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>
38 static inline bool bio_full(struct bio *bio, unsigned len)
40 if (bio->bi_vcnt >= bio->bi_max_vecs)
42 if (bio->bi_iter.bi_size > UINT_MAX - len)
47 static inline struct address_space *faults_disabled_mapping(void)
49 return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
52 static inline void set_fdm_dropped_locks(void)
54 current->faults_disabled_mapping =
55 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
58 static inline bool fdm_dropped_locks(void)
60 return ((unsigned long) current->faults_disabled_mapping) & 1;
67 struct bch_writepage_io {
69 struct bch_inode_info *inode;
72 struct bch_write_op op;
76 struct completion done;
82 struct quota_res quota_res;
86 struct iovec inline_vecs[2];
89 struct bch_write_op op;
97 struct bch_read_bio rbio;
100 /* pagecache_block must be held */
101 static int write_invalidate_inode_pages_range(struct address_space *mapping,
102 loff_t start, loff_t end)
107 * XXX: the way this is currently implemented, we can spin if a process
108 * is continually redirtying a specific page
111 if (!mapping->nrpages)
114 ret = filemap_write_and_wait_range(mapping, start, end);
118 if (!mapping->nrpages)
121 ret = invalidate_inode_pages2_range(mapping,
124 } while (ret == -EBUSY);
131 #ifdef CONFIG_BCACHEFS_QUOTA
133 static void bch2_quota_reservation_put(struct bch_fs *c,
134 struct bch_inode_info *inode,
135 struct quota_res *res)
140 mutex_lock(&inode->ei_quota_lock);
141 BUG_ON(res->sectors > inode->ei_quota_reserved);
143 bch2_quota_acct(c, inode->ei_qid, Q_SPC,
144 -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
145 inode->ei_quota_reserved -= res->sectors;
146 mutex_unlock(&inode->ei_quota_lock);
151 static int bch2_quota_reservation_add(struct bch_fs *c,
152 struct bch_inode_info *inode,
153 struct quota_res *res,
159 mutex_lock(&inode->ei_quota_lock);
160 ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
161 check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
163 inode->ei_quota_reserved += sectors;
164 res->sectors += sectors;
166 mutex_unlock(&inode->ei_quota_lock);
173 static void bch2_quota_reservation_put(struct bch_fs *c,
174 struct bch_inode_info *inode,
175 struct quota_res *res)
179 static int bch2_quota_reservation_add(struct bch_fs *c,
180 struct bch_inode_info *inode,
181 struct quota_res *res,
190 /* i_size updates: */
192 struct inode_new_size {
198 static int inode_set_size(struct bch_inode_info *inode,
199 struct bch_inode_unpacked *bi,
202 struct inode_new_size *s = p;
204 bi->bi_size = s->new_size;
205 if (s->fields & ATTR_ATIME)
206 bi->bi_atime = s->now;
207 if (s->fields & ATTR_MTIME)
208 bi->bi_mtime = s->now;
209 if (s->fields & ATTR_CTIME)
210 bi->bi_ctime = s->now;
215 int __must_check bch2_write_inode_size(struct bch_fs *c,
216 struct bch_inode_info *inode,
217 loff_t new_size, unsigned fields)
219 struct inode_new_size s = {
220 .new_size = new_size,
221 .now = bch2_current_time(c),
225 return bch2_write_inode(c, inode, inode_set_size, &s, fields);
228 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
229 struct quota_res *quota_res, s64 sectors)
234 mutex_lock(&inode->ei_quota_lock);
235 BUG_ON((s64) inode->v.i_blocks + sectors < 0);
236 inode->v.i_blocks += sectors;
238 #ifdef CONFIG_BCACHEFS_QUOTA
239 if (quota_res && sectors > 0) {
240 BUG_ON(sectors > quota_res->sectors);
241 BUG_ON(sectors > inode->ei_quota_reserved);
243 quota_res->sectors -= sectors;
244 inode->ei_quota_reserved -= sectors;
246 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
249 mutex_unlock(&inode->ei_quota_lock);
254 /* stored in page->private: */
256 struct bch_page_sector {
257 /* Uncompressed, fully allocated replicas (or on disk reservation): */
258 unsigned nr_replicas:4;
260 /* Owns PAGE_SECTORS * replicas_reserved sized in memory reservation: */
261 unsigned replicas_reserved:4;
268 SECTOR_DIRTY_RESERVED,
273 struct bch_page_state {
275 atomic_t write_count;
277 struct bch_page_sector s[PAGE_SECTORS];
280 static inline struct bch_page_state *__bch2_page_state(struct page *page)
282 return page_has_private(page)
283 ? (struct bch_page_state *) page_private(page)
287 static inline struct bch_page_state *bch2_page_state(struct page *page)
289 EBUG_ON(!PageLocked(page));
291 return __bch2_page_state(page);
294 /* for newly allocated pages: */
295 static void __bch2_page_state_release(struct page *page)
297 kfree(detach_page_private(page));
300 static void bch2_page_state_release(struct page *page)
302 EBUG_ON(!PageLocked(page));
303 __bch2_page_state_release(page);
306 /* for newly allocated pages: */
307 static struct bch_page_state *__bch2_page_state_create(struct page *page,
310 struct bch_page_state *s;
312 s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
316 spin_lock_init(&s->lock);
317 attach_page_private(page, s);
321 static struct bch_page_state *bch2_page_state_create(struct page *page,
324 return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
327 static unsigned bkey_to_sector_state(const struct bkey *k)
329 if (k->type == KEY_TYPE_reservation)
330 return SECTOR_RESERVED;
331 if (bkey_extent_is_allocation(k))
332 return SECTOR_ALLOCATED;
333 return SECTOR_UNALLOCATED;
336 static void __bch2_page_state_set(struct page *page,
337 unsigned pg_offset, unsigned pg_len,
338 unsigned nr_ptrs, unsigned state)
340 struct bch_page_state *s = bch2_page_state_create(page, __GFP_NOFAIL);
343 BUG_ON(pg_offset >= PAGE_SECTORS);
344 BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
348 for (i = pg_offset; i < pg_offset + pg_len; i++) {
349 s->s[i].nr_replicas = nr_ptrs;
350 s->s[i].state = state;
353 if (i == PAGE_SECTORS)
356 spin_unlock(&s->lock);
359 static int bch2_page_state_set(struct bch_fs *c, subvol_inum inum,
360 struct page **pages, unsigned nr_pages)
362 struct btree_trans trans;
363 struct btree_iter iter;
365 u64 offset = pages[0]->index << PAGE_SECTORS_SHIFT;
370 bch2_trans_init(&trans, c, 0, 0);
372 bch2_trans_begin(&trans);
374 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
378 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
379 SPOS(inum.inum, offset, snapshot),
380 BTREE_ITER_SLOTS, k, ret) {
381 unsigned nr_ptrs = bch2_bkey_nr_ptrs_fully_allocated(k);
382 unsigned state = bkey_to_sector_state(k.k);
384 while (pg_idx < nr_pages) {
385 struct page *page = pages[pg_idx];
386 u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
387 u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
388 unsigned pg_offset = max(bkey_start_offset(k.k), pg_start) - pg_start;
389 unsigned pg_len = min(k.k->p.offset, pg_end) - pg_offset - pg_start;
391 BUG_ON(k.k->p.offset < pg_start);
392 BUG_ON(bkey_start_offset(k.k) > pg_end);
394 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate)
395 __bch2_page_state_set(page, pg_offset, pg_len, nr_ptrs, state);
397 if (k.k->p.offset < pg_end)
402 if (pg_idx == nr_pages)
406 offset = iter.pos.offset;
407 bch2_trans_iter_exit(&trans, &iter);
411 bch2_trans_exit(&trans);
416 static void bch2_bio_page_state_set(struct bio *bio, struct bkey_s_c k)
418 struct bvec_iter iter;
420 unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
