1 // SPDX-License-Identifier: GPL-2.0
5 #include "alloc_foreground.h"
7 #include "btree_update.h"
12 #include "extent_update.h"
24 #include <linux/aio.h>
25 #include <linux/backing-dev.h>
26 #include <linux/falloc.h>
27 #include <linux/migrate.h>
28 #include <linux/mmu_context.h>
29 #include <linux/pagevec.h>
30 #include <linux/rmap.h>
31 #include <linux/sched/signal.h>
32 #include <linux/task_io_accounting_ops.h>
33 #include <linux/uio.h>
34 #include <linux/writeback.h>
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 if (test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags))
296 mutex_lock(&inode->ei_quota_lock);
297 ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
298 check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
300 inode->ei_quota_reserved += sectors;
301 res->sectors += sectors;
303 mutex_unlock(&inode->ei_quota_lock);
310 static void __bch2_quota_reservation_put(struct bch_fs *c,
311 struct bch_inode_info *inode,
312 struct quota_res *res) {}
314 static void bch2_quota_reservation_put(struct bch_fs *c,
315 struct bch_inode_info *inode,
316 struct quota_res *res) {}
318 static int bch2_quota_reservation_add(struct bch_fs *c,
319 struct bch_inode_info *inode,
320 struct quota_res *res,
329 /* i_size updates: */
331 struct inode_new_size {
337 static int inode_set_size(struct bch_inode_info *inode,
338 struct bch_inode_unpacked *bi,
341 struct inode_new_size *s = p;
343 bi->bi_size = s->new_size;
344 if (s->fields & ATTR_ATIME)
345 bi->bi_atime = s->now;
346 if (s->fields & ATTR_MTIME)
347 bi->bi_mtime = s->now;
348 if (s->fields & ATTR_CTIME)
349 bi->bi_ctime = s->now;
354 int __must_check bch2_write_inode_size(struct bch_fs *c,
355 struct bch_inode_info *inode,
356 loff_t new_size, unsigned fields)
358 struct inode_new_size s = {
359 .new_size = new_size,
360 .now = bch2_current_time(c),
364 return bch2_write_inode(c, inode, inode_set_size, &s, fields);
367 static void __i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
368 struct quota_res *quota_res, s64 sectors)
370 bch2_fs_inconsistent_on((s64) inode->v.i_blocks + sectors < 0, c,
371 "inode %lu i_blocks underflow: %llu + %lli < 0 (ondisk %lli)",
372 inode->v.i_ino, (u64) inode->v.i_blocks, sectors,
373 inode->ei_inode.bi_sectors);
374 inode->v.i_blocks += sectors;
376 #ifdef CONFIG_BCACHEFS_QUOTA
378 !test_bit(EI_INODE_SNAPSHOT, &inode->ei_flags) &&
380 BUG_ON(sectors > quota_res->sectors);
381 BUG_ON(sectors > inode->ei_quota_reserved);
383 quota_res->sectors -= sectors;
384 inode->ei_quota_reserved -= sectors;
386 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
391 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
392 struct quota_res *quota_res, s64 sectors)
395 mutex_lock(&inode->ei_quota_lock);
396 __i_sectors_acct(c, inode, quota_res, sectors);
397 mutex_unlock(&inode->ei_quota_lock);
403 /* stored in page->private: */
405 #define BCH_FOLIO_SECTOR_STATE() \
412 enum bch_folio_sector_state {
413 #define x(n) SECTOR_##n,
414 BCH_FOLIO_SECTOR_STATE()
418 const char * const bch2_folio_sector_states[] = {
420 BCH_FOLIO_SECTOR_STATE()
425 static inline enum bch_folio_sector_state
426 folio_sector_dirty(enum bch_folio_sector_state state)
429 case SECTOR_unallocated:
431 case SECTOR_reserved:
432 return SECTOR_dirty_reserved;
438 static inline enum bch_folio_sector_state
439 folio_sector_undirty(enum bch_folio_sector_state state)
443 return SECTOR_unallocated;
444 case SECTOR_dirty_reserved:
445 return SECTOR_reserved;
451 static inline enum bch_folio_sector_state
452 folio_sector_reserve(enum bch_folio_sector_state state)
455 case SECTOR_unallocated:
456 return SECTOR_reserved;
458 return SECTOR_dirty_reserved;
464 struct bch_folio_sector {
465 /* Uncompressed, fully allocated replicas (or on disk reservation): */
466 unsigned nr_replicas:4;
468 /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
469 unsigned replicas_reserved:4;
472 enum bch_folio_sector_state state:8;
477 atomic_t write_count;
479 * Is the sector state up to date with the btree?
480 * (Not the data itself)
483 struct bch_folio_sector s[];
486 static inline void folio_sector_set(struct folio *folio,
488 unsigned i, unsigned n)
493 /* file offset (to folio offset) to bch_folio_sector index */
494 static inline int folio_pos_to_s(struct folio *folio, loff_t pos)
496 u64 f_offset = pos - folio_pos(folio);
497 BUG_ON(pos < folio_pos(folio) || pos >= folio_end_pos(folio));
498 return f_offset >> SECTOR_SHIFT;
501 static inline struct bch_folio *__bch2_folio(struct folio *folio)
503 return folio_has_private(folio)
504 ? (struct bch_folio *) folio_get_private(folio)
508 static inline struct bch_folio *bch2_folio(struct folio *folio)
510 EBUG_ON(!folio_test_locked(folio));
512 return __bch2_folio(folio);
515 /* for newly allocated folios: */
516 static void __bch2_folio_release(struct folio *folio)
518 kfree(folio_detach_private(folio));
521 static void bch2_folio_release(struct folio *folio)
523 EBUG_ON(!folio_test_locked(folio));
524 __bch2_folio_release(folio);
527 /* for newly allocated folios: */
528 static struct bch_folio *__bch2_folio_create(struct folio *folio, gfp_t gfp)
532 s = kzalloc(sizeof(*s) +
533 sizeof(struct bch_folio_sector) *
534 folio_sectors(folio), gfp);
538 spin_lock_init(&s->lock);
539 folio_attach_private(folio, s);
543 static struct bch_folio *bch2_folio_create(struct folio *folio, gfp_t gfp)
545 return bch2_folio(folio) ?: __bch2_folio_create(folio, gfp);
548 static unsigned bkey_to_sector_state(struct bkey_s_c k)
550 if (bkey_extent_is_reservation(k))
551 return SECTOR_reserved;
552 if (bkey_extent_is_allocation(k.k))
553 return SECTOR_allocated;
554 return SECTOR_unallocated;
557 static void __bch2_folio_set(struct folio *folio,
558 unsigned pg_offset, unsigned pg_len,
559 unsigned nr_ptrs, unsigned state)
561 struct bch_folio *s = bch2_folio(folio);
562 unsigned i, sectors = folio_sectors(folio);
564 BUG_ON(pg_offset >= sectors);
565 BUG_ON(pg_offset + pg_len > sectors);
569 for (i = pg_offset; i < pg_offset + pg_len; i++) {
570 s->s[i].nr_replicas = nr_ptrs;
571 folio_sector_set(folio, s, i, state);
577 spin_unlock(&s->lock);
581 * Initialize bch_folio state (allocated/unallocated, nr_replicas) from the
584 static int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
585 struct folio **folios, unsigned nr_folios)
587 struct btree_trans trans;
588 struct btree_iter iter;
591 u64 offset = folio_sector(folios[0]);
594 bool need_set = false;
597 for (folio_idx = 0; folio_idx < nr_folios; folio_idx++) {
598 s = bch2_folio_create(folios[folio_idx], GFP_KERNEL);
602 need_set |= !s->uptodate;
609 bch2_trans_init(&trans, c, 0, 0);
611 bch2_trans_begin(&trans);
613 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
617 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
618 SPOS(inum.inum, offset, snapshot),
619 BTREE_ITER_SLOTS, k, ret) {
620 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
621 unsigned state = bkey_to_sector_state(k);
623 while (folio_idx < nr_folios) {
624 struct folio *folio = folios[folio_idx];
625 u64 folio_start = folio_sector(folio);
626 u64 folio_end = folio_end_sector(folio);
627 unsigned folio_offset = max(bkey_start_offset(k.k), folio_start) - folio_start;
628 unsigned folio_len = min(k.k->p.offset, folio_end) - folio_offset - folio_start;
630 BUG_ON(k.k->p.offset < folio_start);
631 BUG_ON(bkey_start_offset(k.k) > folio_end);
633 if (!bch2_folio(folio)->uptodate)
634 __bch2_folio_set(folio, folio_offset, folio_len, nr_ptrs, state);
636 if (k.k->p.offset < folio_end)
641 if (folio_idx == nr_folios)
645 offset = iter.pos.offset;
646 bch2_trans_iter_exit(&trans, &iter);
648 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
650 bch2_trans_exit(&trans);
655 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
657 struct bvec_iter iter;
659 unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
660 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
661 unsigned state = bkey_to_sector_state(k);
663 bio_for_each_folio(fv, bio, iter)
664 __bch2_folio_set(fv.fv_folio,
670 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
673 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
674 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
675 struct folio_batch fbatch;
681 folio_batch_init(&fbatch);
683 while (filemap_get_folios(inode->v.i_mapping,
684 &index, end_index, &fbatch)) {
685 for (i = 0; i < folio_batch_count(&fbatch); i++) {
686 struct folio *folio = fbatch.folios[i];
687 u64 folio_start = folio_sector(folio);
688 u64 folio_end = folio_end_sector(folio);
689 unsigned folio_offset = max(start, folio_start) - folio_start;
690 unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
693 BUG_ON(end <= folio_start);
696 s = bch2_folio(folio);
700 for (j = folio_offset; j < folio_offset + folio_len; j++)
701 s->s[j].nr_replicas = 0;
702 spin_unlock(&s->lock);
707 folio_batch_release(&fbatch);
712 static void mark_pagecache_reserved(struct bch_inode_info *inode,
715 struct bch_fs *c = inode->v.i_sb->s_fs_info;
716 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
717 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
718 struct folio_batch fbatch;
719 s64 i_sectors_delta = 0;
725 folio_batch_init(&fbatch);
727 while (filemap_get_folios(inode->v.i_mapping,
728 &index, end_index, &fbatch)) {
729 for (i = 0; i < folio_batch_count(&fbatch); i++) {
730 struct folio *folio = fbatch.folios[i];
731 u64 folio_start = folio_sector(folio);
