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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifndef NO_BCACHEFS_FS
   3
   4 #include "bcachefs.h"
   5 #include "alloc_foreground.h"
   6 #include "bkey_on_stack.h"
   7 #include "btree_update.h"
   8 #include "buckets.h"
   9 #include "clock.h"
  10 #include "error.h"
  11 #include "extents.h"
  12 #include "extent_update.h"
  13 #include "fs.h"
  14 #include "fs-io.h"
  15 #include "fsck.h"
  16 #include "inode.h"
  17 #include "journal.h"
  18 #include "io.h"
  19 #include "keylist.h"
  20 #include "quota.h"
  21 #include "reflink.h"
  22
  23 #include <linux/aio.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/falloc.h>
  26 #include <linux/migrate.h>
  27 #include <linux/mmu_context.h>
  28 #include <linux/pagevec.h>
  29 #include <linux/rmap.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/task_io_accounting_ops.h>
  32 #include <linux/uio.h>
  33 #include <linux/writeback.h>
  34
  35 #include <trace/events/bcachefs.h>
  36 #include <trace/events/writeback.h>
  37
  38 struct quota_res {
  39         u64                             sectors;
  40 };
  41
  42 struct bch_writepage_io {
  43         struct closure                  cl;
  44         struct bch_inode_info           *inode;
  45
  46         /* must be last: */
  47         struct bch_write_op             op;
  48 };
  49
  50 struct dio_write {
  51         struct completion               done;
  52         struct kiocb                    *req;
  53         struct mm_struct                *mm;
  54         unsigned                        loop:1,
  55                                         sync:1,
  56                                         free_iov:1;
  57         struct quota_res                quota_res;
  58         u64                             written;
  59
  60         struct iov_iter                 iter;
  61         struct iovec                    inline_vecs[2];
  62
  63         /* must be last: */
  64         struct bch_write_op             op;
  65 };
  66
  67 struct dio_read {
  68         struct closure                  cl;
  69         struct kiocb                    *req;
  70         long                            ret;
  71         struct bch_read_bio             rbio;
  72 };
  73
  74 /* pagecache_block must be held */
  75 static int write_invalidate_inode_pages_range(struct address_space *mapping,
  76                                               loff_t start, loff_t end)
  77 {
  78         int ret;
  79
  80         /*
  81          * XXX: the way this is currently implemented, we can spin if a process
  82          * is continually redirtying a specific page
  83          */
  84         do {
  85                 if (!mapping->nrpages &&
  86                     !mapping->nrexceptional)
  87                         return 0;
  88
  89                 ret = filemap_write_and_wait_range(mapping, start, end);
  90                 if (ret)
  91                         break;
  92
  93                 if (!mapping->nrpages)
  94                         return 0;
  95
  96                 ret = invalidate_inode_pages2_range(mapping,
  97                                 start >> PAGE_SHIFT,
  98                                 end >> PAGE_SHIFT);
  99         } while (ret == -EBUSY);
 100
 101         return ret;
 102 }
 103
 104 /* quotas */
 105
 106 #ifdef CONFIG_BCACHEFS_QUOTA
 107
 108 static void bch2_quota_reservation_put(struct bch_fs *c,
 109                                        struct bch_inode_info *inode,
 110                                        struct quota_res *res)
 111 {
 112         if (!res->sectors)
 113                 return;
 114
 115         mutex_lock(&inode->ei_quota_lock);
 116         BUG_ON(res->sectors > inode->ei_quota_reserved);
 117
 118         bch2_quota_acct(c, inode->ei_qid, Q_SPC,
 119                         -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
 120         inode->ei_quota_reserved -= res->sectors;
 121         mutex_unlock(&inode->ei_quota_lock);
 122
 123         res->sectors = 0;
 124 }
 125
 126 static int bch2_quota_reservation_add(struct bch_fs *c,
 127                                       struct bch_inode_info *inode,
 128                                       struct quota_res *res,
 129                                       unsigned sectors,
 130                                       bool check_enospc)
 131 {
 132         int ret;
 133
 134         mutex_lock(&inode->ei_quota_lock);
 135         ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
 136                               check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
 137         if (likely(!ret)) {
 138                 inode->ei_quota_reserved += sectors;
 139                 res->sectors += sectors;
 140         }
 141         mutex_unlock(&inode->ei_quota_lock);
 142
 143         return ret;
 144 }
 145
 146 #else
 147
 148 static void bch2_quota_reservation_put(struct bch_fs *c,
 149                                        struct bch_inode_info *inode,
 150                                        struct quota_res *res)
 151 {
 152 }
 153
 154 static int bch2_quota_reservation_add(struct bch_fs *c,
 155                                       struct bch_inode_info *inode,
 156                                       struct quota_res *res,
 157                                       unsigned sectors,
 158                                       bool check_enospc)
 159 {
 160         return 0;
 161 }
 162
 163 #endif
 164
 165 /* i_size updates: */
 166
 167 struct inode_new_size {
 168         loff_t          new_size;
 169         u64             now;
 170         unsigned        fields;
 171 };
 172
 173 static int inode_set_size(struct bch_inode_info *inode,
 174                           struct bch_inode_unpacked *bi,
 175                           void *p)
 176 {
 177         struct inode_new_size *s = p;
 178
 179         bi->bi_size = s->new_size;
 180         if (s->fields & ATTR_ATIME)
 181                 bi->bi_atime = s->now;
 182         if (s->fields & ATTR_MTIME)
 183                 bi->bi_mtime = s->now;
 184         if (s->fields & ATTR_CTIME)
 185                 bi->bi_ctime = s->now;
 186
 187         return 0;
 188 }
 189
 190 int __must_check bch2_write_inode_size(struct bch_fs *c,
 191                                        struct bch_inode_info *inode,
 192                                        loff_t new_size, unsigned fields)
 193 {
 194         struct inode_new_size s = {
 195                 .new_size       = new_size,
 196                 .now            = bch2_current_time(c),
 197                 .fields         = fields,
 198         };
 199
 200         return bch2_write_inode(c, inode, inode_set_size, &s, fields);
 201 }
 202
 203 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 204                            struct quota_res *quota_res, s64 sectors)
 205 {
 206         if (!sectors)
 207                 return;
 208
 209         mutex_lock(&inode->ei_quota_lock);
 210 #ifdef CONFIG_BCACHEFS_QUOTA
 211         if (quota_res && sectors > 0) {
 212                 BUG_ON(sectors > quota_res->sectors);
 213                 BUG_ON(sectors > inode->ei_quota_reserved);
 214
 215                 quota_res->sectors -= sectors;
 216                 inode->ei_quota_reserved -= sectors;
 217         } else {
 218                 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
 219         }
 220 #endif
 221         inode->v.i_blocks += sectors;
 222         mutex_unlock(&inode->ei_quota_lock);
 223 }
 224
 225 /* page state: */
 226
 227 /* stored in page->private: */
 228
 229 struct bch_page_sector {
 230         /* Uncompressed, fully allocated replicas: */
 231         unsigned                nr_replicas:3;
 232
 233         /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
 234         unsigned                replicas_reserved:3;
 235
 236         /* i_sectors: */
 237         enum {
 238                 SECTOR_UNALLOCATED,
 239                 SECTOR_RESERVED,
 240                 SECTOR_DIRTY,
 241                 SECTOR_ALLOCATED,
 242         }                       state:2;
 243 };
 244
 245 struct bch_page_state {
 246         spinlock_t              lock;
 247         atomic_t                write_count;
 248         struct bch_page_sector  s[PAGE_SECTORS];
 249 };
 250
 251 static inline struct bch_page_state *__bch2_page_state(struct page *page)
 252 {
 253         return page_has_private(page)
 254                 ? (struct bch_page_state *) page_private(page)
 255                 : NULL;
 256 }
 257
 258 static inline struct bch_page_state *bch2_page_state(struct page *page)
 259 {
 260         EBUG_ON(!PageLocked(page));
 261
 262         return __bch2_page_state(page);
 263 }
 264
 265 /* for newly allocated pages: */
 266 static void __bch2_page_state_release(struct page *page)
 267 {
 268         struct bch_page_state *s = __bch2_page_state(page);
 269
 270         if (!s)
 271                 return;
 272
 273         ClearPagePrivate(page);
 274         set_page_private(page, 0);
 275         put_page(page);
 276         kfree(s);
 277 }
 278
 279 static void bch2_page_state_release(struct page *page)
 280 {
 281         struct bch_page_state *s = bch2_page_state(page);
 282
 283         if (!s)
 284                 return;
 285
 286         ClearPagePrivate(page);
 287         set_page_private(page, 0);
 288         put_page(page);
 289         kfree(s);
 290 }
 291
 292 /* for newly allocated pages: */
 293 static struct bch_page_state *__bch2_page_state_create(struct page *page,
 294                                                        gfp_t gfp)
 295 {
 296         struct bch_page_state *s;
 297
 298         s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
 299         if (!s)
 300                 return NULL;
 301
 302         spin_lock_init(&s->lock);
 303         /*
 304          * migrate_page_move_mapping() assumes that pages with private data
 305          * have their count elevated by 1.