421 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
422 unsigned state = bkey_to_sector_state(k.k);
424 bio_for_each_segment(bv, bio, iter)
425 __bch2_page_state_set(bv.bv_page, bv.bv_offset >> 9,
426 bv.bv_len >> 9, nr_ptrs, state);
429 static void mark_pagecache_unallocated(struct bch_inode_info *inode,
432 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
433 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
442 unsigned nr_pages, i, j;
444 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
446 for (i = 0; i < nr_pages; i++) {
447 struct page *page = pvec.pages[i];
448 u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
449 u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
450 unsigned pg_offset = max(start, pg_start) - pg_start;
451 unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
452 struct bch_page_state *s;
454 BUG_ON(end <= pg_start);
455 BUG_ON(pg_offset >= PAGE_SECTORS);
456 BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
459 s = bch2_page_state(page);
463 for (j = pg_offset; j < pg_offset + pg_len; j++)
464 s->s[j].nr_replicas = 0;
465 spin_unlock(&s->lock);
470 pagevec_release(&pvec);
471 } while (index <= end_index);
474 static void mark_pagecache_reserved(struct bch_inode_info *inode,
477 struct bch_fs *c = inode->v.i_sb->s_fs_info;
478 pgoff_t index = start >> PAGE_SECTORS_SHIFT;
479 pgoff_t end_index = (end - 1) >> PAGE_SECTORS_SHIFT;
481 s64 i_sectors_delta = 0;
489 unsigned nr_pages, i, j;
491 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
493 for (i = 0; i < nr_pages; i++) {
494 struct page *page = pvec.pages[i];
495 u64 pg_start = page->index << PAGE_SECTORS_SHIFT;
496 u64 pg_end = (page->index + 1) << PAGE_SECTORS_SHIFT;
497 unsigned pg_offset = max(start, pg_start) - pg_start;
498 unsigned pg_len = min(end, pg_end) - pg_offset - pg_start;
499 struct bch_page_state *s;
501 BUG_ON(end <= pg_start);
502 BUG_ON(pg_offset >= PAGE_SECTORS);
503 BUG_ON(pg_offset + pg_len > PAGE_SECTORS);
506 s = bch2_page_state(page);
510 for (j = pg_offset; j < pg_offset + pg_len; j++)
511 switch (s->s[j].state) {
512 case SECTOR_UNALLOCATED:
513 s->s[j].state = SECTOR_RESERVED;
516 s->s[j].state = SECTOR_DIRTY_RESERVED;
522 spin_unlock(&s->lock);
527 pagevec_release(&pvec);
528 } while (index <= end_index);
530 i_sectors_acct(c, inode, NULL, i_sectors_delta);
533 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
535 /* XXX: this should not be open coded */
536 return inode->ei_inode.bi_data_replicas
537 ? inode->ei_inode.bi_data_replicas - 1
538 : c->opts.data_replicas;
541 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
542 unsigned nr_replicas)
544 return max(0, (int) nr_replicas -
546 s->replicas_reserved);
549 static int bch2_get_page_disk_reservation(struct bch_fs *c,
550 struct bch_inode_info *inode,
551 struct page *page, bool check_enospc)
553 struct bch_page_state *s = bch2_page_state_create(page, 0);
554 unsigned nr_replicas = inode_nr_replicas(c, inode);
555 struct disk_reservation disk_res = { 0 };
556 unsigned i, disk_res_sectors = 0;
562 for (i = 0; i < ARRAY_SIZE(s->s); i++)
563 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
565 if (!disk_res_sectors)
568 ret = bch2_disk_reservation_get(c, &disk_res,
571 ? BCH_DISK_RESERVATION_NOFAIL
576 for (i = 0; i < ARRAY_SIZE(s->s); i++)
577 s->s[i].replicas_reserved +=
578 sectors_to_reserve(&s->s[i], nr_replicas);
583 struct bch2_page_reservation {
584 struct disk_reservation disk;
585 struct quota_res quota;
588 static void bch2_page_reservation_init(struct bch_fs *c,
589 struct bch_inode_info *inode,
590 struct bch2_page_reservation *res)
592 memset(res, 0, sizeof(*res));
594 res->disk.nr_replicas = inode_nr_replicas(c, inode);
597 static void bch2_page_reservation_put(struct bch_fs *c,
598 struct bch_inode_info *inode,
599 struct bch2_page_reservation *res)
601 bch2_disk_reservation_put(c, &res->disk);
602 bch2_quota_reservation_put(c, inode, &res->quota);
605 static int bch2_page_reservation_get(struct bch_fs *c,
606 struct bch_inode_info *inode, struct page *page,
607 struct bch2_page_reservation *res,
608 unsigned offset, unsigned len, bool check_enospc)
610 struct bch_page_state *s = bch2_page_state_create(page, 0);
611 unsigned i, disk_sectors = 0, quota_sectors = 0;
617 BUG_ON(!s->uptodate);
619 for (i = round_down(offset, block_bytes(c)) >> 9;
620 i < round_up(offset + len, block_bytes(c)) >> 9;
622 disk_sectors += sectors_to_reserve(&s->s[i],
623 res->disk.nr_replicas);
624 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
628 ret = bch2_disk_reservation_add(c, &res->disk,
631 ? BCH_DISK_RESERVATION_NOFAIL
638 ret = bch2_quota_reservation_add(c, inode, &res->quota,
642 struct disk_reservation tmp = {
643 .sectors = disk_sectors
646 bch2_disk_reservation_put(c, &tmp);
647 res->disk.sectors -= disk_sectors;
655 static void bch2_clear_page_bits(struct page *page)
657 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
658 struct bch_fs *c = inode->v.i_sb->s_fs_info;
659 struct bch_page_state *s = bch2_page_state(page);
660 struct disk_reservation disk_res = { 0 };
661 int i, dirty_sectors = 0;
666 EBUG_ON(!PageLocked(page));
667 EBUG_ON(PageWriteback(page));
669 for (i = 0; i < ARRAY_SIZE(s->s); i++) {
670 disk_res.sectors += s->s[i].replicas_reserved;
671 s->s[i].replicas_reserved = 0;
673 switch (s->s[i].state) {
675 s->s[i].state = SECTOR_UNALLOCATED;
678 case SECTOR_DIRTY_RESERVED:
679 s->s[i].state = SECTOR_RESERVED;
686 bch2_disk_reservation_put(c, &disk_res);
688 i_sectors_acct(c, inode, NULL, dirty_sectors);
690 bch2_page_state_release(page);
693 static void bch2_set_page_dirty(struct bch_fs *c,
694 struct bch_inode_info *inode, struct page *page,
695 struct bch2_page_reservation *res,
696 unsigned offset, unsigned len)
698 struct bch_page_state *s = bch2_page_state(page);
699 unsigned i, dirty_sectors = 0;
701 WARN_ON((u64) page_offset(page) + offset + len >
702 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
706 for (i = round_down(offset, block_bytes(c)) >> 9;
707 i < round_up(offset + len, block_bytes(c)) >> 9;
709 unsigned sectors = sectors_to_reserve(&s->s[i],
710 res->disk.nr_replicas);
713 * This can happen if we race with the error path in
714 * bch2_writepage_io_done():
716 sectors = min_t(unsigned, sectors, res->disk.sectors);
718 s->s[i].replicas_reserved += sectors;
719 res->disk.sectors -= sectors;
721 switch (s->s[i].state) {
722 case SECTOR_UNALLOCATED:
723 s->s[i].state = SECTOR_DIRTY;
726 case SECTOR_RESERVED:
727 s->s[i].state = SECTOR_DIRTY_RESERVED;
734 spin_unlock(&s->lock);
736 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
738 if (!PageDirty(page))
739 __set_page_dirty_nobuffers(page);
742 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
744 struct file *file = vmf->vma->vm_file;
745 struct address_space *mapping = file->f_mapping;
746 struct address_space *fdm = faults_disabled_mapping();
747 struct bch_inode_info *inode = file_bch_inode(file);
751 return VM_FAULT_SIGBUS;
755 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
757 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
