {
 659                 struct page *page = list_last_entry(pages, struct page, lru);
 660
 661                 __bch2_page_state_create(page, __GFP_NOFAIL);
 662
 663                 iter->pages[iter->nr_pages++] = page;
 664                 list_del(&page->lru);
 665         }
 666
 667         return 0;
 668 }
 669
 670 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 671 {
 672         struct page *page;
 673         unsigned i;
 674         int ret;
 675
 676         BUG_ON(iter->idx > iter->nr_added);
 677         BUG_ON(iter->nr_added > iter->nr_pages);
 678
 679         if (iter->idx < iter->nr_added)
 680                 goto out;
 681
 682         while (1) {
 683                 if (iter->idx == iter->nr_pages)
 684                         return NULL;
 685
 686                 ret = add_to_page_cache_lru_vec(iter->mapping,
 687                                 iter->pages     + iter->nr_added,
 688                                 iter->nr_pages  - iter->nr_added,
 689                                 iter->offset    + iter->nr_added,
 690                                 GFP_NOFS);
 691                 if (ret > 0)
 692                         break;
 693
 694                 page = iter->pages[iter->nr_added];
 695                 iter->idx++;
 696                 iter->nr_added++;
 697
 698                 __bch2_page_state_release(page);
 699                 put_page(page);
 700         }
 701
 702         iter->nr_added += ret;
 703
 704         for (i = iter->idx; i < iter->nr_added; i++)
 705                 put_page(iter->pages[i]);
 706 out:
 707         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 708
 709         return iter->pages[iter->idx];
 710 }
 711
 712 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 713 {
 714         struct bvec_iter iter;
 715         struct bio_vec bv;
 716         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 717                 ? 0 : bch2_bkey_nr_ptrs_allocated(k);
 718         unsigned state = k.k->type == KEY_TYPE_reservation
 719                 ? SECTOR_RESERVED
 720                 : SECTOR_ALLOCATED;
 721
 722         bio_for_each_segment(bv, bio, iter) {
 723                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 724                 unsigned i;
 725
 726                 for (i = bv.bv_offset >> 9;
 727                      i < (bv.bv_offset + bv.bv_len) >> 9;
 728                      i++) {
 729                         s->s[i].nr_replicas = nr_ptrs;
 730                         s->s[i].state = state;
 731                 }
 732         }
 733 }
 734
 735 static void readpage_bio_extend(struct readpages_iter *iter,
 736                                 struct bio *bio,
 737                                 unsigned sectors_this_extent,
 738                                 bool get_more)
 739 {
 740         while (bio_sectors(bio) < sectors_this_extent &&
 741                bio->bi_vcnt < bio->bi_max_vecs) {
 742                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 743                 struct page *page = readpage_iter_next(iter);
 744                 int ret;
 745
 746                 if (page) {
 747                         if (iter->offset + iter->idx != page_offset)
 748                                 break;
 749
 750                         iter->idx++;
 751                 } else {
 752                         if (!get_more)
 753                                 break;
 754
 755                         page = xa_load(&iter->mapping->i_pages, page_offset);
 756                         if (page && !xa_is_value(page))
 757                                 break;
 758
 759                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 760                         if (!page)
 761                                 break;
 762
 763                         if (!__bch2_page_state_create(page, 0)) {
 764                                 put_page(page);
 765                                 break;
 766                         }
 767
 768                         ret = add_to_page_cache_lru(page, iter->mapping,
 769                                                     page_offset, GFP_NOFS);
 770                         if (ret) {
 771                                 __bch2_page_state_release(page);
 772                                 put_page(page);
 773                                 break;
 774                         }
 775
 776                         put_page(page);
 777                 }
 778
 779                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 780         }
 781 }
 782
 783 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 784                        struct bch_read_bio *rbio, u64 inum,
 785                        struct readpages_iter *readpages_iter)
 786 {
 787         struct bch_fs *c = trans->c;
 788         int flags = BCH_READ_RETRY_IF_STALE|
 789                 BCH_READ_MAY_PROMOTE;
 790         int ret = 0;
 791
 792         rbio->c = c;
 793         rbio->start_time = local_clock();
 794 retry:
 795         while (1) {
 796                 BKEY_PADDED(k) tmp;
 797                 struct bkey_s_c k;
 798                 unsigned bytes, sectors, offset_into_extent;
 799
 800                 bch2_btree_iter_set_pos(iter,
 801                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 802
 803                 k = bch2_btree_iter_peek_slot(iter);
 804                 ret = bkey_err(k);
 805                 if (ret)
 806                         break;
 807
 808                 bkey_reassemble(&tmp.k, k);
 809                 k = bkey_i_to_s_c(&tmp.k);
 810
 811                 offset_into_extent = iter->pos.offset -
 812                         bkey_start_offset(k.k);
 813                 sectors = k.k->size - offset_into_extent;
 814
 815                 ret = bch2_read_indirect_extent(trans,
 816                                         &offset_into_extent, &tmp.k);
 817                 if (ret)
 818                         break;
 819
 820                 sectors = min(sectors, k.k->size - offset_into_extent);
 821
 822                 bch2_trans_unlock(trans);
 823
 824                 if (readpages_iter) {
 825                         bool want_full_extent = false;
 826
 827                         if (bkey_extent_is_data(k.k)) {
 828                                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 829                                 const union bch_extent_entry *i;
 830                                 struct extent_ptr_decoded p;
 831
 832                                 bkey_for_each_ptr_decode(k.k, ptrs, p, i)
 833                                         want_full_extent |= ((p.crc.csum_type != 0) |
 834                                                              (p.crc.compression_type != 0));
 835                         }
 836
 837                         readpage_bio_extend(readpages_iter, &rbio->bio,
 838                                             sectors, want_full_extent);
 839                 }
 840
 841                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 842                 swap(rbio->bio.bi_iter.bi_size, bytes);
 843
 844                 if (rbio->bio.bi_iter.bi_size == bytes)
 845                         flags |= BCH_READ_LAST_FRAGMENT;
 846
 847                 if (bkey_extent_is_allocation(k.k))
 848                         bch2_add_page_sectors(&rbio->bio, k);
 849
 850                 bch2_read_extent(c, rbio, k, offset_into_extent, flags);
 851
 852                 if (flags & BCH_READ_LAST_FRAGMENT)
 853                         return;
 854
 855                 swap(rbio->bio.bi_iter.bi_size, bytes);
 856                 bio_advance(&rbio->bio, bytes);
 857         }
 858
 859         if (ret == -EINTR)
 860                 goto retry;
 861
 862         bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
 863         bio_endio(&rbio->bio);
 864 }
 865
 866 int bch2_readpages(struct file *file, struct address_space *mapping,
 867                    struct list_head *pages, unsigned nr_pages)
 868 {
 869         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 870         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 871         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 872         struct btree_trans trans;
 873         struct btree_iter *iter;
 874         struct page *page;
 875         struct readpages_iter readpages_iter;
 876         int ret;
 877
 878         ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
 879         BUG_ON(ret);
 880
 881         bch2_trans_init(&trans, c, 0, 0);
 882
 883         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 884                                    BTREE_ITER_SLOTS);
 885
 886         if (current->pagecache_lock != &mapping->add_lock)
 887                 pagecache_add_get(&mapping->add_lock);
 888
 889         while ((page = readpage_iter_next(&readpages_iter))) {
 890                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 891                 unsigned n = min_t(unsigned,
 892                                    readpages_iter.nr_pages -
 893                                    readpages_iter.idx,
 894                                    BIO_MAX_PAGES);
 895                 struct bch_read_bio *rbio =
 896                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 897                                   opts);
 898
 899                 readpages_iter.idx++;
 900
 901                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 902                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 903                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 904                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 905
 906                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 907                            &readpages_iter);
 908         }
 909
 910         if (current->pagecache_lock != &mapping->add_lock)
 911                 pagecache_add_put(&mapping->add_lock);
 912
 913         bch2_trans_exit(&trans);
 914         kfree(readpages_iter.pages);
 915
 916         return 0;
 917 }
 918
 919 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 920                              u64 inum, struct page *page)
 921 {
 922         struct btree_trans trans;
 923         struct btree_iter *iter;
 924
 925         bch2_page_state_create(page, __GFP_NOFAIL);
 926
 927         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 928         rbio->bio.bi_iter.bi_sector =
 929                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 930         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 931
 932         bch2_trans_init(&trans, c, 0, 0);
 933         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 934                                    BTREE_ITER_SLOTS);
 935
 936         bchfs_read(&trans, iter, rbio, inum, NULL);
 937
 938         bch2_trans_exit(&trans);
 939 }
 940
 941 int bch2_readpage(struct file *file, struct page *page)
 942 {
 943         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 944         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 945         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 946         struct bch_read_bio *rbio;
 947
 948         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 949         rbio->bio.bi_end_io = bch2_readpages_end_io;
 950
 951         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 952         return 0;
 953 }
 954
 955 static void bch2_read_single_page_end_io(struct bio *bio)
 956 {
 957         complete(bio->bi_private);
 958 }
 959
 960 static int bch2_read_single_page(struct page *page,
 961                                  struct address_space *mapping)
 962 {
 963         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 964         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 965         struct bch_read_bio *rbio;
 966         int ret;
 967         DECLARE_COMPLETION_ONSTACK(done);
 968
 969         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 970                          io_opts(c, &inode->ei_inode));
 971         rbio->bio.bi_private = &done;
 972         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 973
 974         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 975         wait_for_completion(&done);
 976
 977         ret = blk_status_to_errno(rbio->bio.bi_status);
 978         bio_put(&rbio->bio);
 979
 980         if (ret < 0)
 981                 return ret;
 982
 983         SetPageUptodate(page);
 984         return 0;
 985 }
 986
 987 /* writepages: */
 988
 989 struct bch_writepage_state {
 990         struct bch_writepage_io *io;
 991         struct bch_io_opts      opts;
 992 };
 993
 994 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
 995                                                                   struct bch_inode_info *inode)
 996 {
 997         return (struct bch_writepage_state) {
 998                 .opts = io_opts(c, &inode->ei_inode)
 999         };
1000 }
1001
1002 static void bch2_writepage_io_free(struct closure *cl)
1003 {
1004         struct bch_writepage_io *io = container_of(cl,
1005                                         struct bch_writepage_io, cl);
1006
1007         bio_put(&io->op.wbio.bio);
1008 }
1009
1010 static void bch2_writepage_io_done(struct closure *cl)
1011 {
1012         struct bch_writepage_io *io = container_of(cl,
1013                                         struct bch_writepage_io, cl);
1014         struct bch_fs *c = io->op.c;
1015         struct bio *bio = &io->op.wbio.bio;
1016         struct bvec_iter_all iter;
1017         struct bio_vec *bvec;
1018         unsigned i;
1019
1020         if (io->op.error) {
1021                 bio_for_each_segment_all(bvec, bio, iter) {
1022                         struct bch_page_state *s;
1023
1024                         SetPageError(bvec->bv_page);
1025                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1026
1027                         s = __bch2_page_state(bvec->bv_page);
1028                         spin_lock(&s->lock);
1029                         for (i = 0; i < PAGE_SECTORS; i++)
1030                                 s->s[i].nr_replicas = 0;
1031                         spin_unlock(&s->lock);
1032                 }
1033         }
1034
1035         /*
1036          * racing with fallocate can cause us to add fewer sectors than
1037          * expected - but we shouldn't add more sectors than expected:
1038          */
1039         BUG_ON(io->op.i_sectors_delta > 0);
1040
1041         /*
1042          * (error (due to going RO) halfway through a page can screw that up
1043          * slightly)
1044          * XXX wtf?