 306          */
 307         get_page(page);
 308         set_page_private(page, (unsigned long) s);
 309         SetPagePrivate(page);
 310         return s;
 311 }
 312
 313 static struct bch_page_state *bch2_page_state_create(struct page *page,
 314                                                      gfp_t gfp)
 315 {
 316         return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
 317 }
 318
 319 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
 320 {
 321         /* XXX: this should not be open coded */
 322         return inode->ei_inode.bi_data_replicas
 323                 ? inode->ei_inode.bi_data_replicas - 1
 324                 : c->opts.data_replicas;
 325 }
 326
 327 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
 328                                                   unsigned nr_replicas)
 329 {
 330         return max(0, (int) nr_replicas -
 331                    s->nr_replicas -
 332                    s->replicas_reserved);
 333 }
 334
 335 static int bch2_get_page_disk_reservation(struct bch_fs *c,
 336                                 struct bch_inode_info *inode,
 337                                 struct page *page, bool check_enospc)
 338 {
 339         struct bch_page_state *s = bch2_page_state_create(page, 0);
 340         unsigned nr_replicas = inode_nr_replicas(c, inode);
 341         struct disk_reservation disk_res = { 0 };
 342         unsigned i, disk_res_sectors = 0;
 343         int ret;
 344
 345         if (!s)
 346                 return -ENOMEM;
 347
 348         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 349                 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
 350
 351         if (!disk_res_sectors)
 352                 return 0;
 353
 354         ret = bch2_disk_reservation_get(c, &disk_res,
 355                                         disk_res_sectors, 1,
 356                                         !check_enospc
 357                                         ? BCH_DISK_RESERVATION_NOFAIL
 358                                         : 0);
 359         if (unlikely(ret))
 360                 return ret;
 361
 362         for (i = 0; i < ARRAY_SIZE(s->s); i++)
 363                 s->s[i].replicas_reserved +=
 364                         sectors_to_reserve(&s->s[i], nr_replicas);
 365
 366         return 0;
 367 }
 368
 369 struct bch2_page_reservation {
 370         struct disk_reservation disk;
 371         struct quota_res        quota;
 372 };
 373
 374 static void bch2_page_reservation_init(struct bch_fs *c,
 375                         struct bch_inode_info *inode,
 376                         struct bch2_page_reservation *res)
 377 {
 378         memset(res, 0, sizeof(*res));
 379
 380         res->disk.nr_replicas = inode_nr_replicas(c, inode);
 381 }
 382
 383 static void bch2_page_reservation_put(struct bch_fs *c,
 384                         struct bch_inode_info *inode,
 385                         struct bch2_page_reservation *res)
 386 {
 387         bch2_disk_reservation_put(c, &res->disk);
 388         bch2_quota_reservation_put(c, inode, &res->quota);
 389 }
 390
 391 static int bch2_page_reservation_get(struct bch_fs *c,
 392                         struct bch_inode_info *inode, struct page *page,
 393                         struct bch2_page_reservation *res,
 394                         unsigned offset, unsigned len, bool check_enospc)
 395 {
 396         struct bch_page_state *s = bch2_page_state_create(page, 0);
 397         unsigned i, disk_sectors = 0, quota_sectors = 0;
 398         int ret;
 399
 400         if (!s)
 401                 return -ENOMEM;
 402
 403         for (i = round_down(offset, block_bytes(c)) >> 9;
 404              i < round_up(offset + len, block_bytes(c)) >> 9;
 405              i++) {
 406                 disk_sectors += sectors_to_reserve(&s->s[i],
 407                                                 res->disk.nr_replicas);
 408                 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
 409         }
 410
 411         if (disk_sectors) {
 412                 ret = bch2_disk_reservation_add(c, &res->disk,
 413                                                 disk_sectors,
 414                                                 !check_enospc
 415                                                 ? BCH_DISK_RESERVATION_NOFAIL
 416                                                 : 0);
 417                 if (unlikely(ret))
 418                         return ret;
 419         }
 420
 421         if (quota_sectors) {
 422                 ret = bch2_quota_reservation_add(c, inode, &res->quota,
 423                                                  quota_sectors,
 424                                                  check_enospc);
 425                 if (unlikely(ret)) {
 426                         struct disk_reservation tmp = {
 427                                 .sectors = disk_sectors
 428                         };
 429
 430                         bch2_disk_reservation_put(c, &tmp);
 431                         res->disk.sectors -= disk_sectors;
 432                         return ret;
 433                 }
 434         }
 435
 436         return 0;
 437 }
 438
 439 static void bch2_clear_page_bits(struct page *page)
 440 {
 441         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 442         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 443         struct bch_page_state *s = bch2_page_state(page);
 444         struct disk_reservation disk_res = { 0 };
 445         int i, dirty_sectors = 0;
 446
 447         if (!s)
 448                 return;
 449
 450         EBUG_ON(!PageLocked(page));
 451         EBUG_ON(PageWriteback(page));
 452
 453         for (i = 0; i < ARRAY_SIZE(s->s); i++) {
 454                 disk_res.sectors += s->s[i].replicas_reserved;
 455                 s->s[i].replicas_reserved = 0;
 456
 457                 if (s->s[i].state == SECTOR_DIRTY) {
 458                         dirty_sectors++;
 459                         s->s[i].state = SECTOR_UNALLOCATED;
 460                 }
 461         }
 462
 463         bch2_disk_reservation_put(c, &disk_res);
 464
 465         if (dirty_sectors)
 466                 i_sectors_acct(c, inode, NULL, -dirty_sectors);
 467
 468         bch2_page_state_release(page);
 469 }
 470
 471 static void bch2_set_page_dirty(struct bch_fs *c,
 472                         struct bch_inode_info *inode, struct page *page,
 473                         struct bch2_page_reservation *res,
 474                         unsigned offset, unsigned len)
 475 {
 476         struct bch_page_state *s = bch2_page_state(page);
 477         unsigned i, dirty_sectors = 0;
 478
 479         WARN_ON((u64) page_offset(page) + offset + len >
 480                 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
 481
 482         spin_lock(&s->lock);
 483
 484         for (i = round_down(offset, block_bytes(c)) >> 9;
 485              i < round_up(offset + len, block_bytes(c)) >> 9;
 486              i++) {
 487                 unsigned sectors = sectors_to_reserve(&s->s[i],
 488                                                 res->disk.nr_replicas);
 489
 490                 /*
 491                  * This can happen if we race with the error path in
 492                  * bch2_writepage_io_done():
 493                  */
 494                 sectors = min_t(unsigned, sectors, res->disk.sectors);
 495
 496                 s->s[i].replicas_reserved += sectors;
 497                 res->disk.sectors -= sectors;
 498
 499                 if (s->s[i].state == SECTOR_UNALLOCATED)
 500                         dirty_sectors++;
 501
 502                 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
 503         }
 504
 505         spin_unlock(&s->lock);
 506
 507         if (dirty_sectors)
 508                 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
 509
 510         if (!PageDirty(page))
 511                 __set_page_dirty_nobuffers(page);
 512 }
 513
 514 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
 515 {
 516         struct file *file = vmf->vma->vm_file;
 517         struct bch_inode_info *inode = file_bch_inode(file);
 518         int ret;
 519
 520         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 521         ret = filemap_fault(vmf);
 522         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 523
 524         return ret;
 525 }
 526
 527 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
 528 {
 529         struct page *page = vmf->page;
 530         struct file *file = vmf->vma->vm_file;
 531         struct bch_inode_info *inode = file_bch_inode(file);
 532         struct address_space *mapping = file->f_mapping;
 533         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 534         struct bch2_page_reservation res;
 535         unsigned len;
 536         loff_t isize;
 537         int ret = VM_FAULT_LOCKED;
 538
 539         bch2_page_reservation_init(c, inode, &res);
 540
 541         sb_start_pagefault(inode->v.i_sb);
 542         file_update_time(file);
 543
 544         /*
 545          * Not strictly necessary, but helps avoid dio writes livelocking in
 546          * write_invalidate_inode_pages_range() - can drop this if/when we get
 547          * a write_invalidate_inode_pages_range() that works without dropping
 548          * page lock before invalidating page
 549          */
 550         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 551
 552         lock_page(page);
 553         isize = i_size_read(&inode->v);
 554
 555         if (page->mapping != mapping || page_offset(page) >= isize) {
 556                 unlock_page(page);
 557                 ret = VM_FAULT_NOPAGE;
 558                 goto out;
 559         }
 560
 561         len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
 562
 563         if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
 564                 unlock_page(page);
 565                 ret = VM_FAULT_SIGBUS;
 566                 goto out;
 567         }
 568
 569         bch2_set_page_dirty(c, inode, page, &res, 0, len);
 570         bch2_page_reservation_put(c, inode, &res);
 571
 572         wait_for_stable_page(page);
 573 out:
 574         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 575         sb_end_pagefault(inode->v.i_sb);
 576
 577         return ret;
 578 }
 579
 580 void bch2_invalidatepage(struct page *page, unsigned int offset,
 581                          unsigned int length)
 582 {
 583         if (offset || length < PAGE_SIZE)
 584                 return;
 585
 586         bch2_clear_page_bits(page);
 587 }
 588
 589 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
 590 {
 591         if (PageDirty(page))
 592                 return 0;
 593
 594         bch2_clear_page_bits(page);
 595         return 1;
 596 }
 597
 598 #ifdef CONFIG_MIGRATION
 599 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 600                       struct page *page, enum migrate_mode mode)
 601 {
 602         int ret;
 603
 604         EBUG_ON(!PageLocked(page));
 605         EBUG_ON(!PageLocked(newpage));
 606
 607         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 608         if (ret != MIGRATEPAGE_SUCCESS)
 609                 return ret;
 610
 611         if (PagePrivate(page)) {
 612                 ClearPagePrivate(page);
 613                 get_page(newpage);
 614                 set_page_private(newpage, page_private(page));
 615                 set_page_private(page, 0);
 616                 put_page(page);
 617                 SetPagePrivate(newpage);
 618         }
 619
 620         if (mode != MIGRATE_SYNC_NO_COPY)
 621                 migrate_page_copy(newpage, page);
 622         else
 623                 migrate_page_states(newpage, page);
 624         return MIGRATEPAGE_SUCCESS;
 625 }
 626 #endif
 627
 628 /* readpage(s): */
 629
 630 static void bch2_readpages_end_io(struct bio *bio)
 631 {
 632         struct bvec_iter_all iter;
 633         struct bio_vec *bv;
 634
 635         bio_for_each_segment_all(bv, bio, iter) {
 636                 struct page *page = bv->bv_page;
 637
 638                 if (!bio->bi_status) {
 639                         SetPageUptodate(page);
 640                 } else {
 641                         ClearPageUptodate(page);
 642                         SetPageError(page);
 643                 }
 644                 unlock_page(page);
 645         }
 646
 647         bio_put(bio);
 648 }
 649
 650 static inline void page_state_init_for_read(struct page *page)
 651 {
 652         SetPagePrivate(page);
 653         page->private = 0;
 654 }
 655
 656 struct readpages_iter {
 657         struct address_space    *mapping;
 658         struct page             **pages;
 659         unsigned                nr_pages;
 660         unsigned                nr_added;
 661         unsigned                idx;
 662         pgoff_t                 offset;
 663 };
 664
 665 static int readpages_iter_init(struct readpages_iter *iter,
 666                                struct address_space *mapping,
 667                                struct list_head *pages, unsigned nr_pages)
 668 {
 669         memset(iter, 0, sizeof(*iter));
 670
 671         iter->mapping   = mapping;
 672         iter->offset    = list_last_entry(pages, struct page, lru)->index;
 673
 674         iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 675         if (!iter->pages)
 676                 return -ENOMEM;
 677
 678         while (!list_empty(pages)) {
 679                 struct page *page = list_last_entry(pages, struct page, lru);
 680
 681                 __bch2_page_state_create(page, __GFP_NOFAIL);
 682
 683                 iter->pages[iter->nr_pages++] = page;
 684                 list_del(&page->lru);
 685         }
 686
 687         return 0;
 688 }
 689
 690 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
 691 {
 692         struct page *page;
 693         unsigned i;
 694         int ret;
 695
 696         BUG_ON(iter->idx > iter->nr_added);
 697         BUG_ON(iter->nr_added > iter->nr_pages);
 698
 699         if (iter->idx < iter->nr_added)
 700                 goto out;
 701
 702         while (1) {
 703                 if (iter->idx == iter->nr_pages)
 704                         return NULL;
 705
 706                 ret = add_to_page_cache_lru_vec(iter->mapping,
 707                                 iter->pages     + iter->nr_added,
 708                                 iter->nr_pages  - iter->nr_added,
 709                                 iter->offset    + iter->nr_added,
 710                                 GFP_NOFS);
 711                 if (ret > 0)
 712                         break;
 713
 714                 page = iter->pages[iter->nr_added];
 715                 iter->idx++;
 716                 iter->nr_added++;
 717
 718                 __bch2_page_state_release(page);
 719                 put_page(page);
 720         }
 721
 722         iter->nr_added += ret;
 723
 724         for (i = iter->idx; i < iter->nr_added; i++)
 725                 put_page(iter->pages[i]);
 726 out:
 727         EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
 728
 729         return iter->pages[iter->idx];
 730 }
 731
 732 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
 733 {
 734         struct bvec_iter iter;
 735         struct bio_vec bv;
 736         unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
 737                 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
 738         unsigned state = k.k->type == KEY_TYPE_reservation
 739                 ? SECTOR_RESERVED
 740                 : SECTOR_ALLOCATED;
 741
 742         bio_for_each_segment(bv, bio, iter) {
 743                 struct bch_page_state *s = bch2_page_state(bv.bv_page);
 744                 unsigned i;
 745
 746                 for (i = bv.bv_offset >> 9;
 747                      i < (bv.bv_offset + bv.bv_len) >> 9;
 748                      i++) {
 749                         s->s[i].nr_replicas = nr_ptrs;
 750                         s->s[i].state = state;
 751                 }
 752         }
 753 }
 754
 755 static bool extent_partial_reads_expensive(struct bkey_s_c k)
 756 {