760 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
762 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
763 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
765 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
767 /* Signal that lock has been dropped: */
768 set_fdm_dropped_locks();
769 return VM_FAULT_SIGBUS;
772 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
774 ret = filemap_fault(vmf);
775 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
780 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
782 struct page *page = vmf->page;
783 struct file *file = vmf->vma->vm_file;
784 struct bch_inode_info *inode = file_bch_inode(file);
785 struct address_space *mapping = file->f_mapping;
786 struct bch_fs *c = inode->v.i_sb->s_fs_info;
787 struct bch2_page_reservation res;
792 bch2_page_reservation_init(c, inode, &res);
794 sb_start_pagefault(inode->v.i_sb);
795 file_update_time(file);
798 * Not strictly necessary, but helps avoid dio writes livelocking in
799 * write_invalidate_inode_pages_range() - can drop this if/when we get
800 * a write_invalidate_inode_pages_range() that works without dropping
801 * page lock before invalidating page
803 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
806 isize = i_size_read(&inode->v);
808 if (page->mapping != mapping || page_offset(page) >= isize) {
810 ret = VM_FAULT_NOPAGE;
814 len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
816 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
817 if (bch2_page_state_set(c, inode_inum(inode), &page, 1)) {
819 ret = VM_FAULT_SIGBUS;
824 if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
826 ret = VM_FAULT_SIGBUS;
830 bch2_set_page_dirty(c, inode, page, &res, 0, len);
831 bch2_page_reservation_put(c, inode, &res);
833 wait_for_stable_page(page);
834 ret = VM_FAULT_LOCKED;
836 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
837 sb_end_pagefault(inode->v.i_sb);
842 void bch2_invalidatepage(struct page *page, unsigned int offset,
845 if (offset || length < PAGE_SIZE)
848 bch2_clear_page_bits(page);
851 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
856 bch2_clear_page_bits(page);
860 #ifdef CONFIG_MIGRATION
861 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
862 struct page *page, enum migrate_mode mode)
866 EBUG_ON(!PageLocked(page));
867 EBUG_ON(!PageLocked(newpage));
869 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
870 if (ret != MIGRATEPAGE_SUCCESS)
873 if (PagePrivate(page))
874 attach_page_private(newpage, detach_page_private(page));
876 if (mode != MIGRATE_SYNC_NO_COPY)
877 migrate_page_copy(newpage, page);
879 migrate_page_states(newpage, page);
880 return MIGRATEPAGE_SUCCESS;
886 static void bch2_readpages_end_io(struct bio *bio)
888 struct bvec_iter_all iter;
891 bio_for_each_segment_all(bv, bio, iter) {
892 struct page *page = bv->bv_page;
894 if (!bio->bi_status) {
895 SetPageUptodate(page);
897 ClearPageUptodate(page);
906 struct readpages_iter {
907 struct address_space *mapping;
914 static int readpages_iter_init(struct readpages_iter *iter,
915 struct readahead_control *ractl)
917 unsigned i, nr_pages = readahead_count(ractl);
919 memset(iter, 0, sizeof(*iter));
921 iter->mapping = ractl->mapping;
922 iter->offset = readahead_index(ractl);
923 iter->nr_pages = nr_pages;
925 iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
929 nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
930 for (i = 0; i < nr_pages; i++) {
931 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
932 put_page(iter->pages[i]);
938 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
940 if (iter->idx >= iter->nr_pages)
943 EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
945 return iter->pages[iter->idx];
948 static bool extent_partial_reads_expensive(struct bkey_s_c k)
950 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
951 struct bch_extent_crc_unpacked crc;
952 const union bch_extent_entry *i;
954 bkey_for_each_crc(k.k, ptrs, crc, i)
955 if (crc.csum_type || crc.compression_type)
960 static void readpage_bio_extend(struct readpages_iter *iter,
962 unsigned sectors_this_extent,
965 while (bio_sectors(bio) < sectors_this_extent &&
966 bio->bi_vcnt < bio->bi_max_vecs) {
967 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
968 struct page *page = readpage_iter_next(iter);
972 if (iter->offset + iter->idx != page_offset)
980 page = xa_load(&iter->mapping->i_pages, page_offset);
981 if (page && !xa_is_value(page))
984 page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
988 if (!__bch2_page_state_create(page, 0)) {
993 ret = add_to_page_cache_lru(page, iter->mapping,
994 page_offset, GFP_NOFS);
996 __bch2_page_state_release(page);
1004 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
1008 static void bchfs_read(struct btree_trans *trans,
1009 struct bch_read_bio *rbio,
1011 struct readpages_iter *readpages_iter)
1013 struct bch_fs *c = trans->c;
1014 struct btree_iter iter;
1016 int flags = BCH_READ_RETRY_IF_STALE|
1017 BCH_READ_MAY_PROMOTE;
1022 rbio->start_time = local_clock();
1023 rbio->subvol = inum.subvol;
1025 bch2_bkey_buf_init(&sk);
1027 bch2_trans_begin(trans);
1028 iter = (struct btree_iter) { NULL };
1030 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1034 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1035 SPOS(inum.inum, rbio->bio.bi_iter.bi_sector, snapshot),
1039 unsigned bytes, sectors, offset_into_extent;
1040 enum btree_id data_btree = BTREE_ID_extents;
1043 * read_extent -> io_time_reset may cause a transaction restart
1044 * without returning an error, we need to check for that here:
1046 if (!bch2_trans_relock(trans)) {
1051 bch2_btree_iter_set_pos(&iter,
1052 POS(inum.inum, rbio->bio.bi_iter.bi_sector));
1054 k = bch2_btree_iter_peek_slot(&iter);
1059 offset_into_extent = iter.pos.offset -
1060 bkey_start_offset(k.k);
1061 sectors = k.k->size - offset_into_extent;
1063 bch2_bkey_buf_reassemble(&sk, c, k);
1065 ret = bch2_read_indirect_extent(trans, &data_btree,
1066 &offset_into_extent, &sk);
1070 k = bkey_i_to_s_c(sk.k);
1072 sectors = min(sectors, k.k->size - offset_into_extent);
1075 readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
1076 extent_partial_reads_expensive(k));
1078 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
1079 swap(rbio->bio.bi_iter.bi_size, bytes);
1081 if (rbio->bio.bi_iter.bi_size == bytes)
1082 flags |= BCH_READ_LAST_FRAGMENT;
1084 bch2_bio_page_state_set(&rbio->bio, k);
1086 bch2_read_extent(trans, rbio, iter.pos,
1087 data_btree, k, offset_into_extent, flags);
1089 if (flags & BCH_READ_LAST_FRAGMENT)
1092 swap(rbio->bio.bi_iter.bi_size, bytes);
1093 bio_advance(&rbio->bio, bytes);
1095 ret = btree_trans_too_many_iters(trans);
1100 bch2_trans_iter_exit(trans, &iter);
1106 bch_err_inum_ratelimited(c, inum.inum,
1107 "read error %i from btree lookup", ret);
1108 rbio->bio.bi_status = BLK_STS_IOERR;
1109 bio_endio(&rbio->bio);
1112 bch2_bkey_buf_exit(&sk, c);
1115 void bch2_readahead(struct readahead_control *ractl)
1117 struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
1118 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1119 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1120 struct btree_trans trans;