1045            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1046          */
1047
1048         /*
1049          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1050          * before calling end_page_writeback:
1051          */
1052         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1053
1054         bio_for_each_segment_all(bvec, bio, iter) {
1055                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1056
1057                 if (atomic_dec_and_test(&s->write_count))
1058                         end_page_writeback(bvec->bv_page);
1059         }
1060
1061         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1062 }
1063
1064 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1065 {
1066         struct bch_writepage_io *io = w->io;
1067
1068         w->io = NULL;
1069         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1070         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1071 }
1072
1073 /*
1074  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1075  * possible, else allocating a new one:
1076  */
1077 static void bch2_writepage_io_alloc(struct bch_fs *c,
1078                                     struct bch_writepage_state *w,
1079                                     struct bch_inode_info *inode,
1080                                     u64 sector,
1081                                     unsigned nr_replicas)
1082 {
1083         struct bch_write_op *op;
1084
1085         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1086                                               BIO_MAX_PAGES,
1087                                               &c->writepage_bioset),
1088                              struct bch_writepage_io, op.wbio.bio);
1089
1090         closure_init(&w->io->cl, NULL);
1091         w->io->inode            = inode;
1092
1093         op                      = &w->io->op;
1094         bch2_write_op_init(op, c, w->opts);
1095         op->target              = w->opts.foreground_target;
1096         op_journal_seq_set(op, &inode->ei_journal_seq);
1097         op->nr_replicas         = nr_replicas;
1098         op->res.nr_replicas     = nr_replicas;
1099         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1100         op->pos                 = POS(inode->v.i_ino, sector);
1101         op->wbio.bio.bi_iter.bi_sector = sector;
1102 }
1103
1104 static int __bch2_writepage(struct page *page,
1105                             struct writeback_control *wbc,
1106                             void *data)
1107 {
1108         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1109         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1110         struct bch_writepage_state *w = data;
1111         struct bch_page_state *s, orig;
1112         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1113         loff_t i_size = i_size_read(&inode->v);
1114         pgoff_t end_index = i_size >> PAGE_SHIFT;
1115         int ret;
1116
1117         EBUG_ON(!PageUptodate(page));
1118
1119         /* Is the page fully inside i_size? */
1120         if (page->index < end_index)
1121                 goto do_io;
1122
1123         /* Is the page fully outside i_size? (truncate in progress) */
1124         offset = i_size & (PAGE_SIZE - 1);
1125         if (page->index > end_index || !offset) {
1126                 unlock_page(page);
1127                 return 0;
1128         }
1129
1130         /*
1131          * The page straddles i_size.  It must be zeroed out on each and every
1132          * writepage invocation because it may be mmapped.  "A file is mapped
1133          * in multiples of the page size.  For a file that is not a multiple of
1134          * the  page size, the remaining memory is zeroed when mapped, and
1135          * writes to that region are not written out to the file."
1136          */
1137         zero_user_segment(page, offset, PAGE_SIZE);
1138 do_io:
1139         s = bch2_page_state_create(page, __GFP_NOFAIL);
1140
1141         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1142         if (ret) {
1143                 SetPageError(page);
1144                 mapping_set_error(page->mapping, ret);
1145                 unlock_page(page);
1146                 return 0;
1147         }
1148
1149         /* Before unlocking the page, get copy of reservations: */
1150         orig = *s;
1151
1152         for (i = 0; i < PAGE_SECTORS; i++) {
1153                 if (s->s[i].state < SECTOR_DIRTY)
1154                         continue;
1155
1156                 nr_replicas_this_write =
1157                         min_t(unsigned, nr_replicas_this_write,
1158                               s->s[i].nr_replicas +
1159                               s->s[i].replicas_reserved);
1160         }
1161
1162         for (i = 0; i < PAGE_SECTORS; i++) {
1163                 if (s->s[i].state < SECTOR_DIRTY)
1164                         continue;
1165
1166                 s->s[i].nr_replicas = w->opts.compression
1167                         ? 0 : nr_replicas_this_write;
1168
1169                 s->s[i].replicas_reserved = 0;
1170                 s->s[i].state = SECTOR_ALLOCATED;
1171         }
1172
1173         BUG_ON(atomic_read(&s->write_count));
1174         atomic_set(&s->write_count, 1);
1175
1176         BUG_ON(PageWriteback(page));
1177         set_page_writeback(page);
1178
1179         unlock_page(page);
1180
1181         offset = 0;
1182         while (1) {
1183                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1184                 u64 sector;
1185
1186                 while (offset < PAGE_SECTORS &&
1187                        orig.s[offset].state < SECTOR_DIRTY)
1188                         offset++;
1189
1190                 if (offset == PAGE_SECTORS)
1191                         break;
1192
1193                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1194
1195                 while (offset + sectors < PAGE_SECTORS &&
1196                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1197                         sectors++;
1198
1199                 for (i = offset; i < offset + sectors; i++) {
1200                         reserved_sectors += orig.s[i].replicas_reserved;
1201                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1202                 }
1203
1204                 if (w->io &&
1205                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1206                      bio_full(&w->io->op.wbio.bio) ||
1207                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1208                         bch2_writepage_do_io(w);
1209
1210                 if (!w->io)
1211                         bch2_writepage_io_alloc(c, w, inode, sector,
1212                                                 nr_replicas_this_write);
1213
1214                 atomic_inc(&s->write_count);
1215
1216                 BUG_ON(inode != w->io->inode);
1217                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1218                                      sectors << 9, offset << 9));
1219
1220                 /* Check for writing past i_size: */
1221                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1222                         round_up(i_size, block_bytes(c)));
1223
1224                 w->io->op.res.sectors += reserved_sectors;
1225                 w->io->op.i_sectors_delta -= dirty_sectors;
1226                 w->io->op.new_i_size = i_size;
1227
1228                 if (wbc->sync_mode == WB_SYNC_ALL)
1229                         w->io->op.wbio.bio.bi_opf |= REQ_SYNC;
1230
1231                 offset += sectors;
1232         }
1233
1234         if (atomic_dec_and_test(&s->write_count))
1235                 end_page_writeback(page);
1236
1237         return 0;
1238 }
1239
1240 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1241 {
1242         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1243         struct bch_writepage_state w =
1244                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1245         struct blk_plug plug;
1246         int ret;
1247
1248         blk_start_plug(&plug);
1249         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1250         if (w.io)
1251                 bch2_writepage_do_io(&w);
1252         blk_finish_plug(&plug);
1253         return ret;
1254 }
1255
1256 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1257 {
1258         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1259         struct bch_writepage_state w =
1260                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1261         int ret;
1262
1263         ret = __bch2_writepage(page, wbc, &w);
1264         if (w.io)
1265                 bch2_writepage_do_io(&w);
1266
1267         return ret;
1268 }
1269
1270 /* buffered writes: */
1271
1272 int bch2_write_begin(struct file *file, struct address_space *mapping,
1273                      loff_t pos, unsigned len, unsigned flags,
1274                      struct page **pagep, void **fsdata)
1275 {
1276         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1277         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1278         struct bch2_page_reservation *res;
1279         pgoff_t index = pos >> PAGE_SHIFT;
1280         unsigned offset = pos & (PAGE_SIZE - 1);
1281         struct page *page;
1282         int ret = -ENOMEM;
1283
1284         res = kmalloc(sizeof(*res), GFP_KERNEL);
1285         if (!res)
1286                 return -ENOMEM;
1287
1288         bch2_page_reservation_init(c, inode, res);
1289         *fsdata = res;
1290
1291         /* Not strictly necessary - same reason as mkwrite(): */
1292         pagecache_add_get(&mapping->add_lock);
1293
1294         page = grab_cache_page_write_begin(mapping, index, flags);
1295         if (!page)
1296                 goto err_unlock;
1297
1298         if (PageUptodate(page))
1299                 goto out;
1300
1301         /* If we're writing entire page, don't need to read it in first: */
1302         if (len == PAGE_SIZE)
1303                 goto out;
1304
1305         if (!offset && pos + len >= inode->v.i_size) {
1306                 zero_user_segment(page, len, PAGE_SIZE);
1307                 flush_dcache_page(page);
1308                 goto out;
1309         }
1310
1311         if (index > inode->v.i_size >> PAGE_SHIFT) {
1312                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1313                 flush_dcache_page(page);
1314                 goto out;
1315         }
1316 readpage:
1317         ret = bch2_read_single_page(page, mapping);
1318         if (ret)
1319                 goto err;
1320 out:
1321         ret = bch2_page_reservation_get(c, inode, page, res,
1322                                         offset, len, true);
1323         if (ret) {