732 u64 folio_end = folio_end_sector(folio);
733 unsigned folio_offset = max(start, folio_start) - folio_start;
734 unsigned folio_len = min(end, folio_end) - folio_offset - folio_start;
737 BUG_ON(end <= folio_start);
740 s = bch2_folio(folio);
744 for (j = folio_offset; j < folio_offset + folio_len; j++) {
745 i_sectors_delta -= s->s[j].state == SECTOR_dirty;
746 folio_sector_set(folio, s, j, folio_sector_reserve(s->s[j].state));
748 spin_unlock(&s->lock);
753 folio_batch_release(&fbatch);
757 i_sectors_acct(c, inode, NULL, i_sectors_delta);
760 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
762 /* XXX: this should not be open coded */
763 return inode->ei_inode.bi_data_replicas
764 ? inode->ei_inode.bi_data_replicas - 1
765 : c->opts.data_replicas;
768 static inline unsigned sectors_to_reserve(struct bch_folio_sector *s,
769 unsigned nr_replicas)
771 return max(0, (int) nr_replicas -
773 s->replicas_reserved);
776 static int bch2_get_folio_disk_reservation(struct bch_fs *c,
777 struct bch_inode_info *inode,
778 struct folio *folio, bool check_enospc)
780 struct bch_folio *s = bch2_folio_create(folio, 0);
781 unsigned nr_replicas = inode_nr_replicas(c, inode);
782 struct disk_reservation disk_res = { 0 };
783 unsigned i, sectors = folio_sectors(folio), disk_res_sectors = 0;
789 for (i = 0; i < sectors; i++)
790 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
792 if (!disk_res_sectors)
795 ret = bch2_disk_reservation_get(c, &disk_res,
798 ? BCH_DISK_RESERVATION_NOFAIL
803 for (i = 0; i < sectors; i++)
804 s->s[i].replicas_reserved +=
805 sectors_to_reserve(&s->s[i], nr_replicas);
810 struct bch2_folio_reservation {
811 struct disk_reservation disk;
812 struct quota_res quota;
815 static void bch2_folio_reservation_init(struct bch_fs *c,
816 struct bch_inode_info *inode,
817 struct bch2_folio_reservation *res)
819 memset(res, 0, sizeof(*res));
821 res->disk.nr_replicas = inode_nr_replicas(c, inode);
824 static void bch2_folio_reservation_put(struct bch_fs *c,
825 struct bch_inode_info *inode,
826 struct bch2_folio_reservation *res)
828 bch2_disk_reservation_put(c, &res->disk);
829 bch2_quota_reservation_put(c, inode, &res->quota);
832 static int bch2_folio_reservation_get(struct bch_fs *c,
833 struct bch_inode_info *inode,
835 struct bch2_folio_reservation *res,
836 unsigned offset, unsigned len)
838 struct bch_folio *s = bch2_folio_create(folio, 0);
839 unsigned i, disk_sectors = 0, quota_sectors = 0;
845 BUG_ON(!s->uptodate);
847 for (i = round_down(offset, block_bytes(c)) >> 9;
848 i < round_up(offset + len, block_bytes(c)) >> 9;
850 disk_sectors += sectors_to_reserve(&s->s[i],
851 res->disk.nr_replicas);
852 quota_sectors += s->s[i].state == SECTOR_unallocated;
856 ret = bch2_disk_reservation_add(c, &res->disk, disk_sectors, 0);
862 ret = bch2_quota_reservation_add(c, inode, &res->quota,
863 quota_sectors, true);
865 struct disk_reservation tmp = {
866 .sectors = disk_sectors
869 bch2_disk_reservation_put(c, &tmp);
870 res->disk.sectors -= disk_sectors;
878 static void bch2_clear_folio_bits(struct folio *folio)
880 struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
881 struct bch_fs *c = inode->v.i_sb->s_fs_info;
882 struct bch_folio *s = bch2_folio(folio);
883 struct disk_reservation disk_res = { 0 };
884 int i, sectors = folio_sectors(folio), dirty_sectors = 0;
889 EBUG_ON(!folio_test_locked(folio));
890 EBUG_ON(folio_test_writeback(folio));
892 for (i = 0; i < sectors; i++) {
893 disk_res.sectors += s->s[i].replicas_reserved;
894 s->s[i].replicas_reserved = 0;
896 dirty_sectors -= s->s[i].state == SECTOR_dirty;
897 folio_sector_set(folio, s, i, folio_sector_undirty(s->s[i].state));
900 bch2_disk_reservation_put(c, &disk_res);
902 i_sectors_acct(c, inode, NULL, dirty_sectors);
904 bch2_folio_release(folio);
907 static void bch2_set_folio_dirty(struct bch_fs *c,
908 struct bch_inode_info *inode,
910 struct bch2_folio_reservation *res,
911 unsigned offset, unsigned len)
913 struct bch_folio *s = bch2_folio(folio);
914 unsigned i, dirty_sectors = 0;
916 WARN_ON((u64) folio_pos(folio) + offset + len >
917 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
919 BUG_ON(!s->uptodate);
923 for (i = round_down(offset, block_bytes(c)) >> 9;
924 i < round_up(offset + len, block_bytes(c)) >> 9;
926 unsigned sectors = sectors_to_reserve(&s->s[i],
927 res->disk.nr_replicas);
930 * This can happen if we race with the error path in
931 * bch2_writepage_io_done():
933 sectors = min_t(unsigned, sectors, res->disk.sectors);
935 s->s[i].replicas_reserved += sectors;
936 res->disk.sectors -= sectors;
938 dirty_sectors += s->s[i].state == SECTOR_unallocated;
940 folio_sector_set(folio, s, i, folio_sector_dirty(s->s[i].state));
943 spin_unlock(&s->lock);
945 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
947 if (!folio_test_dirty(folio))
948 filemap_dirty_folio(inode->v.i_mapping, folio);
951 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
953 struct file *file = vmf->vma->vm_file;
954 struct address_space *mapping = file->f_mapping;
955 struct address_space *fdm = faults_disabled_mapping();
956 struct bch_inode_info *inode = file_bch_inode(file);
960 return VM_FAULT_SIGBUS;
964 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
966 if (bch2_pagecache_add_tryget(inode))
969 bch2_pagecache_block_put(fdm_host);
971 bch2_pagecache_add_get(inode);
972 bch2_pagecache_add_put(inode);
974 bch2_pagecache_block_get(fdm_host);
976 /* Signal that lock has been dropped: */
977 set_fdm_dropped_locks();
978 return VM_FAULT_SIGBUS;
981 bch2_pagecache_add_get(inode);
983 ret = filemap_fault(vmf);
984 bch2_pagecache_add_put(inode);
989 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
991 struct folio *folio = page_folio(vmf->page);
992 struct file *file = vmf->vma->vm_file;
993 struct bch_inode_info *inode = file_bch_inode(file);
994 struct address_space *mapping = file->f_mapping;
995 struct bch_fs *c = inode->v.i_sb->s_fs_info;
996 struct bch2_folio_reservation res;
1001 bch2_folio_reservation_init(c, inode, &res);
1003 sb_start_pagefault(inode->v.i_sb);
1004 file_update_time(file);
1007 * Not strictly necessary, but helps avoid dio writes livelocking in
1008 * write_invalidate_inode_pages_range() - can drop this if/when we get
1009 * a write_invalidate_inode_pages_range() that works without dropping
1010 * page lock before invalidating page
1012 bch2_pagecache_add_get(inode);
1015 isize = i_size_read(&inode->v);
1017 if (folio->mapping != mapping || folio_pos(folio) >= isize) {
1018 folio_unlock(folio);
1019 ret = VM_FAULT_NOPAGE;
1023 len = min_t(loff_t, folio_size(folio), isize - folio_pos(folio));
1025 if (bch2_folio_set(c, inode_inum(inode), &folio, 1) ?:
1026 bch2_folio_reservation_get(c, inode, folio, &res, 0, len)) {
1027 folio_unlock(folio);
1028 ret = VM_FAULT_SIGBUS;
1032 bch2_set_folio_dirty(c, inode, folio, &res, 0, len);
1033 bch2_folio_reservation_put(c, inode, &res);
1035 folio_wait_stable(folio);
1036 ret = VM_FAULT_LOCKED;
1038 bch2_pagecache_add_put(inode);
1039 sb_end_pagefault(inode->v.i_sb);
1044 void bch2_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1046 if (offset || length < folio_size(folio))
1049 bch2_clear_folio_bits(folio);
1052 bool bch2_release_folio(struct folio *folio, gfp_t gfp_mask)
1054 if (folio_test_dirty(folio) || folio_test_writeback(folio))
1057 bch2_clear_folio_bits(folio);
1063 static void bch2_readpages_end_io(struct bio *bio)
1065 struct bvec_iter_all iter;
1066 struct folio_vec fv;
1068 bio_for_each_folio_all(fv, bio, iter) {
1069 if (!bio->bi_status) {
1070 folio_mark_uptodate(fv.fv_folio);
1072 folio_clear_uptodate(fv.fv_folio);
1073 folio_set_error(fv.fv_folio);
1075 folio_unlock(fv.fv_folio);
1081 struct readpages_iter {
1082 struct address_space *mapping;
1087 static int readpages_iter_init(struct readpages_iter *iter,
1088 struct readahead_control *ractl)
1093 memset(iter, 0, sizeof(*iter));
1095 iter->mapping = ractl->mapping;
1097 ret = filemap_get_contig_folios_d(iter->mapping,
1098 ractl->_index << PAGE_SHIFT,
1099 (ractl->_index + ractl->_nr_pages) << PAGE_SHIFT,
1100 0, mapping_gfp_mask(iter->mapping),
1105 darray_for_each(iter->folios, fi) {
1106 ractl->_nr_pages -= 1U << folio_order(*fi);
1107 __bch2_folio_create(*fi, __GFP_NOFAIL|GFP_KERNEL);
1115 static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
1117 if (iter->idx >= iter->folios.nr)
1119 return iter->folios.data[iter->idx];
1122 static inline void readpage_iter_advance(struct readpages_iter *iter)
1127 static bool extent_partial_reads_expensive(struct bkey_s_c k)
1129 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1130 struct bch_extent_crc_unpacked crc;
1131 const union bch_extent_entry *i;
1133 bkey_for_each_crc(k.k, ptrs, crc, i)
1134 if (crc.csum_type || crc.compression_type)
1139 static int readpage_bio_extend(struct btree_trans *trans,
1140 struct readpages_iter *iter,
1142 unsigned sectors_this_extent,
1145 /* Don't hold btree locks while allocating memory: */
1146 bch2_trans_unlock(trans);
1148 while (bio_sectors(bio) < sectors_this_extent &&
1149 bio->bi_vcnt < bio->bi_max_vecs) {
1150 struct folio *folio = readpage_iter_peek(iter);
1154 readpage_iter_advance(iter);
1156 pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
1161 folio = xa_load(&iter->mapping->i_pages, folio_offset);
1162 if (folio && !xa_is_value(folio))
1165 folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
