1 // SPDX-License-Identifier: GPL-2.0
5 #include "alloc_foreground.h"
7 #include "btree_update.h"
12 #include "extent_update.h"
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>
35 #include <trace/events/bcachefs.h>
36 #include <trace/events/writeback.h>
39 * Use u64 for the end pos and sector helpers because if the folio covers the
40 * max supported range of the mapping, the start offset of the next folio
41 * overflows loff_t. This breaks much of the range based processing in the
42 * buffered write path.
44 static inline u64 folio_end_pos(struct folio *folio)
46 return folio_pos(folio) + folio_size(folio);
49 static inline size_t folio_sectors(struct folio *folio)
51 return PAGE_SECTORS << folio_order(folio);
54 static inline loff_t folio_sector(struct folio *folio)
56 return folio_pos(folio) >> 9;
59 static inline u64 folio_end_sector(struct folio *folio)
61 return folio_end_pos(folio) >> 9;
64 typedef DARRAY(struct folio *) folios;
66 static int filemap_get_contig_folios_d(struct address_space *mapping,
67 loff_t start, u64 end,
68 int fgp_flags, gfp_t gfp,
76 if ((u64) pos >= (u64) start + (1ULL << 20))
77 fgp_flags &= ~FGP_CREAT;
79 ret = darray_make_room_gfp(folios, 1, gfp & GFP_KERNEL);
83 f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
87 BUG_ON(folios->nr && folio_pos(f) != pos);
89 pos = folio_end_pos(f);
90 darray_push(folios, f);
93 if (!folios->nr && !ret && (fgp_flags & FGP_CREAT))
96 return folios->nr ? 0 : ret;
105 static void nocow_flush_endio(struct bio *_bio)
108 struct nocow_flush *bio = container_of(_bio, struct nocow_flush, bio);
110 closure_put(bio->cl);
111 percpu_ref_put(&bio->ca->io_ref);
115 static void bch2_inode_flush_nocow_writes_async(struct bch_fs *c,
116 struct bch_inode_info *inode,
119 struct nocow_flush *bio;
121 struct bch_devs_mask devs;
124 dev = find_first_bit(inode->ei_devs_need_flush.d, BCH_SB_MEMBERS_MAX);
125 if (dev == BCH_SB_MEMBERS_MAX)
128 devs = inode->ei_devs_need_flush;
129 memset(&inode->ei_devs_need_flush, 0, sizeof(inode->ei_devs_need_flush));
131 for_each_set_bit(dev, devs.d, BCH_SB_MEMBERS_MAX) {
133 ca = rcu_dereference(c->devs[dev]);
134 if (ca && !percpu_ref_tryget(&ca->io_ref))
141 bio = container_of(bio_alloc_bioset(ca->disk_sb.bdev, 0,
144 &c->nocow_flush_bioset),
145 struct nocow_flush, bio);
148 bio->bio.bi_end_io = nocow_flush_endio;
149 closure_bio_submit(&bio->bio, cl);
153 static int bch2_inode_flush_nocow_writes(struct bch_fs *c,
154 struct bch_inode_info *inode)
158 closure_init_stack(&cl);
159 bch2_inode_flush_nocow_writes_async(c, inode, &cl);
165 static inline bool bio_full(struct bio *bio, unsigned len)
167 if (bio->bi_vcnt >= bio->bi_max_vecs)
169 if (bio->bi_iter.bi_size > UINT_MAX - len)
174 static inline struct address_space *faults_disabled_mapping(void)
176 return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
179 static inline void set_fdm_dropped_locks(void)
181 current->faults_disabled_mapping =
182 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
185 static inline bool fdm_dropped_locks(void)
187 return ((unsigned long) current->faults_disabled_mapping) & 1;
194 struct bch_writepage_io {
195 struct bch_inode_info *inode;
198 struct bch_write_op op;
203 struct address_space *mapping;
204 struct bch_inode_info *inode;
205 struct mm_struct *mm;
211 struct quota_res quota_res;
214 struct iov_iter iter;
215 struct iovec inline_vecs[2];
218 struct bch_write_op op;
226 struct bch_read_bio rbio;
229 /* pagecache_block must be held */
230 static noinline int write_invalidate_inode_pages_range(struct address_space *mapping,
231 loff_t start, loff_t end)
236 * XXX: the way this is currently implemented, we can spin if a process
237 * is continually redirtying a specific page
240 if (!mapping->nrpages)
243 ret = filemap_write_and_wait_range(mapping, start, end);
247 if (!mapping->nrpages)
250 ret = invalidate_inode_pages2_range(mapping,
253 } while (ret == -EBUSY);
260 #ifdef CONFIG_BCACHEFS_QUOTA
262 static void __bch2_quota_reservation_put(struct bch_fs *c,
263 struct bch_inode_info *inode,
264 struct quota_res *res)
266 BUG_ON(res->sectors > inode->ei_quota_reserved);
268 bch2_quota_acct(c, inode->ei_qid, Q_SPC,
269 -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
270 inode->ei_quota_reserved -= res->sectors;
274 static void bch2_quota_reservation_put(struct bch_fs *c,
275 struct bch_inode_info *inode,
276 struct quota_res *res)
279 mutex_lock(&inode->ei_quota_lock);
280 __bch2_quota_reservation_put(c, inode, res);
281 mutex_unlock(&inode->ei_quota_lock);
285 static int bch2_quota_reservation_add(struct bch_fs *c,
286 struct bch_inode_info *inode,
287 struct quota_res *res,
293 mutex_lock(&inode->ei_quota_lock);
294 ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
295 check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
297 inode->ei_quota_reserved += sectors;
298 res->sectors += sectors;
300 mutex_unlock(&inode->ei_quota_lock);
307 static void __bch2_quota_reservation_put(struct bch_fs *c,
308 struct bch_inode_info *inode,
309 struct quota_res *res) {}
311 static void bch2_quota_reservation_put(struct bch_fs *c,
312 struct bch_inode_info *inode,
313 struct quota_res *res) {}
315 static int bch2_quota_reservation_add(struct bch_fs *c,
316 struct bch_inode_info *inode,
317 struct quota_res *res,
326 /* i_size updates: */
328 struct inode_new_size {
334 static int inode_set_size(struct bch_inode_info *inode,
335 struct bch_inode_unpacked *bi,
338 struct inode_new_size *s = p;
340 bi->bi_size = s->new_size;
341 if (s->fields & ATTR_ATIME)
342 bi->bi_atime = s->now;
343 if (s->fields & ATTR_MTIME)
344 bi->bi_mtime = s->now;
345 if (s->fields & ATTR_CTIME)
346 bi->bi_ctime = s->now;
351 int __must_check bch2_write_inode_size(struct bch_fs *c,
352 struct bch_inode_info *inode,
353 loff_t new_size, unsigned fields)
355 struct inode_new_size s = {
356 .new_size = new_size,
357 .now = bch2_current_time(c),
361 return bch2_write_inode(c, inode, inode_set_size, &s, fields);
364 static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
365 struct quota_res *quota_res, s64 sectors)
367 bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
368 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
369 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
370 inode->ei_inode.bi_sectors);
371 inode->v.i_blocks += sectors;
373 #ifdef CONFIG_BCACHEFS_QUOTA
374 if (quota_res && sectors > 0) {
375 BUG_ON(sectors > quota_res->sectors);
376 BUG_ON(sectors > inode->ei_quota_reserved);
378 quota_res->sectors -= sectors;
379 inode->ei_quota_reserved -= sectors;
381 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
386 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
387 struct quota_res *quota_res, s64 sectors)
390 mutex_lock(&inode->ei_quota_lock);
391 __i_sectors_acct(c, inode, quota_res, sectors);
392 mutex_unlock(&inode->ei_quota_lock);
398 /* stored in page->private: */
400 #define BCH_FOLIO_SECTOR_STATE() \
407 enum bch_folio_sector_state {
408 #define x(n) SECTOR_##n,
409 BCH_FOLIO_SECTOR_STATE()
413 const char * const bch2_folio_sector_states[] = {
415 BCH_FOLIO_SECTOR_STATE()
420 static inline enum bch_folio_sector_state
421 folio_sector_dirty(enum bch_folio_sector_state state)
424 case SECTOR_unallocated:
426 case SECTOR_reserved:
427 return SECTOR_dirty_reserved;
433 static inline enum bch_folio_sector_state
434 folio_sector_undirty(enum bch_folio_sector_state state)
438 return SECTOR_unallocated;
439 case SECTOR_dirty_reserved:
440 return SECTOR_reserved;
446 static inline enum bch_folio_sector_state
447 folio_sector_reserve(enum bch_folio_sector_state state)
450 case SECTOR_unallocated:
451 return SECTOR_reserved;
453 return SECTOR_dirty_reserved;
459 struct bch_folio_sector {
460 /* Uncompressed, fully allocated replicas (or on disk reservation): */
461 unsigned nr_replicas:4;
463 /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
464 unsigned replicas_reserved:4;
467 enum bch_folio_sector_state state:8;
472 atomic_t write_count;
474 * Is the sector state up to date with the btree?