1324                 if (!PageUptodate(page)) {
1325                         /*
1326                          * If the page hasn't been read in, we won't know if we
1327                          * actually need a reservation - we don't actually need
1328                          * to read here, we just need to check if the page is
1329                          * fully backed by uncompressed data:
1330                          */
1331                         goto readpage;
1332                 }
1333
1334                 goto err;
1335         }
1336
1337         *pagep = page;
1338         return 0;
1339 err:
1340         unlock_page(page);
1341         put_page(page);
1342         *pagep = NULL;
1343 err_unlock:
1344         pagecache_add_put(&mapping->add_lock);
1345         kfree(res);
1346         *fsdata = NULL;
1347         return ret;
1348 }
1349
1350 int bch2_write_end(struct file *file, struct address_space *mapping,
1351                    loff_t pos, unsigned len, unsigned copied,
1352                    struct page *page, void *fsdata)
1353 {
1354         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1355         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1356         struct bch2_page_reservation *res = fsdata;
1357         unsigned offset = pos & (PAGE_SIZE - 1);
1358
1359         lockdep_assert_held(&inode->v.i_rwsem);
1360
1361         if (unlikely(copied < len && !PageUptodate(page))) {
1362                 /*
1363                  * The page needs to be read in, but that would destroy
1364                  * our partial write - simplest thing is to just force
1365                  * userspace to redo the write:
1366                  */
1367                 zero_user(page, 0, PAGE_SIZE);
1368                 flush_dcache_page(page);
1369                 copied = 0;
1370         }
1371
1372         spin_lock(&inode->v.i_lock);
1373         if (pos + copied > inode->v.i_size)
1374                 i_size_write(&inode->v, pos + copied);
1375         spin_unlock(&inode->v.i_lock);
1376
1377         if (copied) {
1378                 if (!PageUptodate(page))
1379                         SetPageUptodate(page);
1380
1381                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1382
1383                 inode->ei_last_dirtied = (unsigned long) current;
1384         }
1385
1386         unlock_page(page);
1387         put_page(page);
1388         pagecache_add_put(&mapping->add_lock);
1389
1390         bch2_page_reservation_put(c, inode, res);
1391         kfree(res);
1392
1393         return copied;
1394 }
1395
1396 #define WRITE_BATCH_PAGES       32
1397
1398 static int __bch2_buffered_write(struct bch_inode_info *inode,
1399                                  struct address_space *mapping,
1400                                  struct iov_iter *iter,
1401                                  loff_t pos, unsigned len)
1402 {
1403         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1404         struct page *pages[WRITE_BATCH_PAGES];
1405         struct bch2_page_reservation res;
1406         unsigned long index = pos >> PAGE_SHIFT;
1407         unsigned offset = pos & (PAGE_SIZE - 1);
1408         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1409         unsigned i, reserved = 0, set_dirty = 0;
1410         unsigned copied = 0, nr_pages_copied = 0;
1411         int ret = 0;
1412
1413         BUG_ON(!len);
1414         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1415
1416         bch2_page_reservation_init(c, inode, &res);
1417
1418         for (i = 0; i < nr_pages; i++) {
1419                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1420                 if (!pages[i]) {
1421                         nr_pages = i;
1422                         if (!i) {
1423                                 ret = -ENOMEM;
1424                                 goto out;
1425                         }
1426                         len = min_t(unsigned, len,
1427                                     nr_pages * PAGE_SIZE - offset);
1428                         break;
1429                 }
1430         }
1431
1432         if (offset && !PageUptodate(pages[0])) {
1433                 ret = bch2_read_single_page(pages[0], mapping);
1434                 if (ret)
1435                         goto out;
1436         }
1437
1438         if ((pos + len) & (PAGE_SIZE - 1) &&
1439             !PageUptodate(pages[nr_pages - 1])) {
1440                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1441                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1442                 } else {
1443                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1444                         if (ret)
1445                                 goto out;
1446                 }
1447         }
1448
1449         while (reserved < len) {
1450                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1451                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1452                 unsigned pg_len = min_t(unsigned, len - reserved,
1453                                         PAGE_SIZE - pg_offset);
1454 retry_reservation:
1455                 ret = bch2_page_reservation_get(c, inode, page, &res,
1456                                                 pg_offset, pg_len, true);
1457
1458                 if (ret && !PageUptodate(page)) {
1459                         ret = bch2_read_single_page(page, mapping);
1460                         if (!ret)
1461                                 goto retry_reservation;
1462                 }
1463
1464                 if (ret)
1465                         goto out;
1466
1467                 reserved += pg_len;
1468         }
1469
1470         if (mapping_writably_mapped(mapping))
1471                 for (i = 0; i < nr_pages; i++)
1472                         flush_dcache_page(pages[i]);
1473
1474         while (copied < len) {
1475                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1476                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1477                 unsigned pg_len = min_t(unsigned, len - copied,
1478                                         PAGE_SIZE - pg_offset);
1479                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1480                                                 iter, pg_offset, pg_len);
1481
1482                 if (!pg_copied)
1483                         break;
1484
1485                 flush_dcache_page(page);
1486                 iov_iter_advance(iter, pg_copied);
1487                 copied += pg_copied;
1488         }
1489
1490         if (!copied)
1491                 goto out;
1492
1493         if (copied < len &&
1494             ((offset + copied) & (PAGE_SIZE - 1))) {
1495                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1496
1497                 if (!PageUptodate(page)) {
1498                         zero_user(page, 0, PAGE_SIZE);
1499                         copied -= (offset + copied) & (PAGE_SIZE - 1);
1500                 }
1501         }
1502
1503         spin_lock(&inode->v.i_lock);
1504         if (pos + copied > inode->v.i_size)
1505                 i_size_write(&inode->v, pos + copied);
1506         spin_unlock(&inode->v.i_lock);
1507
1508         while (set_dirty < copied) {
1509                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1510                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1511                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1512                                         PAGE_SIZE - pg_offset);
1513
1514                 if (!PageUptodate(page))
1515                         SetPageUptodate(page);
1516
1517                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1518                 unlock_page(page);
1519                 put_page(page);
1520
1521                 set_dirty += pg_len;
1522         }
1523
1524         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1525         inode->ei_last_dirtied = (unsigned long) current;
1526 out:
1527         for (i = nr_pages_copied; i < nr_pages; i++) {
1528                 unlock_page(pages[i]);
1529                 put_page(pages[i]);
1530         }
1531
1532         bch2_page_reservation_put(c, inode, &res);
1533
1534         return copied ?: ret;
1535 }
1536
1537 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1538 {
1539         struct file *file = iocb->ki_filp;
1540         struct address_space *mapping = file->f_mapping;
1541         struct bch_inode_info *inode = file_bch_inode(file);
1542         loff_t pos = iocb->ki_pos;
1543         ssize_t written = 0;
1544         int ret = 0;
1545
1546         pagecache_add_get(&mapping->add_lock);
1547
1548         do {
1549                 unsigned offset = pos & (PAGE_SIZE - 1);
1550                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1551                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1552 again:
1553                 /*
1554                  * Bring in the user page that we will copy from _first_.
1555                  * Otherwise there's a nasty deadlock on copying from the
1556                  * same page as we're writing to, without it being marked
1557                  * up-to-date.
1558                  *
1559                  * Not only is this an optimisation, but it is also required
1560                  * to check that the address is actually valid, when atomic
1561                  * usercopies are used, below.
1562                  */
1563                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1564                         bytes = min_t(unsigned long, iov_iter_count(iter),
1565                                       PAGE_SIZE - offset);
1566
1567                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1568                                 ret = -EFAULT;
1569                                 break;
1570                         }
1571                 }
1572
1573                 if (unlikely(fatal_signal_pending(current))) {
1574                         ret = -EINTR;
1575                         break;
1576                 }
1577
1578                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1579                 if (unlikely(ret < 0))
1580                         break;
1581
1582                 cond_resched();
1583
1584                 if (unlikely(ret == 0)) {
1585                         /*
1586                          * If we were unable to copy any data at all, we must
1587                          * fall back to a single segment length write.
1588                          *
1589                          * If we didn't fallback here, we could livelock
1590                          * because not all segments in the iov can be copied at
1591                          * once without a pagefault.