 757         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 758         struct bch_extent_crc_unpacked crc;
 759         const union bch_extent_entry *i;
 760
 761         bkey_for_each_crc(k.k, ptrs, crc, i)
 762                 if (crc.csum_type || crc.compression_type)
 763                         return true;
 764         return false;
 765 }
 766
 767 static void readpage_bio_extend(struct readpages_iter *iter,
 768                                 struct bio *bio,
 769                                 unsigned sectors_this_extent,
 770                                 bool get_more)
 771 {
 772         while (bio_sectors(bio) < sectors_this_extent &&
 773                bio->bi_vcnt < bio->bi_max_vecs) {
 774                 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
 775                 struct page *page = readpage_iter_next(iter);
 776                 int ret;
 777
 778                 if (page) {
 779                         if (iter->offset + iter->idx != page_offset)
 780                                 break;
 781
 782                         iter->idx++;
 783                 } else {
 784                         if (!get_more)
 785                                 break;
 786
 787                         page = xa_load(&iter->mapping->i_pages, page_offset);
 788                         if (page && !xa_is_value(page))
 789                                 break;
 790
 791                         page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
 792                         if (!page)
 793                                 break;
 794
 795                         if (!__bch2_page_state_create(page, 0)) {
 796                                 put_page(page);
 797                                 break;
 798                         }
 799
 800                         ret = add_to_page_cache_lru(page, iter->mapping,
 801                                                     page_offset, GFP_NOFS);
 802                         if (ret) {
 803                                 __bch2_page_state_release(page);
 804                                 put_page(page);
 805                                 break;
 806                         }
 807
 808                         put_page(page);
 809                 }
 810
 811                 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
 812         }
 813 }
 814
 815 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
 816                        struct bch_read_bio *rbio, u64 inum,
 817                        struct readpages_iter *readpages_iter)
 818 {
 819         struct bch_fs *c = trans->c;
 820         struct bkey_on_stack sk;
 821         int flags = BCH_READ_RETRY_IF_STALE|
 822                 BCH_READ_MAY_PROMOTE;
 823         int ret = 0;
 824
 825         rbio->c = c;
 826         rbio->start_time = local_clock();
 827
 828         bkey_on_stack_init(&sk);
 829 retry:
 830         while (1) {
 831                 struct bkey_s_c k;
 832                 unsigned bytes, sectors, offset_into_extent;
 833
 834                 bch2_btree_iter_set_pos(iter,
 835                                 POS(inum, rbio->bio.bi_iter.bi_sector));
 836
 837                 k = bch2_btree_iter_peek_slot(iter);
 838                 ret = bkey_err(k);
 839                 if (ret)
 840                         break;
 841
 842                 bkey_on_stack_reassemble(&sk, c, k);
 843                 k = bkey_i_to_s_c(sk.k);
 844
 845                 offset_into_extent = iter->pos.offset -
 846                         bkey_start_offset(k.k);
 847                 sectors = k.k->size - offset_into_extent;
 848
 849                 ret = bch2_read_indirect_extent(trans,
 850                                         &offset_into_extent, &sk);
 851                 if (ret)
 852                         break;
 853
 854                 sectors = min(sectors, k.k->size - offset_into_extent);
 855
 856                 bch2_trans_unlock(trans);
 857
 858                 if (readpages_iter)
 859                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 860                                             extent_partial_reads_expensive(k));
 861
 862                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 863                 swap(rbio->bio.bi_iter.bi_size, bytes);
 864
 865                 if (rbio->bio.bi_iter.bi_size == bytes)
 866                         flags |= BCH_READ_LAST_FRAGMENT;
 867
 868                 if (bkey_extent_is_allocation(k.k))
 869                         bch2_add_page_sectors(&rbio->bio, k);
 870
 871                 bch2_read_extent(trans, rbio, k, offset_into_extent, flags);
 872
 873                 if (flags & BCH_READ_LAST_FRAGMENT)
 874                         break;
 875
 876                 swap(rbio->bio.bi_iter.bi_size, bytes);
 877                 bio_advance(&rbio->bio, bytes);
 878         }
 879
 880         if (ret == -EINTR)
 881                 goto retry;
 882
 883         if (ret) {
 884                 bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
 885                 bio_endio(&rbio->bio);
 886         }
 887
 888         bkey_on_stack_exit(&sk, c);
 889 }
 890
 891 int bch2_readpages(struct file *file, struct address_space *mapping,
 892                    struct list_head *pages, unsigned nr_pages)
 893 {
 894         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 895         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 896         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 897         struct btree_trans trans;
 898         struct btree_iter *iter;
 899         struct page *page;
 900         struct readpages_iter readpages_iter;
 901         int ret;
 902
 903         ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
 904         BUG_ON(ret);
 905
 906         bch2_trans_init(&trans, c, 0, 0);
 907
 908         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 909                                    BTREE_ITER_SLOTS);
 910
 911         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 912
 913         while ((page = readpage_iter_next(&readpages_iter))) {
 914                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 915                 unsigned n = min_t(unsigned,
 916                                    readpages_iter.nr_pages -
 917                                    readpages_iter.idx,
 918                                    BIO_MAX_PAGES);
 919                 struct bch_read_bio *rbio =
 920                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 921                                   opts);
 922
 923                 readpages_iter.idx++;
 924
 925                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 926                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 927                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 928                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 929
 930                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 931                            &readpages_iter);
 932         }
 933
 934         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 935
 936         bch2_trans_exit(&trans);
 937         kfree(readpages_iter.pages);
 938
 939         return 0;
 940 }
 941
 942 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 943                              u64 inum, struct page *page)
 944 {
 945         struct btree_trans trans;
 946         struct btree_iter *iter;
 947
 948         bch2_page_state_create(page, __GFP_NOFAIL);
 949
 950         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 951         rbio->bio.bi_iter.bi_sector =
 952                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 953         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 954
 955         bch2_trans_init(&trans, c, 0, 0);
 956         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 957                                    BTREE_ITER_SLOTS);
 958
 959         bchfs_read(&trans, iter, rbio, inum, NULL);
 960
 961         bch2_trans_exit(&trans);
 962 }
 963
 964 int bch2_readpage(struct file *file, struct page *page)
 965 {
 966         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 967         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 968         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 969         struct bch_read_bio *rbio;
 970
 971         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 972         rbio->bio.bi_end_io = bch2_readpages_end_io;
 973
 974         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 975         return 0;
 976 }
 977
 978 static void bch2_read_single_page_end_io(struct bio *bio)
 979 {
 980         complete(bio->bi_private);
 981 }
 982
 983 static int bch2_read_single_page(struct page *page,
 984                                  struct address_space *mapping)
 985 {
 986         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 987         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 988         struct bch_read_bio *rbio;
 989         int ret;
 990         DECLARE_COMPLETION_ONSTACK(done);
 991
 992         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 993                          io_opts(c, &inode->ei_inode));
 994         rbio->bio.bi_private = &done;
 995         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 996
 997         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 998         wait_for_completion(&done);
 999
1000         ret = blk_status_to_errno(rbio->bio.bi_status);
1001         bio_put(&rbio->bio);
1002
1003         if (ret < 0)
1004                 return ret;
1005
1006         SetPageUptodate(page);
1007         return 0;
1008 }
1009
1010 /* writepages: */
1011
1012 struct bch_writepage_state {
1013         struct bch_writepage_io *io;
1014         struct bch_io_opts      opts;
1015 };
1016
1017 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1018                                                                   struct bch_inode_info *inode)
1019 {
1020         return (struct bch_writepage_state) {
1021                 .opts = io_opts(c, &inode->ei_inode)
1022         };
1023 }
1024
1025 static void bch2_writepage_io_free(struct closure *cl)
1026 {
1027         struct bch_writepage_io *io = container_of(cl,
1028                                         struct bch_writepage_io, cl);
1029
1030         bio_put(&io->op.wbio.bio);
1031 }
1032
1033 static void bch2_writepage_io_done(struct closure *cl)
1034 {
1035         struct bch_writepage_io *io = container_of(cl,
1036                                         struct bch_writepage_io, cl);
1037         struct bch_fs *c = io->op.c;
1038         struct bio *bio = &io->op.wbio.bio;
1039         struct bvec_iter_all iter;
1040         struct bio_vec *bvec;
1041         unsigned i;
1042
1043         if (io->op.error) {
1044                 bio_for_each_segment_all(bvec, bio, iter) {
1045                         struct bch_page_state *s;
1046
1047                         SetPageError(bvec->bv_page);
1048                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1049
1050                         s = __bch2_page_state(bvec->bv_page);
1051                         spin_lock(&s->lock);
1052                         for (i = 0; i < PAGE_SECTORS; i++)
1053                                 s->s[i].nr_replicas = 0;
1054                         spin_unlock(&s->lock);
1055                 }
1056         }
1057
1058         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1059                 bio_for_each_segment_all(bvec, bio, iter) {
1060                         struct bch_page_state *s;
1061
1062                         s = __bch2_page_state(bvec->bv_page);
1063                         spin_lock(&s->lock);
1064                         for (i = 0; i < PAGE_SECTORS; i++)
1065                                 s->s[i].nr_replicas = 0;
1066                         spin_unlock(&s->lock);
1067                 }
1068         }
1069
1070         /*
1071          * racing with fallocate can cause us to add fewer sectors than
1072          * expected - but we shouldn't add more sectors than expected:
1073          */
1074         BUG_ON(io->op.i_sectors_delta > 0);
1075
1076         /*
1077          * (error (due to going RO) halfway through a page can screw that up
1078          * slightly)
1079          * XXX wtf?
1080            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1081          */
1082
1083         /*
1084          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1085          * before calling end_page_writeback:
1086          */
1087         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1088
1089         bio_for_each_segment_all(bvec, bio, iter) {
1090                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1091
1092                 if (atomic_dec_and_test(&s->write_count))
1093                         end_page_writeback(bvec->bv_page);
1094         }
1095
1096         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1097 }
1098
1099 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1100 {
1101         struct bch_writepage_io *io = w->io;
1102
1103         w->io = NULL;
1104         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1105         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1106 }
1107
1108 /*
1109  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1110  * possible, else allocating a new one:
1111  */
1112 static void bch2_writepage_io_alloc(struct bch_fs *c,
1113                                     struct writeback_control *wbc,
1114                                     struct bch_writepage_state *w,
1115                                     struct bch_inode_info *inode,
1116                                     u64 sector,
1117                                     unsigned nr_replicas)
1118 {
1119         struct bch_write_op *op;
1120
1121         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1122                                               BIO_MAX_PAGES,
1123                                               &c->writepage_bioset),
1124                              struct bch_writepage_io, op.wbio.bio);
1125
1126         closure_init(&w->io->cl, NULL);
1127         w->io->inode            = inode;
1128
1129         op                      = &w->io->op;
1130         bch2_write_op_init(op, c, w->opts);
1131         op->target              = w->opts.foreground_target;
1132         op_journal_seq_set(op, &inode->ei_journal_seq);
1133         op->nr_replicas         = nr_replicas;
1134         op->res.nr_replicas     = nr_replicas;
1135         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1136         op->pos                 = POS(inode->v.i_ino, sector);
1137         op->wbio.bio.bi_iter.bi_sector = sector;
1138         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1139 }
1140
1141 static int __bch2_writepage(struct page *page,
1142                             struct writeback_control *wbc,
1143                             void *data)
1144 {
1145         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1146         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1147         struct bch_writepage_state *w = data;
1148         struct bch_page_state *s, orig;
1149         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1150         loff_t i_size = i_size_read(&inode->v);
1151         pgoff_t end_index = i_size >> PAGE_SHIFT;
1152         int ret;
1153
1154         EBUG_ON(!PageUptodate(page));
1155
1156         /* Is the page fully inside i_size? */
1157         if (page->index < end_index)
1158                 goto do_io;
1159
1160         /* Is the page fully outside i_size? (truncate in progress) */
1161         offset = i_size & (PAGE_SIZE - 1);
1162         if (page->index > end_index || !offset) {
1163                 unlock_page(page);
1164                 return 0;
1165         }
1166
1167         /*
1168          * The page straddles i_size.  It must be zeroed out on each and every
1169          * writepage invocation because it may be mmapped.  "A file is mapped
1170          * in multiples of the page size.  For a file that is not a multiple of
1171          * the  page size, the remaining memory is zeroed when mapped, and
1172          * writes to that region are not written out to the file."