1169 if (!__bch2_folio_create(folio, GFP_KERNEL)) {
1174 ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
1176 __bch2_folio_release(folio);
1184 BUG_ON(folio_sector(folio) != bio_end_sector(bio));
1186 BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
1189 return bch2_trans_relock(trans);
1192 static void bchfs_read(struct btree_trans *trans,
1193 struct bch_read_bio *rbio,
1195 struct readpages_iter *readpages_iter)
1197 struct bch_fs *c = trans->c;
1198 struct btree_iter iter;
1200 int flags = BCH_READ_RETRY_IF_STALE|
1201 BCH_READ_MAY_PROMOTE;
1206 rbio->start_time = local_clock();
1207 rbio->subvol = inum.subvol;
1209 bch2_bkey_buf_init(&sk);
1211 bch2_trans_begin(trans);
1212 iter = (struct btree_iter) { NULL };
1214 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1218 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1219 SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1223 unsigned bytes, sectors, offset_into_extent;
1224 enum btree_id data_btree = BTREE_ID_extents;
1227 * read_extent -> io_time_reset may cause a transaction restart
1228 * without returning an error, we need to check for that here:
1230 ret = bch2_trans_relock(trans);
1234 bch2_btree_iter_set_pos(&iter,
1235 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1237 k = bch2_btree_iter_peek_slot(&iter);
1242 offset_into_extent = iter.pos.offset -
1243 bkey_start_offset(k.k);
1244 sectors = k.k->size - offset_into_extent;
1246 bch2_bkey_buf_reassemble(&sk, c, k);
1248 ret = bch2_read_indirect_extent(trans, &data_btree,
1249 &offset_into_extent, &sk);
1253 k = bkey_i_to_s_c(sk.k);
1255 sectors = min(sectors, k.k->size - offset_into_extent);
1257 if (readpages_iter) {
1258 ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
1259 extent_partial_reads_expensive(k));
1264 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1265 swap(rbio->bio.bi_iter.bi_size, bytes);
1267 if (rbio->bio.bi_iter.bi_size == bytes)
1268 flags |= BCH_READ_LAST_FRAGMENT;
1270 bch2_bio_page_state_set(&rbio->bio, k);
1272 bch2_read_extent(trans, rbio, iter.pos,
1273 data_btree, k, offset_into_extent, flags);
1275 if (flags & BCH_READ_LAST_FRAGMENT)
1278 swap(rbio->bio.bi_iter.bi_size, bytes);
1279 bio_advance(&rbio->bio, bytes);
1281 ret = btree_trans_too_many_iters(trans);
1286 bch2_trans_iter_exit(trans, &iter);
1288 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1292 bch_err_inum_offset_ratelimited(c,
1294 iter.pos.offset << 9,
1295 "read error %i from btree lookup", ret);
1296 rbio->bio.bi_status = BLK_STS_IOERR;
1297 bio_endio(&rbio->bio);
1300 bch2_bkey_buf_exit(&sk, c);
1303 void bch2_readahead(struct readahead_control *ractl)
1305 struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1306 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1307 struct bch_io_opts opts;
1308 struct btree_trans trans;
1309 struct folio *folio;
1310 struct readpages_iter readpages_iter;
1313 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1315 ret = readpages_iter_init(&readpages_iter, ractl);
1318 bch2_trans_init(&trans, c, 0, 0);
1320 bch2_pagecache_add_get(inode);
1322 while ((folio = readpage_iter_peek(&readpages_iter))) {
1323 unsigned n = min_t(unsigned,
1324 readpages_iter.folios.nr -
1327 struct bch_read_bio *rbio =
1328 rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
1329 GFP_KERNEL, &c->bio_read),
1332 readpage_iter_advance(&readpages_iter);
1334 rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1335 rbio->bio.bi_end_io = bch2_readpages_end_io;
1336 BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1338 bchfs_read(&trans, rbio, inode_inum(inode),
1340 bch2_trans_unlock(&trans);
1343 bch2_pagecache_add_put(inode);
1345 bch2_trans_exit(&trans);
1346 darray_exit(&readpages_iter.folios);
1349 static void __bchfs_readfolio(struct bch_fs *c, struct bch_read_bio *rbio,
1350 subvol_inum inum, struct folio *folio)
1352 struct btree_trans trans;
1354 bch2_folio_create(folio, __GFP_NOFAIL);
1356 rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
1357 rbio->bio.bi_iter.bi_sector = folio_sector(folio);
1358 BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
1360 bch2_trans_init(&trans, c, 0, 0);
1361 bchfs_read(&trans, rbio, inum, NULL);
1362 bch2_trans_exit(&trans);
1365 static void bch2_read_single_folio_end_io(struct bio *bio)
1367 complete(bio->bi_private);
1370 static int bch2_read_single_folio(struct folio *folio,
1371 struct address_space *mapping)
1373 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1374 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1375 struct bch_read_bio *rbio;
1376 struct bch_io_opts opts;
1378 DECLARE_COMPLETION_ONSTACK(done);
1380 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
1382 rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
1384 rbio->bio.bi_private = &done;
1385 rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
1387 __bchfs_readfolio(c, rbio, inode_inum(inode), folio);
1388 wait_for_completion(&done);
1390 ret = blk_status_to_errno(rbio->bio.bi_status);
1391 bio_put(&rbio->bio);
1396 folio_mark_uptodate(folio);
1400 int bch2_read_folio(struct file *file, struct folio *folio)
1404 ret = bch2_read_single_folio(folio, folio->mapping);
1405 folio_unlock(folio);
1406 return bch2_err_class(ret);
1411 struct bch_writepage_state {
1412 struct bch_writepage_io *io;
1413 struct bch_io_opts opts;
1414 struct bch_folio_sector *tmp;
1415 unsigned tmp_sectors;
1418 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1419 struct bch_inode_info *inode)
1421 struct bch_writepage_state ret = { 0 };
1423 bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
1427 static void bch2_writepage_io_done(struct bch_write_op *op)
1429 struct bch_writepage_io *io =
1430 container_of(op, struct bch_writepage_io, op);
1431 struct bch_fs *c = io->op.c;
1432 struct bio *bio = &io->op.wbio.bio;
1433 struct bvec_iter_all iter;
1434 struct folio_vec fv;
1438 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1440 bio_for_each_folio_all(fv, bio, iter) {
1441 struct bch_folio *s;
1443 folio_set_error(fv.fv_folio);
1444 mapping_set_error(fv.fv_folio->mapping, -EIO);
1446 s = __bch2_folio(fv.fv_folio);
1447 spin_lock(&s->lock);
1448 for (i = 0; i < folio_sectors(fv.fv_folio); i++)
1449 s->s[i].nr_replicas = 0;
1450 spin_unlock(&s->lock);
1454 if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1455 bio_for_each_folio_all(fv, bio, iter) {
1456 struct bch_folio *s;
1458 s = __bch2_folio(fv.fv_folio);
1459 spin_lock(&s->lock);
1460 for (i = 0; i < folio_sectors(fv.fv_folio); i++)
1461 s->s[i].nr_replicas = 0;
1462 spin_unlock(&s->lock);
1467 * racing with fallocate can cause us to add fewer sectors than
1468 * expected - but we shouldn't add more sectors than expected:
1470 WARN_ON_ONCE(io->op.i_sectors_delta > 0);
1473 * (error (due to going RO) halfway through a page can screw that up
1476 BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1480 * PageWriteback is effectively our ref on the inode - fixup i_blocks
1481 * before calling end_page_writeback:
1483 i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1485 bio_for_each_folio_all(fv, bio, iter) {
1486 struct bch_folio *s = __bch2_folio(fv.fv_folio);
1488 if (atomic_dec_and_test(&s->write_count))
1489 folio_end_writeback(fv.fv_folio);
1492 bio_put(&io->op.wbio.bio);
1495 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1497 struct bch_writepage_io *io = w->io;
1500 closure_call(&io->op.cl, bch2_write, NULL, NULL);
1504 * Get a bch_writepage_io and add @page to it - appending to an existing one if
1505 * possible, else allocating a new one:
1507 static void bch2_writepage_io_alloc(struct bch_fs *c,
1508 struct writeback_control *wbc,
1509 struct bch_writepage_state *w,
1510 struct bch_inode_info *inode,
1512 unsigned nr_replicas)
1514 struct bch_write_op *op;
1516 w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
1519 &c->writepage_bioset),
1520 struct bch_writepage_io, op.wbio.bio);
1522 w->io->inode = inode;
1524 bch2_write_op_init(op, c, w->opts);
1525 op->target = w->opts.foreground_target;
1526 op->nr_replicas = nr_replicas;
1527 op->res.nr_replicas = nr_replicas;
1528 op->write_point = writepoint_hashed(inode->ei_last_dirtied);
1529 op->subvol = inode->ei_subvol;
1530 op->pos = POS(inode->v.i_ino, sector);
1531 op->end_io = bch2_writepage_io_done;
1532 op->devs_need_flush = &inode->ei_devs_need_flush;
1533 op->wbio.bio.bi_iter.bi_sector = sector;
1534 op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
1537 static int __bch2_writepage(struct folio *folio,
1538 struct writeback_control *wbc,
1541 struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
1542 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1543 struct bch_writepage_state *w = data;
1544 struct bch_folio *s;
1545 unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
1546 loff_t i_size = i_size_read(&inode->v);
1549 EBUG_ON(!folio_test_uptodate(folio));
1551 /* Is the folio fully inside i_size? */
1552 if (folio_end_pos(folio) <= i_size)
1555 /* Is the folio fully outside i_size? (truncate in progress) */
1556 if (folio_pos(folio) >= i_size) {
1557 folio_unlock(folio);
1562 * The folio straddles i_size. It must be zeroed out on each and every
1563 * writepage invocation because it may be mmapped. "A file is mapped
1564 * in multiples of the folio size. For a file that is not a multiple of
1565 * the folio size, the remaining memory is zeroed when mapped, and
1566 * writes to that region are not written out to the file."