1122 struct readpages_iter readpages_iter;
1125 ret = readpages_iter_init(&readpages_iter, ractl);
1128 bch2_trans_init(&trans, c, 0, 0);
1130 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1132 while ((page = readpage_iter_next(&readpages_iter))) {
1133 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
1134 unsigned n = min_t(unsigned,
1135 readpages_iter.nr_pages -
1138 struct bch_read_bio *rbio =
1139 rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
1142 readpages_iter.idx++;
1144 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
1145 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTORS_SHIFT;
1146 rbio->bio.bi_end_io = bch2_readpages_end_io;
1147 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1149 bchfs_read(&trans, rbio, inode_inum(inode),
1153 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1155 bch2_trans_exit(&trans);
1156 kfree(readpages_iter.pages);
1159 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
1160 subvol_inum inum, struct page *page)
1162 struct btree_trans trans;
1164 bch2_page_state_create(page, __GFP_NOFAIL);
1166 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
1167 rbio->bio.bi_iter.bi_sector =
1168 (sector_t) page->index << PAGE_SECTORS_SHIFT;
1169 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
1171 bch2_trans_init(&trans, c, 0, 0);
1172 bchfs_read(&trans, rbio, inum, NULL);
1173 bch2_trans_exit(&trans);
1176 int bch2_readpage(struct file *file, struct page *page)
1178 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1179 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1180 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1181 struct bch_read_bio *rbio;
1183 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
1184 rbio->bio.bi_end_io = bch2_readpages_end_io;
1186 __bchfs_readpage(c, rbio, inode_inum(inode), page);
1190 static void bch2_read_single_page_end_io(struct bio *bio)
1192 complete(bio->bi_private);
1195 static int bch2_read_single_page(struct page *page,
1196 struct address_space *mapping)
1198 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1199 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1200 struct bch_read_bio *rbio;
1202 DECLARE_COMPLETION_ONSTACK(done);
1204 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
1205 io_opts(c, &inode->ei_inode));
1206 rbio->bio.bi_private = &done;
1207 rbio->bio.bi_end_io = bch2_read_single_page_end_io;
1209 __bchfs_readpage(c, rbio, inode_inum(inode), page);
1210 wait_for_completion(&done);
1212 ret = blk_status_to_errno(rbio->bio.bi_status);
1213 bio_put(&rbio->bio);
1218 SetPageUptodate(page);
1224 struct bch_writepage_state {
1225 struct bch_writepage_io *io;
1226 struct bch_io_opts opts;
1229 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1230 struct bch_inode_info *inode)
1232 return (struct bch_writepage_state) {
1233 .opts = io_opts(c, &inode->ei_inode)
1237 static void bch2_writepage_io_free(struct closure *cl)
1239 struct bch_writepage_io *io = container_of(cl,
1240 struct bch_writepage_io, cl);
1242 bio_put(&io->op.wbio.bio);
1245 static void bch2_writepage_io_done(struct closure *cl)
1247 struct bch_writepage_io *io = container_of(cl,
1248 struct bch_writepage_io, cl);
1249 struct bch_fs *c = io->op.c;
1250 struct bio *bio = &io->op.wbio.bio;
1251 struct bvec_iter_all iter;
1252 struct bio_vec *bvec;
1255 up(&io->op.c->io_in_flight);
1258 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1260 bio_for_each_segment_all(bvec, bio, iter) {
1261 struct bch_page_state *s;
1263 SetPageError(bvec->bv_page);
1264 mapping_set_error(bvec->bv_page->mapping, -EIO);
1266 s = __bch2_page_state(bvec->bv_page);
1267 spin_lock(&s->lock);
1268 for (i = 0; i < PAGE_SECTORS; i++)
1269 s->s[i].nr_replicas = 0;
1270 spin_unlock(&s->lock);
1274 if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1275 bio_for_each_segment_all(bvec, bio, iter) {
1276 struct bch_page_state *s;
1278 s = __bch2_page_state(bvec->bv_page);
1279 spin_lock(&s->lock);
1280 for (i = 0; i < PAGE_SECTORS; i++)
1281 s->s[i].nr_replicas = 0;
1282 spin_unlock(&s->lock);
1287 * racing with fallocate can cause us to add fewer sectors than
1288 * expected - but we shouldn't add more sectors than expected:
1290 WARN_ON(io->op.i_sectors_delta > 0);
1293 * (error (due to going RO) halfway through a page can screw that up
1296 BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1300 * PageWriteback is effectively our ref on the inode - fixup i_blocks
1301 * before calling end_page_writeback:
1303 i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1305 bio_for_each_segment_all(bvec, bio, iter) {
1306 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1308 if (atomic_dec_and_test(&s->write_count))
1309 end_page_writeback(bvec->bv_page);
1312 closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1315 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1317 struct bch_writepage_io *io = w->io;
1319 down(&io->op.c->io_in_flight);
1322 closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1323 continue_at(&io->cl, bch2_writepage_io_done, NULL);
1327 * Get a bch_writepage_io and add @page to it - appending to an existing one if
1328 * possible, else allocating a new one:
1330 static void bch2_writepage_io_alloc(struct bch_fs *c,
1331 struct writeback_control *wbc,
1332 struct bch_writepage_state *w,
1333 struct bch_inode_info *inode,
1335 unsigned nr_replicas)
1337 struct bch_write_op *op;
1339 w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1340 &c->writepage_bioset),
1341 struct bch_writepage_io, op.wbio.bio);
1343 closure_init(&w->io->cl, NULL);
1344 w->io->inode = inode;
1347 bch2_write_op_init(op, c, w->opts);
1348 op->target = w->opts.foreground_target;
1349 op->nr_replicas = nr_replicas;
1350 op->res.nr_replicas = nr_replicas;
1351 op->write_point = writepoint_hashed(inode->ei_last_dirtied);
1352 op->subvol = inode->ei_subvol;
1353 op->pos = POS(inode->v.i_ino, sector);
1354 op->wbio.bio.bi_iter.bi_sector = sector;
1355 op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
1358 static int __bch2_writepage(struct page *page,
1359 struct writeback_control *wbc,
1362 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1363 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1364 struct bch_writepage_state *w = data;
1365 struct bch_page_state *s, orig;
1366 unsigned i, offset, nr_replicas_this_write = U32_MAX;
1367 loff_t i_size = i_size_read(&inode->v);
1368 pgoff_t end_index = i_size >> PAGE_SHIFT;
1371 EBUG_ON(!PageUptodate(page));
1373 /* Is the page fully inside i_size? */
1374 if (page->index < end_index)
1377 /* Is the page fully outside i_size? (truncate in progress) */
1378 offset = i_size & (PAGE_SIZE - 1);
1379 if (page->index > end_index || !offset) {
1385 * The page straddles i_size. It must be zeroed out on each and every
1386 * writepage invocation because it may be mmapped. "A file is mapped
1387 * in multiples of the page size. For a file that is not a multiple of
1388 * the page size, the remaining memory is zeroed when mapped, and
1389 * writes to that region are not written out to the file."
