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                 offset_into_extent = iter->pos.offset -
 843                         bkey_start_offset(k.k);
 844                 sectors = k.k->size - offset_into_extent;
 845
 846                 bkey_on_stack_reassemble(&sk, c, k);
 847
 848                 ret = bch2_read_indirect_extent(trans,
 849                                         &offset_into_extent, &sk);
 850                 if (ret)
 851                         break;
 852
 853                 k = bkey_i_to_s_c(sk.k);
 854
 855                 sectors = min(sectors, k.k->size - offset_into_extent);
 856
 857                 bch2_trans_unlock(trans);
 858
 859                 if (readpages_iter)
 860                         readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
 861                                             extent_partial_reads_expensive(k));
 862
 863                 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
 864                 swap(rbio->bio.bi_iter.bi_size, bytes);
 865
 866                 if (rbio->bio.bi_iter.bi_size == bytes)
 867                         flags |= BCH_READ_LAST_FRAGMENT;
 868
 869                 if (bkey_extent_is_allocation(k.k))
 870                         bch2_add_page_sectors(&rbio->bio, k);
 871
 872                 bch2_read_extent(trans, rbio, k, offset_into_extent, flags);
 873
 874                 if (flags & BCH_READ_LAST_FRAGMENT)
 875                         break;
 876
 877                 swap(rbio->bio.bi_iter.bi_size, bytes);
 878                 bio_advance(&rbio->bio, bytes);
 879         }
 880
 881         if (ret == -EINTR)
 882                 goto retry;
 883
 884         if (ret) {
 885                 bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
 886                 bio_endio(&rbio->bio);
 887         }
 888
 889         bkey_on_stack_exit(&sk, c);
 890 }
 891
 892 int bch2_readpages(struct file *file, struct address_space *mapping,
 893                    struct list_head *pages, unsigned nr_pages)
 894 {
 895         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 896         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 897         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 898         struct btree_trans trans;
 899         struct btree_iter *iter;
 900         struct page *page;
 901         struct readpages_iter readpages_iter;
 902         int ret;
 903
 904         ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
 905         BUG_ON(ret);
 906
 907         bch2_trans_init(&trans, c, 0, 0);
 908
 909         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 910                                    BTREE_ITER_SLOTS);
 911
 912         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
 913
 914         while ((page = readpage_iter_next(&readpages_iter))) {
 915                 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
 916                 unsigned n = min_t(unsigned,
 917                                    readpages_iter.nr_pages -
 918                                    readpages_iter.idx,
 919                                    BIO_MAX_PAGES);
 920                 struct bch_read_bio *rbio =
 921                         rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
 922                                   opts);
 923
 924                 readpages_iter.idx++;
 925
 926                 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
 927                 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
 928                 rbio->bio.bi_end_io = bch2_readpages_end_io;
 929                 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 930
 931                 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
 932                            &readpages_iter);
 933         }
 934
 935         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
 936
 937         bch2_trans_exit(&trans);
 938         kfree(readpages_iter.pages);
 939
 940         return 0;
 941 }
 942
 943 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
 944                              u64 inum, struct page *page)
 945 {
 946         struct btree_trans trans;
 947         struct btree_iter *iter;
 948
 949         bch2_page_state_create(page, __GFP_NOFAIL);
 950
 951         bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
 952         rbio->bio.bi_iter.bi_sector =
 953                 (sector_t) page->index << PAGE_SECTOR_SHIFT;
 954         BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
 955
 956         bch2_trans_init(&trans, c, 0, 0);
 957         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
 958                                    BTREE_ITER_SLOTS);
 959
 960         bchfs_read(&trans, iter, rbio, inum, NULL);
 961
 962         bch2_trans_exit(&trans);
 963 }
 964
 965 int bch2_readpage(struct file *file, struct page *page)
 966 {
 967         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
 968         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 969         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
 970         struct bch_read_bio *rbio;
 971
 972         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
 973         rbio->bio.bi_end_io = bch2_readpages_end_io;
 974
 975         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 976         return 0;
 977 }
 978
 979 static void bch2_read_single_page_end_io(struct bio *bio)
 980 {
 981         complete(bio->bi_private);
 982 }
 983
 984 static int bch2_read_single_page(struct page *page,
 985                                  struct address_space *mapping)
 986 {
 987         struct bch_inode_info *inode = to_bch_ei(mapping->host);
 988         struct bch_fs *c = inode->v.i_sb->s_fs_info;
 989         struct bch_read_bio *rbio;
 990         int ret;
 991         DECLARE_COMPLETION_ONSTACK(done);
 992
 993         rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
 994                          io_opts(c, &inode->ei_inode));
 995         rbio->bio.bi_private = &done;
 996         rbio->bio.bi_end_io = bch2_read_single_page_end_io;
 997
 998         __bchfs_readpage(c, rbio, inode->v.i_ino, page);
 999         wait_for_completion(&done);
1000
1001         ret = blk_status_to_errno(rbio->bio.bi_status);
1002         bio_put(&rbio->bio);
1003
1004         if (ret < 0)
1005                 return ret;
1006
1007         SetPageUptodate(page);
1008         return 0;
1009 }
1010
1011 /* writepages: */
1012
1013 struct bch_writepage_state {
1014         struct bch_writepage_io *io;
1015         struct bch_io_opts      opts;
1016 };
1017
1018 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1019                                                                   struct bch_inode_info *inode)
1020 {
1021         return (struct bch_writepage_state) {
1022                 .opts = io_opts(c, &inode->ei_inode)
1023         };
1024 }
1025
1026 static void bch2_writepage_io_free(struct closure *cl)
1027 {
1028         struct bch_writepage_io *io = container_of(cl,
1029                                         struct bch_writepage_io, cl);
1030
1031         bio_put(&io->op.wbio.bio);
1032 }
1033
1034 static void bch2_writepage_io_done(struct closure *cl)
1035 {
1036         struct bch_writepage_io *io = container_of(cl,
1037                                         struct bch_writepage_io, cl);
1038         struct bch_fs *c = io->op.c;
1039         struct bio *bio = &io->op.wbio.bio;
1040         struct bvec_iter_all iter;
1041         struct bio_vec *bvec;
1042         unsigned i;
1043
1044         if (io->op.error) {
1045                 bio_for_each_segment_all(bvec, bio, iter) {
1046                         struct bch_page_state *s;
1047
1048                         SetPageError(bvec->bv_page);
1049                         mapping_set_error(bvec->bv_page->mapping, -EIO);
1050
1051                         s = __bch2_page_state(bvec->bv_page);
1052                         spin_lock(&s->lock);
1053                         for (i = 0; i < PAGE_SECTORS; i++)
1054                                 s->s[i].nr_replicas = 0;
1055                         spin_unlock(&s->lock);
1056                 }
1057         }
1058
1059         if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1060                 bio_for_each_segment_all(bvec, bio, iter) {
1061                         struct bch_page_state *s;
1062
1063                         s = __bch2_page_state(bvec->bv_page);
1064                         spin_lock(&s->lock);
1065                         for (i = 0; i < PAGE_SECTORS; i++)
1066                                 s->s[i].nr_replicas = 0;
1067                         spin_unlock(&s->lock);
1068                 }
1069         }
1070
1071         /*
1072          * racing with fallocate can cause us to add fewer sectors than
1073          * expected - but we shouldn't add more sectors than expected:
1074          */
1075         BUG_ON(io->op.i_sectors_delta > 0);
1076
1077         /*
1078          * (error (due to going RO) halfway through a page can screw that up
1079          * slightly)
1080          * XXX wtf?