1568 folio_zero_segment(folio,
1569 i_size - folio_pos(folio),
1572 f_sectors = folio_sectors(folio);
1573 s = bch2_folio(folio);
1575 if (f_sectors > w->tmp_sectors) {
1577 w->tmp = kzalloc(sizeof(struct bch_folio_sector) *
1578 f_sectors, __GFP_NOFAIL);
1579 w->tmp_sectors = f_sectors;
1583 * Things get really hairy with errors during writeback:
1585 ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
1588 /* Before unlocking the page, get copy of reservations: */
1589 spin_lock(&s->lock);
1590 memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
1592 for (i = 0; i < f_sectors; i++) {
1593 if (s->s[i].state < SECTOR_dirty)
1596 nr_replicas_this_write =
1597 min_t(unsigned, nr_replicas_this_write,
1598 s->s[i].nr_replicas +
1599 s->s[i].replicas_reserved);
1602 for (i = 0; i < f_sectors; i++) {
1603 if (s->s[i].state < SECTOR_dirty)
1606 s->s[i].nr_replicas = w->opts.compression
1607 ? 0 : nr_replicas_this_write;
1609 s->s[i].replicas_reserved = 0;
1610 folio_sector_set(folio, s, i, SECTOR_allocated);
1612 spin_unlock(&s->lock);
1614 BUG_ON(atomic_read(&s->write_count));
1615 atomic_set(&s->write_count, 1);
1617 BUG_ON(folio_test_writeback(folio));
1618 folio_start_writeback(folio);
1620 folio_unlock(folio);
1624 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1627 while (offset < f_sectors &&
1628 w->tmp[offset].state < SECTOR_dirty)
1631 if (offset == f_sectors)
1634 while (offset + sectors < f_sectors &&
1635 w->tmp[offset + sectors].state >= SECTOR_dirty) {
1636 reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
1637 dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
1642 sector = folio_sector(folio) + offset;
1645 (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1646 bio_full(&w->io->op.wbio.bio, sectors << 9) ||
1647 w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1648 (BIO_MAX_VECS * PAGE_SIZE) ||
1649 bio_end_sector(&w->io->op.wbio.bio) != sector))
1650 bch2_writepage_do_io(w);
1653 bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1654 nr_replicas_this_write);
1656 atomic_inc(&s->write_count);
1658 BUG_ON(inode != w->io->inode);
1659 BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
1660 sectors << 9, offset << 9));
1662 /* Check for writing past i_size: */
1663 WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1664 round_up(i_size, block_bytes(c)) &&
1665 !test_bit(BCH_FS_EMERGENCY_RO, &c->flags),
1666 "writing past i_size: %llu > %llu (unrounded %llu)\n",
1667 bio_end_sector(&w->io->op.wbio.bio) << 9,
1668 round_up(i_size, block_bytes(c)),
1671 w->io->op.res.sectors += reserved_sectors;
1672 w->io->op.i_sectors_delta -= dirty_sectors;
1673 w->io->op.new_i_size = i_size;
1678 if (atomic_dec_and_test(&s->write_count))
1679 folio_end_writeback(folio);
1684 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1686 struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1687 struct bch_writepage_state w =
1688 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1689 struct blk_plug plug;
1692 blk_start_plug(&plug);
1693 ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1695 bch2_writepage_do_io(&w);
1696 blk_finish_plug(&plug);
1698 return bch2_err_class(ret);
1701 /* buffered writes: */
1703 int bch2_write_begin(struct file *file, struct address_space *mapping,
1704 loff_t pos, unsigned len,
1705 struct page **pagep, void **fsdata)
1707 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1708 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1709 struct bch2_folio_reservation *res;
1710 struct folio *folio;
1714 res = kmalloc(sizeof(*res), GFP_KERNEL);
1718 bch2_folio_reservation_init(c, inode, res);
1721 bch2_pagecache_add_get(inode);
1723 folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
1724 FGP_LOCK|FGP_WRITE|FGP_CREAT|FGP_STABLE,
1725 mapping_gfp_mask(mapping));
1729 if (folio_test_uptodate(folio))
1732 offset = pos - folio_pos(folio);
1733 len = min_t(size_t, len, folio_end_pos(folio) - pos);
1735 /* If we're writing entire folio, don't need to read it in first: */
1736 if (!offset && len == folio_size(folio))
1739 if (!offset && pos + len >= inode->v.i_size) {
1740 folio_zero_segment(folio, len, folio_size(folio));
1741 flush_dcache_folio(folio);
1745 if (folio_pos(folio) >= inode->v.i_size) {
1746 folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
1747 flush_dcache_folio(folio);
1751 ret = bch2_read_single_folio(folio, mapping);
1755 ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
1759 ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
1761 if (!folio_test_uptodate(folio)) {
1763 * If the folio hasn't been read in, we won't know if we
1764 * actually need a reservation - we don't actually need
1765 * to read here, we just need to check if the folio is
1766 * fully backed by uncompressed data:
1774 *pagep = &folio->page;
1777 folio_unlock(folio);
1781 bch2_pagecache_add_put(inode);
1784 return bch2_err_class(ret);
1787 int bch2_write_end(struct file *file, struct address_space *mapping,
1788 loff_t pos, unsigned len, unsigned copied,
1789 struct page *page, void *fsdata)
1791 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1792 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1793 struct bch2_folio_reservation *res = fsdata;
1794 struct folio *folio = page_folio(page);
1795 unsigned offset = pos - folio_pos(folio);
1797 lockdep_assert_held(&inode->v.i_rwsem);
1798 BUG_ON(offset + copied > folio_size(folio));
1800 if (unlikely(copied < len && !folio_test_uptodate(folio))) {
1802 * The folio needs to be read in, but that would destroy
1803 * our partial write - simplest thing is to just force
1804 * userspace to redo the write:
1806 folio_zero_range(folio, 0, folio_size(folio));
1807 flush_dcache_folio(folio);
1811 spin_lock(&inode->v.i_lock);
1812 if (pos + copied > inode->v.i_size)
1813 i_size_write(&inode->v, pos + copied);
1814 spin_unlock(&inode->v.i_lock);
1817 if (!folio_test_uptodate(folio))
1818 folio_mark_uptodate(folio);
1820 bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
1822 inode->ei_last_dirtied = (unsigned long) current;
1825 folio_unlock(folio);
1827 bch2_pagecache_add_put(inode);
1829 bch2_folio_reservation_put(c, inode, res);
1835 static noinline void folios_trunc(folios *folios, struct folio **fi)
1837 while (folios->data + folios->nr > fi) {
1838 struct folio *f = darray_pop(folios);
1845 static int __bch2_buffered_write(struct bch_inode_info *inode,
1846 struct address_space *mapping,
1847 struct iov_iter *iter,
1848 loff_t pos, unsigned len)
1850 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1851 struct bch2_folio_reservation res;
1853 struct folio **fi, *f;
1854 unsigned copied = 0, f_offset;
1855 u64 end = pos + len, f_pos;
1856 loff_t last_folio_pos = inode->v.i_size;
1861 bch2_folio_reservation_init(c, inode, &res);
1862 darray_init(&folios);
1864 ret = filemap_get_contig_folios_d(mapping, pos, end,
1865 FGP_LOCK|FGP_WRITE|FGP_STABLE|FGP_CREAT,
1866 mapping_gfp_mask(mapping),
1873 f = darray_first(folios);
1874 if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
1875 ret = bch2_read_single_folio(f, mapping);
1880 f = darray_last(folios);
1881 end = min(end, folio_end_pos(f));
1882 last_folio_pos = folio_pos(f);
1883 if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
1884 if (end >= inode->v.i_size) {
1885 folio_zero_range(f, 0, folio_size(f));
1887 ret = bch2_read_single_folio(f, mapping);
1893 ret = bch2_folio_set(c, inode_inum(inode), folios.data, folios.nr);
1898 f_offset = pos - folio_pos(darray_first(folios));
1899 darray_for_each(folios, fi) {
1900 struct folio *f = *fi;
1901 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1904 * XXX: per POSIX and fstests generic/275, on -ENOSPC we're
1905 * supposed to write as much as we have disk space for.
1907 * On failure here we should still write out a partial page if
1908 * we aren't completely out of disk space - we don't do that
1911 ret = bch2_folio_reservation_get(c, inode, f, &res, f_offset, f_len);
1912 if (unlikely(ret)) {
1913 folios_trunc(&folios, fi);
1917 end = min(end, folio_end_pos(darray_last(folios)));
1921 f_pos = folio_end_pos(f);
1925 if (mapping_writably_mapped(mapping))
1926 darray_for_each(folios, fi)
1927 flush_dcache_folio(*fi);
1930 f_offset = pos - folio_pos(darray_first(folios));
1931 darray_for_each(folios, fi) {
1932 struct folio *f = *fi;
1933 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1934 unsigned f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter);
1937 folios_trunc(&folios, fi);
1941 if (!folio_test_uptodate(f) &&
1942 f_copied != folio_size(f) &&
1943 pos + copied + f_copied < inode->v.i_size) {
1944 folio_zero_range(f, 0, folio_size(f));
1945 folios_trunc(&folios, fi);
1949 flush_dcache_folio(f);
1952 if (f_copied != f_len) {
1953 folios_trunc(&folios, fi + 1);
1957 f_pos = folio_end_pos(f);
1966 spin_lock(&inode->v.i_lock);
1967 if (end > inode->v.i_size)
1968 i_size_write(&inode->v, end);
1969 spin_unlock(&inode->v.i_lock);
1972 f_offset = pos - folio_pos(darray_first(folios));
1973 darray_for_each(folios, fi) {
1974 struct folio *f = *fi;
1975 u64 f_len = min(end, folio_end_pos(f)) - f_pos;
1977 if (!folio_test_uptodate(f))
1978 folio_mark_uptodate(f);
1980 bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
1982 f_pos = folio_end_pos(f);
1986 inode->ei_last_dirtied = (unsigned long) current;
1988 darray_for_each(folios, fi) {
1994 * If the last folio added to the mapping starts beyond current EOF, we
1995 * performed a short write but left around at least one post-EOF folio.
1996 * Clean up the mapping before we return.
1998 if (last_folio_pos >= inode->v.i_size)
1999 truncate_pagecache(&inode->v, inode->v.i_size);
2001 darray_exit(&folios);
2002 bch2_folio_reservation_put(c, inode, &res);
2004 return copied ?: ret;
2007 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
2009 struct file *file = iocb->ki_filp;
2010 struct address_space *mapping = file->f_mapping;
2011 struct bch_inode_info *inode = file_bch_inode(file);
2012 loff_t pos = iocb->ki_pos;
2013 ssize_t written = 0;
2016 bch2_pagecache_add_get(inode);
2019 unsigned offset = pos & (PAGE_SIZE - 1);
2020 unsigned bytes = iov_iter_count(iter);
2023 * Bring in the user page that we will copy from _first_.
2024 * Otherwise there's a nasty deadlock on copying from the
2025 * same page as we're writing to, without it being marked
2028 * Not only is this an optimisation, but it is also required
2029 * to check that the address is actually valid, when atomic
2030 * usercopies are used, below.
2032 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2033 bytes = min_t(unsigned long, iov_iter_count(iter),
2034 PAGE_SIZE - offset);
2036 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
2042 if (unlikely(fatal_signal_pending(current))) {
2047 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
2048 if (unlikely(ret < 0))
2053 if (unlikely(ret == 0)) {
2055 * If we were unable to copy any data at all, we must
2056 * fall back to a single segment length write.
2058 * If we didn't fallback here, we could livelock
2059 * because not all segments in the iov can be copied at
2060 * once without a pagefault.