1391 zero_user_segment(page, offset, PAGE_SIZE);
1393 s = bch2_page_state_create(page, __GFP_NOFAIL);
1396 * Things get really hairy with errors during writeback:
1398 ret = bch2_get_page_disk_reservation(c, inode, page, false);
1401 /* Before unlocking the page, get copy of reservations: */
1402 spin_lock(&s->lock);
1404 spin_unlock(&s->lock);
1406 for (i = 0; i < PAGE_SECTORS; i++) {
1407 if (s->s[i].state < SECTOR_DIRTY)
1410 nr_replicas_this_write =
1411 min_t(unsigned, nr_replicas_this_write,
1412 s->s[i].nr_replicas +
1413 s->s[i].replicas_reserved);
1416 for (i = 0; i < PAGE_SECTORS; i++) {
1417 if (s->s[i].state < SECTOR_DIRTY)
1420 s->s[i].nr_replicas = w->opts.compression
1421 ? 0 : nr_replicas_this_write;
1423 s->s[i].replicas_reserved = 0;
1424 s->s[i].state = SECTOR_ALLOCATED;
1427 BUG_ON(atomic_read(&s->write_count));
1428 atomic_set(&s->write_count, 1);
1430 BUG_ON(PageWriteback(page));
1431 set_page_writeback(page);
1437 unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
1440 while (offset < PAGE_SECTORS &&
1441 orig.s[offset].state < SECTOR_DIRTY)
1444 if (offset == PAGE_SECTORS)
1447 while (offset + sectors < PAGE_SECTORS &&
1448 orig.s[offset + sectors].state >= SECTOR_DIRTY) {
1449 reserved_sectors += orig.s[offset + sectors].replicas_reserved;
1450 dirty_sectors += orig.s[offset + sectors].state == SECTOR_DIRTY;
1455 sector = ((u64) page->index << PAGE_SECTORS_SHIFT) + offset;
1458 (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1459 bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1460 w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1461 (BIO_MAX_VECS * PAGE_SIZE) ||
1462 bio_end_sector(&w->io->op.wbio.bio) != sector))
1463 bch2_writepage_do_io(w);
1466 bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1467 nr_replicas_this_write);
1469 atomic_inc(&s->write_count);
1471 BUG_ON(inode != w->io->inode);
1472 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1473 sectors << 9, offset << 9));
1475 /* Check for writing past i_size: */
1476 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1477 round_up(i_size, block_bytes(c)));
1479 w->io->op.res.sectors += reserved_sectors;
1480 w->io->op.i_sectors_delta -= dirty_sectors;
1481 w->io->op.new_i_size = i_size;
1486 if (atomic_dec_and_test(&s->write_count))
1487 end_page_writeback(page);
1492 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1494 struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1495 struct bch_writepage_state w =
1496 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1497 struct blk_plug plug;
1500 blk_start_plug(&plug);
1501 ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1503 bch2_writepage_do_io(&w);
1504 blk_finish_plug(&plug);
1508 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1510 struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1511 struct bch_writepage_state w =
1512 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1515 ret = __bch2_writepage(page, wbc, &w);
1517 bch2_writepage_do_io(&w);
1522 /* buffered writes: */
1524 int bch2_write_begin(struct file *file, struct address_space *mapping,
1525 loff_t pos, unsigned len, unsigned flags,
1526 struct page **pagep, void **fsdata)
1528 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1529 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1530 struct bch2_page_reservation *res;
1531 pgoff_t index = pos >> PAGE_SHIFT;
1532 unsigned offset = pos & (PAGE_SIZE - 1);
1536 res = kmalloc(sizeof(*res), GFP_KERNEL);
1540 bch2_page_reservation_init(c, inode, res);
1543 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1545 page = grab_cache_page_write_begin(mapping, index, flags);
1549 if (PageUptodate(page))
1552 /* If we're writing entire page, don't need to read it in first: */
1553 if (len == PAGE_SIZE)
1556 if (!offset && pos + len >= inode->v.i_size) {
1557 zero_user_segment(page, len, PAGE_SIZE);
1558 flush_dcache_page(page);
1562 if (index > inode->v.i_size >> PAGE_SHIFT) {
1563 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1564 flush_dcache_page(page);
1568 ret = bch2_read_single_page(page, mapping);
1572 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1573 ret = bch2_page_state_set(c, inode_inum(inode), &page, 1);
1578 ret = bch2_page_reservation_get(c, inode, page, res,
1581 if (!PageUptodate(page)) {
1583 * If the page hasn't been read in, we won't know if we
1584 * actually need a reservation - we don't actually need
1585 * to read here, we just need to check if the page is
1586 * fully backed by uncompressed data:
1601 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1607 int bch2_write_end(struct file *file, struct address_space *mapping,
1608 loff_t pos, unsigned len, unsigned copied,
1609 struct page *page, void *fsdata)
1611 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1612 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1613 struct bch2_page_reservation *res = fsdata;
1614 unsigned offset = pos & (PAGE_SIZE - 1);
1616 lockdep_assert_held(&inode->v.i_rwsem);
1618 if (unlikely(copied < len && !PageUptodate(page))) {
1620 * The page needs to be read in, but that would destroy
1621 * our partial write - simplest thing is to just force
1622 * userspace to redo the write:
1624 zero_user(page, 0, PAGE_SIZE);
1625 flush_dcache_page(page);
1629 spin_lock(&inode->v.i_lock);
1630 if (pos + copied > inode->v.i_size)
1631 i_size_write(&inode->v, pos + copied);
1632 spin_unlock(&inode->v.i_lock);
1635 if (!PageUptodate(page))
1636 SetPageUptodate(page);
1638 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1640 inode->ei_last_dirtied = (unsigned long) current;
1645 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1647 bch2_page_reservation_put(c, inode, res);
1653 #define WRITE_BATCH_PAGES 32
1655 static int __bch2_buffered_write(struct bch_inode_info *inode,
1656 struct address_space *mapping,
1657 struct iov_iter *iter,
1658 loff_t pos, unsigned len)
1660 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1661 struct page *pages[WRITE_BATCH_PAGES];
1662 struct bch2_page_reservation res;
1663 unsigned long index = pos >> PAGE_SHIFT;
1664 unsigned offset = pos & (PAGE_SIZE - 1);
1665 unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1666 unsigned i, reserved = 0, set_dirty = 0;
1667 unsigned copied = 0, nr_pages_copied = 0;
1671 BUG_ON(nr_pages > ARRAY_SIZE(pages));
1673 bch2_page_reservation_init(c, inode, &res);
1675 for (i = 0; i < nr_pages; i++) {