1081            BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1082          */
1083
1084         /*
1085          * PageWriteback is effectively our ref on the inode - fixup i_blocks
1086          * before calling end_page_writeback:
1087          */
1088         i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1089
1090         bio_for_each_segment_all(bvec, bio, iter) {
1091                 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1092
1093                 if (atomic_dec_and_test(&s->write_count))
1094                         end_page_writeback(bvec->bv_page);
1095         }
1096
1097         closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1098 }
1099
1100 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1101 {
1102         struct bch_writepage_io *io = w->io;
1103
1104         w->io = NULL;
1105         closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1106         continue_at(&io->cl, bch2_writepage_io_done, NULL);
1107 }
1108
1109 /*
1110  * Get a bch_writepage_io and add @page to it - appending to an existing one if
1111  * possible, else allocating a new one:
1112  */
1113 static void bch2_writepage_io_alloc(struct bch_fs *c,
1114                                     struct writeback_control *wbc,
1115                                     struct bch_writepage_state *w,
1116                                     struct bch_inode_info *inode,
1117                                     u64 sector,
1118                                     unsigned nr_replicas)
1119 {
1120         struct bch_write_op *op;
1121
1122         w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1123                                               BIO_MAX_PAGES,
1124                                               &c->writepage_bioset),
1125                              struct bch_writepage_io, op.wbio.bio);
1126
1127         closure_init(&w->io->cl, NULL);
1128         w->io->inode            = inode;
1129
1130         op                      = &w->io->op;
1131         bch2_write_op_init(op, c, w->opts);
1132         op->target              = w->opts.foreground_target;
1133         op_journal_seq_set(op, &inode->ei_journal_seq);
1134         op->nr_replicas         = nr_replicas;
1135         op->res.nr_replicas     = nr_replicas;
1136         op->write_point         = writepoint_hashed(inode->ei_last_dirtied);
1137         op->pos                 = POS(inode->v.i_ino, sector);
1138         op->wbio.bio.bi_iter.bi_sector = sector;
1139         op->wbio.bio.bi_opf     = wbc_to_write_flags(wbc);
1140 }
1141
1142 static int __bch2_writepage(struct page *page,
1143                             struct writeback_control *wbc,
1144                             void *data)
1145 {
1146         struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1147         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1148         struct bch_writepage_state *w = data;
1149         struct bch_page_state *s, orig;
1150         unsigned i, offset, nr_replicas_this_write = U32_MAX;
1151         loff_t i_size = i_size_read(&inode->v);
1152         pgoff_t end_index = i_size >> PAGE_SHIFT;
1153         int ret;
1154
1155         EBUG_ON(!PageUptodate(page));
1156
1157         /* Is the page fully inside i_size? */
1158         if (page->index < end_index)
1159                 goto do_io;
1160
1161         /* Is the page fully outside i_size? (truncate in progress) */
1162         offset = i_size & (PAGE_SIZE - 1);
1163         if (page->index > end_index || !offset) {
1164                 unlock_page(page);
1165                 return 0;
1166         }
1167
1168         /*
1169          * The page straddles i_size.  It must be zeroed out on each and every
1170          * writepage invocation because it may be mmapped.  "A file is mapped
1171          * in multiples of the page size.  For a file that is not a multiple of
1172          * the  page size, the remaining memory is zeroed when mapped, and
1173          * writes to that region are not written out to the file."
1174          */
1175         zero_user_segment(page, offset, PAGE_SIZE);
1176 do_io:
1177         s = bch2_page_state_create(page, __GFP_NOFAIL);
1178
1179         ret = bch2_get_page_disk_reservation(c, inode, page, true);
1180         if (ret) {
1181                 SetPageError(page);
1182                 mapping_set_error(page->mapping, ret);
1183                 unlock_page(page);
1184                 return 0;
1185         }
1186
1187         /* Before unlocking the page, get copy of reservations: */
1188         orig = *s;
1189
1190         for (i = 0; i < PAGE_SECTORS; i++) {
1191                 if (s->s[i].state < SECTOR_DIRTY)
1192                         continue;
1193
1194                 nr_replicas_this_write =
1195                         min_t(unsigned, nr_replicas_this_write,
1196                               s->s[i].nr_replicas +
1197                               s->s[i].replicas_reserved);
1198         }
1199
1200         for (i = 0; i < PAGE_SECTORS; i++) {
1201                 if (s->s[i].state < SECTOR_DIRTY)
1202                         continue;
1203
1204                 s->s[i].nr_replicas = w->opts.compression
1205                         ? 0 : nr_replicas_this_write;
1206
1207                 s->s[i].replicas_reserved = 0;
1208                 s->s[i].state = SECTOR_ALLOCATED;
1209         }
1210
1211         BUG_ON(atomic_read(&s->write_count));
1212         atomic_set(&s->write_count, 1);
1213
1214         BUG_ON(PageWriteback(page));
1215         set_page_writeback(page);
1216
1217         unlock_page(page);
1218
1219         offset = 0;
1220         while (1) {
1221                 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1222                 u64 sector;
1223
1224                 while (offset < PAGE_SECTORS &&
1225                        orig.s[offset].state < SECTOR_DIRTY)
1226                         offset++;
1227
1228                 if (offset == PAGE_SECTORS)
1229                         break;
1230
1231                 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1232
1233                 while (offset + sectors < PAGE_SECTORS &&
1234                        orig.s[offset + sectors].state >= SECTOR_DIRTY)
1235                         sectors++;
1236
1237                 for (i = offset; i < offset + sectors; i++) {
1238                         reserved_sectors += orig.s[i].replicas_reserved;
1239                         dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1240                 }
1241
1242                 if (w->io &&
1243                     (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1244                      bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1245                      w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1246                      (BIO_MAX_PAGES * PAGE_SIZE) ||
1247                      bio_end_sector(&w->io->op.wbio.bio) != sector))
1248                         bch2_writepage_do_io(w);
1249
1250                 if (!w->io)
1251                         bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1252                                                 nr_replicas_this_write);
1253
1254                 atomic_inc(&s->write_count);
1255
1256                 BUG_ON(inode != w->io->inode);
1257                 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1258                                      sectors << 9, offset << 9));
1259
1260                 /* Check for writing past i_size: */
1261                 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1262                         round_up(i_size, block_bytes(c)));
1263
1264                 w->io->op.res.sectors += reserved_sectors;
1265                 w->io->op.i_sectors_delta -= dirty_sectors;
1266                 w->io->op.new_i_size = i_size;
1267
1268                 offset += sectors;
1269         }
1270
1271         if (atomic_dec_and_test(&s->write_count))
1272                 end_page_writeback(page);
1273
1274         return 0;
1275 }
1276
1277 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1278 {
1279         struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1280         struct bch_writepage_state w =
1281                 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1282         struct blk_plug plug;
1283         int ret;
1284
1285         blk_start_plug(&plug);
1286         ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1287         if (w.io)
1288                 bch2_writepage_do_io(&w);
1289         blk_finish_plug(&plug);
1290         return ret;
1291 }
1292
1293 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1294 {
1295         struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1296         struct bch_writepage_state w =
1297                 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1298         int ret;
1299
1300         ret = __bch2_writepage(page, wbc, &w);
1301         if (w.io)
1302                 bch2_writepage_do_io(&w);
1303
1304         return ret;
1305 }
1306
1307 /* buffered writes: */
1308
1309 int bch2_write_begin(struct file *file, struct address_space *mapping,
1310                      loff_t pos, unsigned len, unsigned flags,
1311                      struct page **pagep, void **fsdata)
1312 {
1313         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1314         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1315         struct bch2_page_reservation *res;
1316         pgoff_t index = pos >> PAGE_SHIFT;
1317         unsigned offset = pos & (PAGE_SIZE - 1);
1318         struct page *page;
1319         int ret = -ENOMEM;
1320
1321         res = kmalloc(sizeof(*res), GFP_KERNEL);
1322         if (!res)
1323                 return -ENOMEM;
1324
1325         bch2_page_reservation_init(c, inode, res);
1326         *fsdata = res;
1327
1328         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1329
1330         page = grab_cache_page_write_begin(mapping, index, flags);
1331         if (!page)
1332                 goto err_unlock;
1333
1334         if (PageUptodate(page))
1335                 goto out;
1336
1337         /* If we're writing entire page, don't need to read it in first: */
1338         if (len == PAGE_SIZE)
1339                 goto out;
1340
1341         if (!offset && pos + len >= inode->v.i_size) {
1342                 zero_user_segment(page, len, PAGE_SIZE);
1343                 flush_dcache_page(page);
1344                 goto out;
1345         }
1346
1347         if (index > inode->v.i_size >> PAGE_SHIFT) {
1348                 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1349                 flush_dcache_page(page);
1350                 goto out;
1351         }
1352 readpage:
1353         ret = bch2_read_single_page(page, mapping);
1354         if (ret)
1355                 goto err;
1356 out:
1357         ret = bch2_page_reservation_get(c, inode, page, res,
1358                                         offset, len, true);