2062 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2063 iov_iter_single_seg_count(iter));
2070 balance_dirty_pages_ratelimited(mapping);
2071 } while (iov_iter_count(iter));
2073 bch2_pagecache_add_put(inode);
2075 return written ? written : ret;
2078 /* O_DIRECT reads */
2080 static void bio_check_or_release(struct bio *bio, bool check_dirty)
2083 bio_check_pages_dirty(bio);
2085 bio_release_pages(bio, false);
2090 static void bch2_dio_read_complete(struct closure *cl)
2092 struct dio_read *dio = container_of(cl, struct dio_read, cl);
2094 dio->req->ki_complete(dio->req, dio->ret);
2095 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2098 static void bch2_direct_IO_read_endio(struct bio *bio)
2100 struct dio_read *dio = bio->bi_private;
2103 dio->ret = blk_status_to_errno(bio->bi_status);
2105 closure_put(&dio->cl);
2108 static void bch2_direct_IO_read_split_endio(struct bio *bio)
2110 struct dio_read *dio = bio->bi_private;
2111 bool should_dirty = dio->should_dirty;
2113 bch2_direct_IO_read_endio(bio);
2114 bio_check_or_release(bio, should_dirty);
2117 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
2119 struct file *file = req->ki_filp;
2120 struct bch_inode_info *inode = file_bch_inode(file);
2121 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2122 struct bch_io_opts opts;
2123 struct dio_read *dio;
2125 loff_t offset = req->ki_pos;
2126 bool sync = is_sync_kiocb(req);
2130 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2132 if ((offset|iter->count) & (block_bytes(c) - 1))
2135 ret = min_t(loff_t, iter->count,
2136 max_t(loff_t, 0, i_size_read(&inode->v) - offset));
2141 shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
2142 iter->count -= shorten;
2144 bio = bio_alloc_bioset(NULL,
2145 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2148 &c->dio_read_bioset);
2150 bio->bi_end_io = bch2_direct_IO_read_endio;
2152 dio = container_of(bio, struct dio_read, rbio.bio);
2153 closure_init(&dio->cl, NULL);
2156 * this is a _really_ horrible hack just to avoid an atomic sub at the
2160 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
2161 atomic_set(&dio->cl.remaining,
2162 CLOSURE_REMAINING_INITIALIZER -
2164 CLOSURE_DESTRUCTOR);
2166 atomic_set(&dio->cl.remaining,
2167 CLOSURE_REMAINING_INITIALIZER + 1);
2173 * This is one of the sketchier things I've encountered: we have to skip
2174 * the dirtying of requests that are internal from the kernel (i.e. from
2175 * loopback), because we'll deadlock on page_lock.
2177 dio->should_dirty = iter_is_iovec(iter);
2180 while (iter->count) {
2181 bio = bio_alloc_bioset(NULL,
2182 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2186 bio->bi_end_io = bch2_direct_IO_read_split_endio;
2188 bio->bi_opf = REQ_OP_READ|REQ_SYNC;
2189 bio->bi_iter.bi_sector = offset >> 9;
2190 bio->bi_private = dio;
2192 ret = bio_iov_iter_get_pages(bio, iter);
2194 /* XXX: fault inject this path */
2195 bio->bi_status = BLK_STS_RESOURCE;
2200 offset += bio->bi_iter.bi_size;
2202 if (dio->should_dirty)
2203 bio_set_pages_dirty(bio);
2206 closure_get(&dio->cl);
2208 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
2211 iter->count += shorten;
2214 closure_sync(&dio->cl);
2215 closure_debug_destroy(&dio->cl);
2217 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2220 return -EIOCBQUEUED;
2224 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2226 struct file *file = iocb->ki_filp;
2227 struct bch_inode_info *inode = file_bch_inode(file);
2228 struct address_space *mapping = file->f_mapping;
2229 size_t count = iov_iter_count(iter);
2233 return 0; /* skip atime */
2235 if (iocb->ki_flags & IOCB_DIRECT) {
2236 struct blk_plug plug;
2238 if (unlikely(mapping->nrpages)) {
2239 ret = filemap_write_and_wait_range(mapping,
2241 iocb->ki_pos + count - 1);
2246 file_accessed(file);
2248 blk_start_plug(&plug);
2249 ret = bch2_direct_IO_read(iocb, iter);
2250 blk_finish_plug(&plug);
2253 iocb->ki_pos += ret;
2255 bch2_pagecache_add_get(inode);
2256 ret = generic_file_read_iter(iocb, iter);
2257 bch2_pagecache_add_put(inode);
2260 return bch2_err_class(ret);
2263 /* O_DIRECT writes */
2265 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2266 u64 offset, u64 size,
2267 unsigned nr_replicas, bool compressed)
2269 struct btree_trans trans;
2270 struct btree_iter iter;
2272 u64 end = offset + size;
2277 bch2_trans_init(&trans, c, 0, 0);
2279 bch2_trans_begin(&trans);
2281 err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2285 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2286 SPOS(inum.inum, offset, snapshot),
2287 BTREE_ITER_SLOTS, k, err) {
2288 if (bkey_ge(bkey_start_pos(k.k), POS(inum.inum, end)))
2291 if (k.k->p.snapshot != snapshot ||
2292 nr_replicas > bch2_bkey_replicas(c, k) ||
2293 (!compressed && bch2_bkey_sectors_compressed(k))) {
2299 offset = iter.pos.offset;
2300 bch2_trans_iter_exit(&trans, &iter);
2302 if (bch2_err_matches(err, BCH_ERR_transaction_restart))
2304 bch2_trans_exit(&trans);
2306 return err ? false : ret;
2309 static noinline bool bch2_dio_write_check_allocated(struct dio_write *dio)
2311 struct bch_fs *c = dio->op.c;
2312 struct bch_inode_info *inode = dio->inode;
2313 struct bio *bio = &dio->op.wbio.bio;
2315 return bch2_check_range_allocated(c, inode_inum(inode),
2316 dio->op.pos.offset, bio_sectors(bio),
2317 dio->op.opts.data_replicas,
2318 dio->op.opts.compression != 0);
2321 static void bch2_dio_write_loop_async(struct bch_write_op *);
2322 static __always_inline long bch2_dio_write_done(struct dio_write *dio);
2324 static noinline int bch2_dio_write_copy_iov(struct dio_write *dio)
2326 struct iovec *iov = dio->inline_vecs;
2328 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2329 iov = kmalloc_array(dio->iter.nr_segs, sizeof(*iov),
2334 dio->free_iov = true;
2337 memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2338 dio->iter.iov = iov;
2342 static void bch2_dio_write_flush_done(struct closure *cl)
2344 struct dio_write *dio = container_of(cl, struct dio_write, op.cl);
2345 struct bch_fs *c = dio->op.c;
2347 closure_debug_destroy(cl);
2349 dio->op.error = bch2_journal_error(&c->journal);
2351 bch2_dio_write_done(dio);
2354 static noinline void bch2_dio_write_flush(struct dio_write *dio)
2356 struct bch_fs *c = dio->op.c;
2357 struct bch_inode_unpacked inode;
2362 closure_init(&dio->op.cl, NULL);
2364 if (!dio->op.error) {
2365 ret = bch2_inode_find_by_inum(c, inode_inum(dio->inode), &inode);
2367 dio->op.error = ret;
2369 bch2_journal_flush_seq_async(&c->journal, inode.bi_journal_seq, &dio->op.cl);
2370 bch2_inode_flush_nocow_writes_async(c, dio->inode, &dio->op.cl);
2375 closure_sync(&dio->op.cl);
2376 closure_debug_destroy(&dio->op.cl);
2378 continue_at(&dio->op.cl, bch2_dio_write_flush_done, NULL);
2382 static __always_inline long bch2_dio_write_done(struct dio_write *dio)
2384 struct kiocb *req = dio->req;
2385 struct bch_inode_info *inode = dio->inode;
2386 bool sync = dio->sync;
2389 if (unlikely(dio->flush)) {
2390 bch2_dio_write_flush(dio);
2392 return -EIOCBQUEUED;
2395 bch2_pagecache_block_put(inode);
2398 kfree(dio->iter.iov);
2400 ret = dio->op.error ?: ((long) dio->written << 9);
2401 bio_put(&dio->op.wbio.bio);
2403 /* inode->i_dio_count is our ref on inode and thus bch_fs */
2404 inode_dio_end(&inode->v);
2407 ret = bch2_err_class(ret);
2410 req->ki_complete(req, ret);
2416 static __always_inline void bch2_dio_write_end(struct dio_write *dio)
2418 struct bch_fs *c = dio->op.c;
2419 struct kiocb *req = dio->req;
2420 struct bch_inode_info *inode = dio->inode;
2421 struct bio *bio = &dio->op.wbio.bio;
2423 req->ki_pos += (u64) dio->op.written << 9;
2424 dio->written += dio->op.written;
2426 if (dio->extending) {
2427 spin_lock(&inode->v.i_lock);
2428 if (req->ki_pos > inode->v.i_size)
2429 i_size_write(&inode->v, req->ki_pos);
2430 spin_unlock(&inode->v.i_lock);
2433 if (dio->op.i_sectors_delta || dio->quota_res.sectors) {
2434 mutex_lock(&inode->ei_quota_lock);
2435 __i_sectors_acct(c, inode, &dio->quota_res, dio->op.i_sectors_delta);
2436 __bch2_quota_reservation_put(c, inode, &dio->quota_res);
2437 mutex_unlock(&inode->ei_quota_lock);
2440 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF))) {
2441 struct bvec_iter_all iter;
2442 struct folio_vec fv;
2444 bio_for_each_folio_all(fv, bio, iter)
2445 folio_put(fv.fv_folio);
2448 if (unlikely(dio->op.error))
2449 set_bit(EI_INODE_ERROR, &inode->ei_flags);
2452 static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
2454 struct bch_fs *c = dio->op.c;
2455 struct kiocb *req = dio->req;
2456 struct address_space *mapping = dio->mapping;
2457 struct bch_inode_info *inode = dio->inode;
2458 struct bch_io_opts opts;
2459 struct bio *bio = &dio->op.wbio.bio;
2460 unsigned unaligned, iter_count;
2461 bool sync = dio->sync, dropped_locks;
2464 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
2467 iter_count = dio->iter.count;
2469 EBUG_ON(current->faults_disabled_mapping);
2470 current->faults_disabled_mapping = mapping;
2472 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2474 dropped_locks = fdm_dropped_locks();
2476 current->faults_disabled_mapping = NULL;
2479 * If the fault handler returned an error but also signalled
2480 * that it dropped & retook ei_pagecache_lock, we just need to
2481 * re-shoot down the page cache and retry:
2483 if (dropped_locks && ret)
2486 if (unlikely(ret < 0))
2489 if (unlikely(dropped_locks)) {
2490 ret = write_invalidate_inode_pages_range(mapping,
2492 req->ki_pos + iter_count - 1);
2496 if (!bio->bi_iter.bi_size)
2500 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2501 bio->bi_iter.bi_size -= unaligned;
2502 iov_iter_revert(&dio->iter, unaligned);
2504 if (!bio->bi_iter.bi_size) {
2506 * bio_iov_iter_get_pages was only able to get <
2507 * blocksize worth of pages:
2513 bch2_write_op_init(&dio->op, c, opts);
2514 dio->op.end_io = sync
2516 : bch2_dio_write_loop_async;
2517 dio->op.target = dio->op.opts.foreground_target;
2518 dio->op.write_point = writepoint_hashed((unsigned long) current);
2519 dio->op.nr_replicas = dio->op.opts.data_replicas;
2520 dio->op.subvol = inode->ei_subvol;
2521 dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2522 dio->op.devs_need_flush = &inode->ei_devs_need_flush;
2525 dio->op.flags |= BCH_WRITE_SYNC;
2526 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2528 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2529 bio_sectors(bio), true);
2533 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2534 dio->op.opts.data_replicas, 0);
2535 if (unlikely(ret) &&
2536 !bch2_dio_write_check_allocated(dio))
2539 task_io_account_write(bio->bi_iter.bi_size);
2541 if (unlikely(dio->iter.count) &&
2544 bch2_dio_write_copy_iov(dio))
2545 dio->sync = sync = true;
2548 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2551 return -EIOCBQUEUED;
2553 bch2_dio_write_end(dio);