1676 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1683 len = min_t(unsigned, len,
1684 nr_pages * PAGE_SIZE - offset);
1689 if (offset && !PageUptodate(pages[0])) {
1690 ret = bch2_read_single_page(pages[0], mapping);
1695 if ((pos + len) & (PAGE_SIZE - 1) &&
1696 !PageUptodate(pages[nr_pages - 1])) {
1697 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1698 zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1700 ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1706 while (reserved < len) {
1707 unsigned i = (offset + reserved) >> PAGE_SHIFT;
1708 struct page *page = pages[i];
1709 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1710 unsigned pg_len = min_t(unsigned, len - reserved,
1711 PAGE_SIZE - pg_offset);
1713 if (!bch2_page_state_create(page, __GFP_NOFAIL)->uptodate) {
1714 ret = bch2_page_state_set(c, inode_inum(inode),
1715 pages + i, nr_pages - i);
1720 ret = bch2_page_reservation_get(c, inode, page, &res,
1721 pg_offset, pg_len, true);
1728 if (mapping_writably_mapped(mapping))
1729 for (i = 0; i < nr_pages; i++)
1730 flush_dcache_page(pages[i]);
1732 while (copied < len) {
1733 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1734 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1735 unsigned pg_len = min_t(unsigned, len - copied,
1736 PAGE_SIZE - pg_offset);
1737 unsigned pg_copied = copy_page_from_iter_atomic(page,
1738 pg_offset, pg_len,iter);
1743 if (!PageUptodate(page) &&
1744 pg_copied != PAGE_SIZE &&
1745 pos + copied + pg_copied < inode->v.i_size) {
1746 zero_user(page, 0, PAGE_SIZE);
1750 flush_dcache_page(page);
1751 copied += pg_copied;
1753 if (pg_copied != pg_len)
1760 spin_lock(&inode->v.i_lock);
1761 if (pos + copied > inode->v.i_size)
1762 i_size_write(&inode->v, pos + copied);
1763 spin_unlock(&inode->v.i_lock);
1765 while (set_dirty < copied) {
1766 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1767 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1768 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1769 PAGE_SIZE - pg_offset);
1771 if (!PageUptodate(page))
1772 SetPageUptodate(page);
1774 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1778 set_dirty += pg_len;
1781 nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1782 inode->ei_last_dirtied = (unsigned long) current;
1784 for (i = nr_pages_copied; i < nr_pages; i++) {
1785 unlock_page(pages[i]);
1789 bch2_page_reservation_put(c, inode, &res);
1791 return copied ?: ret;
1794 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1796 struct file *file = iocb->ki_filp;
1797 struct address_space *mapping = file->f_mapping;
1798 struct bch_inode_info *inode = file_bch_inode(file);
1799 loff_t pos = iocb->ki_pos;
1800 ssize_t written = 0;
1803 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1806 unsigned offset = pos & (PAGE_SIZE - 1);
1807 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1808 PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1811 * Bring in the user page that we will copy from _first_.
1812 * Otherwise there's a nasty deadlock on copying from the
1813 * same page as we're writing to, without it being marked
1816 * Not only is this an optimisation, but it is also required
1817 * to check that the address is actually valid, when atomic
1818 * usercopies are used, below.
1820 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1821 bytes = min_t(unsigned long, iov_iter_count(iter),
1822 PAGE_SIZE - offset);
1824 if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
1830 if (unlikely(fatal_signal_pending(current))) {
1835 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1836 if (unlikely(ret < 0))
1841 if (unlikely(ret == 0)) {
1843 * If we were unable to copy any data at all, we must
1844 * fall back to a single segment length write.
1846 * If we didn't fallback here, we could livelock
1847 * because not all segments in the iov can be copied at
1848 * once without a pagefault.
1850 bytes = min_t(unsigned long, PAGE_SIZE - offset,
1851 iov_iter_single_seg_count(iter));
1858 balance_dirty_pages_ratelimited(mapping);
1859 } while (iov_iter_count(iter));
1861 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1863 return written ? written : ret;
1866 /* O_DIRECT reads */
1868 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1871 bio_check_pages_dirty(bio);
1873 bio_release_pages(bio, false);
1878 static void bch2_dio_read_complete(struct closure *cl)
1880 struct dio_read *dio = container_of(cl, struct dio_read, cl);
1882 dio->req->ki_complete(dio->req, dio->ret);
1883 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1886 static void bch2_direct_IO_read_endio(struct bio *bio)
1888 struct dio_read *dio = bio->bi_private;
1891 dio->ret = blk_status_to_errno(bio->bi_status);
1893 closure_put(&dio->cl);
1896 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1898 struct dio_read *dio = bio->bi_private;
1899 bool should_dirty = dio->should_dirty;
1901 bch2_direct_IO_read_endio(bio);
1902 bio_check_or_release(bio, should_dirty);
1905 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1907 struct file *file = req->ki_filp;
1908 struct bch_inode_info *inode = file_bch_inode(file);
1909 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1910 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1911 struct dio_read *dio;
1913 loff_t offset = req->ki_pos;
1914 bool sync = is_sync_kiocb(req);
1918 if ((offset|iter->count) & (block_bytes(c) - 1))
1921 ret = min_t(loff_t, iter->count,
1922 max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1927 shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1928 iter->count -= shorten;
1930 bio = bio_alloc_bioset(GFP_KERNEL,
1931 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1932 &c->dio_read_bioset);
1934 bio->bi_end_io = bch2_direct_IO_read_endio;
1936 dio = container_of(bio, struct dio_read, rbio.bio);
1937 closure_init(&dio->cl, NULL);
1940 * this is a _really_ horrible hack just to avoid an atomic sub at the
1944 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1945 atomic_set(&dio->cl.remaining,
1946 CLOSURE_REMAINING_INITIALIZER -
1948 CLOSURE_DESTRUCTOR);
1950 atomic_set(&dio->cl.remaining,
1951 CLOSURE_REMAINING_INITIALIZER + 1);
1957 * This is one of the sketchier things I've encountered: we have to skip
1958 * the dirtying of requests that are internal from the kernel (i.e. from
1959 * loopback), because we'll deadlock on page_lock.