1359         if (ret) {
1360                 if (!PageUptodate(page)) {
1361                         /*
1362                          * If the page hasn't been read in, we won't know if we
1363                          * actually need a reservation - we don't actually need
1364                          * to read here, we just need to check if the page is
1365                          * fully backed by uncompressed data:
1366                          */
1367                         goto readpage;
1368                 }
1369
1370                 goto err;
1371         }
1372
1373         *pagep = page;
1374         return 0;
1375 err:
1376         unlock_page(page);
1377         put_page(page);
1378         *pagep = NULL;
1379 err_unlock:
1380         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1381         kfree(res);
1382         *fsdata = NULL;
1383         return ret;
1384 }
1385
1386 int bch2_write_end(struct file *file, struct address_space *mapping,
1387                    loff_t pos, unsigned len, unsigned copied,
1388                    struct page *page, void *fsdata)
1389 {
1390         struct bch_inode_info *inode = to_bch_ei(mapping->host);
1391         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1392         struct bch2_page_reservation *res = fsdata;
1393         unsigned offset = pos & (PAGE_SIZE - 1);
1394
1395         lockdep_assert_held(&inode->v.i_rwsem);
1396
1397         if (unlikely(copied < len && !PageUptodate(page))) {
1398                 /*
1399                  * The page needs to be read in, but that would destroy
1400                  * our partial write - simplest thing is to just force
1401                  * userspace to redo the write:
1402                  */
1403                 zero_user(page, 0, PAGE_SIZE);
1404                 flush_dcache_page(page);
1405                 copied = 0;
1406         }
1407
1408         spin_lock(&inode->v.i_lock);
1409         if (pos + copied > inode->v.i_size)
1410                 i_size_write(&inode->v, pos + copied);
1411         spin_unlock(&inode->v.i_lock);
1412
1413         if (copied) {
1414                 if (!PageUptodate(page))
1415                         SetPageUptodate(page);
1416
1417                 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1418
1419                 inode->ei_last_dirtied = (unsigned long) current;
1420         }
1421
1422         unlock_page(page);
1423         put_page(page);
1424         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1425
1426         bch2_page_reservation_put(c, inode, res);
1427         kfree(res);
1428
1429         return copied;
1430 }
1431
1432 #define WRITE_BATCH_PAGES       32
1433
1434 static int __bch2_buffered_write(struct bch_inode_info *inode,
1435                                  struct address_space *mapping,
1436                                  struct iov_iter *iter,
1437                                  loff_t pos, unsigned len)
1438 {
1439         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1440         struct page *pages[WRITE_BATCH_PAGES];
1441         struct bch2_page_reservation res;
1442         unsigned long index = pos >> PAGE_SHIFT;
1443         unsigned offset = pos & (PAGE_SIZE - 1);
1444         unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1445         unsigned i, reserved = 0, set_dirty = 0;
1446         unsigned copied = 0, nr_pages_copied = 0;
1447         int ret = 0;
1448
1449         BUG_ON(!len);
1450         BUG_ON(nr_pages > ARRAY_SIZE(pages));
1451
1452         bch2_page_reservation_init(c, inode, &res);
1453
1454         for (i = 0; i < nr_pages; i++) {
1455                 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1456                 if (!pages[i]) {
1457                         nr_pages = i;
1458                         if (!i) {
1459                                 ret = -ENOMEM;
1460                                 goto out;
1461                         }
1462                         len = min_t(unsigned, len,
1463                                     nr_pages * PAGE_SIZE - offset);
1464                         break;
1465                 }
1466         }
1467
1468         if (offset && !PageUptodate(pages[0])) {
1469                 ret = bch2_read_single_page(pages[0], mapping);
1470                 if (ret)
1471                         goto out;
1472         }
1473
1474         if ((pos + len) & (PAGE_SIZE - 1) &&
1475             !PageUptodate(pages[nr_pages - 1])) {
1476                 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1477                         zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1478                 } else {
1479                         ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1480                         if (ret)
1481                                 goto out;
1482                 }
1483         }
1484
1485         while (reserved < len) {
1486                 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1487                 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1488                 unsigned pg_len = min_t(unsigned, len - reserved,
1489                                         PAGE_SIZE - pg_offset);
1490 retry_reservation:
1491                 ret = bch2_page_reservation_get(c, inode, page, &res,
1492                                                 pg_offset, pg_len, true);
1493
1494                 if (ret && !PageUptodate(page)) {
1495                         ret = bch2_read_single_page(page, mapping);
1496                         if (!ret)
1497                                 goto retry_reservation;
1498                 }
1499
1500                 if (ret)
1501                         goto out;
1502
1503                 reserved += pg_len;
1504         }
1505
1506         if (mapping_writably_mapped(mapping))
1507                 for (i = 0; i < nr_pages; i++)
1508                         flush_dcache_page(pages[i]);
1509
1510         while (copied < len) {
1511                 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1512                 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1513                 unsigned pg_len = min_t(unsigned, len - copied,
1514                                         PAGE_SIZE - pg_offset);
1515                 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1516                                                 iter, pg_offset, pg_len);
1517
1518                 if (!pg_copied)
1519                         break;
1520
1521                 if (!PageUptodate(page) &&
1522                     pg_copied != PAGE_SIZE &&
1523                     pos + copied + pg_copied < inode->v.i_size) {
1524                         zero_user(page, 0, PAGE_SIZE);
1525                         break;
1526                 }
1527
1528                 flush_dcache_page(page);
1529                 iov_iter_advance(iter, pg_copied);
1530                 copied += pg_copied;
1531
1532                 if (pg_copied != pg_len)
1533                         break;
1534         }
1535
1536         if (!copied)
1537                 goto out;
1538
1539         spin_lock(&inode->v.i_lock);
1540         if (pos + copied > inode->v.i_size)
1541                 i_size_write(&inode->v, pos + copied);
1542         spin_unlock(&inode->v.i_lock);
1543
1544         while (set_dirty < copied) {
1545                 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1546                 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1547                 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1548                                         PAGE_SIZE - pg_offset);
1549
1550                 if (!PageUptodate(page))
1551                         SetPageUptodate(page);
1552
1553                 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1554                 unlock_page(page);
1555                 put_page(page);
1556
1557                 set_dirty += pg_len;
1558         }
1559
1560         nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1561         inode->ei_last_dirtied = (unsigned long) current;
1562 out:
1563         for (i = nr_pages_copied; i < nr_pages; i++) {
1564                 unlock_page(pages[i]);
1565                 put_page(pages[i]);
1566         }
1567
1568         bch2_page_reservation_put(c, inode, &res);
1569
1570         return copied ?: ret;
1571 }
1572
1573 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1574 {
1575         struct file *file = iocb->ki_filp;
1576         struct address_space *mapping = file->f_mapping;
1577         struct bch_inode_info *inode = file_bch_inode(file);
1578         loff_t pos = iocb->ki_pos;
1579         ssize_t written = 0;
1580         int ret = 0;
1581
1582         bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1583
1584         do {
1585                 unsigned offset = pos & (PAGE_SIZE - 1);
1586                 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1587                               PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1588 again:
1589                 /*
1590                  * Bring in the user page that we will copy from _first_.
1591                  * Otherwise there's a nasty deadlock on copying from the
1592                  * same page as we're writing to, without it being marked
1593                  * up-to-date.
1594                  *
1595                  * Not only is this an optimisation, but it is also required
1596                  * to check that the address is actually valid, when atomic
1597                  * usercopies are used, below.
1598                  */
1599                 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1600                         bytes = min_t(unsigned long, iov_iter_count(iter),
1601                                       PAGE_SIZE - offset);
1602
1603                         if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1604                                 ret = -EFAULT;
1605                                 break;
1606                         }
1607                 }
1608
1609                 if (unlikely(fatal_signal_pending(current))) {
1610                         ret = -EINTR;
1611                         break;
1612                 }
1613
1614                 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1615                 if (unlikely(ret < 0))
1616                         break;
1617
1618                 cond_resched();
1619
1620                 if (unlikely(ret == 0)) {
1621                         /*
1622                          * If we were unable to copy any data at all, we must
1623                          * fall back to a single segment length write.
1624                          *
1625                          * If we didn't fallback here, we could livelock
1626                          * because not all segments in the iov can be copied at
1627                          * once without a pagefault.