2555 if (likely(!dio->iter.count) || dio->op.error)
2558 bio_reset(bio, NULL, REQ_OP_WRITE);
2561 return bch2_dio_write_done(dio);
2563 dio->op.error = ret;
2565 if (!bio_flagged(bio, BIO_NO_PAGE_REF)) {
2566 struct bvec_iter_all iter;
2567 struct folio_vec fv;
2569 bio_for_each_folio_all(fv, bio, iter)
2570 folio_put(fv.fv_folio);
2573 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2577 static noinline __cold void bch2_dio_write_continue(struct dio_write *dio)
2579 struct mm_struct *mm = dio->mm;
2581 bio_reset(&dio->op.wbio.bio, NULL, REQ_OP_WRITE);
2585 bch2_dio_write_loop(dio);
2587 kthread_unuse_mm(mm);
2590 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2592 struct dio_write *dio = container_of(op, struct dio_write, op);
2594 bch2_dio_write_end(dio);
2596 if (likely(!dio->iter.count) || dio->op.error)
2597 bch2_dio_write_done(dio);
2599 bch2_dio_write_continue(dio);
2603 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2605 struct file *file = req->ki_filp;
2606 struct address_space *mapping = file->f_mapping;
2607 struct bch_inode_info *inode = file_bch_inode(file);
2608 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2609 struct dio_write *dio;
2611 bool locked = true, extending;
2615 prefetch((void *) &c->opts + 64);
2616 prefetch(&inode->ei_inode);
2617 prefetch((void *) &inode->ei_inode + 64);
2619 inode_lock(&inode->v);
2621 ret = generic_write_checks(req, iter);
2622 if (unlikely(ret <= 0))
2625 ret = file_remove_privs(file);
2629 ret = file_update_time(file);
2633 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2636 inode_dio_begin(&inode->v);
2637 bch2_pagecache_block_get(inode);
2639 extending = req->ki_pos + iter->count > inode->v.i_size;
2641 inode_unlock(&inode->v);
2645 bio = bio_alloc_bioset(NULL,
2646 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2649 &c->dio_write_bioset);
2650 dio = container_of(bio, struct dio_write, op.wbio.bio);
2652 dio->mapping = mapping;
2654 dio->mm = current->mm;
2656 dio->extending = extending;
2657 dio->sync = is_sync_kiocb(req) || extending;
2658 dio->flush = iocb_is_dsync(req) && !c->opts.journal_flush_disabled;
2659 dio->free_iov = false;
2660 dio->quota_res.sectors = 0;
2665 if (unlikely(mapping->nrpages)) {
2666 ret = write_invalidate_inode_pages_range(mapping,
2668 req->ki_pos + iter->count - 1);
2673 ret = bch2_dio_write_loop(dio);
2676 inode_unlock(&inode->v);
2679 bch2_pagecache_block_put(inode);
2681 inode_dio_end(&inode->v);
2685 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2687 struct file *file = iocb->ki_filp;
2688 struct bch_inode_info *inode = file_bch_inode(file);
2691 if (iocb->ki_flags & IOCB_DIRECT) {
2692 ret = bch2_direct_write(iocb, from);
2696 /* We can write back this queue in page reclaim */
2697 current->backing_dev_info = inode_to_bdi(&inode->v);
2698 inode_lock(&inode->v);
2700 ret = generic_write_checks(iocb, from);
2704 ret = file_remove_privs(file);
2708 ret = file_update_time(file);
2712 ret = bch2_buffered_write(iocb, from);
2713 if (likely(ret > 0))
2714 iocb->ki_pos += ret;
2716 inode_unlock(&inode->v);
2717 current->backing_dev_info = NULL;
2720 ret = generic_write_sync(iocb, ret);
2722 return bch2_err_class(ret);
2728 * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2729 * insert trigger: look up the btree inode instead
2731 static int bch2_flush_inode(struct bch_fs *c,
2732 struct bch_inode_info *inode)
2734 struct bch_inode_unpacked u;
2737 if (c->opts.journal_flush_disabled)
2740 ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u);
2744 return bch2_journal_flush_seq(&c->journal, u.bi_journal_seq) ?:
2745 bch2_inode_flush_nocow_writes(c, inode);
2748 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2750 struct bch_inode_info *inode = file_bch_inode(file);
2751 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2752 int ret, ret2, ret3;
2754 ret = file_write_and_wait_range(file, start, end);
2755 ret2 = sync_inode_metadata(&inode->v, 1);
2756 ret3 = bch2_flush_inode(c, inode);
2758 return bch2_err_class(ret ?: ret2 ?: ret3);
2763 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2767 struct btree_trans trans;
2768 struct btree_iter iter;
2772 bch2_trans_init(&trans, c, 0, 0);
2774 bch2_trans_begin(&trans);
2776 ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2780 for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents, start, end, 0, k, ret)
2781 if (bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k)) {
2786 bch2_trans_iter_exit(&trans, &iter);
2788 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
2791 bch2_trans_exit(&trans);
2795 static int __bch2_truncate_folio(struct bch_inode_info *inode,
2796 pgoff_t index, loff_t start, loff_t end)
2798 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2799 struct address_space *mapping = inode->v.i_mapping;
2800 struct bch_folio *s;
2801 unsigned start_offset = start & (PAGE_SIZE - 1);
2802 unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2804 struct folio *folio;
2805 s64 i_sectors_delta = 0;
2809 folio = filemap_lock_folio(mapping, index);
2812 * XXX: we're doing two index lookups when we end up reading the
2815 ret = range_has_data(c, inode->ei_subvol,
2816 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT)),
2817 POS(inode->v.i_ino, (index << PAGE_SECTORS_SHIFT) + PAGE_SECTORS));
2821 folio = __filemap_get_folio(mapping, index,
2822 FGP_LOCK|FGP_CREAT, GFP_KERNEL);
2823 if (unlikely(!folio)) {
2829 BUG_ON(start >= folio_end_pos(folio));
2830 BUG_ON(end <= folio_pos(folio));
2832 start_offset = max(start, folio_pos(folio)) - folio_pos(folio);
2833 end_offset = min_t(u64, end, folio_end_pos(folio)) - folio_pos(folio);
2835 /* Folio boundary? Nothing to do */
2836 if (start_offset == 0 &&
2837 end_offset == folio_size(folio)) {
2842 s = bch2_folio_create(folio, 0);
2848 if (!folio_test_uptodate(folio)) {
2849 ret = bch2_read_single_folio(folio, mapping);
2854 ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
2858 for (i = round_up(start_offset, block_bytes(c)) >> 9;
2859 i < round_down(end_offset, block_bytes(c)) >> 9;
2861 s->s[i].nr_replicas = 0;
2863 i_sectors_delta -= s->s[i].state == SECTOR_dirty;
2864 folio_sector_set(folio, s, i, SECTOR_unallocated);
2867 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2870 * Caller needs to know whether this folio will be written out by
2871 * writeback - doing an i_size update if necessary - or whether it will
2872 * be responsible for the i_size update.
2874 * Note that we shouldn't ever see a folio beyond EOF, but check and
2875 * warn if so. This has been observed by failure to clean up folios
2876 * after a short write and there's still a chance reclaim will fix
2879 WARN_ON_ONCE(folio_pos(folio) >= inode->v.i_size);
2880 end_pos = folio_end_pos(folio);
2881 if (inode->v.i_size > folio_pos(folio))
2882 end_pos = min_t(u64, inode->v.i_size, end_pos);
2883 ret = s->s[folio_pos_to_s(folio, end_pos - 1)].state >= SECTOR_dirty;
2885 folio_zero_segment(folio, start_offset, end_offset);
2888 * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2890 * XXX: because we aren't currently tracking whether the folio has actual
2891 * data in it (vs. just 0s, or only partially written) this wrong. ick.
2893 BUG_ON(bch2_get_folio_disk_reservation(c, inode, folio, false));
2896 * This removes any writeable userspace mappings; we need to force
2897 * .page_mkwrite to be called again before any mmapped writes, to
2898 * redirty the full page:
2900 folio_mkclean(folio);
2901 filemap_dirty_folio(mapping, folio);
2903 folio_unlock(folio);
2909 static int bch2_truncate_folio(struct bch_inode_info *inode, loff_t from)
2911 return __bch2_truncate_folio(inode, from >> PAGE_SHIFT,
2912 from, ANYSINT_MAX(loff_t));
2915 static int bch2_truncate_folios(struct bch_inode_info *inode,
2916 loff_t start, loff_t end)
2918 int ret = __bch2_truncate_folio(inode, start >> PAGE_SHIFT,
2922 start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2923 ret = __bch2_truncate_folio(inode,
2924 (end - 1) >> PAGE_SHIFT,
2929 static int bch2_extend(struct mnt_idmap *idmap,
2930 struct bch_inode_info *inode,
2931 struct bch_inode_unpacked *inode_u,
2932 struct iattr *iattr)
2934 struct address_space *mapping = inode->v.i_mapping;
2940 * this has to be done _before_ extending i_size:
2942 ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2946 truncate_setsize(&inode->v, iattr->ia_size);
2948 return bch2_setattr_nonsize(idmap, inode, iattr);
2951 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2952 struct bch_inode_unpacked *bi,
2955 bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2959 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2960 struct bch_inode_unpacked *bi, void *p)
2962 u64 *new_i_size = p;
2964 bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2965 bi->bi_size = *new_i_size;
2969 int bch2_truncate(struct mnt_idmap *idmap,
2970 struct bch_inode_info *inode, struct iattr *iattr)
2972 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2973 struct address_space *mapping = inode->v.i_mapping;
2974 struct bch_inode_unpacked inode_u;
2975 u64 new_i_size = iattr->ia_size;
2976 s64 i_sectors_delta = 0;
2980 * If the truncate call with change the size of the file, the
2981 * cmtimes should be updated. If the size will not change, we
2982 * do not need to update the cmtimes.
2984 if (iattr->ia_size != inode->v.i_size) {
2985 if (!(iattr->ia_valid & ATTR_MTIME))
2986 ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2987 if (!(iattr->ia_valid & ATTR_CTIME))
2988 ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2989 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2992 inode_dio_wait(&inode->v);
2993 bch2_pagecache_block_get(inode);
2995 ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
3000 * check this before next assertion; on filesystem error our normal
3001 * invariants are a bit broken (truncate has to truncate the page cache
3002 * before the inode).
3004 ret = bch2_journal_error(&c->journal);
3008 WARN_ONCE(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
3009 inode->v.i_size < inode_u.bi_size,
3010 "truncate spotted in mem i_size < btree i_size: %llu < %llu\n",
3011 (u64) inode->v.i_size, inode_u.bi_size);
3013 if (iattr->ia_size > inode->v.i_size) {
3014 ret = bch2_extend(idmap, inode, &inode_u, iattr);
3018 iattr->ia_valid &= ~ATTR_SIZE;
3020 ret = bch2_truncate_folio(inode, iattr->ia_size);
3021 if (unlikely(ret < 0))
3025 * When extending, we're going to write the new i_size to disk
3026 * immediately so we need to flush anything above the current on disk
3029 * Also, when extending we need to flush the page that i_size currently
3030 * straddles - if it's mapped to userspace, we need to ensure that
3031 * userspace has to redirty it and call .mkwrite -> set_page_dirty
3032 * again to allocate the part of the page that was extended.