1173          */
1174         zero_user_segment(page, offset, PAGE_SIZE);
1175 do_io:
1176         s = bch2_page_state_create(page, __GFP_NOFAIL);
1177
1178         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1179         if (ret) {
1180                 SetPageError(page);
1181                 mapping_set_error(page->mapping, ret);
1182                 unlock_page(page);
1183                 return 0;
1184         }
1185
1186         /* Before unlocking the page, get copy of reservations: */
1187         orig = *s;
1188
1189         for (i = 0; i < PAGE_SECTORS; i++) {
1190                 if (s->s[i].state < SECTOR_DIRTY)
1191                         continue;
1192
1193                 nr_replicas_this_write =
1194                         min_t(unsigned, nr_replicas_this_write,
1195                               s->s[i].nr_replicas +
1196                               s->s[i].replicas_reserved);
1197         }
1198
1199         for (i = 0; i < PAGE_SECTORS; i++) {
1200                 if (s->s[i].state < SECTOR_DIRTY)
1201                         continue;
1202
1203                 s->s[i].nr_replicas = w->opts.compression
1204                         ? 0 : nr_replicas_this_write;
1205
1206                 s->s[i].replicas_reserved = 0;
1207                 s->s[i].state = SECTOR_ALLOCATED;
1208         }
1209
1210         BUG_ON(atomic_read(&s->write_count));
1211         atomic_set(&s->write_count, 1);
1212
1213         BUG_ON(PageWriteback(page));
1214         set_page_writeback(page);
1215
1216         unlock_page(page);
1217
1218         offset = 0;
1219         while (1) {
1220                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1221                 u64 sector;
1222
1223                 while (offset < PAGE_SECTORS &&
1224                        orig.s[offset].state < SECTOR_DIRTY)
1225                         offset++;
1226
1227                 if (offset == PAGE_SECTORS)
1228                         break;
1229
1230                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1231
1232                 while (offset + sectors < PAGE_SECTORS &&
1233                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1234                         sectors++;
1235
1236                 for (i = offset; i < offset + sectors; i++) {
1237                         reserved_sectors += orig.s[i].replicas_reserved;
1238                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1239                 }
1240
1241                 if (w->io &&
1242                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1243                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1244                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1245                      (BIO_MAX_PAGES * PAGE_SIZE) ||
1246                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1247                         bch2_writepage_do_io(w);
1248
1249                 if (!w->io)
1250                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1251                                                 nr_replicas_this_write);
1252
1253                 atomic_inc(&s->write_count);
1254
1255                 BUG_ON(inode != w->io->inode);
1256                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1257                                      sectors << 9, offset << 9));
1258
1259                 /* Check for writing past i_size: */
1260                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1261                         round_up(i_size, block_bytes(c)));
1262
1263                 w->io->op.res.sectors += reserved_sectors;
1264                 w->io->op.i_sectors_delta -= dirty_sectors;
1265                 w->io->op.new_i_size = i_size;
1266
1267                 offset += sectors;
1268         }
1269
1270         if (atomic_dec_and_test(&s->write_count))
1271                 end_page_writeback(page);
1272
1273         return 0;
1274 }
1275
1276 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1277 {
1278         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1279         struct bch_writepage_state w =
1280                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1281         struct blk_plug plug;
1282         int ret;
1283
1284         blk_start_plug(&plug);
1285         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1286         if (w.io)
1287                 bch2_writepage_do_io(&w);
1288         blk_finish_plug(&plug);
1289         return ret;
1290 }
1291
1292 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1293 {
1294         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1295         struct bch_writepage_state w =
1296                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1297         int ret;
1298
1299         ret = __bch2_writepage(page, wbc, &w);
1300         if (w.io)
1301                 bch2_writepage_do_io(&w);
1302
1303         return ret;
1304 }
1305
1306 /* buffered writes: */
1307
1308 int bch2_write_begin(struct file *file, struct address_space *mapping,
1309                      loff_t pos, unsigned len, unsigned flags,
1310                      struct page **pagep, void **fsdata)
1311 {
1312         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1313         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1314         struct bch2_page_reservation *res;
1315         pgoff_t index = pos >> PAGE_SHIFT;
1316         unsigned offset = pos & (PAGE_SIZE - 1);
1317         struct page *page;
1318         int ret = -ENOMEM;
1319
1320         res = kmalloc(sizeof(*res), GFP_KERNEL);
1321         if (!res)
1322                 return -ENOMEM;
1323
1324         bch2_page_reservation_init(c, inode, res);
1325         *fsdata = res;
1326
1327         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1328
1329         page = grab_cache_page_write_begin(mapping, index, flags);
1330         if (!page)
1331                 goto err_unlock;
1332
1333         if (PageUptodate(page))
1334                 goto out;
1335
1336         /* If we're writing entire page, don't need to read it in first: */
1337         if (len == PAGE_SIZE)
1338                 goto out;
1339
1340         if (!offset && pos + len >= inode->v.i_size) {
1341                 zero_user_segment(page, len, PAGE_SIZE);
1342                 flush_dcache_page(page);
1343                 goto out;
1344         }
1345
1346         if (index > inode->v.i_size >> PAGE_SHIFT) {
1347                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1348                 flush_dcache_page(page);
1349                 goto out;
1350         }
1351 readpage:
1352         ret = bch2_read_single_page(page, mapping);
1353         if (ret)
1354                 goto err;
1355 out:
1356         ret = bch2_page_reservation_get(c, inode, page, res,
1357                                         offset, len, true);
1358         if (ret) {
1359                 if (!PageUptodate(page)) {
1360                         /*
1361                          * If the page hasn't been read in, we won't know if we
1362                          * actually need a reservation - we don't actually need
1363                          * to read here, we just need to check if the page is
1364                          * fully backed by uncompressed data:
1365                          */
1366                         goto readpage;
1367                 }
1368
1369                 goto err;
1370         }
1371
1372         *pagep = page;
1373         return 0;
1374 err:
1375         unlock_page(page);
1376         put_page(page);
1377         *pagep = NULL;
1378 err_unlock:
1379         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1380         kfree(res);
1381         *fsdata = NULL;
1382         return ret;
1383 }
1384
1385 int bch2_write_end(struct file *file, struct address_space *mapping,
1386                    loff_t pos, unsigned len, unsigned copied,
1387                    struct page *page, void *fsdata)
1388 {
1389         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1390         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1391         struct bch2_page_reservation *res = fsdata;
1392         unsigned offset = pos & (PAGE_SIZE - 1);
1393
1394         lockdep_assert_held(&inode->v.i_rwsem);
1395
1396         if (unlikely(copied < len && !PageUptodate(page))) {
1397                 /*
1398                  * The page needs to be read in, but that would destroy
1399                  * our partial write - simplest thing is to just force
1400                  * userspace to redo the write:
1401                  */
1402                 zero_user(page, 0, PAGE_SIZE);
1403                 flush_dcache_page(page);
1404                 copied = 0;
1405         }
1406
1407         spin_lock(&inode->v.i_lock);
1408         if (pos + copied > inode->v.i_size)
1409                 i_size_write(&inode->v, pos + copied);
1410         spin_unlock(&inode->v.i_lock);
1411
1412         if (copied) {
1413                 if (!PageUptodate(page))
1414                         SetPageUptodate(page);
1415
1416                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1417
1418                 inode->ei_last_dirtied = (unsigned long) current;
1419         }
1420
1421         unlock_page(page);
1422         put_page(page);
1423         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1424
1425         bch2_page_reservation_put(c, inode, res);
1426         kfree(res);
1427
1428         return copied;
1429 }
1430
1431 #define WRITE_BATCH_PAGES       32
1432
1433 static int __bch2_buffered_write(struct bch_inode_info *inode,
1434                                  struct address_space *mapping,
1435                                  struct iov_iter *iter,
1436                                  loff_t pos, unsigned len)
1437 {
1438         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1439         struct page *pages[WRITE_BATCH_PAGES];
1440         struct bch2_page_reservation res;
1441         unsigned long index = pos >> PAGE_SHIFT;
1442         unsigned offset = pos & (PAGE_SIZE - 1);
1443         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1444         unsigned i, reserved = 0, set_dirty = 0;
1445         unsigned copied = 0, nr_pages_copied = 0;
1446         int ret = 0;
1447
1448         BUG_ON(!len);
1449         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1450
1451         bch2_page_reservation_init(c, inode, &res);
1452
1453         for (i = 0; i < nr_pages; i++) {
1454                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1455                 if (!pages[i]) {
1456                         nr_pages = i;
1457                         if (!i) {
1458                                 ret = -ENOMEM;
1459                                 goto out;
1460                         }
1461                         len = min_t(unsigned, len,
1462                                     nr_pages * PAGE_SIZE - offset);
1463                         break;
1464                 }
1465         }
1466
1467         if (offset && !PageUptodate(pages[0])) {
1468                 ret = bch2_read_single_page(pages[0], mapping);
1469                 if (ret)
1470                         goto out;
1471         }
1472
1473         if ((pos + len) & (PAGE_SIZE - 1) &&
1474             !PageUptodate(pages[nr_pages - 1])) {
1475                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1476                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1477                 } else {
1478                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1479                         if (ret)
1480                                 goto out;
1481                 }
1482         }
1483
1484         while (reserved < len) {
1485                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1486                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1487                 unsigned pg_len = min_t(unsigned, len - reserved,
1488                                         PAGE_SIZE - pg_offset);
1489 retry_reservation:
1490                 ret = bch2_page_reservation_get(c, inode, page, &res,
1491                                                 pg_offset, pg_len, true);
1492
1493                 if (ret && !PageUptodate(page)) {
1494                         ret = bch2_read_single_page(page, mapping);
1495                         if (!ret)
1496                                 goto retry_reservation;
1497                 }
1498
1499                 if (ret)
1500                         goto out;
1501
1502                 reserved += pg_len;
1503         }
1504
1505         if (mapping_writably_mapped(mapping))
1506                 for (i = 0; i < nr_pages; i++)
1507                         flush_dcache_page(pages[i]);
1508
1509         while (copied < len) {
1510                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1511                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1512                 unsigned pg_len = min_t(unsigned, len - copied,
1513                                         PAGE_SIZE - pg_offset);
1514                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1515                                                 iter, pg_offset, pg_len);
1516
1517                 if (!pg_copied)
1518                         break;
1519
1520                 if (!PageUptodate(page) &&
1521                     pg_copied != PAGE_SIZE &&
1522                     pos + copied + pg_copied < inode->v.i_size) {
1523                         zero_user(page, 0, PAGE_SIZE);
1524                         break;
1525                 }
1526
1527                 flush_dcache_page(page);
1528                 iov_iter_advance(iter, pg_copied);
1529                 copied += pg_copied;
1530
1531                 if (pg_copied != pg_len)
1532                         break;
1533         }
1534
1535         if (!copied)
1536                 goto out;
1537
1538         spin_lock(&inode->v.i_lock);
1539         if (pos + copied > inode->v.i_size)
1540                 i_size_write(&inode->v, pos + copied);
1541         spin_unlock(&inode->v.i_lock);
1542
1543         while (set_dirty < copied) {
1544                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1545                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1546                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1547                                         PAGE_SIZE - pg_offset);
1548
1549                 if (!PageUptodate(page))
1550                         SetPageUptodate(page);
1551
1552                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1553                 unlock_page(page);
1554                 put_page(page);
1555
1556                 set_dirty += pg_len;
1557         }
1558
1559         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1560         inode->ei_last_dirtied = (unsigned long) current;
1561 out:
1562         for (i = nr_pages_copied; i < nr_pages; i++) {
1563                 unlock_page(pages[i]);
1564                 put_page(pages[i]);
1565         }
1566
1567         bch2_page_reservation_put(c, inode, &res);
1568
1569         return copied ?: ret;
1570 }
1571
1572 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1573 {
1574         struct file *file = iocb->ki_filp;
1575         struct address_space *mapping = file->f_mapping;
1576         struct bch_inode_info *inode = file_bch_inode(file);
1577         loff_t pos = iocb->ki_pos;
1578         ssize_t written = 0;
1579         int ret = 0;
1580
1581         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1582
1583         do {
1584                 unsigned offset = pos & (PAGE_SIZE - 1);
1585                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1586                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1587 again:
1588                 /*
1589                  * Bring in the user page that we will copy from _first_.