1628                          */
1629                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
1630                                       iov_iter_single_seg_count(iter));
1631                         goto again;
1632                 }
1633                 pos += ret;
1634                 written += ret;
1635                 ret = 0;
1636
1637                 balance_dirty_pages_ratelimited(mapping);
1638         } while (iov_iter_count(iter));
1639
1640         bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1641
1642         return written ? written : ret;
1643 }
1644
1645 /* O_DIRECT reads */
1646
1647 static void bch2_dio_read_complete(struct closure *cl)
1648 {
1649         struct dio_read *dio = container_of(cl, struct dio_read, cl);
1650
1651         dio->req->ki_complete(dio->req, dio->ret, 0);
1652         bio_check_pages_dirty(&dio->rbio.bio);  /* transfers ownership */
1653 }
1654
1655 static void bch2_direct_IO_read_endio(struct bio *bio)
1656 {
1657         struct dio_read *dio = bio->bi_private;
1658
1659         if (bio->bi_status)
1660                 dio->ret = blk_status_to_errno(bio->bi_status);
1661
1662         closure_put(&dio->cl);
1663 }
1664
1665 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1666 {
1667         bch2_direct_IO_read_endio(bio);
1668         bio_check_pages_dirty(bio);     /* transfers ownership */
1669 }
1670
1671 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1672 {
1673         struct file *file = req->ki_filp;
1674         struct bch_inode_info *inode = file_bch_inode(file);
1675         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1676         struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1677         struct dio_read *dio;
1678         struct bio *bio;
1679         loff_t offset = req->ki_pos;
1680         bool sync = is_sync_kiocb(req);
1681         size_t shorten;
1682         ssize_t ret;
1683
1684         if ((offset|iter->count) & (block_bytes(c) - 1))
1685                 return -EINVAL;
1686
1687         ret = min_t(loff_t, iter->count,
1688                     max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1689
1690         if (!ret)
1691                 return ret;
1692
1693         shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1694         iter->count -= shorten;
1695
1696         bio = bio_alloc_bioset(GFP_KERNEL,
1697                                iov_iter_npages(iter, BIO_MAX_PAGES),
1698                                &c->dio_read_bioset);
1699
1700         bio->bi_end_io = bch2_direct_IO_read_endio;
1701
1702         dio = container_of(bio, struct dio_read, rbio.bio);
1703         closure_init(&dio->cl, NULL);
1704
1705         /*
1706          * this is a _really_ horrible hack just to avoid an atomic sub at the
1707          * end:
1708          */
1709         if (!sync) {
1710                 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1711                 atomic_set(&dio->cl.remaining,
1712                            CLOSURE_REMAINING_INITIALIZER -
1713                            CLOSURE_RUNNING +
1714                            CLOSURE_DESTRUCTOR);
1715         } else {
1716                 atomic_set(&dio->cl.remaining,
1717                            CLOSURE_REMAINING_INITIALIZER + 1);
1718         }
1719
1720         dio->req        = req;
1721         dio->ret        = ret;
1722
1723         goto start;
1724         while (iter->count) {
1725                 bio = bio_alloc_bioset(GFP_KERNEL,
1726                                        iov_iter_npages(iter, BIO_MAX_PAGES),
1727                                        &c->bio_read);
1728                 bio->bi_end_io          = bch2_direct_IO_read_split_endio;
1729 start:
1730                 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1731                 bio->bi_iter.bi_sector  = offset >> 9;
1732                 bio->bi_private         = dio;
1733
1734                 ret = bio_iov_iter_get_pages(bio, iter);
1735                 if (ret < 0) {
1736                         /* XXX: fault inject this path */
1737                         bio->bi_status = BLK_STS_RESOURCE;
1738                         bio_endio(bio);
1739                         break;
1740                 }
1741
1742                 offset += bio->bi_iter.bi_size;
1743                 bio_set_pages_dirty(bio);
1744
1745                 if (iter->count)
1746                         closure_get(&dio->cl);
1747
1748                 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1749         }
1750
1751         iter->count += shorten;
1752
1753         if (sync) {
1754                 closure_sync(&dio->cl);
1755                 closure_debug_destroy(&dio->cl);
1756                 ret = dio->ret;
1757                 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1758                 return ret;
1759         } else {
1760                 return -EIOCBQUEUED;
1761         }
1762 }
1763
1764 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1765 {
1766         struct file *file = iocb->ki_filp;
1767         struct bch_inode_info *inode = file_bch_inode(file);
1768         struct address_space *mapping = file->f_mapping;
1769         size_t count = iov_iter_count(iter);
1770         ssize_t ret;
1771
1772         if (!count)
1773                 return 0; /* skip atime */
1774
1775         if (iocb->ki_flags & IOCB_DIRECT) {
1776                 struct blk_plug plug;
1777
1778                 ret = filemap_write_and_wait_range(mapping,
1779                                         iocb->ki_pos,
1780                                         iocb->ki_pos + count - 1);
1781                 if (ret < 0)
1782                         return ret;
1783
1784                 file_accessed(file);
1785
1786                 blk_start_plug(&plug);
1787                 ret = bch2_direct_IO_read(iocb, iter);
1788                 blk_finish_plug(&plug);
1789
1790                 if (ret >= 0)
1791                         iocb->ki_pos += ret;
1792         } else {
1793                 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1794                 ret = generic_file_read_iter(iocb, iter);
1795                 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1796         }
1797
1798         return ret;
1799 }
1800
1801 /* O_DIRECT writes */
1802
1803 static void bch2_dio_write_loop_async(struct bch_write_op *);
1804
1805 static long bch2_dio_write_loop(struct dio_write *dio)
1806 {
1807         bool kthread = (current->flags & PF_KTHREAD) != 0;
1808         struct kiocb *req = dio->req;
1809         struct address_space *mapping = req->ki_filp->f_mapping;
1810         struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1811         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1812         struct bio *bio = &dio->op.wbio.bio;
1813         struct bvec_iter_all iter;
1814         struct bio_vec *bv;
1815         unsigned unaligned;
1816         bool sync = dio->sync;
1817         long ret;
1818
1819         if (dio->loop)
1820                 goto loop;
1821
1822         while (1) {
1823                 if (kthread)
1824                         kthread_use_mm(dio->mm);
1825                 BUG_ON(current->faults_disabled_mapping);
1826                 current->faults_disabled_mapping = mapping;
1827
1828                 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1829
1830                 current->faults_disabled_mapping = NULL;
1831                 if (kthread)
1832                         kthread_unuse_mm(dio->mm);
1833
1834                 if (unlikely(ret < 0))
1835                         goto err;
1836
1837                 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1838                 bio->bi_iter.bi_size -= unaligned;
1839                 iov_iter_revert(&dio->iter, unaligned);
1840
1841                 if (!bio->bi_iter.bi_size) {
1842                         /*
1843                          * bio_iov_iter_get_pages was only able to get <
1844                          * blocksize worth of pages:
1845                          */
1846                         bio_for_each_segment_all(bv, bio, iter)
1847                                 put_page(bv->bv_page);
1848                         ret = -EFAULT;
1849                         goto err;
1850                 }
1851
1852                 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1853                 dio->op.end_io          = bch2_dio_write_loop_async;
1854                 dio->op.target          = dio->op.opts.foreground_target;
1855                 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1856                 dio->op.write_point     = writepoint_hashed((unsigned long) current);
1857                 dio->op.nr_replicas     = dio->op.opts.data_replicas;
1858                 dio->op.pos             = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1859
1860                 if ((req->ki_flags & IOCB_DSYNC) &&
1861                     !c->opts.journal_flush_disabled)
1862                         dio->op.flags |= BCH_WRITE_FLUSH;
1863
1864                 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1865                                                 dio->op.opts.data_replicas, 0);
1866                 if (unlikely(ret) &&
1867                     !bch2_check_range_allocated(c, dio->op.pos,
1868                                 bio_sectors(bio), dio->op.opts.data_replicas))
1869                         goto err;
1870
1871                 task_io_account_write(bio->bi_iter.bi_size);
1872
1873                 if (!dio->sync && !dio->loop && dio->iter.count) {
1874                         struct iovec *iov = dio->inline_vecs;
1875
1876                         if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1877                                 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1878                                               GFP_KERNEL);
1879                                 if (unlikely(!iov)) {
1880                                         dio->sync = sync = true;
1881                                         goto do_io;
1882                                 }
1883
1884                                 dio->free_iov = true;
1885                         }
1886
1887                         memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1888                         dio->iter.iov = iov;
1889                 }
1890 do_io:
1891                 dio->loop = true;
1892                 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1893
1894                 if (sync)
1895                         wait_for_completion(&dio->done);
1896                 else
1897                         return -EIOCBQUEUED;
1898 loop:
1899                 i_sectors_acct(c, inode, &dio->quota_res,
1900                                dio->op.i_sectors_delta);
1901                 req->ki_pos += (u64) dio->op.written << 9;
1902                 dio->written += dio->op.written;
1903
1904                 spin_lock(&inode->v.i_lock);
1905                 if (req->ki_pos > inode->v.i_size)
1906                         i_size_write(&inode->v, req->ki_pos);
1907                 spin_unlock(&inode->v.i_lock);
1908
1909                 bio_for_each_segment_all(bv, bio, iter)
1910                         put_page(bv->bv_page);
1911                 if (!dio->iter.count || dio->op.error)
1912                         break;
1913
1914                 bio_reset(bio);
1915                 reinit_completion(&dio->done);
1916         }
1917
1918         ret = dio->op.error ?: ((long) dio->written << 9);
1919 err:
1920         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1921         bch2_quota_reservation_put(c, inode, &dio->quota_res);
1922
1923         if (dio->free_iov)
1924                 kfree(dio->iter.iov);
1925
1926         bio_put(bio);
1927
1928         /* inode->i_dio_count is our ref on inode and thus bch_fs */
1929         inode_dio_end(&inode->v);
1930
1931         if (!sync) {
1932                 req->ki_complete(req, ret, 0);
1933                 ret = -EIOCBQUEUED;
1934         }
1935         return ret;
1936 }
1937
1938 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1939 {
1940         struct dio_write *dio = container_of(op, struct dio_write, op);
1941
1942         if (dio->sync)
1943                 complete(&dio->done);
1944         else
1945                 bch2_dio_write_loop(dio);
1946 }
1947
1948 static noinline