1592                          */
1593                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1594                                       iov_iter_single_seg_count(iter));
1595                         goto again;
1596                 }
1597                 pos += ret;
1598                 written += ret;
1599
1600                 balance_dirty_pages_ratelimited(mapping);
1601         } while (iov_iter_count(iter));
1602
1603         pagecache_add_put(&mapping->add_lock);
1604
1605         return written ? written : ret;
1606 }
1607
1608 /* O_DIRECT reads */
1609
1610 static void bch2_dio_read_complete(struct closure *cl)
1611 {
1612         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1613
1614         dio->req->ki_complete(dio->req, dio->ret, 0);
1615         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1616 }
1617
1618 static void bch2_direct_IO_read_endio(struct bio *bio)
1619 {
1620         struct dio_read *dio = bio->bi_private;
1621
1622         if (bio->bi_status)
1623                 dio->ret = blk_status_to_errno(bio->bi_status);
1624
1625         closure_put(&dio->cl);
1626 }
1627
1628 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1629 {
1630         bch2_direct_IO_read_endio(bio);
1631         bio_check_pages_dirty(bio);     /* transfers ownership */
1632 }
1633
1634 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1635 {
1636         struct file *file = req->ki_filp;
1637         struct bch_inode_info *inode = file_bch_inode(file);
1638         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1639         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1640         struct dio_read *dio;
1641         struct bio *bio;
1642         loff_t offset = req->ki_pos;
1643         bool sync = is_sync_kiocb(req);
1644         size_t shorten;
1645         ssize_t ret;
1646
1647         if ((offset|iter->count) & (block_bytes(c) - 1))
1648                 return -EINVAL;
1649
1650         ret = min_t(loff_t, iter->count,
1651                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1652
1653         if (!ret)
1654                 return ret;
1655
1656         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1657         iter->count -= shorten;
1658
1659         bio = bio_alloc_bioset(GFP_KERNEL,
1660                                iov_iter_npages(iter, BIO_MAX_PAGES),
1661                                &c->dio_read_bioset);
1662
1663         bio->bi_end_io = bch2_direct_IO_read_endio;
1664
1665         dio = container_of(bio, struct dio_read, rbio.bio);
1666         closure_init(&dio->cl, NULL);
1667
1668         /*
1669          * this is a _really_ horrible hack just to avoid an atomic sub at the
1670          * end:
1671          */
1672         if (!sync) {
1673                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1674                 atomic_set(&dio->cl.remaining,
1675                            CLOSURE_REMAINING_INITIALIZER -
1676                            CLOSURE_RUNNING +
1677                            CLOSURE_DESTRUCTOR);
1678         } else {
1679                 atomic_set(&dio->cl.remaining,
1680                            CLOSURE_REMAINING_INITIALIZER + 1);
1681         }
1682
1683         dio->req        = req;
1684         dio->ret        = ret;
1685
1686         goto start;
1687         while (iter->count) {
1688                 bio = bio_alloc_bioset(GFP_KERNEL,
1689                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1690                                        &c->bio_read);
1691                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1692 start:
1693                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1694                 bio->bi_iter.bi_sector  = offset >> 9;
1695                 bio->bi_private         = dio;
1696
1697                 ret = bio_iov_iter_get_pages(bio, iter);
1698                 if (ret < 0) {
1699                         /* XXX: fault inject this path */
1700                         bio->bi_status = BLK_STS_RESOURCE;
1701                         bio_endio(bio);
1702                         break;
1703                 }
1704
1705                 offset += bio->bi_iter.bi_size;
1706                 bio_set_pages_dirty(bio);
1707
1708                 if (iter->count)
1709                         closure_get(&dio->cl);
1710
1711                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1712         }
1713
1714         iter->count += shorten;
1715
1716         if (sync) {
1717                 closure_sync(&dio->cl);
1718                 closure_debug_destroy(&dio->cl);
1719                 ret = dio->ret;
1720                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1721                 return ret;
1722         } else {
1723                 return -EIOCBQUEUED;
1724         }
1725 }
1726
1727 /* O_DIRECT writes */
1728
1729 static void bch2_dio_write_loop_async(struct closure *);
1730
1731 static long bch2_dio_write_loop(struct dio_write *dio)
1732 {
1733         bool kthread = (current->flags & PF_KTHREAD) != 0;
1734         struct bch_fs *c = dio->op.c;
1735         struct kiocb *req = dio->req;
1736         struct address_space *mapping = req->ki_filp->f_mapping;
1737         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1738         struct bio *bio = &dio->op.wbio.bio;
1739         struct bvec_iter_all iter;
1740         struct bio_vec *bv;
1741         unsigned unaligned;
1742         u64 new_i_size;
1743         loff_t offset;
1744         bool sync;
1745         long ret;
1746
1747         if (dio->loop)
1748                 goto loop;
1749
1750         inode_dio_begin(&inode->v);
1751         __pagecache_block_get(&mapping->add_lock);
1752
1753         /* Write and invalidate pagecache range that we're writing to: */
1754         offset = req->ki_pos + (dio->op.written << 9);
1755         ret = write_invalidate_inode_pages_range(mapping,
1756                                         offset,
1757                                         offset + iov_iter_count(&dio->iter) - 1);
1758         if (unlikely(ret))
1759                 goto err;
1760
1761         while (1) {
1762                 offset = req->ki_pos + (dio->op.written << 9);
1763
1764                 BUG_ON(current->pagecache_lock);
1765                 current->pagecache_lock = &mapping->add_lock;
1766                 if (kthread)
1767                         use_mm(dio->mm);
1768
1769                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1770
1771                 if (kthread)
1772                         unuse_mm(dio->mm);
1773                 current->pagecache_lock = NULL;
1774
1775                 if (unlikely(ret < 0))
1776                         goto err;
1777
1778                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1779                 bio->bi_iter.bi_size -= unaligned;
1780                 iov_iter_revert(&dio->iter, unaligned);
1781
1782                 if (!bio->bi_iter.bi_size) {
1783                         /*
1784                          * bio_iov_iter_get_pages was only able to get <
1785                          * blocksize worth of pages:
1786                          */
1787                         bio_for_each_segment_all(bv, bio, iter)
1788                                 put_page(bv->bv_page);
1789                         ret = -EFAULT;
1790                         goto err;
1791                 }
1792
1793                 /* gup might have faulted pages back in: */
1794                 ret = write_invalidate_inode_pages_range(mapping,
1795                                 offset,
1796                                 offset + bio->bi_iter.bi_size - 1);
1797                 if (unlikely(ret))
1798                         goto err;
1799
1800                 dio->op.pos = POS(inode->v.i_ino, offset >> 9);
1801
1802                 task_io_account_write(bio->bi_iter.bi_size);
1803
1804                 closure_call(&dio->op.cl, bch2_write, NULL, &dio->cl);
1805
1806                 if (!dio->sync && !dio->loop && dio->iter.count) {
1807                         struct iovec *iov = dio->inline_vecs;
1808
1809                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1810                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1811                                               GFP_KERNEL);
1812                                 if (unlikely(!iov)) {
1813                                         dio->op.error = -ENOMEM;
1814                                         goto err_wait_io;
1815                                 }
1816
1817                                 dio->free_iov = true;
1818                         }
1819
1820                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1821                         dio->iter.iov = iov;
1822                 }
1823 err_wait_io:
1824                 dio->loop = true;
1825
1826                 if (!dio->sync) {
1827                         continue_at(&dio->cl, bch2_dio_write_loop_async, NULL);
1828                         return -EIOCBQUEUED;
1829                 }
1830
1831                 closure_sync(&dio->cl);
1832 loop:
1833                 i_sectors_acct(c, inode, &dio->quota_res,
1834                                dio->op.i_sectors_delta);
1835                 dio->op.i_sectors_delta = 0;
1836
1837                 new_i_size = req->ki_pos + ((u64) dio->op.written << 9);
1838
1839                 spin_lock(&inode->v.i_lock);
1840                 if (new_i_size > inode->v.i_size)
1841                         i_size_write(&inode->v, new_i_size);
1842                 spin_unlock(&inode->v.i_lock);
1843
1844                 bio_for_each_segment_all(bv, bio, iter)
1845                         put_page(bv->bv_page);
1846                 if (!dio->iter.count || dio->op.error)
1847                         break;
1848                 bio_reset(bio);
1849         }
1850
1851         ret = dio->op.error ?: ((long) dio->op.written << 9);
1852 err:
1853         __pagecache_block_put(&mapping->add_lock);
1854         bch2_disk_reservation_put(c, &dio->op.res);
1855         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1856
1857         if (dio->free_iov)
1858                 kfree(dio->iter.iov);
1859
1860         closure_debug_destroy(&dio->cl);
1861
1862         sync = dio->sync;
1863         bio_put(bio);
1864
1865         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1866         inode_dio_end(&inode->v);
1867
1868         if (!sync) {
1869                 req->ki_complete(req, ret, 0);
1870                 ret = -EIOCBQUEUED;
1871         }
1872         return ret;
1873 }
1874
1875 static void bch2_dio_write_loop_async(struct closure *cl)
1876 {
1877         struct dio_write *dio = container_of(cl, struct dio_write, cl);
1878
1879         bch2_dio_write_loop(dio);
1880 }
1881
1882 static int bch2_direct_IO_write(struct kiocb *req,
1883                                 struct iov_iter *iter,
1884                                 bool swap)
1885 {
1886         struct file *file = req->ki_filp;
1887         struct bch_inode_info *inode = file_bch_inode(file);
1888         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1889         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1890         struct dio_write *dio;
1891         struct bio *bio;
1892         ssize_t ret;
1893
1894         lockdep_assert_held(&inode->v.i_rwsem);
1895
1896         if (unlikely(!iter->count))
1897                 return 0;
1898
1899         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1900                 return -EINVAL;
1901
1902         bio = bio_alloc_bioset(GFP_KERNEL,
1903                                iov_iter_npages(iter, BIO_MAX_PAGES),
1904                                &c->dio_write_bioset);
1905         dio = container_of(bio, struct dio_write, op.wbio.bio);
1906         closure_init(&dio->cl, NULL);
1907         dio->req                = req;
1908         dio->mm                 = current->mm;
1909         dio->loop               = false;
1910         dio->sync               = is_sync_kiocb(req) ||
1911                 req->ki_pos + iter->count > inode->v.i_size;
1912         dio->free_iov           = false;
1913         dio->quota_res.sectors  = 0;