1590                  * Otherwise there's a nasty deadlock on copying from the
1591                  * same page as we're writing to, without it being marked
1592                  * up-to-date.
1593                  *
1594                  * Not only is this an optimisation, but it is also required
1595                  * to check that the address is actually valid, when atomic
1596                  * usercopies are used, below.
1597                  */
1598                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1599                         bytes = min_t(unsigned long, iov_iter_count(iter),
1600                                       PAGE_SIZE - offset);
1601
1602                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1603                                 ret = -EFAULT;
1604                                 break;
1605                         }
1606                 }
1607
1608                 if (unlikely(fatal_signal_pending(current))) {
1609                         ret = -EINTR;
1610                         break;
1611                 }
1612
1613                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1614                 if (unlikely(ret < 0))
1615                         break;
1616
1617                 cond_resched();
1618
1619                 if (unlikely(ret == 0)) {
1620                         /*
1621                          * If we were unable to copy any data at all, we must
1622                          * fall back to a single segment length write.
1623                          *
1624                          * If we didn't fallback here, we could livelock
1625                          * because not all segments in the iov can be copied at
1626                          * once without a pagefault.
1627                          */
1628                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1629                                       iov_iter_single_seg_count(iter));
1630                         goto again;
1631                 }
1632                 pos += ret;
1633                 written += ret;
1634                 ret = 0;
1635
1636                 balance_dirty_pages_ratelimited(mapping);
1637         } while (iov_iter_count(iter));
1638
1639         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1640
1641         return written ? written : ret;
1642 }
1643
1644 /* O_DIRECT reads */
1645
1646 static void bch2_dio_read_complete(struct closure *cl)
1647 {
1648         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1649
1650         dio->req->ki_complete(dio->req, dio->ret, 0);
1651         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1652 }
1653
1654 static void bch2_direct_IO_read_endio(struct bio *bio)
1655 {
1656         struct dio_read *dio = bio->bi_private;
1657
1658         if (bio->bi_status)
1659                 dio->ret = blk_status_to_errno(bio->bi_status);
1660
1661         closure_put(&dio->cl);
1662 }
1663
1664 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1665 {
1666         bch2_direct_IO_read_endio(bio);
1667         bio_check_pages_dirty(bio);     /* transfers ownership */
1668 }
1669
1670 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1671 {
1672         struct file *file = req->ki_filp;
1673         struct bch_inode_info *inode = file_bch_inode(file);
1674         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1675         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1676         struct dio_read *dio;
1677         struct bio *bio;
1678         loff_t offset = req->ki_pos;
1679         bool sync = is_sync_kiocb(req);
1680         size_t shorten;
1681         ssize_t ret;
1682
1683         if ((offset|iter->count) & (block_bytes(c) - 1))
1684                 return -EINVAL;
1685
1686         ret = min_t(loff_t, iter->count,
1687                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1688
1689         if (!ret)
1690                 return ret;
1691
1692         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1693         iter->count -= shorten;
1694
1695         bio = bio_alloc_bioset(GFP_KERNEL,
1696                                iov_iter_npages(iter, BIO_MAX_PAGES),
1697                                &c->dio_read_bioset);
1698
1699         bio->bi_end_io = bch2_direct_IO_read_endio;
1700
1701         dio = container_of(bio, struct dio_read, rbio.bio);
1702         closure_init(&dio->cl, NULL);
1703
1704         /*
1705          * this is a _really_ horrible hack just to avoid an atomic sub at the
1706          * end:
1707          */
1708         if (!sync) {
1709                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1710                 atomic_set(&dio->cl.remaining,
1711                            CLOSURE_REMAINING_INITIALIZER -
1712                            CLOSURE_RUNNING +
1713                            CLOSURE_DESTRUCTOR);
1714         } else {
1715                 atomic_set(&dio->cl.remaining,
1716                            CLOSURE_REMAINING_INITIALIZER + 1);
1717         }
1718
1719         dio->req        = req;
1720         dio->ret        = ret;
1721
1722         goto start;
1723         while (iter->count) {
1724                 bio = bio_alloc_bioset(GFP_KERNEL,
1725                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1726                                        &c->bio_read);
1727                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1728 start:
1729                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1730                 bio->bi_iter.bi_sector  = offset >> 9;
1731                 bio->bi_private         = dio;
1732
1733                 ret = bio_iov_iter_get_pages(bio, iter);
1734                 if (ret < 0) {
1735                         /* XXX: fault inject this path */
1736                         bio->bi_status = BLK_STS_RESOURCE;
1737                         bio_endio(bio);
1738                         break;
1739                 }
1740
1741                 offset += bio->bi_iter.bi_size;
1742                 bio_set_pages_dirty(bio);
1743
1744                 if (iter->count)
1745                         closure_get(&dio->cl);
1746
1747                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1748         }
1749
1750         iter->count += shorten;
1751
1752         if (sync) {
1753                 closure_sync(&dio->cl);
1754                 closure_debug_destroy(&dio->cl);
1755                 ret = dio->ret;
1756                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1757                 return ret;
1758         } else {
1759                 return -EIOCBQUEUED;
1760         }
1761 }
1762
1763 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1764 {
1765         struct file *file = iocb->ki_filp;
1766         struct bch_inode_info *inode = file_bch_inode(file);
1767         struct address_space *mapping = file->f_mapping;
1768         size_t count = iov_iter_count(iter);
1769         ssize_t ret;
1770
1771         if (!count)
1772                 return 0; /* skip atime */
1773
1774         if (iocb->ki_flags & IOCB_DIRECT) {
1775                 struct blk_plug plug;
1776
1777                 ret = filemap_write_and_wait_range(mapping,
1778                                         iocb->ki_pos,
1779                                         iocb->ki_pos + count - 1);
1780                 if (ret < 0)
1781                         return ret;
1782
1783                 file_accessed(file);
1784
1785                 blk_start_plug(&plug);
1786                 ret = bch2_direct_IO_read(iocb, iter);
1787                 blk_finish_plug(&plug);
1788
1789                 if (ret >= 0)
1790                         iocb->ki_pos += ret;
1791         } else {
1792                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1793                 ret = generic_file_read_iter(iocb, iter);
1794                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1795         }
1796
1797         return ret;
1798 }
1799
1800 /* O_DIRECT writes */
1801
1802 static void bch2_dio_write_loop_async(struct bch_write_op *);
1803
1804 static long bch2_dio_write_loop(struct dio_write *dio)
1805 {
1806         bool kthread = (current->flags & PF_KTHREAD) != 0;
1807         struct kiocb *req = dio->req;
1808         struct address_space *mapping = req->ki_filp->f_mapping;
1809         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1810         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1811         struct bio *bio = &dio->op.wbio.bio;
1812         struct bvec_iter_all iter;
1813         struct bio_vec *bv;
1814         unsigned unaligned;
1815         bool sync = dio->sync;
1816         long ret;
1817
1818         if (dio->loop)
1819                 goto loop;
1820
1821         while (1) {
1822                 if (kthread)
1823                         kthread_use_mm(dio->mm);
1824                 BUG_ON(current->faults_disabled_mapping);
1825                 current->faults_disabled_mapping = mapping;
1826
1827                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1828
1829                 current->faults_disabled_mapping = NULL;
1830                 if (kthread)
1831                         kthread_unuse_mm(dio->mm);
1832
1833                 if (unlikely(ret < 0))
1834                         goto err;
1835
1836                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1837                 bio->bi_iter.bi_size -= unaligned;
1838                 iov_iter_revert(&dio->iter, unaligned);
1839
1840                 if (!bio->bi_iter.bi_size) {
1841                         /*
1842                          * bio_iov_iter_get_pages was only able to get <
1843                          * blocksize worth of pages:
1844                          */
1845                         bio_for_each_segment_all(bv, bio, iter)
1846                                 put_page(bv->bv_page);
1847                         ret = -EFAULT;
1848                         goto err;
1849                 }
1850
1851                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1852                 dio->op.end_io          = bch2_dio_write_loop_async;
1853                 dio->op.target          = dio->op.opts.foreground_target;
1854                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1855                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1856                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1857                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1858
1859                 if ((req->ki_flags & IOCB_DSYNC) &&
1860                     !c->opts.journal_flush_disabled)
1861                         dio->op.flags |= BCH_WRITE_FLUSH;
1862
1863                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1864                                                 dio->op.opts.data_replicas, 0);
1865                 if (unlikely(ret) &&
1866                     !bch2_check_range_allocated(c, dio->op.pos,
1867                                 bio_sectors(bio), dio->op.opts.data_replicas))
1868                         goto err;
1869
1870                 task_io_account_write(bio->bi_iter.bi_size);
1871
1872                 if (!dio->sync && !dio->loop && dio->iter.count) {
1873                         struct iovec *iov = dio->inline_vecs;
1874
1875                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1876                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1877                                               GFP_KERNEL);
1878                                 if (unlikely(!iov)) {
1879                                         dio->sync = sync = true;
1880                                         goto do_io;
1881                                 }
1882
1883                                 dio->free_iov = true;
1884                         }
1885
1886                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1887                         dio->iter.iov = iov;
1888                 }
1889 do_io:
1890                 dio->loop = true;
1891                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1892
1893                 if (sync)
1894                         wait_for_completion(&dio->done);
1895                 else
1896                         return -EIOCBQUEUED;
1897 loop:
1898                 i_sectors_acct(c, inode, &dio->quota_res,
1899                                dio->op.i_sectors_delta);
1900                 req->ki_pos += (u64) dio->op.written << 9;
1901                 dio->written += dio->op.written;
1902
1903                 spin_lock(&inode->v.i_lock);
1904                 if (req->ki_pos > inode->v.i_size)
1905                         i_size_write(&inode->v, req->ki_pos);
1906                 spin_unlock(&inode->v.i_lock);
1907
1908                 bio_for_each_segment_all(bv, bio, iter)
1909                         put_page(bv->bv_page);
1910                 if (!dio->iter.count || dio->op.error)
1911                         break;
1912
1913                 bio_reset(bio);
1914                 reinit_completion(&dio->done);
1915         }
1916
1917         ret = dio->op.error ?: ((long) dio->written << 9);
1918 err:
1919         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1920         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1921
1922         if (dio->free_iov)
1923                 kfree(dio->iter.iov);
1924
1925         bio_put(bio);
1926
1927         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1928         inode_dio_end(&inode->v);
1929
1930         if (!sync) {
1931                 req->ki_complete(req, ret, 0);
1932                 ret = -EIOCBQUEUED;
1933         }