1949 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1950 {
1951         struct file *file = req->ki_filp;
1952         struct address_space *mapping = file->f_mapping;
1953         struct bch_inode_info *inode = file_bch_inode(file);
1954         struct bch_fs *c = inode->v.i_sb->s_fs_info;
1955         struct dio_write *dio;
1956         struct bio *bio;
1957         bool locked = true, extending;
1958         ssize_t ret;
1959
1960         prefetch(&c->opts);
1961         prefetch((void *) &c->opts + 64);
1962         prefetch(&inode->ei_inode);
1963         prefetch((void *) &inode->ei_inode + 64);
1964
1965         inode_lock(&inode->v);
1966
1967         ret = generic_write_checks(req, iter);
1968         if (unlikely(ret <= 0))
1969                 goto err;
1970
1971         ret = file_remove_privs(file);
1972         if (unlikely(ret))
1973                 goto err;
1974
1975         ret = file_update_time(file);
1976         if (unlikely(ret))
1977                 goto err;
1978
1979         if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1980                 goto err;
1981
1982         inode_dio_begin(&inode->v);
1983         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
1984
1985         extending = req->ki_pos + iter->count > inode->v.i_size;
1986         if (!extending) {
1987                 inode_unlock(&inode->v);
1988                 locked = false;
1989         }
1990
1991         bio = bio_alloc_bioset(GFP_KERNEL,
1992                                iov_iter_npages(iter, BIO_MAX_PAGES),
1993                                &c->dio_write_bioset);
1994         dio = container_of(bio, struct dio_write, op.wbio.bio);
1995         init_completion(&dio->done);
1996         dio->req                = req;
1997         dio->mm                 = current->mm;
1998         dio->loop               = false;
1999         dio->sync               = is_sync_kiocb(req) || extending;
2000         dio->free_iov           = false;
2001         dio->quota_res.sectors  = 0;
2002         dio->written            = 0;
2003         dio->iter               = *iter;
2004
2005         ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2006                                          iter->count >> 9, true);
2007         if (unlikely(ret))
2008                 goto err_put_bio;
2009
2010         ret = write_invalidate_inode_pages_range(mapping,
2011                                         req->ki_pos,
2012                                         req->ki_pos + iter->count - 1);
2013         if (unlikely(ret))
2014                 goto err_put_bio;
2015
2016         ret = bch2_dio_write_loop(dio);
2017 err:
2018         if (locked)
2019                 inode_unlock(&inode->v);
2020         return ret;
2021 err_put_bio:
2022         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2023         bch2_quota_reservation_put(c, inode, &dio->quota_res);
2024         bio_put(bio);
2025         inode_dio_end(&inode->v);
2026         goto err;
2027 }
2028
2029 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2030 {
2031         struct file *file = iocb->ki_filp;
2032         struct bch_inode_info *inode = file_bch_inode(file);
2033         ssize_t ret;
2034
2035         if (iocb->ki_flags & IOCB_DIRECT)
2036                 return bch2_direct_write(iocb, from);
2037
2038         /* We can write back this queue in page reclaim */
2039         current->backing_dev_info = inode_to_bdi(&inode->v);
2040         inode_lock(&inode->v);
2041
2042         ret = generic_write_checks(iocb, from);
2043         if (ret <= 0)
2044                 goto unlock;
2045
2046         ret = file_remove_privs(file);
2047         if (ret)
2048                 goto unlock;
2049
2050         ret = file_update_time(file);
2051         if (ret)
2052                 goto unlock;
2053
2054         ret = bch2_buffered_write(iocb, from);
2055         if (likely(ret > 0))
2056                 iocb->ki_pos += ret;
2057 unlock:
2058         inode_unlock(&inode->v);
2059         current->backing_dev_info = NULL;
2060
2061         if (ret > 0)
2062                 ret = generic_write_sync(iocb, ret);
2063
2064         return ret;
2065 }
2066
2067 /* fsync: */
2068
2069 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2070 {
2071         struct bch_inode_info *inode = file_bch_inode(file);
2072         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2073         int ret, ret2;
2074
2075         ret = file_write_and_wait_range(file, start, end);
2076         if (ret)
2077                 return ret;
2078
2079         if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2080                 goto out;
2081
2082         ret = sync_inode_metadata(&inode->v, 1);
2083         if (ret)
2084                 return ret;
2085 out:
2086         if (!c->opts.journal_flush_disabled)
2087                 ret = bch2_journal_flush_seq(&c->journal,
2088                                              inode->ei_journal_seq);
2089         ret2 = file_check_and_advance_wb_err(file);
2090
2091         return ret ?: ret2;
2092 }
2093
2094 /* truncate: */
2095
2096 static inline int range_has_data(struct bch_fs *c,
2097                                   struct bpos start,
2098                                   struct bpos end)
2099 {
2100         struct btree_trans trans;
2101         struct btree_iter *iter;
2102         struct bkey_s_c k;
2103         int ret = 0;
2104
2105         bch2_trans_init(&trans, c, 0, 0);
2106
2107         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2108                 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2109                         break;
2110
2111                 if (bkey_extent_is_data(k.k)) {
2112                         ret = 1;
2113                         break;
2114                 }
2115         }
2116
2117         return bch2_trans_exit(&trans) ?: ret;
2118 }
2119
2120 static int __bch2_truncate_page(struct bch_inode_info *inode,
2121                                 pgoff_t index, loff_t start, loff_t end)
2122 {
2123         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2124         struct address_space *mapping = inode->v.i_mapping;
2125         struct bch_page_state *s;
2126         unsigned start_offset = start & (PAGE_SIZE - 1);
2127         unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2128         unsigned i;
2129         struct page *page;
2130         int ret = 0;
2131
2132         /* Page boundary? Nothing to do */
2133         if (!((index == start >> PAGE_SHIFT && start_offset) ||
2134               (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2135                 return 0;
2136
2137         /* Above i_size? */
2138         if (index << PAGE_SHIFT >= inode->v.i_size)
2139                 return 0;
2140
2141         page = find_lock_page(mapping, index);
2142         if (!page) {
2143                 /*
2144                  * XXX: we're doing two index lookups when we end up reading the
2145                  * page
2146                  */
2147                 ret = range_has_data(c,
2148                                 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2149                                 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2150                 if (ret <= 0)
2151                         return ret;
2152
2153                 page = find_or_create_page(mapping, index, GFP_KERNEL);
2154                 if (unlikely(!page)) {
2155                         ret = -ENOMEM;
2156                         goto out;
2157                 }
2158         }
2159
2160         s = bch2_page_state_create(page, 0);
2161         if (!s) {
2162                 ret = -ENOMEM;
2163                 goto unlock;
2164         }
2165
2166         if (!PageUptodate(page)) {
2167                 ret = bch2_read_single_page(page, mapping);
2168                 if (ret)
2169                         goto unlock;
2170         }
2171
2172         if (index != start >> PAGE_SHIFT)
2173                 start_offset = 0;
2174         if (index != end >> PAGE_SHIFT)
2175                 end_offset = PAGE_SIZE;
2176
2177         for (i = round_up(start_offset, block_bytes(c)) >> 9;
2178              i < round_down(end_offset, block_bytes(c)) >> 9;
2179              i++) {
2180                 s->s[i].nr_replicas     = 0;
2181                 s->s[i].state           = SECTOR_UNALLOCATED;
2182         }
2183
2184         zero_user_segment(page, start_offset, end_offset);
2185
2186         /*
2187          * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2188          *
2189          * XXX: because we aren't currently tracking whether the page has actual
2190          * data in it (vs. just 0s, or only partially written) this wrong. ick.
2191          */
2192         ret = bch2_get_page_disk_reservation(c, inode, page, false);
2193         BUG_ON(ret);
2194
2195         /*
2196          * This removes any writeable userspace mappings; we need to force
2197          * .page_mkwrite to be called again before any mmapped writes, to
2198          * redirty the full page:
2199          */
2200         page_mkclean(page);
2201         __set_page_dirty_nobuffers(page);
2202 unlock:
2203         unlock_page(page);
2204         put_page(page);
2205 out:
2206         return ret;
2207 }
2208
2209 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2210 {
2211         return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2212                                     from, round_up(from, PAGE_SIZE));
2213 }
2214
2215 static int bch2_extend(struct bch_inode_info *inode,
2216                        struct bch_inode_unpacked *inode_u,
2217                        struct iattr *iattr)
2218 {
2219         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2220         struct address_space *mapping = inode->v.i_mapping;
2221         int ret;
2222
2223         /*
2224          * sync appends:
2225          *
2226          * this has to be done _before_ extending i_size:
2227          */
2228         ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2229         if (ret)
2230                 return ret;
2231
2232         truncate_setsize(&inode->v, iattr->ia_size);
2233         setattr_copy(&inode->v, iattr);
2234
2235         mutex_lock(&inode->ei_update_lock);
2236         ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2237                                     ATTR_MTIME|ATTR_CTIME);
2238         mutex_unlock(&inode->ei_update_lock);
2239
2240         return ret;
2241 }
2242
2243 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2244                                    struct bch_inode_unpacked *bi,
2245                                    void *p)
2246 {
2247         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2248
2249         bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2250         bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2251         return 0;
2252 }
2253
2254 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2255                                   struct bch_inode_unpacked *bi, void *p)
2256 {
2257         u64 *new_i_size = p;
2258
2259         bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2260         bi->bi_size = *new_i_size;
2261         return 0;
2262 }
2263
2264 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2265 {
2266         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2267         struct address_space *mapping = inode->v.i_mapping;
2268         struct bch_inode_unpacked inode_u;
2269         struct btree_trans trans;
2270         struct btree_iter *iter;
2271         u64 new_i_size = iattr->ia_size;
2272         s64 i_sectors_delta = 0;
2273         int ret = 0;
2274
2275         inode_dio_wait(&inode->v);
2276         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2277
2278         /*
2279          * fetch current on disk i_size: inode is locked, i_size can only
2280          * increase underneath us:
2281          */
2282         bch2_trans_init(&trans, c, 0, 0);
2283         iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2284         ret = PTR_ERR_OR_ZERO(iter);
2285         bch2_trans_exit(&trans);
2286
2287         if (ret)
2288                 goto err;
2289
2290         /*
2291          * check this before next assertion; on filesystem error our normal
2292          * invariants are a bit broken (truncate has to truncate the page cache
2293          * before the inode).