1961 dio->should_dirty = iter_is_iovec(iter);
1964 while (iter->count) {
1965 bio = bio_alloc_bioset(GFP_KERNEL,
1966 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
1968 bio->bi_end_io = bch2_direct_IO_read_split_endio;
1970 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1971 bio->bi_iter.bi_sector = offset >> 9;
1972 bio->bi_private = dio;
1974 ret = bio_iov_iter_get_pages(bio, iter);
1976 /* XXX: fault inject this path */
1977 bio->bi_status = BLK_STS_RESOURCE;
1982 offset += bio->bi_iter.bi_size;
1984 if (dio->should_dirty)
1985 bio_set_pages_dirty(bio);
1988 closure_get(&dio->cl);
1990 bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
1993 iter->count += shorten;
1996 closure_sync(&dio->cl);
1997 closure_debug_destroy(&dio->cl);
1999 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
2002 return -EIOCBQUEUED;
2006 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
2008 struct file *file = iocb->ki_filp;
2009 struct bch_inode_info *inode = file_bch_inode(file);
2010 struct address_space *mapping = file->f_mapping;
2011 size_t count = iov_iter_count(iter);
2015 return 0; /* skip atime */
2017 if (iocb->ki_flags & IOCB_DIRECT) {
2018 struct blk_plug plug;
2020 ret = filemap_write_and_wait_range(mapping,
2022 iocb->ki_pos + count - 1);
2026 file_accessed(file);
2028 blk_start_plug(&plug);
2029 ret = bch2_direct_IO_read(iocb, iter);
2030 blk_finish_plug(&plug);
2033 iocb->ki_pos += ret;
2035 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
2036 ret = generic_file_read_iter(iocb, iter);
2037 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
2043 /* O_DIRECT writes */
2045 static bool bch2_check_range_allocated(struct bch_fs *c, subvol_inum inum,
2046 u64 offset, u64 size,
2047 unsigned nr_replicas, bool compressed)
2049 struct btree_trans trans;
2050 struct btree_iter iter;
2052 u64 end = offset + size;
2057 bch2_trans_init(&trans, c, 0, 0);
2059 bch2_trans_begin(&trans);
2061 err = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2065 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
2066 SPOS(inum.inum, offset, snapshot),
2067 BTREE_ITER_SLOTS, k, err) {
2068 if (bkey_cmp(bkey_start_pos(k.k), POS(inum.inum, end)) >= 0)
2071 if (k.k->p.snapshot != snapshot ||
2072 nr_replicas > bch2_bkey_replicas(c, k) ||
2073 (!compressed && bch2_bkey_sectors_compressed(k))) {
2079 offset = iter.pos.offset;
2080 bch2_trans_iter_exit(&trans, &iter);
2084 bch2_trans_exit(&trans);
2086 return err ? false : ret;
2089 static void bch2_dio_write_loop_async(struct bch_write_op *);
2091 static long bch2_dio_write_loop(struct dio_write *dio)
2093 bool kthread = (current->flags & PF_KTHREAD) != 0;
2094 struct kiocb *req = dio->req;
2095 struct address_space *mapping = req->ki_filp->f_mapping;
2096 struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
2097 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2098 struct bio *bio = &dio->op.wbio.bio;
2099 struct bvec_iter_all iter;
2101 unsigned unaligned, iter_count;
2102 bool sync = dio->sync, dropped_locks;
2108 down(&c->io_in_flight);
2111 iter_count = dio->iter.count;
2113 if (kthread && dio->mm)
2114 kthread_use_mm(dio->mm);
2115 BUG_ON(current->faults_disabled_mapping);
2116 current->faults_disabled_mapping = mapping;
2118 ret = bio_iov_iter_get_pages(bio, &dio->iter);
2120 dropped_locks = fdm_dropped_locks();
2122 current->faults_disabled_mapping = NULL;
2123 if (kthread && dio->mm)
2124 kthread_unuse_mm(dio->mm);
2127 * If the fault handler returned an error but also signalled
2128 * that it dropped & retook ei_pagecache_lock, we just need to
2129 * re-shoot down the page cache and retry:
2131 if (dropped_locks && ret)
2134 if (unlikely(ret < 0))
2137 if (unlikely(dropped_locks)) {
2138 ret = write_invalidate_inode_pages_range(mapping,
2140 req->ki_pos + iter_count - 1);
2144 if (!bio->bi_iter.bi_size)
2148 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
2149 bio->bi_iter.bi_size -= unaligned;
2150 iov_iter_revert(&dio->iter, unaligned);
2152 if (!bio->bi_iter.bi_size) {
2154 * bio_iov_iter_get_pages was only able to get <
2155 * blocksize worth of pages:
2161 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
2162 dio->op.end_io = bch2_dio_write_loop_async;
2163 dio->op.target = dio->op.opts.foreground_target;
2164 dio->op.write_point = writepoint_hashed((unsigned long) current);
2165 dio->op.nr_replicas = dio->op.opts.data_replicas;
2166 dio->op.subvol = inode->ei_subvol;
2167 dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
2169 if ((req->ki_flags & IOCB_DSYNC) &&
2170 !c->opts.journal_flush_disabled)
2171 dio->op.flags |= BCH_WRITE_FLUSH;
2172 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
2174 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
2175 dio->op.opts.data_replicas, 0);
2176 if (unlikely(ret) &&
2177 !bch2_check_range_allocated(c, inode_inum(inode),
2178 dio->op.pos.offset, bio_sectors(bio),
2179 dio->op.opts.data_replicas,
2180 dio->op.opts.compression != 0))
2183 task_io_account_write(bio->bi_iter.bi_size);
2185 if (!dio->sync && !dio->loop && dio->iter.count) {
2186 struct iovec *iov = dio->inline_vecs;
2188 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
2189 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
2191 if (unlikely(!iov)) {
2192 dio->sync = sync = true;
2196 dio->free_iov = true;
2199 memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
2200 dio->iter.iov = iov;
2204 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
2207 wait_for_completion(&dio->done);
2209 return -EIOCBQUEUED;
2211 i_sectors_acct(c, inode, &dio->quota_res,
2212 dio->op.i_sectors_delta);
2213 req->ki_pos += (u64) dio->op.written << 9;
2214 dio->written += dio->op.written;
2216 spin_lock(&inode->v.i_lock);
2217 if (req->ki_pos > inode->v.i_size)
2218 i_size_write(&inode->v, req->ki_pos);
2219 spin_unlock(&inode->v.i_lock);
2221 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2222 bio_for_each_segment_all(bv, bio, iter)
2223 put_page(bv->bv_page);
2226 if (dio->op.error) {
2227 set_bit(EI_INODE_ERROR, &inode->ei_flags);
2231 if (!dio->iter.count)
2235 reinit_completion(&dio->done);
2238 ret = dio->op.error ?: ((long) dio->written << 9);
2240 up(&c->io_in_flight);
2241 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2242 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2245 kfree(dio->iter.iov);
2247 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
2248 bio_for_each_segment_all(bv, bio, iter)
2249 put_page(bv->bv_page);
2252 /* inode->i_dio_count is our ref on inode and thus bch_fs */
2253 inode_dio_end(&inode->v);
2256 req->ki_complete(req, ret);
2262 static void bch2_dio_write_loop_async(struct bch_write_op *op)
2264 struct dio_write *dio = container_of(op, struct dio_write, op);
2267 complete(&dio->done);
2269 bch2_dio_write_loop(dio);
2273 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
2275 struct file *file = req->ki_filp;
2276 struct address_space *mapping = file->f_mapping;
2277 struct bch_inode_info *inode = file_bch_inode(file);
2278 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2279 struct dio_write *dio;
2281 bool locked = true, extending;
2285 prefetch((void *) &c->opts + 64);
2286 prefetch(&inode->ei_inode);
2287 prefetch((void *) &inode->ei_inode + 64);
2289 inode_lock(&inode->v);
2291 ret = generic_write_checks(req, iter);
2292 if (unlikely(ret <= 0))
2295 ret = file_remove_privs(file);
2299 ret = file_update_time(file);
2303 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2306 inode_dio_begin(&inode->v);
2307 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2309 extending = req->ki_pos + iter->count > inode->v.i_size;
2311 inode_unlock(&inode->v);
2315 bio = bio_alloc_bioset(GFP_KERNEL,
2316 bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
2317 &c->dio_write_bioset);
2318 dio = container_of(bio, struct dio_write, op.wbio.bio);
2319 init_completion(&dio->done);
2321 dio->mm = current->mm;
2323 dio->sync = is_sync_kiocb(req) || extending;
2324 dio->free_iov = false;
2325 dio->quota_res.sectors = 0;
2329 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2330 iter->count >> 9, true);