475 * (Not the data itself)
478 struct bch_folio_sector s[];
481 static inline void folio_sector_set(struct folio *folio,
483 unsigned i, unsigned n)
488 static inline struct bch_folio *__bch2_folio(struct folio *folio)
490 return folio_has_private(folio)
491 ? (struct bch_folio *) folio_get_private(folio)
495 static inline struct bch_folio *bch2_folio(struct folio *folio)
497 EBUG_ON(!folio_test_locked(folio));
499 return __bch2_folio(folio);
502 /* for newly allocated folios: */
503 static void __bch2_folio_release(struct folio *folio)
505 kfree(folio_detach_private(folio));
508 static void bch2_folio_release(struct folio *folio)
510 EBUG_ON(!folio_test_locked(folio));
511 __bch2_folio_release(folio);
514 /* for newly allocated folios: */
515 static struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
519 s = kzalloc(sizeof(*s) +
520 sizeof(struct bch_folio_sector) *
521 folio_sectors(folio), GFP_NOFS|gfp);
525 spin_lock_init(&s->lock);
526 folio_attach_private(folio, s);
530 static struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
532 return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
535 static unsigned bkey_to_sector_state(struct bkey_s_c k)
537 if (bkey_extent_is_reservation(k))
538 return SECTOR_reserved;
539 if (bkey_extent_is_allocation(k.k))
540 return SECTOR_allocated;
541 return SECTOR_unallocated;
544 static void __bch2_folio_set(struct folio *folio,
545 unsigned pg_offset, unsigned pg_len,
546 unsigned nr_ptrs, unsigned state)
548 struct bch_folio *s = bch2_folio_create(folio, __GFP_NOFAIL);
549 unsigned i, sectors = folio_sectors(folio);
551 BUG_ON(pg_offset >= sectors);
552 BUG_ON(pg_offset + pg_len > sectors);
556 for (i = pg_offset; i < pg_offset + pg_len; i++) {
557 s->s[i].nr_replicas = nr_ptrs;
558 folio_sector_set(folio, s, i, state);
564 spin_unlock(&s->lock);
568 * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
571 static int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
572 struct folio **folios, unsigned nr_folios)
574 struct btree_trans trans;
575 struct btree_iter iter;
577 u64 offset = folio_sector(folios[0]);
578 unsigned folio_idx = 0;
582 bch2_trans_init(&trans, c, 0, 0);
584 bch2_trans_begin(&trans);
586 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
590 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
591 SPOS(inum.inum, offset, snapshot),
592 BTREE_ITER_SLOTS, k, ret) {
593 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
594 unsigned state = bkey_to_sector_state(k);
596 while (folio_idx < nr_folios) {
597 struct folio *folio = folios[folio_idx];
598 u64 folio_start = folio_sector(folio);
599 u64 folio_end = folio_end_sector(folio);
600 unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start;
601 unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start;
603 BUG_ON(k.k->p.offset < folio_start);
604 BUG_ON(bkey_start_offset(k.k) > folio_end);
606 if (!bch2_folio_create(folio, __GFP_NOFAIL)->uptodate)
607 __bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
609 if (k.k->p.offset < folio_end)
614 if (folio_idx == nr_folios)
618 offset = iter.pos.offset;
619 bch2_trans_iter_exit(&trans, &iter);
621 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
623 bch2_trans_exit(&trans);
628 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
630 struct bvec_iter iter;
632 unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
633 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
634 unsigned state = bkey_to_sector_state(k);
636 bio_for_each_folio(fv, bio, iter)
637 __bch2_folio_set(fv.fv_folio,
643 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
646 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
647 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
648 struct folio_batch fbatch;
654 folio_batch_init(&fbatch);
656 while (filemap_get_folios(inode->v.i_mapping,
657 &index, end_index, &fbatch)) {
658 for (i = 0; i < folio_batch_count(&fbatch); i++) {
659 struct folio *folio = fbatch.folios[i];
660 u64 folio_start = folio_sector(folio);
661 u64 folio_end = folio_end_sector(folio);
662 unsigned folio_offset = max(start, folio_start) - folio_start;
663 unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
666 BUG_ON(end <= folio_start);
669 s = bch2_folio(folio);
673 for (j = folio_offset; j < folio_offset + folio_len; j++)
674 s->s[j].nr_replicas = 0;
675 spin_unlock(&s->lock);
680 folio_batch_release(&fbatch);
685 static void mark_pagecache_reserved(struct bch_inode_info *inode,
688 struct bch_fs *c = inode->v.i_sb->s_fs_info;
689 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
690 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
691 struct folio_batch fbatch;
692 s64 i_sectors_delta = 0;
698 folio_batch_init(&fbatch);
700 while (filemap_get_folios(inode->v.i_mapping,
701 &index, end_index, &fbatch)) {
702 for (i = 0; i < folio_batch_count(&fbatch); i++) {
703 struct folio *folio = fbatch.folios[i];
704 u64 folio_start = folio_sector(folio);
705 u64 folio_end = folio_end_sector(folio);
706 unsigned folio_offset = max(start, folio_start) - folio_start;
707 unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
710 BUG_ON(end <= folio_start);
713 s = bch2_folio(folio);
717 for (j = folio_offset; j < folio_offset + folio_len; j++) {
718 i_sectors_delta -= s->s[j].state == SECTOR_dirty;
719 folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state));
721 spin_unlock(&s->lock);
726 folio_batch_release(&fbatch);
730 i_sectors_acct(c, inode, NULL, i_sectors_delta);
733 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
735 /* XXX: this should not be open coded */
736 return inode->ei_inode.bi_data_replicas
737 ? inode->ei_inode.bi_data_replicas - 1
738 : c->opts.data_replicas;
741 static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
742 unsigned nr_replicas)
744 return max(0, (int) nr_replicas -
746 s->replicas_reserved);
749 static int bch2_get_folio_disk_reservation(struct bch_fs *c,
750 struct bch_inode_info *inode,
751 struct folio *folio, bool check_enospc)
753 struct bch_folio *s = bch2_folio_create(folio, 0);
754 unsigned nr_replicas = inode_nr_replicas(c, inode);
755 struct disk_reservation disk_res = { 0 };
756 unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
762 for (i = 0; i < sectors; i++)
763 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
765 if (!disk_res_sectors)
768 ret = bch2_disk_reservation_get(c, &disk_res,
771 ? BCH_DISK_RESERVATION_NOFAIL
776 for (i = 0; i < sectors; i++)
777 s->s[i].replicas_reserved +=
778 sectors_to_reserve(&s->s[i], nr_replicas);
783 struct bch2_folio_reservation {
784 struct disk_reservation disk;
785 struct quota_res quota;
788 static void bch2_folio_reservation_init(struct bch_fs *c,
789 struct bch_inode_info *inode,
790 struct bch2_folio_reservation *res)
792 memset(res, 0, sizeof(*res));
794 res->disk.nr_replicas = inode_nr_replicas(c, inode);
797 static void bch2_folio_reservation_put(struct bch_fs *c,
798 struct bch_inode_info *inode,
799 struct bch2_folio_reservation *res)
801 bch2_disk_reservation_put(c, &res->disk);
802 bch2_quota_reservation_put(c, inode, &res->quota);
805 static int bch2_folio_reservation_get(struct bch_fs *c,
806 struct bch_inode_info *inode,
808 struct bch2_folio_reservation *res,
809 unsigned offset, unsigned len)
811 struct bch_folio *s = bch2_folio_create(folio, 0);
812 unsigned i, disk_sectors = 0, quota_sectors = 0;
818 BUG_ON(!s->uptodate);
820 for (i = round_down(offset, block_bytes(c)) >> 9;
821 i < round_up(offset + len, block_bytes(c)) >> 9;
823 disk_sectors += sectors_to_reserve(&s->s[i],
824 res->disk.nr_replicas);
825 quota_sectors += s->s[i].state == SECTOR_unallocated;
829 ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
835 ret = bch2_quota_reservation_add(c, inode, &res->quota,
836 quota_sectors, true);
838 struct disk_reservation tmp = {
839 .sectors = disk_sectors
842 bch2_disk_reservation_put(c, &tmp);
843 res->disk.sectors -= disk_sectors;
851 static void bch2_clear_folio_bits(struct folio *folio)
853 struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
854 struct bch_fs *c = inode->v.i_sb->s_fs_info;
855 struct bch_folio *s = bch2_folio(folio);
856 struct disk_reservation disk_res = { 0 };
857 int i, sectors = folio_sectors(folio), dirty_sectors = 0;
862 EBUG_ON(!folio_test_locked(folio));
863 EBUG_ON(folio_test_writeback(folio));
865 for (i = 0; i < sectors; i++) {
866 disk_res.sectors += s->s[i].replicas_reserved;
867 s->s[i].replicas_reserved = 0;
869 dirty_sectors -= s->s[i].state == SECTOR_dirty;
870 folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
873 bch2_disk_reservation_put(c, &disk_res);
875 i_sectors_acct(c, inode, NULL, dirty_sectors);
877 bch2_folio_release(folio);
880 static void bch2_set_folio_dirty(struct bch_fs *c,
881 struct bch_inode_info *inode,
883 struct bch2_folio_reservation *res,
884 unsigned offset, unsigned len)
886 struct bch_folio *s = bch2_folio(folio);
887 unsigned i, dirty_sectors = 0;
889 WARN_ON((u64) folio_pos(folio) + offset + len >
890 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
892 BUG_ON(!s->uptodate);
896 for (i = round_down(offset, block_bytes(c)) >> 9;
897 i < round_up(offset + len, block_bytes(c)) >> 9;
899 unsigned sectors = sectors_to_reserve(&s->s[i],
900 res->disk.nr_replicas);
903 * This can happen if we race with the error path in
904 * bch2_writepage_io_done():
906 sectors = min_t(unsigned, sectors, res->disk.sectors);
908 s->s[i].replicas_reserved += sectors;
909 res->disk.sectors -= sectors;
911 dirty_sectors += s->s[i].state == SECTOR_unallocated;
913 folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
916 spin_unlock(&s->lock);
918 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
920 if (!folio_test_dirty(folio))
921 filemap_dirty_folio(inode->v.i_mapping, folio);
924 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
926 struct file *file = vmf->vma->vm_file;
927 struct address_space *mapping = file->f_mapping;
928 struct address_space *fdm = faults_disabled_mapping();
929 struct bch_inode_info *inode = file_bch_inode(file);
933 return VM_FAULT_SIGBUS;
937 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
939 if (bch2_pagecache_add_tryget(inode))
942 bch2_pagecache_block_put(fdm_host);
944 bch2_pagecache_add_get(inode);
945 bch2_pagecache_add_put(inode);
947 bch2_pagecache_block_get(fdm_host);
949 /* Signal that lock has been dropped: */
950 set_fdm_dropped_locks();
951 return VM_FAULT_SIGBUS;
954 bch2_pagecache_add_get(inode);
956 ret = filemap_fault(vmf);
957 bch2_pagecache_add_put(inode);
962 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
964 struct folio *folio = page_folio(vmf->page);
965 struct file *file = vmf->vma->vm_file;
966 struct bch_inode_info *inode = file_bch_inode(file);
967 struct address_space *mapping = file->f_mapping;
968 struct bch_fs *c = inode->v.i_sb->s_fs_info;
969 struct bch2_folio_reservation res;
974 bch2_folio_reservation_init(c, inode, &res);
976 sb_start_pagefault(inode->v.i_sb);
977 file_update_time(file);
980 * Not strictly necessary, but helps avoid dio writes livelocking in
981 * write_invalidate_inode_pages_range() - can drop this if/when we get
982 * a write_invalidate_inode_pages_range() that works without dropping
983 * page lock before invalidating page
985 bch2_pagecache_add_get(inode);
988 isize = i_size_read(&inode->v);
990 if (folio->mapping != mapping || folio_pos(folio) >= isize) {
992 ret = VM_FAULT_NOPAGE;
996 len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
998 if (!bch2_folio_create(folio, __GFP_NOFAIL)->uptodate) {
999 if (bch2_folio_set(c, inode_inum(inode), &folio, 1)) {
1000 folio_unlock(folio);
1001 ret = VM_FAULT_SIGBUS;
1006 if (bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
1007 folio_unlock(folio);
1008 ret = VM_FAULT_SIGBUS;
1012 bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