2294          */
2295         ret = bch2_journal_error(&c->journal);
2296         if (ret)
2297                 goto err;
2298
2299         BUG_ON(inode->v.i_size < inode_u.bi_size);
2300
2301         if (iattr->ia_size > inode->v.i_size) {
2302                 ret = bch2_extend(inode, &inode_u, iattr);
2303                 goto err;
2304         }
2305
2306         ret = bch2_truncate_page(inode, iattr->ia_size);
2307         if (unlikely(ret))
2308                 goto err;
2309
2310         /*
2311          * When extending, we're going to write the new i_size to disk
2312          * immediately so we need to flush anything above the current on disk
2313          * i_size first:
2314          *
2315          * Also, when extending we need to flush the page that i_size currently
2316          * straddles - if it's mapped to userspace, we need to ensure that
2317          * userspace has to redirty it and call .mkwrite -> set_page_dirty
2318          * again to allocate the part of the page that was extended.
2319          */
2320         if (iattr->ia_size > inode_u.bi_size)
2321                 ret = filemap_write_and_wait_range(mapping,
2322                                 inode_u.bi_size,
2323                                 iattr->ia_size - 1);
2324         else if (iattr->ia_size & (PAGE_SIZE - 1))
2325                 ret = filemap_write_and_wait_range(mapping,
2326                                 round_down(iattr->ia_size, PAGE_SIZE),
2327                                 iattr->ia_size - 1);
2328         if (ret)
2329                 goto err;
2330
2331         mutex_lock(&inode->ei_update_lock);
2332         ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2333                                &new_i_size, 0);
2334         mutex_unlock(&inode->ei_update_lock);
2335
2336         if (unlikely(ret))
2337                 goto err;
2338
2339         truncate_setsize(&inode->v, iattr->ia_size);
2340
2341         ret = bch2_fpunch(c, inode->v.i_ino,
2342                         round_up(iattr->ia_size, block_bytes(c)) >> 9,
2343                         U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2344         i_sectors_acct(c, inode, NULL, i_sectors_delta);
2345
2346         if (unlikely(ret))
2347                 goto err;
2348
2349         setattr_copy(&inode->v, iattr);
2350
2351         mutex_lock(&inode->ei_update_lock);
2352         ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2353                                ATTR_MTIME|ATTR_CTIME);
2354         mutex_unlock(&inode->ei_update_lock);
2355 err:
2356         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2357         return ret;
2358 }
2359
2360 /* fallocate: */
2361
2362 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2363 {
2364         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2365         u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2366         u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2367         int ret = 0;
2368
2369         inode_lock(&inode->v);
2370         inode_dio_wait(&inode->v);
2371         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2372
2373         ret = __bch2_truncate_page(inode,
2374                                    offset >> PAGE_SHIFT,
2375                                    offset, offset + len);
2376         if (unlikely(ret))
2377                 goto err;
2378
2379         if (offset >> PAGE_SHIFT !=
2380             (offset + len) >> PAGE_SHIFT) {
2381                 ret = __bch2_truncate_page(inode,
2382                                            (offset + len) >> PAGE_SHIFT,
2383                                            offset, offset + len);
2384                 if (unlikely(ret))
2385                         goto err;
2386         }
2387
2388         truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2389
2390         if (discard_start < discard_end) {
2391                 s64 i_sectors_delta = 0;
2392
2393                 ret = bch2_fpunch(c, inode->v.i_ino,
2394                                   discard_start, discard_end,
2395                                   &inode->ei_journal_seq,
2396                                   &i_sectors_delta);
2397                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2398         }
2399 err:
2400         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2401         inode_unlock(&inode->v);
2402
2403         return ret;
2404 }
2405
2406 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2407                                    loff_t offset, loff_t len,
2408                                    bool insert)
2409 {
2410         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2411         struct address_space *mapping = inode->v.i_mapping;
2412         struct bkey_on_stack copy;
2413         struct btree_trans trans;
2414         struct btree_iter *src, *dst;
2415         loff_t shift, new_size;
2416         u64 src_start;
2417         int ret;
2418
2419         if ((offset | len) & (block_bytes(c) - 1))
2420                 return -EINVAL;
2421
2422         bkey_on_stack_init(&copy);
2423         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2424
2425         /*
2426          * We need i_mutex to keep the page cache consistent with the extents
2427          * btree, and the btree consistent with i_size - we don't need outside
2428          * locking for the extents btree itself, because we're using linked
2429          * iterators
2430          */
2431         inode_lock(&inode->v);
2432         inode_dio_wait(&inode->v);
2433         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2434
2435         if (insert) {
2436                 ret = -EFBIG;
2437                 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2438                         goto err;
2439
2440                 ret = -EINVAL;
2441                 if (offset >= inode->v.i_size)
2442                         goto err;
2443
2444                 src_start       = U64_MAX;
2445                 shift           = len;
2446         } else {
2447                 ret = -EINVAL;
2448                 if (offset + len >= inode->v.i_size)
2449                         goto err;
2450
2451                 src_start       = offset + len;
2452                 shift           = -len;
2453         }
2454
2455         new_size = inode->v.i_size + shift;
2456
2457         ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2458         if (ret)
2459                 goto err;
2460
2461         if (insert) {
2462                 i_size_write(&inode->v, new_size);
2463                 mutex_lock(&inode->ei_update_lock);
2464                 ret = bch2_write_inode_size(c, inode, new_size,
2465                                             ATTR_MTIME|ATTR_CTIME);
2466                 mutex_unlock(&inode->ei_update_lock);
2467         } else {
2468                 s64 i_sectors_delta = 0;
2469
2470                 ret = bch2_fpunch(c, inode->v.i_ino,
2471                                   offset >> 9, (offset + len) >> 9,
2472                                   &inode->ei_journal_seq,
2473                                   &i_sectors_delta);
2474                 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2475
2476                 if (ret)
2477                         goto err;
2478         }
2479
2480         src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2481                         POS(inode->v.i_ino, src_start >> 9),
2482                         BTREE_ITER_INTENT);
2483         BUG_ON(IS_ERR_OR_NULL(src));
2484
2485         dst = bch2_trans_copy_iter(&trans, src);
2486         BUG_ON(IS_ERR_OR_NULL(dst));
2487
2488         while (1) {
2489                 struct disk_reservation disk_res =
2490                         bch2_disk_reservation_init(c, 0);
2491                 struct bkey_i delete;
2492                 struct bkey_s_c k;
2493                 struct bpos next_pos;
2494                 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2495                 struct bpos atomic_end;
2496                 unsigned trigger_flags = 0;
2497
2498                 k = insert
2499                         ? bch2_btree_iter_peek_prev(src)
2500                         : bch2_btree_iter_peek(src);
2501                 if ((ret = bkey_err(k)))
2502                         goto bkey_err;
2503
2504                 if (!k.k || k.k->p.inode != inode->v.i_ino)
2505                         break;
2506
2507                 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2508
2509                 if (insert &&
2510                     bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2511                         break;
2512 reassemble:
2513                 bkey_on_stack_reassemble(&copy, c, k);
2514
2515                 if (insert &&
2516                     bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2517                         bch2_cut_front(move_pos, copy.k);
2518
2519                 copy.k->k.p.offset += shift >> 9;
2520                 bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
2521
2522                 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2523                 if (ret)
2524                         goto bkey_err;
2525
2526                 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2527                         if (insert) {
2528                                 move_pos = atomic_end;
2529                                 move_pos.offset -= shift >> 9;
2530                                 goto reassemble;
2531                         } else {
2532                                 bch2_cut_back(atomic_end, copy.k);
2533                         }
2534                 }
2535
2536                 bkey_init(&delete.k);
2537                 delete.k.p = copy.k->k.p;
2538                 delete.k.size = copy.k->k.size;
2539                 delete.k.p.offset -= shift >> 9;
2540
2541                 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2542
2543                 if (copy.k->k.size == k.k->size) {
2544                         /*
2545                          * If we're moving the entire extent, we can skip
2546                          * running triggers:
2547                          */
2548                         trigger_flags |= BTREE_TRIGGER_NORUN;
2549                 } else {
2550                         /* We might end up splitting compressed extents: */
2551                         unsigned nr_ptrs =
2552                                 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2553
2554                         ret = bch2_disk_reservation_get(c, &disk_res,
2555                                         copy.k->k.size, nr_ptrs,
2556                                         BCH_DISK_RESERVATION_NOFAIL);
2557                         BUG_ON(ret);
2558                 }
2559
2560                 bch2_btree_iter_set_pos(src, bkey_start_pos(&delete.k));
2561
2562                 ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
2563                         bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2564                         bch2_trans_commit(&trans, &disk_res,
2565                                           &inode->ei_journal_seq,