1934         return ret;
1935 }
1936
1937 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1938 {
1939         struct dio_write *dio = container_of(op, struct dio_write, op);
1940
1941         if (dio->sync)
1942                 complete(&dio->done);
1943         else
1944                 bch2_dio_write_loop(dio);
1945 }
1946
1947 static noinline
1948 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1949 {
1950         struct file *file = req->ki_filp;
1951         struct address_space *mapping = file->f_mapping;
1952         struct bch_inode_info *inode = file_bch_inode(file);
1953         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1954         struct dio_write *dio;
1955         struct bio *bio;
1956         bool locked = true, extending;
1957         ssize_t ret;
1958
1959         prefetch(&c->opts);
1960         prefetch((void *) &c->opts + 64);
1961         prefetch(&inode->ei_inode);
1962         prefetch((void *) &inode->ei_inode + 64);
1963
1964         inode_lock(&inode->v);
1965
1966         ret = generic_write_checks(req, iter);
1967         if (unlikely(ret <= 0))
1968                 goto err;
1969
1970         ret = file_remove_privs(file);
1971         if (unlikely(ret))
1972                 goto err;
1973
1974         ret = file_update_time(file);
1975         if (unlikely(ret))
1976                 goto err;
1977
1978         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1979                 goto err;
1980
1981         inode_dio_begin(&inode->v);
1982         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
1983
1984         extending = req->ki_pos + iter->count > inode->v.i_size;
1985         if (!extending) {
1986                 inode_unlock(&inode->v);
1987                 locked = false;
1988         }
1989
1990         bio = bio_alloc_bioset(GFP_KERNEL,
1991                                iov_iter_npages(iter, BIO_MAX_PAGES),
1992                                &c->dio_write_bioset);
1993         dio = container_of(bio, struct dio_write, op.wbio.bio);
1994         init_completion(&dio->done);
1995         dio->req                = req;
1996         dio->mm                 = current->mm;
1997         dio->loop               = false;
1998         dio->sync               = is_sync_kiocb(req) || extending;
1999         dio->free_iov           = false;
2000         dio->quota_res.sectors  = 0;
2001         dio->written            = 0;
2002         dio->iter               = *iter;
2003
2004         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2005                                          iter->count >> 9, true);
2006         if (unlikely(ret))
2007                 goto err_put_bio;
2008
2009         ret = write_invalidate_inode_pages_range(mapping,
2010                                         req->ki_pos,
2011                                         req->ki_pos + iter->count - 1);
2012         if (unlikely(ret))
2013                 goto err_put_bio;
2014
2015         ret = bch2_dio_write_loop(dio);
2016 err:
2017         if (locked)
2018                 inode_unlock(&inode->v);
2019         return ret;
2020 err_put_bio:
2021         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2022         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2023         bio_put(bio);
2024         inode_dio_end(&inode->v);
2025         goto err;
2026 }
2027
2028 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2029 {
2030         struct file *file = iocb->ki_filp;
2031         struct bch_inode_info *inode = file_bch_inode(file);
2032         ssize_t ret;
2033
2034         if (iocb->ki_flags & IOCB_DIRECT)
2035                 return bch2_direct_write(iocb, from);
2036
2037         /* We can write back this queue in page reclaim */
2038         current->backing_dev_info = inode_to_bdi(&inode->v);
2039         inode_lock(&inode->v);
2040
2041         ret = generic_write_checks(iocb, from);
2042         if (ret <= 0)
2043                 goto unlock;
2044
2045         ret = file_remove_privs(file);
2046         if (ret)
2047                 goto unlock;
2048
2049         ret = file_update_time(file);
2050         if (ret)
2051                 goto unlock;
2052
2053         ret = bch2_buffered_write(iocb, from);
2054         if (likely(ret > 0))
2055                 iocb->ki_pos += ret;
2056 unlock:
2057         inode_unlock(&inode->v);
2058         current->backing_dev_info = NULL;
2059
2060         if (ret > 0)
2061                 ret = generic_write_sync(iocb, ret);
2062
2063         return ret;
2064 }
2065
2066 /* fsync: */
2067
2068 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2069 {
2070         struct bch_inode_info *inode = file_bch_inode(file);
2071         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2072         int ret, ret2;
2073
2074         ret = file_write_and_wait_range(file, start, end);
2075         if (ret)
2076                 return ret;
2077
2078         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2079                 goto out;
2080
2081         ret = sync_inode_metadata(&inode->v, 1);
2082         if (ret)
2083                 return ret;
2084 out:
2085         if (!c->opts.journal_flush_disabled)
2086                 ret = bch2_journal_flush_seq(&c->journal,
2087                                              inode->ei_journal_seq);
2088         ret2 = file_check_and_advance_wb_err(file);
2089
2090         return ret ?: ret2;
2091 }
2092
2093 /* truncate: */
2094
2095 static inline int range_has_data(struct bch_fs *c,
2096                                   struct bpos start,
2097                                   struct bpos end)
2098 {
2099         struct btree_trans trans;
2100         struct btree_iter *iter;
2101         struct bkey_s_c k;
2102         int ret = 0;
2103
2104         bch2_trans_init(&trans, c, 0, 0);
2105
2106         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2107                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2108                         break;
2109
2110                 if (bkey_extent_is_data(k.k)) {
2111                         ret = 1;
2112                         break;
2113                 }
2114         }
2115
2116         return bch2_trans_exit(&trans) ?: ret;
2117 }
2118
2119 static int __bch2_truncate_page(struct bch_inode_info *inode,
2120                                 pgoff_t index, loff_t start, loff_t end)
2121 {
2122         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2123         struct address_space *mapping = inode->v.i_mapping;
2124         struct bch_page_state *s;
2125         unsigned start_offset = start & (PAGE_SIZE - 1);
2126         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2127         unsigned i;
2128         struct page *page;
2129         int ret = 0;
2130
2131         /* Page boundary? Nothing to do */
2132         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2133               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2134                 return 0;
2135
2136         /* Above i_size? */
2137         if (index << PAGE_SHIFT >= inode->v.i_size)
2138                 return 0;
2139
2140         page = find_lock_page(mapping, index);
2141         if (!page) {
2142                 /*
2143                  * XXX: we're doing two index lookups when we end up reading the
2144                  * page
2145                  */
2146                 ret = range_has_data(c,
2147                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2148                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2149                 if (ret <= 0)
2150                         return ret;
2151
2152                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2153                 if (unlikely(!page)) {
2154                         ret = -ENOMEM;
2155                         goto out;
2156                 }
2157         }
2158
2159         s = bch2_page_state_create(page, 0);
2160         if (!s) {
2161                 ret = -ENOMEM;
2162                 goto unlock;
2163         }
2164
2165         if (!PageUptodate(page)) {
2166                 ret = bch2_read_single_page(page, mapping);
2167                 if (ret)
2168                         goto unlock;
2169         }
2170
2171         if (index != start >> PAGE_SHIFT)
2172                 start_offset = 0;
2173         if (index != end >> PAGE_SHIFT)
2174                 end_offset = PAGE_SIZE;
2175
2176         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2177              i < round_down(end_offset, block_bytes(c)) >> 9;
2178              i++) {
2179                 s->s[i].nr_replicas     = 0;
2180                 s->s[i].state           = SECTOR_UNALLOCATED;
2181         }
2182
2183         zero_user_segment(page, start_offset, end_offset);
2184
2185         /*
2186          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2187          *
2188          * XXX: because we aren't currently tracking whether the page has actual
2189          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2190          */
2191         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2192         BUG_ON(ret);
2193
2194         /*
2195          * This removes any writeable userspace mappings; we need to force
2196          * .page_mkwrite to be called again before any mmapped writes, to
2197          * redirty the full page:
2198          */
2199         page_mkclean(page);
2200         __set_page_dirty_nobuffers(page);
2201 unlock:
2202         unlock_page(page);
2203         put_page(page);
2204 out:
2205         return ret;
2206 }
2207
2208 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2209 {
2210         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2211                                     from, round_up(from, PAGE_SIZE));
2212 }
2213
2214 static int bch2_extend(struct bch_inode_info *inode,
2215                        struct bch_inode_unpacked *inode_u,
2216                        struct iattr *iattr)
2217 {
2218         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2219         struct address_space *mapping = inode->v.i_mapping;
2220         int ret;
2221
2222         /*
2223          * sync appends:
2224          *
2225          * this has to be done _before_ extending i_size:
2226          */
2227         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2228         if (ret)
2229                 return ret;
2230
2231         truncate_setsize(&inode->v, iattr->ia_size);
2232         setattr_copy(&inode->v, iattr);
2233
2234         mutex_lock(&inode->ei_update_lock);
2235         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2236                                     ATTR_MTIME|ATTR_CTIME);
2237         mutex_unlock(&inode->ei_update_lock);
2238
2239         return ret;
2240 }
2241
2242 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2243                                    struct bch_inode_unpacked *bi,
2244                                    void *p)
2245 {
2246         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2247
2248         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2249         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2250         return 0;
2251 }
2252
2253 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2254                                   struct bch_inode_unpacked *bi, void *p)
2255 {
2256         u64 *new_i_size = p;
2257
2258         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2259         bi->bi_size = *new_i_size;
2260         return 0;
2261 }
2262
2263 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2264 {
2265         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2266         struct address_space *mapping = inode->v.i_mapping;
2267         struct bch_inode_unpacked inode_u;
2268         struct btree_trans trans;
2269         struct btree_iter *iter;
2270         u64 new_i_size = iattr->ia_size;
2271         s64 i_sectors_delta = 0;
2272         int ret = 0;
2273
2274         inode_dio_wait(&inode->v);
2275         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2276
2277         /*
2278          * fetch current on disk i_size: inode is locked, i_size can only
2279          * increase underneath us:
2280          */
2281         bch2_trans_init(&trans, c, 0, 0);
2282         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2283         ret = PTR_ERR_OR_ZERO(iter);
2284         bch2_trans_exit(&trans);
2285
2286         if (ret)
2287                 goto err;
2288
2289         /*
2290          * check this before next assertion; on filesystem error our normal
2291          * invariants are a bit broken (truncate has to truncate the page cache
2292          * before the inode).
2293          */
2294         ret = bch2_journal_error(&c->journal);
2295         if (ret)
2296                 goto err;
2297
2298         BUG_ON(inode->v.i_size < inode_u.bi_size);
2299
2300         if (iattr->ia_size > inode->v.i_size) {
2301                 ret = bch2_extend(inode, &inode_u, iattr);
2302                 goto err;
2303         }
2304
2305         ret = bch2_truncate_page(inode, iattr->ia_size);
2306         if (unlikely(ret))
2307                 goto err;
2308
2309         /*
2310          * When extending, we're going to write the new i_size to disk
2311          * immediately so we need to flush anything above the current on disk
2312          * i_size first:
2313          *
2314          * Also, when extending we need to flush the page that i_size currently
2315          * straddles - if it's mapped to userspace, we need to ensure that
2316          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2317          * again to allocate the part of the page that was extended.