1013 bch2_folio_reservation_put(c, inode, &res);
1015 folio_wait_stable(folio);
1016 ret = VM_FAULT_LOCKED;
1018 bch2_pagecache_add_put(inode);
1019 sb_end_pagefault(inode->v.i_sb);
1024 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1026 if (offset || length < folio_size(folio))
1029 bch2_clear_folio_bits(folio);
1032 bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
1034 if (folio_test_dirty(folio) || folio_test_writeback(folio))
1037 bch2_clear_folio_bits(folio);
1043 static void bch2_readpages_end_io(struct bio *bio)
1045 struct bvec_iter_all iter;
1046 struct folio_vec fv;
1048 bio_for_each_folio_all(fv, bio, iter) {
1049 if (!bio->bi_status) {
1050 folio_mark_uptodate(fv.fv_folio);
1052 folio_clear_uptodate(fv.fv_folio);
1053 folio_set_error(fv.fv_folio);
1055 folio_unlock(fv.fv_folio);
1061 struct readpages_iter {
1062 struct address_space *mapping;
1067 static int readpages_iter_init(struct readpages_iter *iter,
1068 struct readahead_control *ractl)
1073 memset(iter, 0, sizeof(*iter));
1075 iter->mapping = ractl->mapping;
1077 ret = filemap_get_contig_folios_d(iter->mapping,
1078 ractl->_index << PAGE_SHIFT,
1079 (ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
1080 0, mapping_gfp_mask(iter->mapping),
1085 darray_for_each(iter->folios, fi) {
1086 ractl->_nr_pages -= 1U << folio_order(*fi);
1087 __bch2_folio_create(*fi, __GFP_NOFAIL);
1095 static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
1097 if (iter->idx >= iter->folios.nr)
1099 return iter->folios.data[iter->idx];
1102 static inline void readpage_iter_advance(struct readpages_iter *iter)
1107 static bool extent_partial_reads_expensive(struct bkey_s_c k)
1109 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1110 struct bch_extent_crc_unpacked crc;
1111 const union bch_extent_entry *i;
1113 bkey_for_each_crc(k.k, ptrs, crc, i)
1114 if (crc.csum_type || crc.compression_type)
1119 static void readpage_bio_extend(struct readpages_iter *iter,
1121 unsigned sectors_this_extent,
1124 while (bio_sectors(bio) < sectors_this_extent &&
1125 bio->bi_vcnt < bio->bi_max_vecs) {
1126 struct folio *folio = readpage_iter_peek(iter);
1130 readpage_iter_advance(iter);
1132 pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
1137 folio = xa_load(&iter->mapping->i_pages, folio_offset);
1138 if (folio && !xa_is_value(folio))
1141 folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
1145 if (!__bch2_folio_create(folio, 0)) {
1150 ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_NOFS);
1152 __bch2_folio_release(folio);
1160 BUG_ON(folio_sector(folio) != bio_end_sector(bio));
1162 BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
1166 static void bchfs_read(struct btree_trans *trans,
1167 struct bch_read_bio *rbio,
1169 struct readpages_iter *readpages_iter)
1171 struct bch_fs *c = trans->c;
1172 struct btree_iter iter;
1174 int flags = BCH_READ_RETRY_IF_STALE|
1175 BCH_READ_MAY_PROMOTE;
1180 rbio->start_time = local_clock();
1181 rbio->subvol = inum.subvol;
1183 bch2_bkey_buf_init(&sk);
1185 bch2_trans_begin(trans);
1186 iter = (struct btree_iter) { NULL };
1188 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1192 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1193 SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1197 unsigned bytes, sectors, offset_into_extent;
1198 enum btree_id data_btree = BTREE_ID_extents;
1201 * read_extent -> io_time_reset may cause a transaction restart
1202 * without returning an error, we need to check for that here:
1204 ret = bch2_trans_relock(trans);
1208 bch2_btree_iter_set_pos(&iter,
1209 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1211 k = bch2_btree_iter_peek_slot(&iter);
1216 offset_into_extent = iter.pos.offset -
1217 bkey_start_offset(k.k);
1218 sectors = k.k->size - offset_into_extent;
1220 bch2_bkey_buf_reassemble(&sk, c, k);
1222 ret = bch2_read_indirect_extent(trans, &data_btree,
1223 &offset_into_extent, &sk);
1227 k = bkey_i_to_s_c(sk.k);
1229 sectors = min(sectors, k.k->size - offset_into_extent);
1232 readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
1233 extent_partial_reads_expensive(k));
1235 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1236 swap(rbio->bio.bi_iter.bi_size, bytes);
1238 if (rbio->bio.bi_iter.bi_size == bytes)
1239 flags |= BCH_READ_LAST_FRAGMENT;
1241 bch2_bio_page_state_set(&rbio->bio, k);
1243 bch2_read_extent(trans, rbio, iter.pos,
1244 data_btree, k, offset_into_extent, flags);
1246 if (flags & BCH_READ_LAST_FRAGMENT)
1249 swap(rbio->bio.bi_iter.bi_size, bytes);
1250 bio_advance(&rbio->bio, bytes);
1252 ret = btree_trans_too_many_iters(trans);
1257 bch2_trans_iter_exit(trans, &iter);
1259 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1263 bch_err_inum_offset_ratelimited(c,
1265 iter.pos.offset << 9,
1266 "read error %i from btree lookup", ret);
1267 rbio->bio.bi_status = BLK_STS_IOERR;
1268 bio_endio(&rbio->bio);
1271 bch2_bkey_buf_exit(&sk, c);
1274 void bch2_readahead(struct readahead_control *ractl)
1276 struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1277 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1278 struct bch_io_opts opts;
1279 struct btree_trans trans;
1280 struct folio *folio;
1281 struct readpages_iter readpages_iter;
1284 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1286 ret = readpages_iter_init(&readpages_iter, ractl);
1289 bch2_trans_init(&trans, c, 0, 0);
1291 bch2_pagecache_add_get(inode);
1293 while ((folio = readpage_iter_peek(&readpages_iter))) {
1294 unsigned n = min_t(unsigned,
1295 readpages_iter.folios.nr -
1298 struct bch_read_bio *rbio =
1299 rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
1300 GFP_NOFS, &c->bio_read),
1303 readpage_iter_advance(&readpages_iter);
1305 rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1306 rbio->bio.bi_end_io = bch2_readpages_end_io;
1307 BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1309 bchfs_read(&trans, rbio, inode_inum(inode),
1313 bch2_pagecache_add_put(inode);
1315 bch2_trans_exit(&trans);
1316 darray_exit(&readpages_iter.folios);
1319 static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
1320 subvol_inum inum, struct folio *folio)
1322 struct btree_trans trans;
1324 bch2_folio_create(folio, __GFP_NOFAIL);
1326 rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
1327 rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1328 BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1330 bch2_trans_init(&trans, c, 0, 0);
1331 bchfs_read(&trans, rbio, inum, NULL);
1332 bch2_trans_exit(&trans);
1335 static void bch2_read_single_folio_end_io(struct bio *bio)
1337 complete(bio->bi_private);
1340 static int bch2_read_single_folio(struct folio *folio,
1341 struct address_space *mapping)
1343 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1344 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1345 struct bch_read_bio *rbio;
1346 struct bch_io_opts opts;
1348 DECLARE_COMPLETION_ONSTACK(done);
1350 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1352 rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS, &c->bio_read),
1354 rbio->bio.bi_private = &done;
1355 rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
1357 __bchfs_readfolio(c, rbio, inode_inum(inode), folio);
1358 wait_for_completion(&done);
1360 ret = blk_status_to_errno(rbio->bio.bi_status);
1361 bio_put(&rbio->bio);
1366 folio_mark_uptodate(folio);
1370 int bch2_read_folio(struct file *file, struct folio *folio)
1374 ret = bch2_read_single_folio(folio, folio->mapping);
1375 folio_unlock(folio);
1376 return bch2_err_class(ret);
1381 struct bch_writepage_state {
1382 struct bch_writepage_io *io;
1383 struct bch_io_opts opts;
1384 struct bch_folio_sector *tmp;
1385 unsigned tmp_sectors;
1388 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1389 struct bch_inode_info *inode)
1391 struct bch_writepage_state ret = { 0 };
1393 bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
1397 static void bch2_writepage_io_done(struct bch_write_op *op)
1399 struct bch_writepage_io *io =
1400 container_of(op, struct bch_writepage_io, op);
1401 struct bch_fs *c = io->op.c;
1402 struct bio *bio = &io->op.wbio.bio;
1403 struct bvec_iter_all iter;
1404 struct folio_vec fv;
1408 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1410 bio_for_each_folio_all(fv, bio, iter) {
1411 struct bch_folio *s;
1413 folio_set_error(fv.fv_folio);
1414 mapping_set_error(fv.fv_folio->mapping, -EIO);
1416 s = __bch2_folio(fv.fv_folio);
1417 spin_lock(&s->lock);
1418 for (i = 0; i < folio_sectors(fv.fv_folio); i++)
1419 s->s[i].nr_replicas = 0;
1420 spin_unlock(&s->lock);
1424 if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1425 bio_for_each_folio_all(fv, bio, iter) {
1426 struct bch_folio *s;
1428 s = __bch2_folio(fv.fv_folio);
1429 spin_lock(&s->lock);
1430 for (i = 0; i < folio_sectors(fv.fv_folio); i++)
1431 s->s[i].nr_replicas = 0;
1432 spin_unlock(&s->lock);
1437 * racing with fallocate can cause us to add fewer sectors than
1438 * expected - but we shouldn't add more sectors than expected:
1440 WARN_ON_ONCE(io->op.i_sectors_delta > 0);
1443 * (error (due to going RO) halfway through a page can screw that up
1446 BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1450 * PageWriteback is effectively our ref on the inode - fixup i_blocks
1451 * before calling end_page_writeback:
1453 i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1455 bio_for_each_folio_all(fv, bio, iter) {
1456 struct bch_folio *s = __bch2_folio(fv.fv_folio);
1458 if (atomic_dec_and_test(&s->write_count))
1459 folio_end_writeback(fv.fv_folio);
1462 bio_put(&io->op.wbio.bio);
1465 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1467 struct bch_writepage_io *io = w->io;
1470 closure_call(&io->op.cl, bch2_write, NULL, NULL);
1474 * Get a bch_writepage_io and add @page to it - appending to an existing one if
1475 * possible, else allocating a new one:
1477 static void bch2_writepage_io_alloc(struct bch_fs *c,
1478 struct writeback_control *wbc,
1479 struct bch_writepage_state *w,
1480 struct bch_inode_info *inode,
1482 unsigned nr_replicas)
1484 struct bch_write_op *op;
1486 w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1489 &c->writepage_bioset),
1490 struct bch_writepage_io, op.wbio.bio);
1492 w->io->inode = inode;
1494 bch2_write_op_init(op, c, w->opts);
1495 op->target = w->opts.foreground_target;
1496 op->nr_replicas = nr_replicas;
1497 op->res.nr_replicas = nr_replicas;
1498 op->write_point = writepoint_hashed(inode->ei_last_dirtied);
1499 op->subvol = inode->ei_subvol;
1500 op->pos = POS(inode->v.i_ino, sector);
1501 op->end_io = bch2_writepage_io_done;
1502 op->devs_need_flush = &inode->ei_devs_need_flush;
1503 op->wbio.bio.bi_iter.bi_sector = sector;
1504 op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
1507 static int __bch2_writepage(struct page *_page,
1508 struct writeback_control *wbc,
1511 struct folio *folio = page_folio(_page);
1512 struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
1513 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1514 struct bch_writepage_state *w = data;
1515 struct bch_folio *s;
1516 unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
1517 loff_t i_size = i_size_read(&inode->v);
1520 EBUG_ON(!folio_test_uptodate(folio));
1522 /* Is the folio fully inside i_size? */
1523 if (folio_end_pos(folio) <= i_size)
1526 /* Is the folio fully outside i_size? (truncate in progress) */
1527 if (folio_pos(folio) >= i_size) {
1528 folio_unlock(folio);
1533 * The folio straddles i_size. It must be zeroed out on each and every
1534 * writepage invocation because it may be mmapped. "A file is mapped
1535 * in multiples of the folio size. For a file that is not a multiple of
1536 * the folio size, the remaining memory is zeroed when mapped, and
1537 * writes to that region are not written out to the file."