1914         dio->iter               = *iter;
1915
1916         bch2_write_op_init(&dio->op, c, opts);
1917         dio->op.target          = opts.foreground_target;
1918         op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1919         dio->op.write_point     = writepoint_hashed((unsigned long) current);
1920         dio->op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1921
1922         if ((req->ki_flags & IOCB_DSYNC) &&
1923             !c->opts.journal_flush_disabled)
1924                 dio->op.flags |= BCH_WRITE_FLUSH;
1925
1926         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
1927                                          iter->count >> 9, true);
1928         if (unlikely(ret))
1929                 goto err;
1930
1931         dio->op.nr_replicas     = dio->op.opts.data_replicas;
1932
1933         ret = bch2_disk_reservation_get(c, &dio->op.res, iter->count >> 9,
1934                                         dio->op.opts.data_replicas, 0);
1935         if (unlikely(ret) &&
1936             !bch2_check_range_allocated(c, POS(inode->v.i_ino,
1937                                                req->ki_pos >> 9),
1938                                         iter->count >> 9,
1939                                         dio->op.opts.data_replicas))
1940                 goto err;
1941
1942         return bch2_dio_write_loop(dio);
1943 err:
1944         bch2_disk_reservation_put(c, &dio->op.res);
1945         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1946         closure_debug_destroy(&dio->cl);
1947         bio_put(bio);
1948         return ret;
1949 }
1950
1951 ssize_t bch2_direct_IO(struct kiocb *req, struct iov_iter *iter)
1952 {
1953         struct blk_plug plug;
1954         ssize_t ret;
1955
1956         blk_start_plug(&plug);
1957         ret = iov_iter_rw(iter) == WRITE
1958                 ? bch2_direct_IO_write(req, iter, false)
1959                 : bch2_direct_IO_read(req, iter);
1960         blk_finish_plug(&plug);
1961
1962         return ret;
1963 }
1964
1965 static ssize_t
1966 bch2_direct_write(struct kiocb *iocb, struct iov_iter *iter)
1967 {
1968         return bch2_direct_IO_write(iocb, iter, true);
1969 }
1970
1971 static ssize_t __bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1972 {
1973         struct file *file = iocb->ki_filp;
1974         struct bch_inode_info *inode = file_bch_inode(file);
1975         ssize_t ret;
1976
1977         /* We can write back this queue in page reclaim */
1978         current->backing_dev_info = inode_to_bdi(&inode->v);
1979         ret = file_remove_privs(file);
1980         if (ret)
1981                 goto out;
1982
1983         ret = file_update_time(file);
1984         if (ret)
1985                 goto out;
1986
1987         ret = iocb->ki_flags & IOCB_DIRECT
1988                 ? bch2_direct_write(iocb, from)
1989                 : bch2_buffered_write(iocb, from);
1990
1991         if (likely(ret > 0))
1992                 iocb->ki_pos += ret;
1993 out:
1994         current->backing_dev_info = NULL;
1995         return ret;
1996 }
1997
1998 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
1999 {
2000         struct bch_inode_info *inode = file_bch_inode(iocb->ki_filp);
2001         bool direct = iocb->ki_flags & IOCB_DIRECT;
2002         ssize_t ret;
2003
2004         inode_lock(&inode->v);
2005         ret = generic_write_checks(iocb, from);
2006         if (ret > 0)
2007                 ret = __bch2_write_iter(iocb, from);
2008         inode_unlock(&inode->v);
2009
2010         if (ret > 0 && !direct)
2011                 ret = generic_write_sync(iocb, ret);
2012
2013         return ret;
2014 }
2015
2016 /* fsync: */
2017
2018 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2019 {
2020         struct bch_inode_info *inode = file_bch_inode(file);
2021         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2022         int ret, ret2;
2023
2024         ret = file_write_and_wait_range(file, start, end);
2025         if (ret)
2026                 return ret;
2027
2028         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2029                 goto out;
2030
2031         ret = sync_inode_metadata(&inode->v, 1);
2032         if (ret)
2033                 return ret;
2034 out:
2035         if (!c->opts.journal_flush_disabled)
2036                 ret = bch2_journal_flush_seq(&c->journal,
2037                                              inode->ei_journal_seq);
2038         ret2 = file_check_and_advance_wb_err(file);
2039
2040         return ret ?: ret2;
2041 }
2042
2043 /* truncate: */
2044
2045 static inline int range_has_data(struct bch_fs *c,
2046                                   struct bpos start,
2047                                   struct bpos end)
2048 {
2049         struct btree_trans trans;
2050         struct btree_iter *iter;
2051         struct bkey_s_c k;
2052         int ret = 0;
2053
2054         bch2_trans_init(&trans, c, 0, 0);
2055
2056         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2057                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2058                         break;
2059
2060                 if (bkey_extent_is_data(k.k)) {
2061                         ret = 1;
2062                         break;
2063                 }
2064         }
2065
2066         return bch2_trans_exit(&trans) ?: ret;
2067 }
2068
2069 static int __bch2_truncate_page(struct bch_inode_info *inode,
2070                                 pgoff_t index, loff_t start, loff_t end)
2071 {
2072         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2073         struct address_space *mapping = inode->v.i_mapping;
2074         struct bch_page_state *s;
2075         unsigned start_offset = start & (PAGE_SIZE - 1);
2076         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2077         unsigned i;
2078         struct page *page;
2079         int ret = 0;
2080
2081         /* Page boundary? Nothing to do */
2082         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2083               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2084                 return 0;
2085
2086         /* Above i_size? */
2087         if (index << PAGE_SHIFT >= inode->v.i_size)
2088                 return 0;
2089
2090         page = find_lock_page(mapping, index);
2091         if (!page) {
2092                 /*
2093                  * XXX: we're doing two index lookups when we end up reading the
2094                  * page
2095                  */
2096                 ret = range_has_data(c,
2097                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2098                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2099                 if (ret <= 0)
2100                         return ret;
2101
2102                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2103                 if (unlikely(!page)) {
2104                         ret = -ENOMEM;
2105                         goto out;
2106                 }
2107         }
2108
2109         s = bch2_page_state_create(page, 0);
2110         if (!s) {
2111                 ret = -ENOMEM;
2112                 goto unlock;
2113         }
2114
2115         if (!PageUptodate(page)) {
2116                 ret = bch2_read_single_page(page, mapping);
2117                 if (ret)
2118                         goto unlock;
2119         }
2120
2121         if (index != start >> PAGE_SHIFT)
2122                 start_offset = 0;
2123         if (index != end >> PAGE_SHIFT)
2124                 end_offset = PAGE_SIZE;
2125
2126         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2127              i < round_down(end_offset, block_bytes(c)) >> 9;
2128              i++) {
2129                 s->s[i].nr_replicas     = 0;
2130                 s->s[i].state           = SECTOR_UNALLOCATED;
2131         }
2132
2133         zero_user_segment(page, start_offset, end_offset);
2134
2135         /*
2136          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2137          *
2138          * XXX: because we aren't currently tracking whether the page has actual
2139          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2140          */
2141         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2142         BUG_ON(ret);
2143
2144         __set_page_dirty_nobuffers(page);
2145 unlock:
2146         unlock_page(page);
2147         put_page(page);
2148 out:
2149         return ret;
2150 }
2151
2152 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2153 {
2154         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2155                                     from, round_up(from, PAGE_SIZE));
2156 }
2157
2158 static int bch2_extend(struct bch_inode_info *inode,
2159                        struct bch_inode_unpacked *inode_u,
2160                        struct iattr *iattr)
2161 {
2162         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2163         struct address_space *mapping = inode->v.i_mapping;
2164         int ret;
2165
2166         /*
2167          * sync appends:
2168          *
2169          * this has to be done _before_ extending i_size:
2170          */
2171         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2172         if (ret)
2173                 return ret;
2174
2175         truncate_setsize(&inode->v, iattr->ia_size);
2176         setattr_copy(&inode->v, iattr);
2177
2178         mutex_lock(&inode->ei_update_lock);
2179         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2180                                     ATTR_MTIME|ATTR_CTIME);
2181         mutex_unlock(&inode->ei_update_lock);
2182
2183         return ret;
2184 }
2185
2186 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2187                                    struct bch_inode_unpacked *bi,
2188                                    void *p)
2189 {
2190         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2191
2192         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2193         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2194         return 0;
2195 }
2196
2197 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2198                                   struct bch_inode_unpacked *bi, void *p)
2199 {
2200         u64 *new_i_size = p;
2201
2202         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2203         bi->bi_size = *new_i_size;
2204         return 0;
2205 }
2206
2207 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2208 {
2209         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2210         struct address_space *mapping = inode->v.i_mapping;
2211         struct bch_inode_unpacked inode_u;
2212         struct btree_trans trans;
2213         struct btree_iter *iter;
2214         u64 new_i_size = iattr->ia_size;
2215         s64 i_sectors_delta = 0;
2216         int ret = 0;
2217
2218         inode_dio_wait(&inode->v);
2219         pagecache_block_get(&mapping->add_lock);
2220
2221         /*
2222          * fetch current on disk i_size: inode is locked, i_size can only
2223          * increase underneath us:
2224          */
2225         bch2_trans_init(&trans, c, 0, 0);
2226         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2227         ret = PTR_ERR_OR_ZERO(iter);
2228         bch2_trans_exit(&trans);
2229
2230         if (ret)
2231                 goto err;
2232
2233         BUG_ON(inode->v.i_size < inode_u.bi_size);
2234
2235         if (iattr->ia_size > inode->v.i_size) {
2236                 ret = bch2_extend(inode, &inode_u, iattr);
2237                 goto err;
2238         }
2239
2240         ret = bch2_truncate_page(inode, iattr->ia_size);
2241         if (unlikely(ret))
2242                 goto err;
2243
2244         /*
2245          * When extending, we're going to write the new i_size to disk
2246          * immediately so we need to flush anything above the current on disk
2247          * i_size first:
2248          *
2249          * Also, when extending we need to flush the page that i_size currently
2250          * straddles - if it's mapped to userspace, we need to ensure that
2251          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2252          * again to allocate the part of the page that was extended.