2318          */
2319         if (iattr->ia_size > inode_u.bi_size)
2320                 ret = filemap_write_and_wait_range(mapping,
2321                                 inode_u.bi_size,
2322                                 iattr->ia_size - 1);
2323         else if (iattr->ia_size & (PAGE_SIZE - 1))
2324                 ret = filemap_write_and_wait_range(mapping,
2325                                 round_down(iattr->ia_size, PAGE_SIZE),
2326                                 iattr->ia_size - 1);
2327         if (ret)
2328                 goto err;
2329
2330         mutex_lock(&inode->ei_update_lock);
2331         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2332                                &new_i_size, 0);
2333         mutex_unlock(&inode->ei_update_lock);
2334
2335         if (unlikely(ret))
2336                 goto err;
2337
2338         truncate_setsize(&inode->v, iattr->ia_size);
2339
2340         ret = bch2_fpunch(c, inode->v.i_ino,
2341                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2342                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2343         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2344
2345         if (unlikely(ret))
2346                 goto err;
2347
2348         setattr_copy(&inode->v, iattr);
2349
2350         mutex_lock(&inode->ei_update_lock);
2351         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2352                                ATTR_MTIME|ATTR_CTIME);
2353         mutex_unlock(&inode->ei_update_lock);
2354 err:
2355         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2356         return ret;
2357 }
2358
2359 /* fallocate: */
2360
2361 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2362 {
2363         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2364         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2365         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2366         int ret = 0;
2367
2368         inode_lock(&inode->v);
2369         inode_dio_wait(&inode->v);
2370         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2371
2372         ret = __bch2_truncate_page(inode,
2373                                    offset >> PAGE_SHIFT,
2374                                    offset, offset + len);
2375         if (unlikely(ret))
2376                 goto err;
2377
2378         if (offset >> PAGE_SHIFT !=
2379             (offset + len) >> PAGE_SHIFT) {
2380                 ret = __bch2_truncate_page(inode,
2381                                            (offset + len) >> PAGE_SHIFT,
2382                                            offset, offset + len);
2383                 if (unlikely(ret))
2384                         goto err;
2385         }
2386
2387         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2388
2389         if (discard_start < discard_end) {
2390                 s64 i_sectors_delta = 0;
2391
2392                 ret = bch2_fpunch(c, inode->v.i_ino,
2393                                   discard_start, discard_end,
2394                                   &inode->ei_journal_seq,
2395                                   &i_sectors_delta);
2396                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2397         }
2398 err:
2399         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2400         inode_unlock(&inode->v);
2401
2402         return ret;
2403 }
2404
2405 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2406                                    loff_t offset, loff_t len,
2407                                    bool insert)
2408 {
2409         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2410         struct address_space *mapping = inode->v.i_mapping;
2411         struct bkey_on_stack copy;
2412         struct btree_trans trans;
2413         struct btree_iter *src, *dst;
2414         loff_t shift, new_size;
2415         u64 src_start;
2416         int ret;
2417
2418         if ((offset | len) & (block_bytes(c) - 1))
2419                 return -EINVAL;
2420
2421         bkey_on_stack_init(&copy);
2422         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2423
2424         /*
2425          * We need i_mutex to keep the page cache consistent with the extents
2426          * btree, and the btree consistent with i_size - we don't need outside
2427          * locking for the extents btree itself, because we're using linked
2428          * iterators
2429          */
2430         inode_lock(&inode->v);
2431         inode_dio_wait(&inode->v);
2432         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2433
2434         if (insert) {
2435                 ret = -EFBIG;
2436                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2437                         goto err;
2438
2439                 ret = -EINVAL;
2440                 if (offset >= inode->v.i_size)
2441                         goto err;
2442
2443                 src_start       = U64_MAX;
2444                 shift           = len;
2445         } else {
2446                 ret = -EINVAL;
2447                 if (offset + len >= inode->v.i_size)
2448                         goto err;
2449
2450                 src_start       = offset + len;
2451                 shift           = -len;
2452         }
2453
2454         new_size = inode->v.i_size + shift;
2455
2456         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2457         if (ret)
2458                 goto err;
2459
2460         if (insert) {
2461                 i_size_write(&inode->v, new_size);
2462                 mutex_lock(&inode->ei_update_lock);
2463                 ret = bch2_write_inode_size(c, inode, new_size,
2464                                             ATTR_MTIME|ATTR_CTIME);
2465                 mutex_unlock(&inode->ei_update_lock);
2466         } else {
2467                 s64 i_sectors_delta = 0;
2468
2469                 ret = bch2_fpunch(c, inode->v.i_ino,
2470                                   offset >> 9, (offset + len) >> 9,
2471                                   &inode->ei_journal_seq,
2472                                   &i_sectors_delta);
2473                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2474
2475                 if (ret)
2476                         goto err;
2477         }
2478
2479         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2480                         POS(inode->v.i_ino, src_start >> 9),
2481                         BTREE_ITER_INTENT);
2482         BUG_ON(IS_ERR_OR_NULL(src));
2483
2484         dst = bch2_trans_copy_iter(&trans, src);
2485         BUG_ON(IS_ERR_OR_NULL(dst));
2486
2487         while (1) {
2488                 struct disk_reservation disk_res =
2489                         bch2_disk_reservation_init(c, 0);
2490                 struct bkey_i delete;
2491                 struct bkey_s_c k;
2492                 struct bpos next_pos;
2493                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2494                 struct bpos atomic_end;
2495                 unsigned trigger_flags = 0;
2496
2497                 k = insert
2498                         ? bch2_btree_iter_peek_prev(src)
2499                         : bch2_btree_iter_peek(src);
2500                 if ((ret = bkey_err(k)))
2501                         goto bkey_err;
2502
2503                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2504                         break;
2505
2506                 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2507
2508                 if (insert &&
2509                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2510                         break;
2511 reassemble:
2512                 bkey_on_stack_reassemble(&copy, c, k);
2513
2514                 if (insert &&
2515                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2516                         bch2_cut_front(move_pos, copy.k);
2517
2518                 copy.k->k.p.offset += shift >> 9;
2519                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2520
2521                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2522                 if (ret)
2523                         goto bkey_err;
2524
2525                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2526                         if (insert) {
2527                                 move_pos = atomic_end;
2528                                 move_pos.offset -= shift >> 9;
2529                                 goto reassemble;
2530                         } else {
2531                                 bch2_cut_back(atomic_end, copy.k);
2532                         }
2533                 }
2534
2535                 bkey_init(&delete.k);
2536                 delete.k.p = copy.k->k.p;
2537                 delete.k.size = copy.k->k.size;
2538                 delete.k.p.offset -= shift >> 9;
2539
2540                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2541
2542                 if (copy.k->k.size == k.k->size) {
2543                         /*
2544                          * If we're moving the entire extent, we can skip
2545                          * running triggers:
2546                          */
2547                         trigger_flags |= BTREE_TRIGGER_NORUN;
2548                 } else {
2549                         /* We might end up splitting compressed extents: */
2550                         unsigned nr_ptrs =
2551                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2552
2553                         ret = bch2_disk_reservation_get(c, &disk_res,
2554                                         copy.k->k.size, nr_ptrs,
2555                                         BCH_DISK_RESERVATION_NOFAIL);
2556                         BUG_ON(ret);
2557                 }
2558
2559                 bch2_btree_iter_set_pos(src, bkey_start_pos(&delete.k));
2560
2561                 ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
2562                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2563                         bch2_trans_commit(&trans, &disk_res,
2564                                           &inode->ei_journal_seq,
2565                                           BTREE_INSERT_NOFAIL);
2566                 bch2_disk_reservation_put(c, &disk_res);
2567 bkey_err:
2568                 if (!ret)
2569                         bch2_btree_iter_set_pos(src, next_pos);
2570
2571                 if (ret == -EINTR)
2572                         ret = 0;
2573                 if (ret)
2574                         goto err;
2575
2576                 bch2_trans_cond_resched(&trans);
2577         }
2578         bch2_trans_unlock(&trans);
2579
2580         if (!insert) {
2581                 i_size_write(&inode->v, new_size);
2582                 mutex_lock(&inode->ei_update_lock);
2583                 ret = bch2_write_inode_size(c, inode, new_size,
2584                                             ATTR_MTIME|ATTR_CTIME);
2585                 mutex_unlock(&inode->ei_update_lock);
2586         }
2587 err:
2588         bch2_trans_exit(&trans);
2589         bkey_on_stack_exit(&copy, c);
2590         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2591         inode_unlock(&inode->v);
2592         return ret;
2593 }
2594
2595 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2596                             loff_t offset, loff_t len)
2597 {
2598         struct address_space *mapping = inode->v.i_mapping;
2599         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2600         struct btree_trans trans;
2601         struct btree_iter *iter;
2602         struct bpos end_pos;
2603         loff_t end              = offset + len;
2604         loff_t block_start      = round_down(offset,    block_bytes(c));
2605         loff_t block_end        = round_up(end,         block_bytes(c));
2606         unsigned sectors;
2607         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2608         int ret;
2609
2610         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2611
2612         inode_lock(&inode->v);
2613         inode_dio_wait(&inode->v);
2614         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2615
2616         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2617                 ret = inode_newsize_ok(&inode->v, end);
2618                 if (ret)
2619                         goto err;
2620         }
2621
2622         if (mode & FALLOC_FL_ZERO_RANGE) {
2623                 ret = __bch2_truncate_page(inode,
2624                                            offset >> PAGE_SHIFT,
2625                                            offset, end);
2626
2627                 if (!ret &&
2628                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2629                         ret = __bch2_truncate_page(inode,
2630                                                    end >> PAGE_SHIFT,
2631                                                    offset, end);
2632
2633                 if (unlikely(ret))
2634                         goto err;
2635
2636                 truncate_pagecache_range(&inode->v, offset, end - 1);
2637         }
2638
2639         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2640                         POS(inode->v.i_ino, block_start >> 9),
2641                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2642         end_pos = POS(inode->v.i_ino, block_end >> 9);
2643
2644         while (bkey_cmp(iter->pos, end_pos) < 0) {
2645                 s64 i_sectors_delta = 0;
2646                 struct disk_reservation disk_res = { 0 };
2647                 struct quota_res quota_res = { 0 };
2648                 struct bkey_i_reservation reservation;
2649                 struct bkey_s_c k;
2650
2651                 bch2_trans_begin(&trans);
2652
2653                 k = bch2_btree_iter_peek_slot(iter);
2654                 if ((ret = bkey_err(k)))
2655                         goto bkey_err;
2656
2657                 /* already reserved */
2658                 if (k.k->type == KEY_TYPE_reservation &&
2659                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2660                         bch2_btree_iter_next_slot(iter);
2661                         continue;
2662                 }
2663
2664                 if (bkey_extent_is_data(k.k) &&
2665                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2666                         bch2_btree_iter_next_slot(iter);
2667                         continue;
2668                 }
2669
2670                 bkey_reservation_init(&reservation.k_i);
2671                 reservation.k.type      = KEY_TYPE_reservation;
2672                 reservation.k.p         = k.k->p;
2673                 reservation.k.size      = k.k->size;
2674
2675                 bch2_cut_front(iter->pos,       &reservation.k_i);
2676                 bch2_cut_back(end_pos,          &reservation.k_i);
2677
2678                 sectors = reservation.k.size;
2679                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2680
2681                 if (!bkey_extent_is_allocation(k.k)) {
2682                         ret = bch2_quota_reservation_add(c, inode,
2683                                         &quota_res,
2684                                         sectors, true);
2685                         if (unlikely(ret))
2686                                 goto bkey_err;
2687                 }
2688
2689                 if (reservation.v.nr_replicas < replicas ||
2690                     bch2_bkey_sectors_compressed(k)) {
2691                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2692                                                         replicas, 0);
2693                         if (unlikely(ret))
2694                                 goto bkey_err;
2695
2696                         reservation.v.nr_replicas = disk_res.nr_replicas;
2697                 }
2698
2699                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2700                                 &disk_res, &inode->ei_journal_seq,
2701                                 0, &i_sectors_delta);
2702                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2703 bkey_err:
2704                 bch2_quota_reservation_put(c, inode, &quota_res);
2705                 bch2_disk_reservation_put(c, &disk_res);
2706                 if (ret == -EINTR)
2707                         ret = 0;
2708                 if (ret)
2709                         goto err;
2710         }
2711
2712         /*
2713          * Do we need to extend the file?