1539 folio_zero_segment(folio,
1540 i_size - folio_pos(folio),
1543 f_sectors = folio_sectors(folio);
1544 s = bch2_folio_create(folio, __GFP_NOFAIL);
1546 if (f_sectors > w->tmp_sectors) {
1548 w->tmp = kzalloc(sizeof(struct bch_folio_sector) *
1549 f_sectors, __GFP_NOFAIL);
1550 w->tmp_sectors = f_sectors;
1554 * Things get really hairy with errors during writeback:
1556 ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
1559 /* Before unlocking the page, get copy of reservations: */
1560 spin_lock(&s->lock);
1561 memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
1563 for (i = 0; i < f_sectors; i++) {
1564 if (s->s[i].state < SECTOR_dirty)
1567 nr_replicas_this_write =
1568 min_t(unsigned, nr_replicas_this_write,
1569 s->s[i].nr_replicas +
1570 s->s[i].replicas_reserved);
1573 for (i = 0; i < f_sectors; i++) {
1574 if (s->s[i].state < SECTOR_dirty)
1577 s->s[i].nr_replicas = w->opts.compression
1578 ? 0 : nr_replicas_this_write;
1580 s->s[i].replicas_reserved = 0;
1581 folio_sector_set(folio, s, i, SECTOR_allocated);
1583 spin_unlock(&s->lock);
1585 BUG_ON(atomic_read(&s->write_count));
1586 atomic_set(&s->write_count, 1);
1588 BUG_ON(folio_test_writeback(folio));
1589 folio_start_writeback(folio);
1591 folio_unlock(folio);
1595 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1598 while (offset < f_sectors &&
1599 w->tmp[offset].state < SECTOR_dirty)
1602 if (offset == f_sectors)
1605 while (offset + sectors < f_sectors &&
1606 w->tmp[offset + sectors].state >= SECTOR_dirty) {
1607 reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
1608 dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
1613 sector = folio_sector(folio) + offset;
1616 (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1617 bio_full(&w->io->op.wbio.bio, sectors << 9) ||
1618 w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1619 (BIO_MAX_VECS * PAGE_SIZE) ||
1620 bio_end_sector(&w->io->op.wbio.bio) != sector))
1621 bch2_writepage_do_io(w);
1624 bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1625 nr_replicas_this_write);
1627 atomic_inc(&s->write_count);
1629 BUG_ON(inode != w->io->inode);
1630 BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
1631 sectors << 9, offset << 9));
1633 /* Check for writing past i_size: */
1634 WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1635 round_up(i_size, block_bytes(c)) &&
1636 !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
1637 "writing past i_size: %llu > %llu (unrounded %llu)\n",
1638 bio_end_sector(&w->io->op.wbio.bio) << 9,
1639 round_up(i_size, block_bytes(c)),
1642 w->io->op.res.sectors += reserved_sectors;
1643 w->io->op.i_sectors_delta -= dirty_sectors;
1644 w->io->op.new_i_size = i_size;
1649 if (atomic_dec_and_test(&s->write_count))
1650 folio_end_writeback(folio);
1655 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1657 struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1658 struct bch_writepage_state w =
1659 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1660 struct blk_plug plug;
1663 blk_start_plug(&plug);
1664 ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1666 bch2_writepage_do_io(&w);
1667 blk_finish_plug(&plug);
1669 return bch2_err_class(ret);
1672 /* buffered writes: */
1674 int bch2_write_begin(struct file *file, struct address_space *mapping,
1675 loff_t pos, unsigned len,
1676 struct page **pagep, void **fsdata)
1678 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1679 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1680 struct bch2_folio_reservation *res;
1681 struct folio *folio;
1685 res = kmalloc(sizeof(*res), GFP_KERNEL);
1689 bch2_folio_reservation_init(c, inode, res);
1692 bch2_pagecache_add_get(inode);
1694 folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
1695 FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
1696 mapping_gfp_mask(mapping));
1700 if (folio_test_uptodate(folio))
1703 offset = pos - folio_pos(folio);
1704 len = min_t(size_t, len, folio_end_pos(folio) - pos);
1706 /* If we're writing entire folio, don't need to read it in first: */
1707 if (!offset && len == folio_size(folio))
1710 if (!offset && pos + len >= inode->v.i_size) {
1711 folio_zero_segment(folio, len, folio_size(folio));
1712 flush_dcache_folio(folio);
1716 if (folio_pos(folio) >= inode->v.i_size) {
1717 folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
1718 flush_dcache_folio(folio);
1722 ret = bch2_read_single_folio(folio, mapping);
1726 if (!bch2_folio_create(folio, __GFP_NOFAIL)->uptodate) {
1727 ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
1732 ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
1734 if (!folio_test_uptodate(folio)) {
1736 * If the folio hasn't been read in, we won't know if we
1737 * actually need a reservation - we don't actually need
1738 * to read here, we just need to check if the folio is
1739 * fully backed by uncompressed data:
1747 *pagep = &folio->page;
1750 folio_unlock(folio);
1754 bch2_pagecache_add_put(inode);
1757 return bch2_err_class(ret);
1760 int bch2_write_end(struct file *file, struct address_space *mapping,
1761 loff_t pos, unsigned len, unsigned copied,
1762 struct page *page, void *fsdata)
1764 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1765 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1766 struct bch2_folio_reservation *res = fsdata;
1767 struct folio *folio = page_folio(page);
1768 unsigned offset = pos - folio_pos(folio);
1770 lockdep_assert_held(&inode->v.i_rwsem);
1771 BUG_ON(offset + copied > folio_size(folio));
1773 if (unlikely(copied < len && !folio_test_uptodate(folio))) {
1775 * The folio needs to be read in, but that would destroy
1776 * our partial write - simplest thing is to just force
1777 * userspace to redo the write:
1779 folio_zero_range(folio, 0, folio_size(folio));
1780 flush_dcache_folio(folio);
1784 spin_lock(&inode->v.i_lock);
1785 if (pos + copied > inode->v.i_size)
1786 i_size_write(&inode->v, pos + copied);
1787 spin_unlock(&inode->v.i_lock);
1790 if (!folio_test_uptodate(folio))
1791 folio_mark_uptodate(folio);
1793 bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
1795 inode->ei_last_dirtied = (unsigned long) current;
1798 folio_unlock(folio);
1800 bch2_pagecache_add_put(inode);
1802 bch2_folio_reservation_put(c, inode, res);
1808 static noinline void folios_trunc(folios *folios, struct folio **fi)
1810 while (folios->data + folios->nr > fi) {
1811 struct folio *f = darray_pop(folios);
1818 static int __bch2_buffered_write(struct bch_inode_info *inode,
1819 struct address_space *mapping,
1820 struct iov_iter *iter,
1821 loff_t pos, unsigned len)
1823 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1824 struct bch2_folio_reservation res;
1826 struct folio **fi, *f;
1827 unsigned copied = 0, f_offset;
1828 u64 end = pos + len, f_pos;
1829 loff_t last_folio_pos = inode->v.i_size;
1834 bch2_folio_reservation_init(c, inode, &res);
1835 darray_init(&folios);
1837 ret = filemap_get_contig_folios_d(mapping, pos, end,
1838 FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
1839 mapping_gfp_mask(mapping),
1846 f = darray_first(folios);
1847 if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
1848 ret = bch2_read_single_folio(f, mapping);
1853 f = darray_last(folios);
1854 end = min(end, folio_end_pos(f));
1855 last_folio_pos = folio_pos(f);
1856 if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
1857 if (end >= inode->v.i_size) {
1858 folio_zero_range(f, 0, folio_size(f));
1860 ret = bch2_read_single_folio(f, mapping);
1867 f_offset = pos - folio_pos(darray_first(folios));
1868 darray_for_each(folios, fi) {
1869 struct folio *f = *fi;
1870 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1872 if (!bch2_folio_create(f, __GFP_NOFAIL)->uptodate) {
1873 ret = bch2_folio_set(c, inode_inum(inode), fi,
1874 folios.data + folios.nr - fi);
1880 * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
1881 * supposed to write as much as we have disk space for.
1883 * On failure here we should still write out a partial page if
1884 * we aren't completely out of disk space - we don't do that
1887 ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
1888 if (unlikely(ret)) {
1889 folios_trunc(&folios, fi);
1893 end = min(end, folio_end_pos(darray_last(folios)));
1897 f_pos = folio_end_pos(f);
1901 if (mapping_writably_mapped(mapping))
1902 darray_for_each(folios, fi)
1903 flush_dcache_folio(*fi);
1906 f_offset = pos - folio_pos(darray_first(folios));
1907 darray_for_each(folios, fi) {
1908 struct folio *f = *fi;
1909 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1910 unsigned f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter);
1913 folios_trunc(&folios, fi);
1917 if (!folio_test_uptodate(f) &&
1918 f_copied != folio_size(f) &&
1919 pos + copied + f_copied < inode->v.i_size) {
1920 folio_zero_range(f, 0, folio_size(f));
1921 folios_trunc(&folios, fi);
1925 flush_dcache_folio(f);
1928 if (f_copied != f_len) {
1929 folios_trunc(&folios, fi + 1);
1933 f_pos = folio_end_pos(f);
1942 spin_lock(&inode->v.i_lock);
1943 if (end > inode->v.i_size)
1944 i_size_write(&inode->v, end);
1945 spin_unlock(&inode->v.i_lock);
1948 f_offset = pos - folio_pos(darray_first(folios));
1949 darray_for_each(folios, fi) {
1950 struct folio *f = *fi;
1951 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1953 if (!folio_test_uptodate(f))
1954 folio_mark_uptodate(f);
1956 bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
1958 f_pos = folio_end_pos(f);
1962 inode->ei_last_dirtied = (unsigned long) current;
1964 darray_for_each(folios, fi) {
1970 * If the last folio added to the mapping starts beyond current EOF, we
1971 * performed a short write but left around at least one post-EOF folio.
1972 * Clean up the mapping before we return.
1974 if (last_folio_pos >= inode->v.i_size)
1975 truncate_pagecache(&inode->v, inode->v.i_size);
1977 darray_exit(&folios);
1978 bch2_folio_reservation_put(c, inode, &res);
1980 return copied ?: ret;
1983 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1985 struct file *file = iocb->ki_filp;
1986 struct address_space *mapping = file->f_mapping;
1987 struct bch_inode_info *inode = file_bch_inode(file);
1988 loff_t pos = iocb->ki_pos;
1989 ssize_t written = 0;
1992 bch2_pagecache_add_get(inode);
1995 unsigned offset = pos & (PAGE_SIZE - 1);
1996 unsigned bytes = iov_iter_count(iter);
1999 * Bring in the user page that we will copy from _first_.