2253          */
2254         if (iattr->ia_size > inode_u.bi_size)
2255                 ret = filemap_write_and_wait_range(mapping,
2256                                 inode_u.bi_size,
2257                                 iattr->ia_size - 1);
2258         else if (iattr->ia_size & (PAGE_SIZE - 1))
2259                 ret = filemap_write_and_wait_range(mapping,
2260                                 round_down(iattr->ia_size, PAGE_SIZE),
2261                                 iattr->ia_size - 1);
2262         if (ret)
2263                 goto err;
2264
2265         mutex_lock(&inode->ei_update_lock);
2266         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2267                                &new_i_size, 0);
2268         mutex_unlock(&inode->ei_update_lock);
2269
2270         if (unlikely(ret))
2271                 goto err;
2272
2273         truncate_setsize(&inode->v, iattr->ia_size);
2274
2275         ret = bch2_fpunch(c, inode->v.i_ino,
2276                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2277                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2278         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2279
2280         if (unlikely(ret))
2281                 goto err;
2282
2283         setattr_copy(&inode->v, iattr);
2284
2285         mutex_lock(&inode->ei_update_lock);
2286         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2287                                ATTR_MTIME|ATTR_CTIME);
2288         mutex_unlock(&inode->ei_update_lock);
2289 err:
2290         pagecache_block_put(&mapping->add_lock);
2291         return ret;
2292 }
2293
2294 /* fallocate: */
2295
2296 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2297 {
2298         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2299         struct address_space *mapping = inode->v.i_mapping;
2300         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2301         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2302         int ret = 0;
2303
2304         inode_lock(&inode->v);
2305         inode_dio_wait(&inode->v);
2306         pagecache_block_get(&mapping->add_lock);
2307
2308         ret = __bch2_truncate_page(inode,
2309                                    offset >> PAGE_SHIFT,
2310                                    offset, offset + len);
2311         if (unlikely(ret))
2312                 goto err;
2313
2314         if (offset >> PAGE_SHIFT !=
2315             (offset + len) >> PAGE_SHIFT) {
2316                 ret = __bch2_truncate_page(inode,
2317                                            (offset + len) >> PAGE_SHIFT,
2318                                            offset, offset + len);
2319                 if (unlikely(ret))
2320                         goto err;
2321         }
2322
2323         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2324
2325         if (discard_start < discard_end) {
2326                 s64 i_sectors_delta = 0;
2327
2328                 ret = bch2_fpunch(c, inode->v.i_ino,
2329                                   discard_start, discard_end,
2330                                   &inode->ei_journal_seq,
2331                                   &i_sectors_delta);
2332                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2333         }
2334 err:
2335         pagecache_block_put(&mapping->add_lock);
2336         inode_unlock(&inode->v);
2337
2338         return ret;
2339 }
2340
2341 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2342                                    loff_t offset, loff_t len,
2343                                    bool insert)
2344 {
2345         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2346         struct address_space *mapping = inode->v.i_mapping;
2347         struct btree_trans trans;
2348         struct btree_iter *src, *dst, *del = NULL;
2349         loff_t shift, new_size;
2350         u64 src_start;
2351         int ret;
2352
2353         if ((offset | len) & (block_bytes(c) - 1))
2354                 return -EINVAL;
2355
2356         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2357
2358         /*
2359          * We need i_mutex to keep the page cache consistent with the extents
2360          * btree, and the btree consistent with i_size - we don't need outside
2361          * locking for the extents btree itself, because we're using linked
2362          * iterators
2363          */
2364         inode_lock(&inode->v);
2365         inode_dio_wait(&inode->v);
2366         pagecache_block_get(&mapping->add_lock);
2367
2368         if (insert) {
2369                 ret = -EFBIG;
2370                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2371                         goto err;
2372
2373                 ret = -EINVAL;
2374                 if (offset >= inode->v.i_size)
2375                         goto err;
2376
2377                 src_start       = U64_MAX;
2378                 shift           = len;
2379         } else {
2380                 ret = -EINVAL;
2381                 if (offset + len >= inode->v.i_size)
2382                         goto err;
2383
2384                 src_start       = offset + len;
2385                 shift           = -len;
2386         }
2387
2388         new_size = inode->v.i_size + shift;
2389
2390         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2391         if (ret)
2392                 goto err;
2393
2394         if (insert) {
2395                 i_size_write(&inode->v, new_size);
2396                 mutex_lock(&inode->ei_update_lock);
2397                 ret = bch2_write_inode_size(c, inode, new_size,
2398                                             ATTR_MTIME|ATTR_CTIME);
2399                 mutex_unlock(&inode->ei_update_lock);
2400         } else {
2401                 s64 i_sectors_delta = 0;
2402
2403                 ret = bch2_fpunch(c, inode->v.i_ino,
2404                                   offset >> 9, (offset + len) >> 9,
2405                                   &inode->ei_journal_seq,
2406                                   &i_sectors_delta);
2407                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2408
2409                 if (ret)
2410                         goto err;
2411         }
2412
2413         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2414                         POS(inode->v.i_ino, src_start >> 9),
2415                         BTREE_ITER_INTENT);
2416         BUG_ON(IS_ERR_OR_NULL(src));
2417
2418         dst = bch2_trans_copy_iter(&trans, src);
2419         BUG_ON(IS_ERR_OR_NULL(dst));
2420
2421         while (1) {
2422                 struct disk_reservation disk_res =
2423                         bch2_disk_reservation_init(c, 0);
2424                 BKEY_PADDED(k) copy;
2425                 struct bkey_i delete;
2426                 struct bkey_s_c k;
2427                 struct bpos next_pos;
2428                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2429                 struct bpos atomic_end;
2430                 unsigned commit_flags = BTREE_INSERT_NOFAIL|
2431                         BTREE_INSERT_ATOMIC|
2432                         BTREE_INSERT_USE_RESERVE;
2433
2434                 k = insert
2435                         ? bch2_btree_iter_peek_prev(src)
2436                         : bch2_btree_iter_peek(src);
2437                 if ((ret = bkey_err(k)))
2438                         goto bkey_err;
2439
2440                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2441                         break;
2442
2443                 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2444
2445                 if (insert &&
2446                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2447                         break;
2448 reassemble:
2449                 bkey_reassemble(&copy.k, k);
2450
2451                 if (insert &&
2452                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0) {
2453                         bch2_cut_front(move_pos, &copy.k);
2454                         bch2_btree_iter_set_pos(src, bkey_start_pos(&copy.k.k));
2455                 }
2456
2457                 copy.k.k.p.offset += shift >> 9;
2458                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k.k));
2459
2460                 ret = bch2_extent_atomic_end(dst, &copy.k, &atomic_end);
2461                 if (ret)
2462                         goto bkey_err;
2463
2464                 if (bkey_cmp(atomic_end, copy.k.k.p)) {
2465                         if (insert) {
2466                                 move_pos = atomic_end;
2467                                 move_pos.offset -= shift >> 9;
2468                                 goto reassemble;
2469                         } else {
2470                                 bch2_cut_back(atomic_end, &copy.k.k);
2471                         }
2472                 }
2473
2474                 bkey_init(&delete.k);
2475                 delete.k.p = src->pos;
2476                 bch2_key_resize(&delete.k, copy.k.k.size);
2477
2478                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2479
2480                 /*
2481                  * If the new and old keys overlap (because we're moving an
2482                  * extent that's bigger than the amount we're collapsing by),
2483                  * we need to trim the delete key here so they don't overlap
2484                  * because overlaps on insertions aren't handled before
2485                  * triggers are run, so the overwrite will get double counted
2486                  * by the triggers machinery:
2487                  */
2488                 if (insert &&
2489                     bkey_cmp(bkey_start_pos(&copy.k.k), delete.k.p) < 0) {
2490                         bch2_cut_back(bkey_start_pos(&copy.k.k), &delete.k);
2491                 } else if (!insert &&
2492                            bkey_cmp(copy.k.k.p,
2493                                     bkey_start_pos(&delete.k)) > 0) {
2494                         bch2_cut_front(copy.k.k.p, &delete);
2495
2496                         del = bch2_trans_copy_iter(&trans, src);
2497                         BUG_ON(IS_ERR_OR_NULL(del));
2498
2499                         bch2_btree_iter_set_pos(del,
2500                                 bkey_start_pos(&delete.k));
2501                 }
2502
2503                 bch2_trans_update(&trans, dst, &copy.k);
2504                 bch2_trans_update(&trans, del ?: src, &delete);
2505
2506                 if (copy.k.k.size == k.k->size) {
2507                         /*
2508                          * If we're moving the entire extent, we can skip
2509                          * running triggers:
2510                          */
2511                         commit_flags |= BTREE_INSERT_NOMARK;
2512                 } else {
2513                         /* We might end up splitting compressed extents: */
2514                         unsigned nr_ptrs =
2515                                 bch2_bkey_nr_dirty_ptrs(bkey_i_to_s_c(&copy.k));
2516
2517                         ret = bch2_disk_reservation_get(c, &disk_res,
2518                                         copy.k.k.size, nr_ptrs,
2519                                         BCH_DISK_RESERVATION_NOFAIL);