2566                                           BTREE_INSERT_NOFAIL);
2567                 bch2_disk_reservation_put(c, &disk_res);
2568 bkey_err:
2569                 if (!ret)
2570                         bch2_btree_iter_set_pos(src, next_pos);
2571
2572                 if (ret == -EINTR)
2573                         ret = 0;
2574                 if (ret)
2575                         goto err;
2576
2577                 bch2_trans_cond_resched(&trans);
2578         }
2579         bch2_trans_unlock(&trans);
2580
2581         if (!insert) {
2582                 i_size_write(&inode->v, new_size);
2583                 mutex_lock(&inode->ei_update_lock);
2584                 ret = bch2_write_inode_size(c, inode, new_size,
2585                                             ATTR_MTIME|ATTR_CTIME);
2586                 mutex_unlock(&inode->ei_update_lock);
2587         }
2588 err:
2589         bch2_trans_exit(&trans);
2590         bkey_on_stack_exit(&copy, c);
2591         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2592         inode_unlock(&inode->v);
2593         return ret;
2594 }
2595
2596 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2597                             loff_t offset, loff_t len)
2598 {
2599         struct address_space *mapping = inode->v.i_mapping;
2600         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2601         struct btree_trans trans;
2602         struct btree_iter *iter;
2603         struct bpos end_pos;
2604         loff_t end              = offset + len;
2605         loff_t block_start      = round_down(offset,    block_bytes(c));
2606         loff_t block_end        = round_up(end,         block_bytes(c));
2607         unsigned sectors;
2608         unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2609         int ret;
2610
2611         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2612
2613         inode_lock(&inode->v);
2614         inode_dio_wait(&inode->v);
2615         bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2616
2617         if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2618                 ret = inode_newsize_ok(&inode->v, end);
2619                 if (ret)
2620                         goto err;
2621         }
2622
2623         if (mode & FALLOC_FL_ZERO_RANGE) {
2624                 ret = __bch2_truncate_page(inode,
2625                                            offset >> PAGE_SHIFT,
2626                                            offset, end);
2627
2628                 if (!ret &&
2629                     offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2630                         ret = __bch2_truncate_page(inode,
2631                                                    end >> PAGE_SHIFT,
2632                                                    offset, end);
2633
2634                 if (unlikely(ret))
2635                         goto err;
2636
2637                 truncate_pagecache_range(&inode->v, offset, end - 1);
2638         }
2639
2640         iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2641                         POS(inode->v.i_ino, block_start >> 9),
2642                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2643         end_pos = POS(inode->v.i_ino, block_end >> 9);
2644
2645         while (bkey_cmp(iter->pos, end_pos) < 0) {
2646                 s64 i_sectors_delta = 0;
2647                 struct disk_reservation disk_res = { 0 };
2648                 struct quota_res quota_res = { 0 };
2649                 struct bkey_i_reservation reservation;
2650                 struct bkey_s_c k;
2651
2652                 bch2_trans_begin(&trans);
2653
2654                 k = bch2_btree_iter_peek_slot(iter);
2655                 if ((ret = bkey_err(k)))
2656                         goto bkey_err;
2657
2658                 /* already reserved */
2659                 if (k.k->type == KEY_TYPE_reservation &&
2660                     bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2661                         bch2_btree_iter_next_slot(iter);
2662                         continue;
2663                 }
2664
2665                 if (bkey_extent_is_data(k.k) &&
2666                     !(mode & FALLOC_FL_ZERO_RANGE)) {
2667                         bch2_btree_iter_next_slot(iter);
2668                         continue;
2669                 }
2670
2671                 bkey_reservation_init(&reservation.k_i);
2672                 reservation.k.type      = KEY_TYPE_reservation;
2673                 reservation.k.p         = k.k->p;
2674                 reservation.k.size      = k.k->size;
2675
2676                 bch2_cut_front(iter->pos,       &reservation.k_i);
2677                 bch2_cut_back(end_pos,          &reservation.k_i);
2678
2679                 sectors = reservation.k.size;
2680                 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2681
2682                 if (!bkey_extent_is_allocation(k.k)) {
2683                         ret = bch2_quota_reservation_add(c, inode,
2684                                         &quota_res,
2685                                         sectors, true);
2686                         if (unlikely(ret))
2687                                 goto bkey_err;
2688                 }
2689
2690                 if (reservation.v.nr_replicas < replicas ||
2691                     bch2_bkey_sectors_compressed(k)) {
2692                         ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2693                                                         replicas, 0);
2694                         if (unlikely(ret))
2695                                 goto bkey_err;
2696
2697                         reservation.v.nr_replicas = disk_res.nr_replicas;
2698                 }
2699
2700                 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2701                                 &disk_res, &inode->ei_journal_seq,
2702                                 0, &i_sectors_delta);
2703                 i_sectors_acct(c, inode, &quota_res, i_sectors_delta);
2704 bkey_err:
2705                 bch2_quota_reservation_put(c, inode, &quota_res);
2706                 bch2_disk_reservation_put(c, &disk_res);
2707                 if (ret == -EINTR)
2708                         ret = 0;
2709                 if (ret)
2710                         goto err;
2711         }
2712
2713         /*
2714          * Do we need to extend the file?
2715          *
2716          * If we zeroed up to the end of the file, we dropped whatever writes
2717          * were going to write out the current i_size, so we have to extend
2718          * manually even if FL_KEEP_SIZE was set:
2719          */
2720         if (end >= inode->v.i_size &&
2721             (!(mode & FALLOC_FL_KEEP_SIZE) ||
2722              (mode & FALLOC_FL_ZERO_RANGE))) {
2723                 struct btree_iter *inode_iter;
2724                 struct bch_inode_unpacked inode_u;
2725
2726                 do {
2727                         bch2_trans_begin(&trans);
2728                         inode_iter = bch2_inode_peek(&trans, &inode_u,
2729                                                      inode->v.i_ino, 0);
2730                         ret = PTR_ERR_OR_ZERO(inode_iter);
2731                 } while (ret == -EINTR);
2732
2733                 bch2_trans_unlock(&trans);
2734
2735                 if (ret)
2736                         goto err;
2737
2738                 /*
2739                  * Sync existing appends before extending i_size,
2740                  * as in bch2_extend():
2741                  */
2742                 ret = filemap_write_and_wait_range(mapping,
2743                                         inode_u.bi_size, S64_MAX);
2744                 if (ret)
2745                         goto err;
2746
2747                 if (mode & FALLOC_FL_KEEP_SIZE)
2748                         end = inode->v.i_size;
2749                 else
2750                         i_size_write(&inode->v, end);
2751
2752                 mutex_lock(&inode->ei_update_lock);
2753                 ret = bch2_write_inode_size(c, inode, end, 0);
2754                 mutex_unlock(&inode->ei_update_lock);
2755         }
2756 err:
2757         bch2_trans_exit(&trans);
2758         bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2759         inode_unlock(&inode->v);
2760         return ret;
2761 }
2762
2763 long bch2_fallocate_dispatch(struct file *file, int mode,
2764                              loff_t offset, loff_t len)
2765 {
2766         struct bch_inode_info *inode = file_bch_inode(file);
2767         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2768         long ret;
2769
2770         if (!percpu_ref_tryget(&c->writes))
2771                 return -EROFS;
2772
2773         if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2774                 ret = bchfs_fallocate(inode, mode, offset, len);
2775         else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2776                 ret = bchfs_fpunch(inode, offset, len);
2777         else if (mode == FALLOC_FL_INSERT_RANGE)
2778                 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2779         else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2780                 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2781         else
2782                 ret = -EOPNOTSUPP;
2783
2784         percpu_ref_put(&c->writes);
2785
2786         return ret;
2787 }
2788
2789 static void mark_range_unallocated(struct bch_inode_info *inode,
2790                                    loff_t start, loff_t end)
2791 {
2792         pgoff_t index = start >> PAGE_SHIFT;
2793         pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2794         struct pagevec pvec;
2795
2796         pagevec_init(&pvec);
2797
2798         do {
2799                 unsigned nr_pages, i, j;
2800
2801                 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2802                                                 &index, end_index);
2803                 if (nr_pages == 0)
2804                         break;
2805
2806                 for (i = 0; i < nr_pages; i++) {
2807                         struct page *page = pvec.pages[i];
2808                         struct bch_page_state *s;
2809
2810                         lock_page(page);
2811                         s = bch2_page_state(page);
2812
2813                         if (s) {
2814                                 spin_lock(&s->lock);
2815                                 for (j = 0; j < PAGE_SECTORS; j++)
2816                                         s->s[j].nr_replicas = 0;
2817                                 spin_unlock(&s->lock);
2818                         }
2819
2820                         unlock_page(page);
2821                 }
2822                 pagevec_release(&pvec);
2823         } while (index <= end_index);
2824 }
2825
2826 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2827                              struct file *file_dst, loff_t pos_dst,
2828                              loff_t len, unsigned remap_flags)
2829 {
2830         struct bch_inode_info *src = file_bch_inode(file_src);
2831         struct bch_inode_info *dst = file_bch_inode(file_dst);
2832         struct bch_fs *c = src->v.i_sb->s_fs_info;