2000 * Otherwise there's a nasty deadlock on copying from the
2001 * same page as we're writing to, without it being marked
2004 * Not only is this an optimisation, but it is also required
2005 * to check that the address is actually valid, when atomic
2006 * usercopies are used, below.
2008 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2009 bytes = min_t(unsigned long, iov_iter_count(iter),
2010 PAGE_SIZE - offset);
2012 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2018 if (unlikely(fatal_signal_pending(current))) {
2023 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
2024 if (unlikely(ret < 0))
2029 if (unlikely(ret == 0)) {
2031 * If we were unable to copy any data at all, we must
2032 * fall back to a single segment length write.
2034 * If we didn't fallback here, we could livelock
2035 * because not all segments in the iov can be copied at
2036 * once without a pagefault.
2038 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2039 iov_iter_single_seg_count(iter));
2046 balance_dirty_pages_ratelimited(mapping);
2047 } while (iov_iter_count(iter));
2049 bch2_pagecache_add_put(inode);
2051 return written ? written : ret;
2054 /* O_DIRECT reads */
2056 static void bio_check_or_release(struct bio *bio, bool check_dirty)
2059 bio_check_pages_dirty(bio);
2061 bio_release_pages(bio, false);
2066 static void bch2_dio_read_complete(struct closure *cl)
2068 struct dio_read *dio = container_of(cl, struct dio_read, cl);
2070 dio->req->ki_complete(dio->req, dio->ret);
2071 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2074 static void bch2_direct_IO_read_endio(struct bio *bio)
2076 struct dio_read *dio = bio->bi_private;
2079 dio->ret = blk_status_to_errno(bio->bi_status);
2081 closure_put(&dio->cl);
2084 static void bch2_direct_IO_read_split_endio(struct bio *bio)
2086 struct dio_read *dio = bio->bi_private;
2087 bool should_dirty = dio->should_dirty;
2089 bch2_direct_IO_read_endio(bio);
2090 bio_check_or_release(bio, should_dirty);
2093 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
2095 struct file *file = req->ki_filp;
2096 struct bch_inode_info *inode = file_bch_inode(file);
2097 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2098 struct bch_io_opts opts;
2099 struct dio_read *dio;
2101 loff_t offset = req->ki_pos;
2102 bool sync = is_sync_kiocb(req);
2106 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2108 if ((offset|iter->count) & (block_bytes(c) - 1))
2111 ret = min_t(loff_t, iter->count,
2112 max_t(loff_t, 0, i_size_read(&inode->v) - offset));
2117 shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
2118 iter->count -= shorten;
2120 bio = bio_alloc_bioset(NULL,
2121 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2124 &c->dio_read_bioset);
2126 bio->bi_end_io = bch2_direct_IO_read_endio;
2128 dio = container_of(bio, struct dio_read, rbio.bio);
2129 closure_init(&dio->cl, NULL);
2132 * this is a _really_ horrible hack just to avoid an atomic sub at the
2136 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
2137 atomic_set(&dio->cl.remaining,
2138 CLOSURE_REMAINING_INITIALIZER -
2140 CLOSURE_DESTRUCTOR);
2142 atomic_set(&dio->cl.remaining,
2143 CLOSURE_REMAINING_INITIALIZER + 1);
2149 * This is one of the sketchier things I've encountered: we have to skip
2150 * the dirtying of requests that are internal from the kernel (i.e. from
2151 * loopback), because we'll deadlock on page_lock.
2153 dio->should_dirty = iter_is_iovec(iter);
2156 while (iter->count) {
2157 bio = bio_alloc_bioset(NULL,
2158 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2162 bio->bi_end_io = bch2_direct_IO_read_split_endio;
2164 bio->bi_opf = REQ_OP_READ|REQ_SYNC;
2165 bio->bi_iter.bi_sector = offset >> 9;
2166 bio->bi_private = dio;
2168 ret = bio_iov_iter_get_pages(bio, iter);
2170 /* XXX: fault inject this path */
2171 bio->bi_status = BLK_STS_RESOURCE;
2176 offset += bio->bi_iter.bi_size;
2178 if (dio->should_dirty)
2179 bio_set_pages_dirty(bio);
2182 closure_get(&dio->cl);
2184 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
2187 iter->count += shorten;
2190 closure_sync(&dio->cl);
2191 closure_debug_destroy(&dio->cl);
2193 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2196 return -EIOCBQUEUED;
2200 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2202 struct file *file = iocb->ki_filp;
2203 struct bch_inode_info *inode = file_bch_inode(file);
2204 struct address_space *mapping = file->f_mapping;
2205 size_t count = iov_iter_count(iter);
2209 return 0; /* skip atime */
2211 if (iocb->ki_flags & IOCB_DIRECT) {
2212 struct blk_plug plug;
2214 if (unlikely(mapping->nrpages)) {
2215 ret = filemap_write_and_wait_range(mapping,
2217 iocb->ki_pos + count - 1);
2222 file_accessed(file);
2224 blk_start_plug(&plug);
2225 ret = bch2_direct_IO_read(iocb, iter);
2226 blk_finish_plug(&plug);
2229 iocb->ki_pos += ret;
2231 bch2_pagecache_add_get(inode);
2232 ret = generic_file_read_iter(iocb, iter);
2233 bch2_pagecache_add_put(inode);
2236 return bch2_err_class(ret);
2239 /* O_DIRECT writes */
2241 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2242 u64 offset, u64 size,
2243 unsigned nr_replicas, bool compressed)
2245 struct btree_trans trans;
2246 struct btree_iter iter;
2248 u64 end = offset + size;
2253 bch2_trans_init(&trans, c, 0, 0);
2255 bch2_trans_begin(&trans);
2257 err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2261 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2262 SPOS(inum.inum, offset, snapshot),
2263 BTREE_ITER_SLOTS, k, err) {
2264 if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
2267 if (k.k->p.snapshot != snapshot ||
2268 nr_replicas > bch2_bkey_replicas(c, k) ||
2269 (!compressed && bch2_bkey_sectors_compressed(k))) {
2275 offset = iter.pos.offset;
2276 bch2_trans_iter_exit(&trans, &iter);
2278 if (bch2_err_matches(err, BCH_ERR_transaction_restart))
2280 bch2_trans_exit(&trans);
2282 return err ? false : ret;
2285 static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
2287 struct bch_fs *c = dio->op.c;
2288 struct bch_inode_info *inode = dio->inode;
2289 struct bio *bio = &dio->op.wbio.bio;
2291 return bch2_check_range_allocated(c, inode_inum(inode),
2292 dio->op.pos.offset, bio_sectors(bio),
2293 dio->op.opts.data_replicas,
2294 dio->op.opts.compression != 0);
2297 static void bch2_dio_write_loop_async(struct bch_write_op *);
2298 static __always_inline long bch2_dio_write_done(struct dio_write *dio);
2300 static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
2302 struct iovec *iov = dio->inline_vecs;
2304 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2305 iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
2310 dio->free_iov = true;
2313 memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2314 dio->iter.iov = iov;
2318 static void bch2_dio_write_flush_done(struct closure *cl)
2320 struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
2321 struct bch_fs *c = dio->op.c;
2323 closure_debug_destroy(cl);
2325 dio->op.error = bch2_journal_error(&c->journal);
2327 bch2_dio_write_done(dio);
2330 static noinline void bch2_dio_write_flush(struct dio_write *dio)
2332 struct bch_fs *c = dio->op.c;
2333 struct bch_inode_unpacked inode;
2338 closure_init(&dio->op.cl, NULL);
2340 if (!dio->op.error) {
2341 ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
2343 dio->op.error = ret;
2345 bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl);
2346 bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
2351 closure_sync(&dio->op.cl);
2352 closure_debug_destroy(&dio->op.cl);
2354 continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
2358 static __always_inline long bch2_dio_write_done(struct dio_write *dio)
2360 struct kiocb *req = dio->req;
2361 struct bch_inode_info *inode = dio->inode;
2362 bool sync = dio->sync;
2365 if (unlikely(dio->flush)) {
2366 bch2_dio_write_flush(dio);
2368 return -EIOCBQUEUED;
2371 bch2_pagecache_block_put(inode);
2374 kfree(dio->iter.iov);
2376 ret = dio->op.error ?: ((long) dio->written << 9);
2377 bio_put(&dio->op.wbio.bio);
2379 /* inode->i_dio_count is our ref on inode and thus bch_fs */
2380 inode_dio_end(&inode->v);
2383 ret = bch2_err_class(ret);
2386 req->ki_complete(req, ret);
2392 static __always_inline void bch2_dio_write_end(struct dio_write *dio)
2394 struct bch_fs *c = dio->op.c;
2395 struct kiocb *req = dio->req;
2396 struct bch_inode_info *inode = dio->inode;
2397 struct bio *bio = &dio->op.wbio.bio;
2399 req->ki_pos += (u64) dio->op.written << 9;
2400 dio->written += dio->op.written;
2402 if (dio->extending) {
2403 spin_lock(&inode->v.i_lock);
2404 if (req->ki_pos > inode->v.i_size)
2405 i_size_write(&inode->v, req->ki_pos);
2406 spin_unlock(&inode->v.i_lock);
2409 if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
2410 mutex_lock(&inode->ei_quota_lock);
2411 __i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
2412 __bch2_quota_reservation_put(c, inode, &dio->quota_res);
2413 mutex_unlock(&inode->ei_quota_lock);
2416 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF))) {
2417 struct bvec_iter_all iter;
2418 struct folio_vec fv;
2420 bio_for_each_folio_all(fv, bio, iter)
2421 folio_put(fv.fv_folio);
2424 if (unlikely(dio->op.error))
2425 set_bit(EI_INODE_ERROR, &inode->ei_flags);
2428 static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
2430 struct bch_fs *c = dio->op.c;
2431 struct kiocb *req = dio->req;
2432 struct address_space *mapping = dio->mapping;
2433 struct bch_inode_info *inode = dio->inode;
2434 struct bch_io_opts opts;
2435 struct bio *bio = &dio->op.wbio.bio;
2436 unsigned unaligned, iter_count;
2437 bool sync = dio->sync, dropped_locks;
2440 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2443 iter_count = dio->iter.count;
2445 EBUG_ON(current->faults_disabled_mapping);
2446 current->faults_disabled_mapping = mapping;
2448 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2450 dropped_locks = fdm_dropped_locks();
2452 current->faults_disabled_mapping = NULL;
2455 * If the fault handler returned an error but also signalled
2456 * that it dropped & retook ei_pagecache_lock, we just need to
2457 * re-shoot down the page cache and retry:
2459 if (dropped_locks && ret)
2462 if (unlikely(ret < 0))
2465 if (unlikely(dropped_locks)) {
2466 ret = write_invalidate_inode_pages_range(mapping,
2468 req->ki_pos + iter_count - 1);
2472 if (!bio->bi_iter.bi_size)
2476 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2477 bio->bi_iter.bi_size -= unaligned;
2478 iov_iter_revert(&dio->iter, unaligned);
2480 if (!bio->bi_iter.bi_size) {
2482 * bio_iov_iter_get_pages was only able to get <
2483 * blocksize worth of pages:
2489 bch2_write_op_init(&dio->op, c, opts);
2490 dio->op.end_io = sync
2492 : bch2_dio_write_loop_async;
2493 dio->op.target = dio->op.opts.foreground_target;
2494 dio->op.write_point = writepoint_hashed((unsigned long) current);