2520                         BUG_ON(ret);
2521                 }
2522
2523                 ret = bch2_trans_commit(&trans, &disk_res,
2524                                         &inode->ei_journal_seq,
2525                                         commit_flags);
2526                 bch2_disk_reservation_put(c, &disk_res);
2527 bkey_err:
2528                 if (del)
2529                         bch2_trans_iter_put(&trans, del);
2530                 del = NULL;
2531
2532                 if (!ret)
2533                         bch2_btree_iter_set_pos(src, next_pos);
2534
2535                 if (ret == -EINTR)
2536                         ret = 0;
2537                 if (ret)
2538                         goto err;
2539
2540                 bch2_trans_cond_resched(&trans);
2541         }
2542         bch2_trans_unlock(&trans);
2543
2544         if (!insert) {
2545                 i_size_write(&inode->v, new_size);
2546                 mutex_lock(&inode->ei_update_lock);
2547                 ret = bch2_write_inode_size(c, inode, new_size,
2548                                             ATTR_MTIME|ATTR_CTIME);
2549                 mutex_unlock(&inode->ei_update_lock);
2550         }
2551 err:
2552         bch2_trans_exit(&trans);
2553         pagecache_block_put(&mapping->add_lock);
2554         inode_unlock(&inode->v);
2555         return ret;
2556 }
2557
2558 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2559                             loff_t offset, loff_t len)
2560 {
2561         struct address_space *mapping = inode->v.i_mapping;
2562         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2563         struct btree_trans trans;
2564         struct btree_iter *iter;
2565         struct bpos end_pos;
2566         loff_t end              = offset + len;
2567         loff_t block_start      = round_down(offset,    block_bytes(c));
2568         loff_t block_end        = round_up(end,         block_bytes(c));
2569         unsigned sectors;
2570         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2571         int ret;
2572
2573         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2574
2575         inode_lock(&inode->v);
2576         inode_dio_wait(&inode->v);
2577         pagecache_block_get(&mapping->add_lock);
2578
2579         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2580                 ret = inode_newsize_ok(&inode->v, end);
2581                 if (ret)
2582                         goto err;
2583         }
2584
2585         if (mode & FALLOC_FL_ZERO_RANGE) {
2586                 ret = __bch2_truncate_page(inode,
2587                                            offset >> PAGE_SHIFT,
2588                                            offset, end);
2589
2590                 if (!ret &&
2591                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2592                         ret = __bch2_truncate_page(inode,
2593                                                    end >> PAGE_SHIFT,
2594                                                    offset, end);
2595
2596                 if (unlikely(ret))
2597                         goto err;
2598
2599                 truncate_pagecache_range(&inode->v, offset, end - 1);
2600         }
2601
2602         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2603                         POS(inode->v.i_ino, block_start >> 9),
2604                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2605         end_pos = POS(inode->v.i_ino, block_end >> 9);
2606
2607         while (bkey_cmp(iter->pos, end_pos) < 0) {
2608                 s64 i_sectors_delta = 0;
2609                 struct disk_reservation disk_res = { 0 };
2610                 struct quota_res quota_res = { 0 };
2611                 struct bkey_i_reservation reservation;
2612                 struct bkey_s_c k;
2613
2614                 k = bch2_btree_iter_peek_slot(iter);
2615                 if ((ret = bkey_err(k)))
2616                         goto bkey_err;
2617
2618                 /* already reserved */
2619                 if (k.k->type == KEY_TYPE_reservation &&
2620                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2621                         bch2_btree_iter_next_slot(iter);
2622                         continue;
2623                 }
2624
2625                 if (bkey_extent_is_data(k.k) &&
2626                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2627                         bch2_btree_iter_next_slot(iter);
2628                         continue;
2629                 }
2630
2631                 bkey_reservation_init(&reservation.k_i);
2632                 reservation.k.type      = KEY_TYPE_reservation;
2633                 reservation.k.p         = k.k->p;
2634                 reservation.k.size      = k.k->size;
2635
2636                 bch2_cut_front(iter->pos, &reservation.k_i);
2637                 bch2_cut_back(end_pos, &reservation.k);
2638
2639                 sectors = reservation.k.size;
2640                 reservation.v.nr_replicas = bch2_bkey_nr_dirty_ptrs(k);
2641
2642                 if (!bkey_extent_is_allocation(k.k)) {
2643                         ret = bch2_quota_reservation_add(c, inode,
2644                                         &quota_res,
2645                                         sectors, true);
2646                         if (unlikely(ret))
2647                                 goto bkey_err;
2648                 }
2649
2650                 if (reservation.v.nr_replicas < replicas ||
2651                     bch2_extent_is_compressed(k)) {
2652                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2653                                                         replicas, 0);
2654                         if (unlikely(ret))
2655                                 goto bkey_err;
2656
2657                         reservation.v.nr_replicas = disk_res.nr_replicas;
2658                 }
2659
2660                 bch2_trans_begin_updates(&trans);
2661
2662                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2663                                 &disk_res, &inode->ei_journal_seq,
2664                                 0, &i_sectors_delta);
2665                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2666 bkey_err:
2667                 bch2_quota_reservation_put(c, inode, &quota_res);
2668                 bch2_disk_reservation_put(c, &disk_res);
2669                 if (ret == -EINTR)
2670                         ret = 0;
2671                 if (ret)
2672                         goto err;
2673         }
2674
2675         /*
2676          * Do we need to extend the file?
2677          *
2678          * If we zeroed up to the end of the file, we dropped whatever writes
2679          * were going to write out the current i_size, so we have to extend
2680          * manually even if FL_KEEP_SIZE was set:
2681          */
2682         if (end >= inode->v.i_size &&
2683             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2684              (mode & FALLOC_FL_ZERO_RANGE))) {
2685                 struct btree_iter *inode_iter;
2686                 struct bch_inode_unpacked inode_u;
2687
2688                 do {
2689                         bch2_trans_begin(&trans);
2690                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2691                                                      inode->v.i_ino, 0);
2692                         ret = PTR_ERR_OR_ZERO(inode_iter);
2693                 } while (ret == -EINTR);
2694
2695                 bch2_trans_unlock(&trans);
2696
2697                 if (ret)
2698                         goto err;
2699
2700                 /*
2701                  * Sync existing appends before extending i_size,
2702                  * as in bch2_extend():
2703                  */
2704                 ret = filemap_write_and_wait_range(mapping,
2705                                         inode_u.bi_size, S64_MAX);
2706                 if (ret)
2707                         goto err;
2708
2709                 if (mode & FALLOC_FL_KEEP_SIZE)
2710                         end = inode->v.i_size;
2711                 else
2712                         i_size_write(&inode->v, end);
2713
2714                 mutex_lock(&inode->ei_update_lock);
2715                 ret = bch2_write_inode_size(c, inode, end, 0);
2716                 mutex_unlock(&inode->ei_update_lock);
2717         }
2718 err:
2719         bch2_trans_exit(&trans);
2720         pagecache_block_put(&mapping->add_lock);
2721         inode_unlock(&inode->v);
2722         return ret;
2723 }
2724
2725 long bch2_fallocate_dispatch(struct file *file, int mode,
2726                              loff_t offset, loff_t len)
2727 {
2728         struct bch_inode_info *inode = file_bch_inode(file);
2729
2730         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2731                 return bchfs_fallocate(inode, mode, offset, len);
2732
2733         if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2734                 return bchfs_fpunch(inode, offset, len);
2735
2736         if (mode == FALLOC_FL_INSERT_RANGE)
2737                 return bchfs_fcollapse_finsert(inode, offset, len, true);
2738
2739         if (mode == FALLOC_FL_COLLAPSE_RANGE)
2740                 return bchfs_fcollapse_finsert(inode, offset, len, false);
2741
2742         return -EOPNOTSUPP;
2743 }
2744
2745 static void mark_range_unallocated(struct bch_inode_info *inode,
2746                                    loff_t start, loff_t end)
2747 {
2748         pgoff_t index = start >> PAGE_SHIFT;
2749         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2750         struct pagevec pvec;
2751
2752         pagevec_init(&pvec);
2753
2754         do {
2755                 unsigned nr_pages, i, j;
2756
2757                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2758                                                 &index, end_index);
2759                 if (nr_pages == 0)
2760                         break;
2761
2762                 for (i = 0; i < nr_pages; i++) {
2763                         struct page *page = pvec.pages[i];
2764                         struct bch_page_state *s;
2765
2766                         lock_page(page);
2767                         s = bch2_page_state(page);
2768
2769                         if (s) {
2770                                 spin_lock(&s->lock);
2771                                 for (j = 0; j < PAGE_SECTORS; j++)
2772                                         s->s[j].nr_replicas = 0;
2773                                 spin_unlock(&s->lock);
2774                         }
2775
2776                         unlock_page(page);
2777                 }
2778                 pagevec_release(&pvec);
2779         } while (index <= end_index);
2780 }
2781
2782 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2783                              struct file *file_dst, loff_t pos_dst,
2784                              loff_t len, unsigned remap_flags)
2785 {
2786         struct bch_inode_info *src = file_bch_inode(file_src);
2787         struct bch_inode_info *dst = file_bch_inode(file_dst);
2788         struct bch_fs *c = src->v.i_sb->s_fs_info;
2789         s64 i_sectors_delta = 0;
2790         loff_t ret = 0;