2833         s64 i_sectors_delta = 0;
2834         u64 aligned_len;
2835         loff_t ret = 0;
2836
2837         if (!c->opts.reflink)
2838                 return -EOPNOTSUPP;
2839
2840         if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2841                 return -EINVAL;
2842
2843         if (remap_flags & REMAP_FILE_DEDUP)
2844                 return -EOPNOTSUPP;
2845
2846         if ((pos_src & (block_bytes(c) - 1)) ||
2847             (pos_dst & (block_bytes(c) - 1)))
2848                 return -EINVAL;
2849
2850         if (src == dst &&
2851             abs(pos_src - pos_dst) < len)
2852                 return -EINVAL;
2853
2854         bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2855
2856         file_update_time(file_dst);
2857
2858         inode_dio_wait(&src->v);
2859         inode_dio_wait(&dst->v);
2860
2861         ret = generic_remap_file_range_prep(file_src, pos_src,
2862                                             file_dst, pos_dst,
2863                                             &len, remap_flags);
2864         if (ret < 0 || len == 0)
2865                 goto err;
2866
2867         aligned_len = round_up((u64) len, block_bytes(c));
2868
2869         ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2870                                 pos_dst, pos_dst + len - 1);
2871         if (ret)
2872                 goto err;
2873
2874         mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2875
2876         ret = bch2_remap_range(c,
2877                                POS(dst->v.i_ino, pos_dst >> 9),
2878                                POS(src->v.i_ino, pos_src >> 9),
2879                                aligned_len >> 9,
2880                                &dst->ei_journal_seq,
2881                                pos_dst + len, &i_sectors_delta);
2882         if (ret < 0)
2883                 goto err;
2884
2885         /*
2886          * due to alignment, we might have remapped slightly more than requsted
2887          */
2888         ret = min((u64) ret << 9, (u64) len);
2889
2890         /* XXX get a quota reservation */
2891         i_sectors_acct(c, dst, NULL, i_sectors_delta);
2892
2893         spin_lock(&dst->v.i_lock);
2894         if (pos_dst + ret > dst->v.i_size)
2895                 i_size_write(&dst->v, pos_dst + ret);
2896         spin_unlock(&dst->v.i_lock);
2897 err:
2898         bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2899
2900         return ret;
2901 }
2902
2903 /* fseek: */
2904
2905 static int page_data_offset(struct page *page, unsigned offset)
2906 {
2907         struct bch_page_state *s = bch2_page_state(page);
2908         unsigned i;
2909
2910         if (s)
2911                 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2912                         if (s->s[i].state >= SECTOR_DIRTY)
2913                                 return i << 9;
2914
2915         return -1;
2916 }
2917
2918 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2919                                        loff_t start_offset,
2920                                        loff_t end_offset)
2921 {
2922         struct address_space *mapping = vinode->i_mapping;
2923         struct page *page;
2924         pgoff_t start_index     = start_offset >> PAGE_SHIFT;
2925         pgoff_t end_index       = end_offset >> PAGE_SHIFT;
2926         pgoff_t index           = start_index;
2927         loff_t ret;
2928         int offset;
2929
2930         while (index <= end_index) {
2931                 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2932                         lock_page(page);
2933
2934                         offset = page_data_offset(page,
2935                                         page->index == start_index
2936                                         ? start_offset & (PAGE_SIZE - 1)
2937                                         : 0);
2938                         if (offset >= 0) {
2939                                 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2940                                             offset,
2941                                             start_offset, end_offset);
2942                                 unlock_page(page);
2943                                 put_page(page);
2944                                 return ret;
2945                         }
2946
2947                         unlock_page(page);
2948                         put_page(page);
2949                 } else {
2950                         break;
2951                 }
2952         }
2953
2954         return end_offset;
2955 }
2956
2957 static loff_t bch2_seek_data(struct file *file, u64 offset)
2958 {
2959         struct bch_inode_info *inode = file_bch_inode(file);
2960         struct bch_fs *c = inode->v.i_sb->s_fs_info;
2961         struct btree_trans trans;
2962         struct btree_iter *iter;
2963         struct bkey_s_c k;
2964         u64 isize, next_data = MAX_LFS_FILESIZE;
2965         int ret;
2966
2967         isize = i_size_read(&inode->v);
2968         if (offset >= isize)
2969                 return -ENXIO;
2970
2971         bch2_trans_init(&trans, c, 0, 0);
2972
2973         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2974                            POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2975                 if (k.k->p.inode != inode->v.i_ino) {
2976                         break;
2977                 } else if (bkey_extent_is_data(k.k)) {
2978                         next_data = max(offset, bkey_start_offset(k.k) << 9);
2979                         break;
2980                 } else if (k.k->p.offset >> 9 > isize)
2981                         break;
2982         }
2983
2984         ret = bch2_trans_exit(&trans) ?: ret;
2985         if (ret)
2986                 return ret;
2987
2988         if (next_data > offset)
2989                 next_data = bch2_seek_pagecache_data(&inode->v,
2990                                                      offset, next_data);
2991
2992         if (next_data >= isize)
2993                 return -ENXIO;
2994
2995         return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2996 }
2997
2998 static int __page_hole_offset(struct page *page, unsigned offset)
2999 {
3000         struct bch_page_state *s = bch2_page_state(page);
3001         unsigned i;
3002
3003         if (!s)
3004                 return 0;
3005
3006         for (i = offset >> 9; i < PAGE_SECTORS; i++)
3007                 if (s->s[i].state < SECTOR_DIRTY)
3008                         return i << 9;
3009
3010         return -1;
3011 }
3012
3013 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3014 {
3015         pgoff_t index = offset >> PAGE_SHIFT;
3016         struct page *page;
3017         int pg_offset;
3018         loff_t ret = -1;
3019
3020         page = find_lock_entry(mapping, index);
3021         if (!page || xa_is_value(page))
3022                 return offset;
3023
3024         pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3025         if (pg_offset >= 0)
3026                 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3027
3028         unlock_page(page);
3029
3030         return ret;
3031 }
3032
3033 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3034                                        loff_t start_offset,
3035                                        loff_t end_offset)
3036 {
3037         struct address_space *mapping = vinode->i_mapping;
3038         loff_t offset = start_offset, hole;
3039
3040         while (offset < end_offset) {
3041                 hole = page_hole_offset(mapping, offset);
3042                 if (hole >= 0 && hole <= end_offset)
3043                         return max(start_offset, hole);
3044
3045                 offset += PAGE_SIZE;
3046                 offset &= PAGE_MASK;
3047         }
3048
3049         return end_offset;
3050 }
3051
3052 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3053 {
3054         struct bch_inode_info *inode = file_bch_inode(file);
3055         struct bch_fs *c = inode->v.i_sb->s_fs_info;
3056         struct btree_trans trans;
3057         struct btree_iter *iter;
3058         struct bkey_s_c k;
3059         u64 isize, next_hole = MAX_LFS_FILESIZE;
3060         int ret;
3061
3062         isize = i_size_read(&inode->v);
3063         if (offset >= isize)
3064                 return -ENXIO;
3065
3066         bch2_trans_init(&trans, c, 0, 0);
3067
3068         for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3069                            POS(inode->v.i_ino, offset >> 9),
3070                            BTREE_ITER_SLOTS, k, ret) {
3071                 if (k.k->p.inode != inode->v.i_ino) {
3072                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3073                                         offset, MAX_LFS_FILESIZE);
3074                         break;
3075                 } else if (!bkey_extent_is_data(k.k)) {
3076                         next_hole = bch2_seek_pagecache_hole(&inode->v,
3077                                         max(offset, bkey_start_offset(k.k) << 9),
3078                                         k.k->p.offset << 9);
3079
3080                         if (next_hole < k.k->p.offset << 9)
3081                                 break;
3082                 } else {
3083                         offset = max(offset, bkey_start_offset(k.k) << 9);
3084                 }
3085         }
3086
3087         ret = bch2_trans_exit(&trans) ?: ret;
3088         if (ret)
3089                 return ret;
3090
3091         if (next_hole > isize)
3092                 next_hole = isize;
3093
3094         return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3095 }
3096
3097 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3098 {
3099         switch (whence) {
3100         case SEEK_SET:
3101         case SEEK_CUR:
3102         case SEEK_END:
3103                 return generic_file_llseek(file, offset, whence);
3104         case SEEK_DATA:
3105                 return bch2_seek_data(file, offset);
3106         case SEEK_HOLE:
3107                 return bch2_seek_hole(file, offset);
3108         }
3109
3110         return -EINVAL;
3111 }
3112
3113 void bch2_fs_fsio_exit(struct bch_fs *c)
3114 {
3115         bioset_exit(&c->dio_write_bioset);
3116         bioset_exit(&c->dio_read_bioset);
3117         bioset_exit(&c->writepage_bioset);
3118 }
3119
3120 int bch2_fs_fsio_init(struct bch_fs *c)
3121 {
3122         int ret = 0;
3123
3124         pr_verbose_init(c->opts, "");
3125
3126         if (bioset_init(&c->writepage_bioset,
3127                         4, offsetof(struct bch_writepage_io, op.wbio.bio),
3128                         BIOSET_NEED_BVECS) ||
3129             bioset_init(&c->dio_read_bioset,
3130                         4, offsetof(struct dio_read, rbio.bio),
3131                         BIOSET_NEED_BVECS) ||
3132             bioset_init(&c->dio_write_bioset,
3133                         4, offsetof(struct dio_write, op.wbio.bio),
3134                         BIOSET_NEED_BVECS))
3135                 ret = -ENOMEM;
3136
3137         pr_verbose_init(c->opts, "ret %i", ret);
3138         return ret;
3139 }
3140
3141 #endif /* NO_BCACHEFS_FS */