2495 dio->op.nr_replicas = dio->op.opts.data_replicas;
2496 dio->op.subvol = inode->ei_subvol;
2497 dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2498 dio->op.devs_need_flush = &inode->ei_devs_need_flush;
2501 dio->op.flags |= BCH_WRITE_SYNC;
2502 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2504 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2505 bio_sectors(bio), true);
2509 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2510 dio->op.opts.data_replicas, 0);
2511 if (unlikely(ret) &&
2512 !bch2_dio_write_check_allocated(dio))
2515 task_io_account_write(bio->bi_iter.bi_size);
2517 if (unlikely(dio->iter.count) &&
2520 bch2_dio_write_copy_iov(dio))
2521 dio->sync = sync = true;
2524 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2527 return -EIOCBQUEUED;
2529 bch2_dio_write_end(dio);
2531 if (likely(!dio->iter.count) || dio->op.error)
2534 bio_reset(bio, NULL, REQ_OP_WRITE);
2537 return bch2_dio_write_done(dio);
2539 dio->op.error = ret;
2541 if (!bio_flagged(bio, BIO_NO_PAGE_REF)) {
2542 struct bvec_iter_all iter;
2543 struct folio_vec fv;
2545 bio_for_each_folio_all(fv, bio, iter)
2546 folio_put(fv.fv_folio);
2549 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2553 static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
2555 struct mm_struct *mm = dio->mm;
2557 bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
2561 bch2_dio_write_loop(dio);
2563 kthread_unuse_mm(mm);
2566 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2568 struct dio_write *dio = container_of(op, struct dio_write, op);
2570 bch2_dio_write_end(dio);
2572 if (likely(!dio->iter.count) || dio->op.error)
2573 bch2_dio_write_done(dio);
2575 bch2_dio_write_continue(dio);
2579 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2581 struct file *file = req->ki_filp;
2582 struct address_space *mapping = file->f_mapping;
2583 struct bch_inode_info *inode = file_bch_inode(file);
2584 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2585 struct dio_write *dio;
2587 bool locked = true, extending;
2591 prefetch((void *) &c->opts + 64);
2592 prefetch(&inode->ei_inode);
2593 prefetch((void *) &inode->ei_inode + 64);
2595 inode_lock(&inode->v);
2597 ret = generic_write_checks(req, iter);
2598 if (unlikely(ret <= 0))
2601 ret = file_remove_privs(file);
2605 ret = file_update_time(file);
2609 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2612 inode_dio_begin(&inode->v);
2613 bch2_pagecache_block_get(inode);
2615 extending = req->ki_pos + iter->count > inode->v.i_size;
2617 inode_unlock(&inode->v);
2621 bio = bio_alloc_bioset(NULL,
2622 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2625 &c->dio_write_bioset);
2626 dio = container_of(bio, struct dio_write, op.wbio.bio);
2628 dio->mapping = mapping;
2630 dio->mm = current->mm;
2632 dio->extending = extending;
2633 dio->sync = is_sync_kiocb(req) || extending;
2634 dio->flush = iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
2635 dio->free_iov = false;
2636 dio->quota_res.sectors = 0;
2641 if (unlikely(mapping->nrpages)) {
2642 ret = write_invalidate_inode_pages_range(mapping,
2644 req->ki_pos + iter->count - 1);
2649 ret = bch2_dio_write_loop(dio);
2652 inode_unlock(&inode->v);
2655 bch2_pagecache_block_put(inode);
2657 inode_dio_end(&inode->v);
2661 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2663 struct file *file = iocb->ki_filp;
2664 struct bch_inode_info *inode = file_bch_inode(file);
2667 if (iocb->ki_flags & IOCB_DIRECT) {
2668 ret = bch2_direct_write(iocb, from);
2672 /* We can write back this queue in page reclaim */
2673 current->backing_dev_info = inode_to_bdi(&inode->v);
2674 inode_lock(&inode->v);
2676 ret = generic_write_checks(iocb, from);
2680 ret = file_remove_privs(file);
2684 ret = file_update_time(file);
2688 ret = bch2_buffered_write(iocb, from);
2689 if (likely(ret > 0))
2690 iocb->ki_pos += ret;
2692 inode_unlock(&inode->v);
2693 current->backing_dev_info = NULL;
2696 ret = generic_write_sync(iocb, ret);
2698 return bch2_err_class(ret);
2704 * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2705 * insert trigger: look up the btree inode instead
2707 static int bch2_flush_inode(struct bch_fs *c,
2708 struct bch_inode_info *inode)
2710 struct bch_inode_unpacked u;
2713 if (c->opts.journal_flush_disabled)
2716 ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u);
2720 return bch2_journal_flush_seq(&c->journal, u.bi_journal_seq) ?:
2721 bch2_inode_flush_nocow_writes(c, inode);
2724 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2726 struct bch_inode_info *inode = file_bch_inode(file);
2727 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2728 int ret, ret2, ret3;
2730 ret = file_write_and_wait_range(file, start, end);
2731 ret2 = sync_inode_metadata(&inode->v, 1);
2732 ret3 = bch2_flush_inode(c, inode);
2734 return bch2_err_class(ret ?: ret2 ?: ret3);
2739 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2743 struct btree_trans trans;
2744 struct btree_iter iter;
2748 bch2_trans_init(&trans, c, 0, 0);
2750 bch2_trans_begin(&trans);
2752 ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2756 for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents, start, end, 0, k, ret)
2757 if (bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k)) {
2762 bch2_trans_iter_exit(&trans, &iter);
2764 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
2767 bch2_trans_exit(&trans);
2771 static int __bch2_truncate_folio(struct bch_inode_info *inode,
2772 pgoff_t index, loff_t start, loff_t end)
2774 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2775 struct address_space *mapping = inode->v.i_mapping;
2776 struct bch_folio *s;
2777 unsigned start_offset = start & (PAGE_SIZE - 1);
2778 unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2780 struct folio *folio;
2781 s64 i_sectors_delta = 0;
2785 folio = filemap_lock_folio(mapping, index);
2788 * XXX: we're doing two index lookups when we end up reading the
2791 ret = range_has_data(c, inode->ei_subvol,
2792 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT)),
2793 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT) + PAGE_SECTORS));
2797 folio = __filemap_get_folio(mapping, index,
2798 FGP_LOCK|FGP_CREAT, GFP_KERNEL);
2799 if (unlikely(!folio)) {
2805 BUG_ON(start >= folio_end_pos(folio));
2806 BUG_ON(end <= folio_pos(folio));
2808 start_offset = max(start, folio_pos(folio)) - folio_pos(folio);
2809 end_offset = min_t(u64, end, folio_end_pos(folio)) - folio_pos(folio);
2811 /* Folio boundary? Nothing to do */
2812 if (start_offset == 0 &&
2813 end_offset == folio_size(folio)) {
2818 s = bch2_folio_create(folio, 0);
2824 if (!folio_test_uptodate(folio)) {
2825 ret = bch2_read_single_folio(folio, mapping);
2831 ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
2836 for (i = round_up(start_offset, block_bytes(c)) >> 9;
2837 i < round_down(end_offset, block_bytes(c)) >> 9;
2839 s->s[i].nr_replicas = 0;
2841 i_sectors_delta -= s->s[i].state == SECTOR_dirty;
2842 folio_sector_set(folio, s, i, SECTOR_unallocated);
2845 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2848 * Caller needs to know whether this folio will be written out by
2849 * writeback - doing an i_size update if necessary - or whether it will
2850 * be responsible for the i_size update.
2852 * Note that we shouldn't ever see a folio beyond EOF, but check and
2853 * warn if so. This has been observed by failure to clean up folios
2854 * after a short write and there's still a chance reclaim will fix
2857 WARN_ON_ONCE(folio_pos(folio) >= inode->v.i_size);
2858 end_pos = folio_end_pos(folio);
2859 if (inode->v.i_size > folio_pos(folio))
2860 end_pos = min_t(u64, inode->v.i_size, end_pos);
2861 ret = s->s[(end_pos - folio_pos(folio) - 1) >> 9].state >= SECTOR_dirty;
2863 folio_zero_segment(folio, start_offset, end_offset);
2866 * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2868 * XXX: because we aren't currently tracking whether the folio has actual
2869 * data in it (vs. just 0s, or only partially written) this wrong. ick.
2871 BUG_ON(bch2_get_folio_disk_reservation(c, inode, folio, false));
2874 * This removes any writeable userspace mappings; we need to force
2875 * .page_mkwrite to be called again before any mmapped writes, to
2876 * redirty the full page:
2878 folio_mkclean(folio);
2879 filemap_dirty_folio(mapping, folio);
2881 folio_unlock(folio);
2887 static int bch2_truncate_folio(struct bch_inode_info *inode, loff_t from)
2889 return __bch2_truncate_folio(inode, from >> PAGE_SHIFT,
2890 from, ANYSINT_MAX(loff_t));
2893 static int bch2_truncate_folios(struct bch_inode_info *inode,
2894 loff_t start, loff_t end)
2896 int ret = __bch2_truncate_folio(inode, start >> PAGE_SHIFT,
2900 start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2901 ret = __bch2_truncate_folio(inode,
2902 (end - 1) >> PAGE_SHIFT,
2907 static int bch2_extend(struct user_namespace *mnt_userns,
2908 struct bch_inode_info *inode,
2909 struct bch_inode_unpacked *inode_u,
2910 struct iattr *iattr)
2912 struct address_space *mapping = inode->v.i_mapping;
2918 * this has to be done _before_ extending i_size:
2920 ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2924 truncate_setsize(&inode->v, iattr->ia_size);
2926 return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2929 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2930 struct bch_inode_unpacked *bi,
2933 bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2937 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2938 struct bch_inode_unpacked *bi, void *p)
2940 u64 *new_i_size = p;
2942 bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2943 bi->bi_size = *new_i_size;
2947 int bch2_truncate(struct user_namespace *mnt_userns,
2948 struct bch_inode_info *inode, struct iattr *iattr)
2950 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2951 struct address_space *mapping = inode->v.i_mapping;
2952 struct bch_inode_unpacked inode_u;
2953 u64 new_i_size = iattr->ia_size;
2954 s64 i_sectors_delta = 0;
2958 * If the truncate call with change the size of the file, the
2959 * cmtimes should be updated. If the size will not change, we
2960 * do not need to update the cmtimes.
2962 if (iattr->ia_size != inode->v.i_size) {
2963 if (!(iattr->ia_valid & ATTR_MTIME))
2964 ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2965 if (!(iattr->ia_valid & ATTR_CTIME))
2966 ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2967 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2970 inode_dio_wait(&inode->v);
2971 bch2_pagecache_block_get(inode);
2973 ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
2978 * check this before next assertion; on filesystem error our normal
2979 * invariants are a bit broken (truncate has to truncate the page cache
2980 * before the inode).