3034 if (iattr->ia_size > inode_u.bi_size)
3035 ret = filemap_write_and_wait_range(mapping,
3037 iattr->ia_size - 1);
3038 else if (iattr->ia_size & (PAGE_SIZE - 1))
3039 ret = filemap_write_and_wait_range(mapping,
3040 round_down(iattr->ia_size, PAGE_SIZE),
3041 iattr->ia_size - 1);
3045 mutex_lock(&inode->ei_update_lock);
3046 ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
3048 mutex_unlock(&inode->ei_update_lock);
3053 truncate_setsize(&inode->v, iattr->ia_size);
3055 ret = bch2_fpunch(c, inode_inum(inode),
3056 round_up(iattr->ia_size, block_bytes(c)) >> 9,
3057 U64_MAX, &i_sectors_delta);
3058 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3060 bch2_fs_inconsistent_on(!inode->v.i_size && inode->v.i_blocks &&
3061 !bch2_journal_error(&c->journal), c,
3062 "inode %lu truncated to 0 but i_blocks %llu (ondisk %lli)",
3063 inode->v.i_ino, (u64) inode->v.i_blocks,
3064 inode->ei_inode.bi_sectors);
3068 mutex_lock(&inode->ei_update_lock);
3069 ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
3070 mutex_unlock(&inode->ei_update_lock);
3072 ret = bch2_setattr_nonsize(idmap, inode, iattr);
3074 bch2_pagecache_block_put(inode);
3075 return bch2_err_class(ret);
3080 static int inode_update_times_fn(struct bch_inode_info *inode,
3081 struct bch_inode_unpacked *bi, void *p)
3083 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3085 bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
3089 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
3091 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3092 u64 end = offset + len;
3093 u64 block_start = round_up(offset, block_bytes(c));
3094 u64 block_end = round_down(end, block_bytes(c));
3095 bool truncated_last_page;
3098 ret = bch2_truncate_folios(inode, offset, end);
3099 if (unlikely(ret < 0))
3102 truncated_last_page = ret;
3104 truncate_pagecache_range(&inode->v, offset, end - 1);
3106 if (block_start < block_end) {
3107 s64 i_sectors_delta = 0;
3109 ret = bch2_fpunch(c, inode_inum(inode),
3110 block_start >> 9, block_end >> 9,
3112 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3115 mutex_lock(&inode->ei_update_lock);
3116 if (end >= inode->v.i_size && !truncated_last_page) {
3117 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
3118 ATTR_MTIME|ATTR_CTIME);
3120 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3121 ATTR_MTIME|ATTR_CTIME);
3123 mutex_unlock(&inode->ei_update_lock);
3128 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
3129 loff_t offset, loff_t len,
3132 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3133 struct address_space *mapping = inode->v.i_mapping;
3134 struct bkey_buf copy;
3135 struct btree_trans trans;
3136 struct btree_iter src, dst, del;
3137 loff_t shift, new_size;
3141 if ((offset | len) & (block_bytes(c) - 1))
3145 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
3148 if (offset >= inode->v.i_size)
3151 src_start = U64_MAX;
3154 if (offset + len >= inode->v.i_size)
3157 src_start = offset + len;
3161 new_size = inode->v.i_size + shift;
3163 ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
3168 i_size_write(&inode->v, new_size);
3169 mutex_lock(&inode->ei_update_lock);
3170 ret = bch2_write_inode_size(c, inode, new_size,
3171 ATTR_MTIME|ATTR_CTIME);
3172 mutex_unlock(&inode->ei_update_lock);
3174 s64 i_sectors_delta = 0;
3176 ret = bch2_fpunch(c, inode_inum(inode),
3177 offset >> 9, (offset + len) >> 9,
3179 i_sectors_acct(c, inode, NULL, i_sectors_delta);
3185 bch2_bkey_buf_init(©);
3186 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
3187 bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
3188 POS(inode->v.i_ino, src_start >> 9),
3190 bch2_trans_copy_iter(&dst, &src);
3191 bch2_trans_copy_iter(&del, &src);
3194 bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
3195 struct disk_reservation disk_res =
3196 bch2_disk_reservation_init(c, 0);
3197 struct bkey_i delete;
3199 struct bpos next_pos;
3200 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
3201 struct bpos atomic_end;
3202 unsigned trigger_flags = 0;
3205 bch2_trans_begin(&trans);
3207 ret = bch2_subvolume_get_snapshot(&trans,
3208 inode->ei_subvol, &snapshot);
3212 bch2_btree_iter_set_snapshot(&src, snapshot);
3213 bch2_btree_iter_set_snapshot(&dst, snapshot);
3214 bch2_btree_iter_set_snapshot(&del, snapshot);
3216 bch2_trans_begin(&trans);
3219 ? bch2_btree_iter_peek_prev(&src)
3220 : bch2_btree_iter_peek_upto(&src, POS(inode->v.i_ino, U64_MAX));
3221 if ((ret = bkey_err(k)))
3224 if (!k.k || k.k->p.inode != inode->v.i_ino)
3228 bkey_le(k.k->p, POS(inode->v.i_ino, offset >> 9)))
3231 bch2_bkey_buf_reassemble(©, c, k);
3234 bkey_lt(bkey_start_pos(k.k), move_pos))
3235 bch2_cut_front(move_pos, copy.k);
3237 copy.k->k.p.offset += shift >> 9;
3238 bch2_btree_iter_set_pos(&dst, bkey_start_pos(©.k->k));
3240 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
3244 if (!bkey_eq(atomic_end, copy.k->k.p)) {
3246 move_pos = atomic_end;
3247 move_pos.offset -= shift >> 9;
3250 bch2_cut_back(atomic_end, copy.k);
3254 bkey_init(&delete.k);
3255 delete.k.p = copy.k->k.p;
3256 delete.k.size = copy.k->k.size;
3257 delete.k.p.offset -= shift >> 9;
3258 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
3260 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
3262 if (copy.k->k.size != k.k->size) {
3263 /* We might end up splitting compressed extents: */
3265 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
3267 ret = bch2_disk_reservation_get(c, &disk_res,
3268 copy.k->k.size, nr_ptrs,
3269 BCH_DISK_RESERVATION_NOFAIL);
3273 ret = bch2_btree_iter_traverse(&del) ?:
3274 bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
3275 bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
3276 bch2_trans_commit(&trans, &disk_res, NULL,
3277 BTREE_INSERT_NOFAIL);
3278 bch2_disk_reservation_put(c, &disk_res);
3281 bch2_btree_iter_set_pos(&src, next_pos);
3283 bch2_trans_iter_exit(&trans, &del);
3284 bch2_trans_iter_exit(&trans, &dst);
3285 bch2_trans_iter_exit(&trans, &src);
3286 bch2_trans_exit(&trans);
3287 bch2_bkey_buf_exit(©, c);
3292 mutex_lock(&inode->ei_update_lock);
3294 i_size_write(&inode->v, new_size);
3295 ret = bch2_write_inode_size(c, inode, new_size,
3296 ATTR_MTIME|ATTR_CTIME);
3298 /* We need an inode update to update bi_journal_seq for fsync: */
3299 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
3300 ATTR_MTIME|ATTR_CTIME);
3302 mutex_unlock(&inode->ei_update_lock);
3306 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
3307 u64 start_sector, u64 end_sector)
3309 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3310 struct btree_trans trans;
3311 struct btree_iter iter;
3312 struct bpos end_pos = POS(inode->v.i_ino, end_sector);
3313 struct bch_io_opts opts;
3316 bch2_inode_opts_get(&opts, c, &inode->ei_inode);
3317 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
3319 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3320 POS(inode->v.i_ino, start_sector),
3321 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
3323 while (!ret && bkey_lt(iter.pos, end_pos)) {
3324 s64 i_sectors_delta = 0;
3325 struct quota_res quota_res = { 0 };
3330 bch2_trans_begin(&trans);
3332 ret = bch2_subvolume_get_snapshot(&trans,
3333 inode->ei_subvol, &snapshot);
3337 bch2_btree_iter_set_snapshot(&iter, snapshot);
3339 k = bch2_btree_iter_peek_slot(&iter);
3340 if ((ret = bkey_err(k)))
3343 /* already reserved */
3344 if (bkey_extent_is_reservation(k) &&
3345 bch2_bkey_nr_ptrs_fully_allocated(k) >= opts.data_replicas) {
3346 bch2_btree_iter_advance(&iter);
3350 if (bkey_extent_is_data(k.k) &&
3351 !(mode & FALLOC_FL_ZERO_RANGE)) {
3352 bch2_btree_iter_advance(&iter);
3357 * XXX: for nocow mode, we should promote shared extents to
3361 sectors = bpos_min(k.k->p, end_pos).offset - iter.pos.offset;
3363 if (!bkey_extent_is_allocation(k.k)) {
3364 ret = bch2_quota_reservation_add(c, inode,
3371 ret = bch2_extent_fallocate(&trans, inode_inum(inode), &iter,
3372 sectors, opts, &i_sectors_delta,
3373 writepoint_hashed((unsigned long) current));
3377 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3379 bch2_quota_reservation_put(c, inode, "a_res);
3380 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3384 bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3385 mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3387 if (bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)) {
3388 struct quota_res quota_res = { 0 };
3389 s64 i_sectors_delta = 0;
3391 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3392 end_sector, &i_sectors_delta);
3393 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3394 bch2_quota_reservation_put(c, inode, "a_res);
3397 bch2_trans_iter_exit(&trans, &iter);
3398 bch2_trans_exit(&trans);
3402 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3403 loff_t offset, loff_t len)
3405 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3406 u64 end = offset + len;
3407 u64 block_start = round_down(offset, block_bytes(c));
3408 u64 block_end = round_up(end, block_bytes(c));
3409 bool truncated_last_page = false;
3412 if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3413 ret = inode_newsize_ok(&inode->v, end);
3418 if (mode & FALLOC_FL_ZERO_RANGE) {
3419 ret = bch2_truncate_folios(inode, offset, end);
3420 if (unlikely(ret < 0))
3423 truncated_last_page = ret;
3425 truncate_pagecache_range(&inode->v, offset, end - 1);
3427 block_start = round_up(offset, block_bytes(c));
3428 block_end = round_down(end, block_bytes(c));
3431 ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3434 * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3435 * so that the VFS cache i_size is consistent with the btree i_size:
3438 !(bch2_err_matches(ret, ENOSPC) && (mode & FALLOC_FL_ZERO_RANGE)))
3441 if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3442 end = inode->v.i_size;
3444 if (end >= inode->v.i_size &&