2714          *
2715          * If we zeroed up to the end of the file, we dropped whatever writes
2716          * were going to write out the current i_size, so we have to extend
2717          * manually even if FL_KEEP_SIZE was set:
2718          */
2719         if (end >= inode->v.i_size &&
2720             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2721              (mode & FALLOC_FL_ZERO_RANGE))) {
2722                 struct btree_iter *inode_iter;
2723                 struct bch_inode_unpacked inode_u;
2724
2725                 do {
2726                         bch2_trans_begin(&trans);
2727                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2728                                                      inode->v.i_ino, 0);
2729                         ret = PTR_ERR_OR_ZERO(inode_iter);
2730                 } while (ret == -EINTR);
2731
2732                 bch2_trans_unlock(&trans);
2733
2734                 if (ret)
2735                         goto err;
2736
2737                 /*
2738                  * Sync existing appends before extending i_size,
2739                  * as in bch2_extend():
2740                  */
2741                 ret = filemap_write_and_wait_range(mapping,
2742                                         inode_u.bi_size, S64_MAX);
2743                 if (ret)
2744                         goto err;
2745
2746                 if (mode & FALLOC_FL_KEEP_SIZE)
2747                         end = inode->v.i_size;
2748                 else
2749                         i_size_write(&inode->v, end);
2750
2751                 mutex_lock(&inode->ei_update_lock);
2752                 ret = bch2_write_inode_size(c, inode, end, 0);
2753                 mutex_unlock(&inode->ei_update_lock);
2754         }
2755 err:
2756         bch2_trans_exit(&trans);
2757         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2758         inode_unlock(&inode->v);
2759         return ret;
2760 }
2761
2762 long bch2_fallocate_dispatch(struct file *file, int mode,
2763                              loff_t offset, loff_t len)
2764 {
2765         struct bch_inode_info *inode = file_bch_inode(file);
2766         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2767         long ret;
2768
2769         if (!percpu_ref_tryget(&c->writes))
2770                 return -EROFS;
2771
2772         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2773                 ret = bchfs_fallocate(inode, mode, offset, len);
2774         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2775                 ret = bchfs_fpunch(inode, offset, len);
2776         else if (mode == FALLOC_FL_INSERT_RANGE)
2777                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2778         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2779                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2780         else
2781                 ret = -EOPNOTSUPP;
2782
2783         percpu_ref_put(&c->writes);
2784
2785         return ret;
2786 }
2787
2788 static void mark_range_unallocated(struct bch_inode_info *inode,
2789                                    loff_t start, loff_t end)
2790 {
2791         pgoff_t index = start >> PAGE_SHIFT;
2792         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2793         struct pagevec pvec;
2794
2795         pagevec_init(&pvec);
2796
2797         do {
2798                 unsigned nr_pages, i, j;
2799
2800                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2801                                                 &index, end_index);
2802                 if (nr_pages == 0)
2803                         break;
2804
2805                 for (i = 0; i < nr_pages; i++) {
2806                         struct page *page = pvec.pages[i];
2807                         struct bch_page_state *s;
2808
2809                         lock_page(page);
2810                         s = bch2_page_state(page);
2811
2812                         if (s) {
2813                                 spin_lock(&s->lock);
2814                                 for (j = 0; j < PAGE_SECTORS; j++)
2815                                         s->s[j].nr_replicas = 0;
2816                                 spin_unlock(&s->lock);
2817                         }
2818
2819                         unlock_page(page);
2820                 }
2821                 pagevec_release(&pvec);
2822         } while (index <= end_index);
2823 }
2824
2825 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2826                              struct file *file_dst, loff_t pos_dst,
2827                              loff_t len, unsigned remap_flags)
2828 {
2829         struct bch_inode_info *src = file_bch_inode(file_src);
2830         struct bch_inode_info *dst = file_bch_inode(file_dst);
2831         struct bch_fs *c = src->v.i_sb->s_fs_info;
2832         s64 i_sectors_delta = 0;
2833         u64 aligned_len;
2834         loff_t ret = 0;
2835
2836         if (!c->opts.reflink)
2837                 return -EOPNOTSUPP;
2838
2839         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2840                 return -EINVAL;
2841
2842         if (remap_flags & REMAP_FILE_DEDUP)
2843                 return -EOPNOTSUPP;
2844
2845         if ((pos_src & (block_bytes(c) - 1)) ||
2846             (pos_dst & (block_bytes(c) - 1)))
2847                 return -EINVAL;
2848
2849         if (src == dst &&
2850             abs(pos_src - pos_dst) < len)
2851                 return -EINVAL;
2852
2853         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2854
2855         file_update_time(file_dst);
2856
2857         inode_dio_wait(&src->v);
2858         inode_dio_wait(&dst->v);
2859
2860         ret = generic_remap_file_range_prep(file_src, pos_src,
2861                                             file_dst, pos_dst,
2862                                             &len, remap_flags);
2863         if (ret < 0 || len == 0)
2864                 goto err;
2865
2866         aligned_len = round_up((u64) len, block_bytes(c));
2867
2868         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2869                                 pos_dst, pos_dst + len - 1);
2870         if (ret)
2871                 goto err;
2872
2873         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2874
2875         ret = bch2_remap_range(c,
2876                                POS(dst->v.i_ino, pos_dst >> 9),
2877                                POS(src->v.i_ino, pos_src >> 9),
2878                                aligned_len >> 9,
2879                                &dst->ei_journal_seq,
2880                                pos_dst + len, &i_sectors_delta);
2881         if (ret < 0)
2882                 goto err;
2883
2884         /*
2885          * due to alignment, we might have remapped slightly more than requsted
2886          */
2887         ret = min((u64) ret << 9, (u64) len);
2888
2889         /* XXX get a quota reservation */
2890         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2891
2892         spin_lock(&dst->v.i_lock);
2893         if (pos_dst + ret > dst->v.i_size)
2894                 i_size_write(&dst->v, pos_dst + ret);
2895         spin_unlock(&dst->v.i_lock);
2896 err:
2897         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2898
2899         return ret;
2900 }
2901
2902 /* fseek: */
2903
2904 static int page_data_offset(struct page *page, unsigned offset)
2905 {
2906         struct bch_page_state *s = bch2_page_state(page);
2907         unsigned i;
2908
2909         if (s)
2910                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2911                         if (s->s[i].state >= SECTOR_DIRTY)
2912                                 return i << 9;
2913
2914         return -1;
2915 }
2916
2917 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2918                                        loff_t start_offset,
2919                                        loff_t end_offset)
2920 {
2921         struct address_space *mapping = vinode->i_mapping;
2922         struct page *page;
2923         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2924         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2925         pgoff_t index           = start_index;
2926         loff_t ret;
2927         int offset;
2928
2929         while (index <= end_index) {
2930                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2931                         lock_page(page);
2932
2933                         offset = page_data_offset(page,
2934                                         page->index == start_index
2935                                         ? start_offset & (PAGE_SIZE - 1)
2936                                         : 0);
2937                         if (offset >= 0) {
2938                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2939                                             offset,
2940                                             start_offset, end_offset);
2941                                 unlock_page(page);
2942                                 put_page(page);
2943                                 return ret;
2944                         }
2945
2946                         unlock_page(page);
2947                         put_page(page);
2948                 } else {
2949                         break;
2950                 }
2951         }
2952
2953         return end_offset;
2954 }
2955
2956 static loff_t bch2_seek_data(struct file *file, u64 offset)
2957 {
2958         struct bch_inode_info *inode = file_bch_inode(file);
2959         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2960         struct btree_trans trans;
2961         struct btree_iter *iter;
2962         struct bkey_s_c k;
2963         u64 isize, next_data = MAX_LFS_FILESIZE;
2964         int ret;
2965
2966         isize = i_size_read(&inode->v);
2967         if (offset >= isize)
2968                 return -ENXIO;
2969
2970         bch2_trans_init(&trans, c, 0, 0);
2971
2972         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2973                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2974                 if (k.k->p.inode != inode->v.i_ino) {
2975                         break;
2976                 } else if (bkey_extent_is_data(k.k)) {
2977                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2978                         break;
2979                 } else if (k.k->p.offset >> 9 > isize)
2980                         break;
2981         }
2982
2983         ret = bch2_trans_exit(&trans) ?: ret;
2984         if (ret)
2985                 return ret;
2986
2987         if (next_data > offset)
2988                 next_data = bch2_seek_pagecache_data(&inode->v,
2989                                                      offset, next_data);
2990
2991         if (next_data >= isize)
2992                 return -ENXIO;
2993
2994         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2995 }
2996
2997 static int __page_hole_offset(struct page *page, unsigned offset)
2998 {
2999         struct bch_page_state *s = bch2_page_state(page);
3000         unsigned i;
3001
3002         if (!s)
3003                 return 0;
3004
3005         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3006                 if (s->s[i].state < SECTOR_DIRTY)
3007                         return i << 9;
3008
3009         return -1;
3010 }
3011
3012 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3013 {
3014         pgoff_t index = offset >> PAGE_SHIFT;
3015         struct page *page;
3016         int pg_offset;
3017         loff_t ret = -1;
3018
3019         page = find_lock_entry(mapping, index);
3020         if (!page || xa_is_value(page))
3021                 return offset;
3022
3023         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3024         if (pg_offset >= 0)
3025                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3026
3027         unlock_page(page);
3028
3029         return ret;
3030 }
3031
3032 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3033                                        loff_t start_offset,
3034                                        loff_t end_offset)
3035 {
3036         struct address_space *mapping = vinode->i_mapping;
3037         loff_t offset = start_offset, hole;
3038
3039         while (offset < end_offset) {
3040                 hole = page_hole_offset(mapping, offset);
3041                 if (hole >= 0 && hole <= end_offset)
3042                         return max(start_offset, hole);
3043
3044                 offset += PAGE_SIZE;
3045                 offset &= PAGE_MASK;
3046         }
3047
3048         return end_offset;
3049 }
3050
3051 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3052 {
3053         struct bch_inode_info *inode = file_bch_inode(file);
3054         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3055         struct btree_trans trans;
3056         struct btree_iter *iter;
3057         struct bkey_s_c k;
3058         u64 isize, next_hole = MAX_LFS_FILESIZE;
3059         int ret;
3060
3061         isize = i_size_read(&inode->v);
3062         if (offset >= isize)
3063                 return -ENXIO;
3064
3065         bch2_trans_init(&trans, c, 0, 0);
3066
3067         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3068                            POS(inode->v.i_ino, offset >> 9),
3069                            BTREE_ITER_SLOTS, k, ret) {
3070                 if (k.k->p.inode != inode->v.i_ino) {
3071                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3072                                         offset, MAX_LFS_FILESIZE);
3073                         break;
3074                 } else if (!bkey_extent_is_data(k.k)) {
3075                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3076                                         max(offset, bkey_start_offset(k.k) << 9),
3077                                         k.k->p.offset << 9);
3078
3079                         if (next_hole < k.k->p.offset << 9)
3080                                 break;
3081                 } else {
3082                         offset = max(offset, bkey_start_offset(k.k) << 9);
3083                 }
3084         }
3085
3086         ret = bch2_trans_exit(&trans) ?: ret;
3087         if (ret)
3088                 return ret;
3089
3090         if (next_hole > isize)
3091                 next_hole = isize;
3092
3093         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3094 }
3095
3096 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3097 {
3098         switch (whence) {
3099         case SEEK_SET:
3100         case SEEK_CUR:
3101         case SEEK_END:
3102                 return generic_file_llseek(file, offset, whence);
3103         case SEEK_DATA:
3104                 return bch2_seek_data(file, offset);
3105         case SEEK_HOLE:
3106                 return bch2_seek_hole(file, offset);
3107         }
3108
3109         return -EINVAL;
3110 }
3111
3112 void bch2_fs_fsio_exit(struct bch_fs *c)
3113 {
3114         bioset_exit(&c->dio_write_bioset);
3115         bioset_exit(&c->dio_read_bioset);
3116         bioset_exit(&c->writepage_bioset);
3117 }
3118
3119 int bch2_fs_fsio_init(struct bch_fs *c)
3120 {
3121         int ret = 0;
3122
3123         pr_verbose_init(c->opts, "");
3124
3125         if (bioset_init(&c->writepage_bioset,
3126                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3127                         BIOSET_NEED_BVECS) ||
3128             bioset_init(&c->dio_read_bioset,
3129                         4, offsetof(struct dio_read, rbio.bio),
3130                         BIOSET_NEED_BVECS) ||
3131             bioset_init(&c->dio_write_bioset,
3132                         4, offsetof(struct dio_write, op.wbio.bio),
3133                         BIOSET_NEED_BVECS))
3134                 ret = -ENOMEM;
3135
3136         pr_verbose_init(c->opts, "ret %i", ret);
3137         return ret;
3138 }
3139
3140 #endif /* NO_BCACHEFS_FS */