2982 ret = bch2_journal_error(&c->journal);
2986 WARN_ONCE(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2987 inode->v.i_size < inode_u.bi_size,
2988 "truncate spotted in mem i_size < btree i_size: %llu < %llu\n",
2989 (u64) inode->v.i_size, inode_u.bi_size);
2991 if (iattr->ia_size > inode->v.i_size) {
2992 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2996 iattr->ia_valid &= ~ATTR_SIZE;
2998 ret = bch2_truncate_folio(inode, iattr->ia_size);
2999 if (unlikely(ret < 0))
3003 * When extending, we're going to write the new i_size to disk
3004 * immediately so we need to flush anything above the current on disk
3007 * Also, when extending we need to flush the page that i_size currently
3008 * straddles - if it's mapped to userspace, we need to ensure that
3009 * userspace has to redirty it and call .mkwrite -> set_page_dirty
3010 * again to allocate the part of the page that was extended.
3012 if (iattr->ia_size > inode_u.bi_size)
3013 ret = filemap_write_and_wait_range(mapping,
3015 iattr->ia_size - 1);
3016 else if (iattr->ia_size & (PAGE_SIZE - 1))
3017 ret = filemap_write_and_wait_range(mapping,
3018 round_down(iattr->ia_size, PAGE_SIZE),
3019 iattr->ia_size - 1);
3023 mutex_lock(&inode->ei_update_lock);
3024 ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
3026 mutex_unlock(&inode->ei_update_lock);
3031 truncate_setsize(&inode->v, iattr->ia_size);
3033 ret = bch2_fpunch(c, inode_inum(inode),
3034 round_up(iattr->ia_size, block_bytes(c)) >> 9,
3035 U64_MAX, &i_sectors_delta);
3036 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3038 bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
3039 !bch2_journal_error(&c->journal), c,
3040 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
3041 inode->v.i_ino, (u64) inode->v.i_blocks,
3042 inode->ei_inode.bi_sectors);
3046 mutex_lock(&inode->ei_update_lock);
3047 ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
3048 mutex_unlock(&inode->ei_update_lock);
3050 ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
3052 bch2_pagecache_block_put(inode);
3053 return bch2_err_class(ret);
3058 static int inode_update_times_fn(struct bch_inode_info *inode,
3059 struct bch_inode_unpacked *bi, void *p)
3061 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3063 bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
3067 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
3069 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3070 u64 end = offset + len;
3071 u64 block_start = round_up(offset, block_bytes(c));
3072 u64 block_end = round_down(end, block_bytes(c));
3073 bool truncated_last_page;
3076 ret = bch2_truncate_folios(inode, offset, end);
3077 if (unlikely(ret < 0))
3080 truncated_last_page = ret;
3082 truncate_pagecache_range(&inode->v, offset, end - 1);
3084 if (block_start < block_end) {
3085 s64 i_sectors_delta = 0;
3087 ret = bch2_fpunch(c, inode_inum(inode),
3088 block_start >> 9, block_end >> 9,
3090 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3093 mutex_lock(&inode->ei_update_lock);
3094 if (end >= inode->v.i_size && !truncated_last_page) {
3095 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
3096 ATTR_MTIME|ATTR_CTIME);
3098 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3099 ATTR_MTIME|ATTR_CTIME);
3101 mutex_unlock(&inode->ei_update_lock);
3106 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
3107 loff_t offset, loff_t len,
3110 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3111 struct address_space *mapping = inode->v.i_mapping;
3112 struct bkey_buf copy;
3113 struct btree_trans trans;
3114 struct btree_iter src, dst, del;
3115 loff_t shift, new_size;
3119 if ((offset | len) & (block_bytes(c) - 1))
3123 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
3126 if (offset >= inode->v.i_size)
3129 src_start = U64_MAX;
3132 if (offset + len >= inode->v.i_size)
3135 src_start = offset + len;
3139 new_size = inode->v.i_size + shift;
3141 ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
3146 i_size_write(&inode->v, new_size);
3147 mutex_lock(&inode->ei_update_lock);
3148 ret = bch2_write_inode_size(c, inode, new_size,
3149 ATTR_MTIME|ATTR_CTIME);
3150 mutex_unlock(&inode->ei_update_lock);
3152 s64 i_sectors_delta = 0;
3154 ret = bch2_fpunch(c, inode_inum(inode),
3155 offset >> 9, (offset + len) >> 9,
3157 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3163 bch2_bkey_buf_init(©);
3164 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
3165 bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
3166 POS(inode->v.i_ino, src_start >> 9),
3168 bch2_trans_copy_iter(&dst, &src);
3169 bch2_trans_copy_iter(&del, &src);
3172 bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
3173 struct disk_reservation disk_res =
3174 bch2_disk_reservation_init(c, 0);
3175 struct bkey_i delete;
3177 struct bpos next_pos;
3178 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
3179 struct bpos atomic_end;
3180 unsigned trigger_flags = 0;
3183 bch2_trans_begin(&trans);
3185 ret = bch2_subvolume_get_snapshot(&trans,
3186 inode->ei_subvol, &snapshot);
3190 bch2_btree_iter_set_snapshot(&src, snapshot);
3191 bch2_btree_iter_set_snapshot(&dst, snapshot);
3192 bch2_btree_iter_set_snapshot(&del, snapshot);
3194 bch2_trans_begin(&trans);
3197 ? bch2_btree_iter_peek_prev(&src)
3198 : bch2_btree_iter_peek_upto(&src, POS(inode->v.i_ino, U64_MAX));
3199 if ((ret = bkey_err(k)))
3202 if (!k.k || k.k->p.inode != inode->v.i_ino)
3206 bkey_le(k.k->p, POS(inode->v.i_ino, offset >> 9)))
3209 bch2_bkey_buf_reassemble(©, c, k);
3212 bkey_lt(bkey_start_pos(k.k), move_pos))
3213 bch2_cut_front(move_pos, copy.k);
3215 copy.k->k.p.offset += shift >> 9;
3216 bch2_btree_iter_set_pos(&dst, bkey_start_pos(©.k->k));
3218 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
3222 if (!bkey_eq(atomic_end, copy.k->k.p)) {
3224 move_pos = atomic_end;
3225 move_pos.offset -= shift >> 9;
3228 bch2_cut_back(atomic_end, copy.k);
3232 bkey_init(&delete.k);
3233 delete.k.p = copy.k->k.p;
3234 delete.k.size = copy.k->k.size;
3235 delete.k.p.offset -= shift >> 9;
3236 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
3238 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
3240 if (copy.k->k.size != k.k->size) {
3241 /* We might end up splitting compressed extents: */
3243 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
3245 ret = bch2_disk_reservation_get(c, &disk_res,
3246 copy.k->k.size, nr_ptrs,
3247 BCH_DISK_RESERVATION_NOFAIL);
3251 ret = bch2_btree_iter_traverse(&del) ?:
3252 bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
3253 bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
3254 bch2_trans_commit(&trans, &disk_res, NULL,
3255 BTREE_INSERT_NOFAIL);
3256 bch2_disk_reservation_put(c, &disk_res);
3259 bch2_btree_iter_set_pos(&src, next_pos);
3261 bch2_trans_iter_exit(&trans, &del);
3262 bch2_trans_iter_exit(&trans, &dst);
3263 bch2_trans_iter_exit(&trans, &src);
3264 bch2_trans_exit(&trans);
3265 bch2_bkey_buf_exit(©, c);
3270 mutex_lock(&inode->ei_update_lock);
3272 i_size_write(&inode->v, new_size);
3273 ret = bch2_write_inode_size(c, inode, new_size,
3274 ATTR_MTIME|ATTR_CTIME);
3276 /* We need an inode update to update bi_journal_seq for fsync: */
3277 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3278 ATTR_MTIME|ATTR_CTIME);
3280 mutex_unlock(&inode->ei_update_lock);
3284 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
3285 u64 start_sector, u64 end_sector)
3287 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3288 struct btree_trans trans;
3289 struct btree_iter iter;
3290 struct bpos end_pos = POS(inode->v.i_ino, end_sector);
3291 struct bch_io_opts opts;
3294 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
3295 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
3297 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3298 POS(inode->v.i_ino, start_sector),
3299 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
3301 while (!ret && bkey_lt(iter.pos, end_pos)) {
3302 s64 i_sectors_delta = 0;
3303 struct quota_res quota_res = { 0 };
3308 bch2_trans_begin(&trans);
3310 ret = bch2_subvolume_get_snapshot(&trans,
3311 inode->ei_subvol, &snapshot);
3315 bch2_btree_iter_set_snapshot(&iter, snapshot);
3317 k = bch2_btree_iter_peek_slot(&iter);
3318 if ((ret = bkey_err(k)))
3321 /* already reserved */
3322 if (bkey_extent_is_reservation(k) &&
3323 bch2_bkey_nr_ptrs_fully_allocated(k) >= opts.data_replicas) {
3324 bch2_btree_iter_advance(&iter);
3328 if (bkey_extent_is_data(k.k) &&
3329 !(mode & FALLOC_FL_ZERO_RANGE)) {
3330 bch2_btree_iter_advance(&iter);
3335 * XXX: for nocow mode, we should promote shared extents to
3339 sectors = bpos_min(k.k->p, end_pos).offset - iter.pos.offset;
3341 if (!bkey_extent_is_allocation(k.k)) {
3342 ret = bch2_quota_reservation_add(c, inode,
3349 ret = bch2_extent_fallocate(&trans, inode_inum(inode), &iter,
3350 sectors, opts, &i_sectors_delta,
3351 writepoint_hashed((unsigned long) current));
3355 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3357 bch2_quota_reservation_put(c, inode, "a_res);
3358 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3362 bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3363 mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3365 if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) {
3366 struct quota_res quota_res = { 0 };
3367 s64 i_sectors_delta = 0;
3369 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3370 end_sector, &i_sectors_delta);
3371 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3372 bch2_quota_reservation_put(c, inode, "a_res);
3375 bch2_trans_iter_exit(&trans, &iter);
3376 bch2_trans_exit(&trans);
3380 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3381 loff_t offset, loff_t len)
3383 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3384 u64 end = offset + len;
3385 u64 block_start = round_down(offset, block_bytes(c));
3386 u64 block_end = round_up(end, block_bytes(c));
3387 bool truncated_last_page = false;
3390 if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3391 ret = inode_newsize_ok(&inode->v, end);
3396 if (mode & FALLOC_FL_ZERO_RANGE) {
3397 ret = bch2_truncate_folios(inode, offset, end);
3398 if (unlikely(ret < 0))
3401 truncated_last_page = ret;
3403 truncate_pagecache_range(&inode->v, offset, end - 1);
3405 block_start = round_up(offset, block_bytes(c));