3445 (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3446 !(mode & FALLOC_FL_KEEP_SIZE))) {
3447 spin_lock(&inode->v.i_lock);
3448 i_size_write(&inode->v, end);
3449 spin_unlock(&inode->v.i_lock);
3451 mutex_lock(&inode->ei_update_lock);
3452 ret2 = bch2_write_inode_size(c, inode, end, 0);
3453 mutex_unlock(&inode->ei_update_lock);
3459 long bch2_fallocate_dispatch(struct file *file, int mode,
3460 loff_t offset, loff_t len)
3462 struct bch_inode_info *inode = file_bch_inode(file);
3463 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3466 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fallocate))
3469 inode_lock(&inode->v);
3470 inode_dio_wait(&inode->v);
3471 bch2_pagecache_block_get(inode);
3473 ret = file_modified(file);
3477 if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3478 ret = bchfs_fallocate(inode, mode, offset, len);
3479 else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3480 ret = bchfs_fpunch(inode, offset, len);
3481 else if (mode == FALLOC_FL_INSERT_RANGE)
3482 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3483 else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3484 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3488 bch2_pagecache_block_put(inode);
3489 inode_unlock(&inode->v);
3490 bch2_write_ref_put(c, BCH_WRITE_REF_fallocate);
3492 return bch2_err_class(ret);
3496 * Take a quota reservation for unallocated blocks in a given file range
3497 * Does not check pagecache
3499 static int quota_reserve_range(struct bch_inode_info *inode,
3500 struct quota_res *res,
3503 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3504 struct btree_trans trans;
3505 struct btree_iter iter;
3508 u64 sectors = end - start;
3512 bch2_trans_init(&trans, c, 0, 0);
3514 bch2_trans_begin(&trans);
3516 ret = bch2_subvolume_get_snapshot(&trans, inode->ei_subvol, &snapshot);
3520 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
3521 SPOS(inode->v.i_ino, pos, snapshot), 0);
3523 while (!(ret = btree_trans_too_many_iters(&trans)) &&
3524 (k = bch2_btree_iter_peek_upto(&iter, POS(inode->v.i_ino, end - 1))).k &&
3525 !(ret = bkey_err(k))) {
3526 if (bkey_extent_is_allocation(k.k)) {
3527 u64 s = min(end, k.k->p.offset) -
3528 max(start, bkey_start_offset(k.k));
3529 BUG_ON(s > sectors);
3532 bch2_btree_iter_advance(&iter);
3534 pos = iter.pos.offset;
3535 bch2_trans_iter_exit(&trans, &iter);
3537 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3540 bch2_trans_exit(&trans);
3545 return bch2_quota_reservation_add(c, inode, res, sectors, true);
3548 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3549 struct file *file_dst, loff_t pos_dst,
3550 loff_t len, unsigned remap_flags)
3552 struct bch_inode_info *src = file_bch_inode(file_src);
3553 struct bch_inode_info *dst = file_bch_inode(file_dst);
3554 struct bch_fs *c = src->v.i_sb->s_fs_info;
3555 struct quota_res quota_res = { 0 };
3556 s64 i_sectors_delta = 0;
3560 if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3563 if (remap_flags & REMAP_FILE_DEDUP)
3566 if ((pos_src & (block_bytes(c) - 1)) ||
3567 (pos_dst & (block_bytes(c) - 1)))
3571 abs(pos_src - pos_dst) < len)
3574 bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3576 inode_dio_wait(&src->v);
3577 inode_dio_wait(&dst->v);
3579 ret = generic_remap_file_range_prep(file_src, pos_src,
3582 if (ret < 0 || len == 0)
3585 aligned_len = round_up((u64) len, block_bytes(c));
3587 ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3588 pos_dst, pos_dst + len - 1);
3592 ret = quota_reserve_range(dst, "a_res, pos_dst >> 9,
3593 (pos_dst + aligned_len) >> 9);
3597 file_update_time(file_dst);
3599 mark_pagecache_unallocated(src, pos_src >> 9,
3600 (pos_src + aligned_len) >> 9);
3602 ret = bch2_remap_range(c,
3603 inode_inum(dst), pos_dst >> 9,
3604 inode_inum(src), pos_src >> 9,
3606 pos_dst + len, &i_sectors_delta);
3611 * due to alignment, we might have remapped slightly more than requsted
3613 ret = min((u64) ret << 9, (u64) len);
3615 i_sectors_acct(c, dst, "a_res, i_sectors_delta);
3617 spin_lock(&dst->v.i_lock);
3618 if (pos_dst + ret > dst->v.i_size)
3619 i_size_write(&dst->v, pos_dst + ret);
3620 spin_unlock(&dst->v.i_lock);
3622 if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3623 IS_SYNC(file_inode(file_dst)))
3624 ret = bch2_flush_inode(c, dst);
3626 bch2_quota_reservation_put(c, dst, "a_res);
3627 bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3629 return bch2_err_class(ret);
3634 static int folio_data_offset(struct folio *folio, loff_t pos)
3636 struct bch_folio *s = bch2_folio(folio);
3637 unsigned i, sectors = folio_sectors(folio);
3640 for (i = folio_pos_to_s(folio, pos); i < sectors; i++)
3641 if (s->s[i].state >= SECTOR_dirty)
3642 return i << SECTOR_SHIFT;
3647 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3648 loff_t start_offset,
3651 struct folio_batch fbatch;
3652 pgoff_t start_index = start_offset >> PAGE_SHIFT;
3653 pgoff_t end_index = end_offset >> PAGE_SHIFT;
3654 pgoff_t index = start_index;
3659 folio_batch_init(&fbatch);
3661 while (filemap_get_folios(vinode->i_mapping,
3662 &index, end_index, &fbatch)) {
3663 for (i = 0; i < folio_batch_count(&fbatch); i++) {
3664 struct folio *folio = fbatch.folios[i];
3667 offset = folio_data_offset(folio,
3668 max(folio_pos(folio), start_offset));
3670 ret = clamp(folio_pos(folio) + offset,
3671 start_offset, end_offset);
3672 folio_unlock(folio);
3673 folio_batch_release(&fbatch);
3676 folio_unlock(folio);
3678 folio_batch_release(&fbatch);
3685 static loff_t bch2_seek_data(struct file *file, u64 offset)
3687 struct bch_inode_info *inode = file_bch_inode(file);
3688 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3689 struct btree_trans trans;
3690 struct btree_iter iter;
3692 subvol_inum inum = inode_inum(inode);
3693 u64 isize, next_data = MAX_LFS_FILESIZE;
3697 isize = i_size_read(&inode->v);
3698 if (offset >= isize)
3701 bch2_trans_init(&trans, c, 0, 0);
3703 bch2_trans_begin(&trans);
3705 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3709 for_each_btree_key_upto_norestart(&trans, iter, BTREE_ID_extents,
3710 SPOS(inode->v.i_ino, offset >> 9, snapshot),
3711 POS(inode->v.i_ino, U64_MAX),
3713 if (bkey_extent_is_data(k.k)) {
3714 next_data = max(offset, bkey_start_offset(k.k) << 9);
3716 } else if (k.k->p.offset >> 9 > isize)
3719 bch2_trans_iter_exit(&trans, &iter);
3721 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3724 bch2_trans_exit(&trans);
3728 if (next_data > offset)
3729 next_data = bch2_seek_pagecache_data(&inode->v,
3732 if (next_data >= isize)
3735 return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3738 static bool folio_hole_offset(struct address_space *mapping, loff_t *offset)
3740 struct folio *folio;
3741 struct bch_folio *s;
3742 unsigned i, sectors;
3745 folio = filemap_lock_folio(mapping, *offset >> PAGE_SHIFT);
3749 s = bch2_folio(folio);
3753 sectors = folio_sectors(folio);
3754 for (i = folio_pos_to_s(folio, *offset); i < sectors; i++)
3755 if (s->s[i].state < SECTOR_dirty) {
3756 *offset = max(*offset,
3757 folio_pos(folio) + (i << SECTOR_SHIFT));
3761 *offset = folio_end_pos(folio);
3764 folio_unlock(folio);
3768 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3769 loff_t start_offset,
3772 struct address_space *mapping = vinode->i_mapping;
3773 loff_t offset = start_offset;
3775 while (offset < end_offset &&
3776 !folio_hole_offset(mapping, &offset))
3779 return min(offset, end_offset);
3782 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3784 struct bch_inode_info *inode = file_bch_inode(file);
3785 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3786 struct btree_trans trans;
3787 struct btree_iter iter;
3789 subvol_inum inum = inode_inum(inode);
3790 u64 isize, next_hole = MAX_LFS_FILESIZE;
3794 isize = i_size_read(&inode->v);
3795 if (offset >= isize)
3798 bch2_trans_init(&trans, c, 0, 0);
3800 bch2_trans_begin(&trans);
3802 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3806 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3807 SPOS(inode->v.i_ino, offset >> 9, snapshot),
3808 BTREE_ITER_SLOTS, k, ret) {
3809 if (k.k->p.inode != inode->v.i_ino) {
3810 next_hole = bch2_seek_pagecache_hole(&inode->v,
3811 offset, MAX_LFS_FILESIZE);
3813 } else if (!bkey_extent_is_data(k.k)) {
3814 next_hole = bch2_seek_pagecache_hole(&inode->v,
3815 max(offset, bkey_start_offset(k.k) << 9),
3816 k.k->p.offset << 9);
3818 if (next_hole < k.k->p.offset << 9)
3821 offset = max(offset, bkey_start_offset(k.k) << 9);
3824 bch2_trans_iter_exit(&trans, &iter);
3826 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
3829 bch2_trans_exit(&trans);
3833 if (next_hole > isize)
3836 return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3839 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3847 ret = generic_file_llseek(file, offset, whence);
3850 ret = bch2_seek_data(file, offset);
3853 ret = bch2_seek_hole(file, offset);
3860 return bch2_err_class(ret);
3863 void bch2_fs_fsio_exit(struct bch_fs *c)
3865 bioset_exit(&c->nocow_flush_bioset);
3866 bioset_exit(&c->dio_write_bioset);
3867 bioset_exit(&c->dio_read_bioset);
3868 bioset_exit(&c->writepage_bioset);
3871 int bch2_fs_fsio_init(struct bch_fs *c)
3875 pr_verbose_init(c->opts, "");
3877 if (bioset_init(&c->writepage_bioset,
3878 4, offsetof(struct bch_writepage_io, op.wbio.bio),
3880 return -BCH_ERR_ENOMEM_writepage_bioset_init;
3882 if (bioset_init(&c->dio_read_bioset,
3883 4, offsetof(struct dio_read, rbio.bio),
3885 return -BCH_ERR_ENOMEM_dio_read_bioset_init;
3887 if (bioset_init(&c->dio_write_bioset,
3888 4, offsetof(struct dio_write, op.wbio.bio),
3890 return -BCH_ERR_ENOMEM_dio_write_bioset_init;
3892 if (bioset_init(&c->nocow_flush_bioset,
3893 1, offsetof(struct nocow_flush, bio), 0))
3894 return -BCH_ERR_ENOMEM_nocow_flush_bioset_init;
3896 pr_verbose_init(c->opts, "ret %i", ret);
3900 #endif /* NO_BCACHEFS_FS */