2791         loff_t aligned_len;
2792
2793         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2794                 return -EINVAL;
2795
2796         if (remap_flags & REMAP_FILE_DEDUP)
2797                 return -EOPNOTSUPP;
2798
2799         if ((pos_src & (block_bytes(c) - 1)) ||
2800             (pos_dst & (block_bytes(c) - 1)))
2801                 return -EINVAL;
2802
2803         if (src == dst &&
2804             abs(pos_src - pos_dst) < len)
2805                 return -EINVAL;
2806
2807         bch2_lock_inodes(INODE_LOCK, src, dst);
2808
2809         file_update_time(file_dst);
2810
2811         inode_dio_wait(&src->v);
2812         inode_dio_wait(&dst->v);
2813
2814         __pagecache_block_get(&src->v.i_mapping->add_lock);
2815         __pagecache_block_get(&dst->v.i_mapping->add_lock);
2816
2817         ret = generic_remap_file_range_prep(file_src, pos_src,
2818                                             file_dst, pos_dst,
2819                                             &len, remap_flags);
2820         if (ret < 0 || len == 0)
2821                 goto err;
2822
2823         aligned_len = round_up(len, block_bytes(c));
2824
2825         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2826                                 pos_dst, pos_dst + aligned_len);
2827         if (ret)
2828                 goto err;
2829
2830         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2831
2832         ret = bch2_remap_range(c,
2833                                POS(dst->v.i_ino, pos_dst >> 9),
2834                                POS(src->v.i_ino, pos_src >> 9),
2835                                aligned_len >> 9,
2836                                &dst->ei_journal_seq,
2837                                pos_dst + len, &i_sectors_delta);
2838         if (ret < 0)
2839                 goto err;
2840
2841         ret <<= 9;
2842         /*
2843          * due to alignment, we might have remapped slightly more than requsted
2844          */
2845         ret = min(ret, len);
2846
2847         /* XXX get a quota reservation */
2848         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2849
2850         spin_lock(&dst->v.i_lock);
2851         if (pos_dst + len > dst->v.i_size)
2852                 i_size_write(&dst->v, pos_dst + len);
2853         spin_unlock(&dst->v.i_lock);
2854 err:
2855         __pagecache_block_put(&dst->v.i_mapping->add_lock);
2856         __pagecache_block_put(&src->v.i_mapping->add_lock);
2857
2858         bch2_unlock_inodes(INODE_LOCK, src, dst);
2859
2860         return ret;
2861 }
2862
2863 /* fseek: */
2864
2865 static int page_data_offset(struct page *page, unsigned offset)
2866 {
2867         struct bch_page_state *s = bch2_page_state(page);
2868         unsigned i;
2869
2870         if (s)
2871                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2872                         if (s->s[i].state >= SECTOR_DIRTY)
2873                                 return i << 9;
2874
2875         return -1;
2876 }
2877
2878 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2879                                        loff_t start_offset,
2880                                        loff_t end_offset)
2881 {
2882         struct address_space *mapping = vinode->i_mapping;
2883         struct page *page;
2884         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2885         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2886         pgoff_t index           = start_index;
2887         loff_t ret;
2888         int offset;
2889
2890         while (index <= end_index) {
2891                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2892                         lock_page(page);
2893
2894                         offset = page_data_offset(page,
2895                                         page->index == start_index
2896                                         ? start_offset & (PAGE_SIZE - 1)
2897                                         : 0);
2898                         if (offset >= 0) {
2899                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2900                                             offset,
2901                                             start_offset, end_offset);
2902                                 unlock_page(page);
2903                                 put_page(page);
2904                                 return ret;
2905                         }
2906
2907                         unlock_page(page);
2908                         put_page(page);
2909                 } else {
2910                         break;
2911                 }
2912         }
2913
2914         return end_offset;
2915 }
2916
2917 static loff_t bch2_seek_data(struct file *file, u64 offset)
2918 {
2919         struct bch_inode_info *inode = file_bch_inode(file);
2920         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2921         struct btree_trans trans;
2922         struct btree_iter *iter;
2923         struct bkey_s_c k;
2924         u64 isize, next_data = MAX_LFS_FILESIZE;
2925         int ret;
2926
2927         isize = i_size_read(&inode->v);
2928         if (offset >= isize)
2929                 return -ENXIO;
2930
2931         bch2_trans_init(&trans, c, 0, 0);
2932
2933         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2934                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2935                 if (k.k->p.inode != inode->v.i_ino) {
2936                         break;
2937                 } else if (bkey_extent_is_data(k.k)) {
2938                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2939                         break;
2940                 } else if (k.k->p.offset >> 9 > isize)
2941                         break;
2942         }
2943
2944         ret = bch2_trans_exit(&trans) ?: ret;
2945         if (ret)
2946                 return ret;
2947
2948         if (next_data > offset)
2949                 next_data = bch2_seek_pagecache_data(&inode->v,
2950                                                      offset, next_data);
2951
2952         if (next_data >= isize)
2953                 return -ENXIO;
2954
2955         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2956 }
2957
2958 static int __page_hole_offset(struct page *page, unsigned offset)
2959 {
2960         struct bch_page_state *s = bch2_page_state(page);
2961         unsigned i;
2962
2963         if (!s)
2964                 return 0;
2965
2966         for (i = offset >> 9; i < PAGE_SECTORS; i++)
2967                 if (s->s[i].state < SECTOR_DIRTY)
2968                         return i << 9;
2969
2970         return -1;
2971 }
2972
2973 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
2974 {
2975         pgoff_t index = offset >> PAGE_SHIFT;
2976         struct page *page;
2977         int pg_offset;
2978         loff_t ret = -1;
2979
2980         page = find_lock_entry(mapping, index);
2981         if (!page || xa_is_value(page))
2982                 return offset;
2983
2984         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
2985         if (pg_offset >= 0)
2986                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
2987
2988         unlock_page(page);
2989
2990         return ret;
2991 }
2992
2993 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
2994                                        loff_t start_offset,
2995                                        loff_t end_offset)
2996 {
2997         struct address_space *mapping = vinode->i_mapping;
2998         loff_t offset = start_offset, hole;
2999
3000         while (offset < end_offset) {
3001                 hole = page_hole_offset(mapping, offset);
3002                 if (hole >= 0 && hole <= end_offset)
3003                         return max(start_offset, hole);
3004
3005                 offset += PAGE_SIZE;
3006                 offset &= PAGE_MASK;
3007         }
3008
3009         return end_offset;
3010 }
3011
3012 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3013 {
3014         struct bch_inode_info *inode = file_bch_inode(file);
3015         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3016         struct btree_trans trans;
3017         struct btree_iter *iter;
3018         struct bkey_s_c k;
3019         u64 isize, next_hole = MAX_LFS_FILESIZE;
3020         int ret;
3021
3022         isize = i_size_read(&inode->v);
3023         if (offset >= isize)
3024                 return -ENXIO;
3025
3026         bch2_trans_init(&trans, c, 0, 0);
3027
3028         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3029                            POS(inode->v.i_ino, offset >> 9),
3030                            BTREE_ITER_SLOTS, k, ret) {
3031                 if (k.k->p.inode != inode->v.i_ino) {
3032                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3033                                         offset, MAX_LFS_FILESIZE);
3034                         break;
3035                 } else if (!bkey_extent_is_data(k.k)) {
3036                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3037                                         max(offset, bkey_start_offset(k.k) << 9),
3038                                         k.k->p.offset << 9);
3039
3040                         if (next_hole < k.k->p.offset << 9)
3041                                 break;
3042                 } else {
3043                         offset = max(offset, bkey_start_offset(k.k) << 9);
3044                 }
3045         }
3046
3047         ret = bch2_trans_exit(&trans) ?: ret;
3048         if (ret)
3049                 return ret;
3050
3051         if (next_hole > isize)
3052                 next_hole = isize;
3053
3054         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3055 }
3056
3057 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3058 {
3059         switch (whence) {
3060         case SEEK_SET:
3061         case SEEK_CUR:
3062         case SEEK_END:
3063                 return generic_file_llseek(file, offset, whence);
3064         case SEEK_DATA:
3065                 return bch2_seek_data(file, offset);
3066         case SEEK_HOLE:
3067                 return bch2_seek_hole(file, offset);
3068         }
3069
3070         return -EINVAL;
3071 }
3072
3073 void bch2_fs_fsio_exit(struct bch_fs *c)
3074 {
3075         bioset_exit(&c->dio_write_bioset);
3076         bioset_exit(&c->dio_read_bioset);
3077         bioset_exit(&c->writepage_bioset);
3078 }
3079
3080 int bch2_fs_fsio_init(struct bch_fs *c)
3081 {
3082         int ret = 0;
3083
3084         pr_verbose_init(c->opts, "");
3085
3086         if (bioset_init(&c->writepage_bioset,
3087                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3088                         BIOSET_NEED_BVECS) ||
3089             bioset_init(&c->dio_read_bioset,
3090                         4, offsetof(struct dio_read, rbio.bio),
3091                         BIOSET_NEED_BVECS) ||
3092             bioset_init(&c->dio_write_bioset,
3093                         4, offsetof(struct dio_write, op.wbio.bio),
3094                         BIOSET_NEED_BVECS))
3095                 ret = -ENOMEM;
3096
3097         pr_verbose_init(c->opts, "ret %i", ret);
3098         return ret;
3099 }
3100
3101 #endif /* NO_BCACHEFS_FS */