3406 block_end = round_down(end, block_bytes(c));
3409 ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3412 * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3413 * so that the VFS cache i_size is consistent with the btree i_size:
3416 !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)))
3419 if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3420 end = inode->v.i_size;
3422 if (end >= inode->v.i_size &&
3423 (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3424 !(mode & FALLOC_FL_KEEP_SIZE))) {
3425 spin_lock(&inode->v.i_lock);
3426 i_size_write(&inode->v, end);
3427 spin_unlock(&inode->v.i_lock);
3429 mutex_lock(&inode->ei_update_lock);
3430 ret2 = bch2_write_inode_size(c, inode, end, 0);
3431 mutex_unlock(&inode->ei_update_lock);
3437 long bch2_fallocate_dispatch(struct file *file, int mode,
3438 loff_t offset, loff_t len)
3440 struct bch_inode_info *inode = file_bch_inode(file);
3441 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3444 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fallocate))
3447 inode_lock(&inode->v);
3448 inode_dio_wait(&inode->v);
3449 bch2_pagecache_block_get(inode);
3451 ret = file_modified(file);
3455 if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3456 ret = bchfs_fallocate(inode, mode, offset, len);
3457 else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3458 ret = bchfs_fpunch(inode, offset, len);
3459 else if (mode == FALLOC_FL_INSERT_RANGE)
3460 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3461 else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3462 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3466 bch2_pagecache_block_put(inode);
3467 inode_unlock(&inode->v);
3468 bch2_write_ref_put(c, BCH_WRITE_REF_fallocate);
3470 return bch2_err_class(ret);
3474 * Take a quota reservation for unallocated blocks in a given file range
3475 * Does not check pagecache
3477 static int quota_reserve_range(struct bch_inode_info *inode,
3478 struct quota_res *res,
3481 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3482 struct btree_trans trans;
3483 struct btree_iter iter;
3486 u64 sectors = end - start;
3490 bch2_trans_init(&trans, c, 0, 0);
3492 bch2_trans_begin(&trans);
3494 ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot);
3498 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3499 SPOS(inode->v.i_ino, pos, snapshot), 0);
3501 while (!(ret = btree_trans_too_many_iters(&trans)) &&
3502 (k = bch2_btree_iter_peek_upto(&iter, POS(inode->v.i_ino, end - 1))).k &&
3503 !(ret = bkey_err(k))) {
3504 if (bkey_extent_is_allocation(k.k)) {
3505 u64 s = min(end, k.k->p.offset) -
3506 max(start, bkey_start_offset(k.k));
3507 BUG_ON(s > sectors);
3510 bch2_btree_iter_advance(&iter);
3512 pos = iter.pos.offset;
3513 bch2_trans_iter_exit(&trans, &iter);
3515 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3518 bch2_trans_exit(&trans);
3523 return bch2_quota_reservation_add(c, inode, res, sectors, true);
3526 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3527 struct file *file_dst, loff_t pos_dst,
3528 loff_t len, unsigned remap_flags)
3530 struct bch_inode_info *src = file_bch_inode(file_src);
3531 struct bch_inode_info *dst = file_bch_inode(file_dst);
3532 struct bch_fs *c = src->v.i_sb->s_fs_info;
3533 struct quota_res quota_res = { 0 };
3534 s64 i_sectors_delta = 0;
3538 if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3541 if (remap_flags & REMAP_FILE_DEDUP)
3544 if ((pos_src & (block_bytes(c) - 1)) ||
3545 (pos_dst & (block_bytes(c) - 1)))
3549 abs(pos_src - pos_dst) < len)
3552 bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3554 inode_dio_wait(&src->v);
3555 inode_dio_wait(&dst->v);
3557 ret = generic_remap_file_range_prep(file_src, pos_src,
3560 if (ret < 0 || len == 0)
3563 aligned_len = round_up((u64) len, block_bytes(c));
3565 ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3566 pos_dst, pos_dst + len - 1);
3570 ret = quota_reserve_range(dst, "a_res, pos_dst >> 9,
3571 (pos_dst + aligned_len) >> 9);
3575 file_update_time(file_dst);
3577 mark_pagecache_unallocated(src, pos_src >> 9,
3578 (pos_src + aligned_len) >> 9);
3580 ret = bch2_remap_range(c,
3581 inode_inum(dst), pos_dst >> 9,
3582 inode_inum(src), pos_src >> 9,
3584 pos_dst + len, &i_sectors_delta);
3589 * due to alignment, we might have remapped slightly more than requsted
3591 ret = min((u64) ret << 9, (u64) len);
3593 i_sectors_acct(c, dst, "a_res, i_sectors_delta);
3595 spin_lock(&dst->v.i_lock);
3596 if (pos_dst + ret > dst->v.i_size)
3597 i_size_write(&dst->v, pos_dst + ret);
3598 spin_unlock(&dst->v.i_lock);
3600 if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3601 IS_SYNC(file_inode(file_dst)))
3602 ret = bch2_flush_inode(c, dst);
3604 bch2_quota_reservation_put(c, dst, "a_res);
3605 bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3607 return bch2_err_class(ret);
3612 static int folio_data_offset(struct folio *folio, unsigned offset)
3614 struct bch_folio *s = bch2_folio(folio);
3615 unsigned i, sectors = folio_sectors(folio);
3618 for (i = offset >> 9; i < sectors; i++)
3619 if (s->s[i].state >= SECTOR_dirty)
3625 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3626 loff_t start_offset,
3629 struct folio_batch fbatch;
3630 pgoff_t start_index = start_offset >> PAGE_SHIFT;
3631 pgoff_t end_index = end_offset >> PAGE_SHIFT;
3632 pgoff_t index = start_index;
3637 folio_batch_init(&fbatch);
3639 while (filemap_get_folios(vinode->i_mapping,
3640 &index, end_index, &fbatch)) {
3641 for (i = 0; i < folio_batch_count(&fbatch); i++) {
3642 struct folio *folio = fbatch.folios[i];
3645 offset = folio_data_offset(folio,
3646 max(folio_pos(folio), start_offset) -
3649 ret = clamp(folio_pos(folio) + offset,
3650 start_offset, end_offset);
3651 folio_unlock(folio);
3652 folio_batch_release(&fbatch);
3655 folio_unlock(folio);
3657 folio_batch_release(&fbatch);
3664 static loff_t bch2_seek_data(struct file *file, u64 offset)
3666 struct bch_inode_info *inode = file_bch_inode(file);
3667 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3668 struct btree_trans trans;
3669 struct btree_iter iter;
3671 subvol_inum inum = inode_inum(inode);
3672 u64 isize, next_data = MAX_LFS_FILESIZE;
3676 isize = i_size_read(&inode->v);
3677 if (offset >= isize)
3680 bch2_trans_init(&trans, c, 0, 0);
3682 bch2_trans_begin(&trans);
3684 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3688 for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents,
3689 SPOS(inode->v.i_ino, offset >> 9, snapshot),
3690 POS(inode->v.i_ino, U64_MAX),
3692 if (bkey_extent_is_data(k.k)) {
3693 next_data = max(offset, bkey_start_offset(k.k) << 9);
3695 } else if (k.k->p.offset >> 9 > isize)
3698 bch2_trans_iter_exit(&trans, &iter);
3700 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3703 bch2_trans_exit(&trans);
3707 if (next_data > offset)
3708 next_data = bch2_seek_pagecache_data(&inode->v,
3711 if (next_data >= isize)
3714 return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3717 static bool folio_hole_offset(struct address_space *mapping, loff_t *offset)
3719 struct folio *folio;
3720 struct bch_folio *s;
3721 unsigned i, sectors, f_offset;
3724 folio = filemap_lock_folio(mapping, *offset >> PAGE_SHIFT);
3728 s = bch2_folio(folio);
3732 sectors = folio_sectors(folio);
3733 f_offset = *offset - folio_pos(folio);
3735 for (i = f_offset >> 9; i < sectors; i++)
3736 if (s->s[i].state < SECTOR_dirty) {
3737 *offset = max(*offset, folio_pos(folio) + (i << 9));
3741 *offset = folio_end_pos(folio);
3744 folio_unlock(folio);
3748 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3749 loff_t start_offset,
3752 struct address_space *mapping = vinode->i_mapping;
3753 loff_t offset = start_offset;
3755 while (offset < end_offset &&
3756 !folio_hole_offset(mapping, &offset))
3759 return min(offset, end_offset);
3762 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3764 struct bch_inode_info *inode = file_bch_inode(file);
3765 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3766 struct btree_trans trans;
3767 struct btree_iter iter;
3769 subvol_inum inum = inode_inum(inode);
3770 u64 isize, next_hole = MAX_LFS_FILESIZE;
3774 isize = i_size_read(&inode->v);
3775 if (offset >= isize)
3778 bch2_trans_init(&trans, c, 0, 0);
3780 bch2_trans_begin(&trans);
3782 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3786 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3787 SPOS(inode->v.i_ino, offset >> 9, snapshot),
3788 BTREE_ITER_SLOTS, k, ret) {
3789 if (k.k->p.inode != inode->v.i_ino) {
3790 next_hole = bch2_seek_pagecache_hole(&inode->v,
3791 offset, MAX_LFS_FILESIZE);
3793 } else if (!bkey_extent_is_data(k.k)) {
3794 next_hole = bch2_seek_pagecache_hole(&inode->v,
3795 max(offset, bkey_start_offset(k.k) << 9),
3796 k.k->p.offset << 9);
3798 if (next_hole < k.k->p.offset << 9)
3801 offset = max(offset, bkey_start_offset(k.k) << 9);
3804 bch2_trans_iter_exit(&trans, &iter);
3806 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3809 bch2_trans_exit(&trans);
3813 if (next_hole > isize)
3816 return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3819 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3827 ret = generic_file_llseek(file, offset, whence);
3830 ret = bch2_seek_data(file, offset);
3833 ret = bch2_seek_hole(file, offset);
3840 return bch2_err_class(ret);
3843 void bch2_fs_fsio_exit(struct bch_fs *c)
3845 bioset_exit(&c->nocow_flush_bioset);
3846 bioset_exit(&c->dio_write_bioset);
3847 bioset_exit(&c->dio_read_bioset);
3848 bioset_exit(&c->writepage_bioset);
3851 int bch2_fs_fsio_init(struct bch_fs *c)
3855 pr_verbose_init(c->opts, "");
3857 if (bioset_init(&c->writepage_bioset,
3858 4, offsetof(struct bch_writepage_io, op.wbio.bio),
3860 return -BCH_ERR_ENOMEM_writepage_bioset_init;
3862 if (bioset_init(&c->dio_read_bioset,
3863 4, offsetof(struct dio_read, rbio.bio),
3865 return -BCH_ERR_ENOMEM_dio_read_bioset_init;
3867 if (bioset_init(&c->dio_write_bioset,
3868 4, offsetof(struct dio_write, op.wbio.bio),
3870 return -BCH_ERR_ENOMEM_dio_write_bioset_init;
3872 if (bioset_init(&c->nocow_flush_bioset,
3873 1, offsetof(struct nocow_flush, bio), 0))
3874 return -BCH_ERR_ENOMEM_nocow_flush_bioset_init;
3876 pr_verbose_init(c->opts, "ret %i", ret);
3880 #endif /* NO_BCACHEFS_FS */