2334 ret = write_invalidate_inode_pages_range(mapping,
2336 req->ki_pos + iter->count - 1);
2340 ret = bch2_dio_write_loop(dio);
2343 inode_unlock(&inode->v);
2346 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2347 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2349 inode_dio_end(&inode->v);
2353 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2355 struct file *file = iocb->ki_filp;
2356 struct bch_inode_info *inode = file_bch_inode(file);
2359 if (iocb->ki_flags & IOCB_DIRECT)
2360 return bch2_direct_write(iocb, from);
2362 /* We can write back this queue in page reclaim */
2363 current->backing_dev_info = inode_to_bdi(&inode->v);
2364 inode_lock(&inode->v);
2366 ret = generic_write_checks(iocb, from);
2370 ret = file_remove_privs(file);
2374 ret = file_update_time(file);
2378 ret = bch2_buffered_write(iocb, from);
2379 if (likely(ret > 0))
2380 iocb->ki_pos += ret;
2382 inode_unlock(&inode->v);
2383 current->backing_dev_info = NULL;
2386 ret = generic_write_sync(iocb, ret);
2394 * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
2395 * insert trigger: look up the btree inode instead
2397 static int bch2_flush_inode(struct bch_fs *c, subvol_inum inum)
2399 struct bch_inode_unpacked inode;
2402 if (c->opts.journal_flush_disabled)
2405 ret = bch2_inode_find_by_inum(c, inum, &inode);
2409 return bch2_journal_flush_seq(&c->journal, inode.bi_journal_seq);
2412 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2414 struct bch_inode_info *inode = file_bch_inode(file);
2415 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2416 int ret, ret2, ret3;
2418 ret = file_write_and_wait_range(file, start, end);
2419 ret2 = sync_inode_metadata(&inode->v, 1);
2420 ret3 = bch2_flush_inode(c, inode_inum(inode));
2422 return ret ?: ret2 ?: ret3;
2427 static inline int range_has_data(struct bch_fs *c, u32 subvol,
2431 struct btree_trans trans;
2432 struct btree_iter iter;
2436 bch2_trans_init(&trans, c, 0, 0);
2438 bch2_trans_begin(&trans);
2440 ret = bch2_subvolume_get_snapshot(&trans, subvol, &start.snapshot);
2444 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2445 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2448 if (bkey_extent_is_data(k.k)) {
2454 bch2_trans_iter_exit(&trans, &iter);
2459 bch2_trans_exit(&trans);
2463 static int __bch2_truncate_page(struct bch_inode_info *inode,
2464 pgoff_t index, loff_t start, loff_t end)
2466 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2467 struct address_space *mapping = inode->v.i_mapping;
2468 struct bch_page_state *s;
2469 unsigned start_offset = start & (PAGE_SIZE - 1);
2470 unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2473 s64 i_sectors_delta = 0;
2476 /* Page boundary? Nothing to do */
2477 if (!((index == start >> PAGE_SHIFT && start_offset) ||
2478 (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2482 if (index << PAGE_SHIFT >= inode->v.i_size)
2485 page = find_lock_page(mapping, index);
2488 * XXX: we're doing two index lookups when we end up reading the
2491 ret = range_has_data(c, inode->ei_subvol,
2492 POS(inode->v.i_ino, index << PAGE_SECTORS_SHIFT),
2493 POS(inode->v.i_ino, (index + 1) << PAGE_SECTORS_SHIFT));
2497 page = find_or_create_page(mapping, index, GFP_KERNEL);
2498 if (unlikely(!page)) {
2504 s = bch2_page_state_create(page, 0);
2510 if (!PageUptodate(page)) {
2511 ret = bch2_read_single_page(page, mapping);
2516 if (index != start >> PAGE_SHIFT)
2518 if (index != end >> PAGE_SHIFT)
2519 end_offset = PAGE_SIZE;
2521 for (i = round_up(start_offset, block_bytes(c)) >> 9;
2522 i < round_down(end_offset, block_bytes(c)) >> 9;
2524 s->s[i].nr_replicas = 0;
2525 if (s->s[i].state == SECTOR_DIRTY)
2527 s->s[i].state = SECTOR_UNALLOCATED;
2530 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2533 * Caller needs to know whether this page will be written out by
2534 * writeback - doing an i_size update if necessary - or whether it will
2535 * be responsible for the i_size update:
2537 ret = s->s[(min_t(u64, inode->v.i_size - (index << PAGE_SHIFT),
2538 PAGE_SIZE) - 1) >> 9].state >= SECTOR_DIRTY;
2540 zero_user_segment(page, start_offset, end_offset);
2543 * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2545 * XXX: because we aren't currently tracking whether the page has actual
2546 * data in it (vs. just 0s, or only partially written) this wrong. ick.
2548 BUG_ON(bch2_get_page_disk_reservation(c, inode, page, false));
2551 * This removes any writeable userspace mappings; we need to force
2552 * .page_mkwrite to be called again before any mmapped writes, to
2553 * redirty the full page:
2556 __set_page_dirty_nobuffers(page);
2564 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2566 return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2567 from, round_up(from, PAGE_SIZE));
2570 static int bch2_truncate_pages(struct bch_inode_info *inode,
2571 loff_t start, loff_t end)
2573 int ret = __bch2_truncate_page(inode, start >> PAGE_SHIFT,
2577 start >> PAGE_SHIFT != end >> PAGE_SHIFT)
2578 ret = __bch2_truncate_page(inode,
2584 static int bch2_extend(struct user_namespace *mnt_userns,
2585 struct bch_inode_info *inode,
2586 struct bch_inode_unpacked *inode_u,
2587 struct iattr *iattr)
2589 struct address_space *mapping = inode->v.i_mapping;
2595 * this has to be done _before_ extending i_size:
2597 ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2601 truncate_setsize(&inode->v, iattr->ia_size);
2603 return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2606 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2607 struct bch_inode_unpacked *bi,
2610 bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2614 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2615 struct bch_inode_unpacked *bi, void *p)
2617 u64 *new_i_size = p;
2619 bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2620 bi->bi_size = *new_i_size;
2624 int bch2_truncate(struct user_namespace *mnt_userns,
2625 struct bch_inode_info *inode, struct iattr *iattr)
2627 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2628 struct address_space *mapping = inode->v.i_mapping;
2629 struct bch_inode_unpacked inode_u;
2630 u64 new_i_size = iattr->ia_size;
2631 s64 i_sectors_delta = 0;
2635 * If the truncate call with change the size of the file, the
2636 * cmtimes should be updated. If the size will not change, we
2637 * do not need to update the cmtimes.
2639 if (iattr->ia_size != inode->v.i_size) {
2640 if (!(iattr->ia_valid & ATTR_MTIME))
2641 ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2642 if (!(iattr->ia_valid & ATTR_CTIME))
2643 ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2644 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2647 inode_dio_wait(&inode->v);
2648 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2650 ret = bch2_inode_find_by_inum(c, inode_inum(inode), &inode_u);
2655 * check this before next assertion; on filesystem error our normal
2656 * invariants are a bit broken (truncate has to truncate the page cache
2657 * before the inode).
2659 ret = bch2_journal_error(&c->journal);
2663 WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2664 inode->v.i_size < inode_u.bi_size);
2666 if (iattr->ia_size > inode->v.i_size) {
2667 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2671 iattr->ia_valid &= ~ATTR_SIZE;
2673 ret = bch2_truncate_page(inode, iattr->ia_size);
2674 if (unlikely(ret < 0))
2678 * When extending, we're going to write the new i_size to disk
2679 * immediately so we need to flush anything above the current on disk
2682 * Also, when extending we need to flush the page that i_size currently
2683 * straddles - if it's mapped to userspace, we need to ensure that
2684 * userspace has to redirty it and call .mkwrite -> set_page_dirty
2685 * again to allocate the part of the page that was extended.
2687 if (iattr->ia_size > inode_u.bi_size)
2688 ret = filemap_write_and_wait_range(mapping,
2690 iattr->ia_size - 1);
2691 else if (iattr->ia_size & (PAGE_SIZE - 1))
2692 ret = filemap_write_and_wait_range(mapping,
2693 round_down(iattr->ia_size, PAGE_SIZE),
2694 iattr->ia_size - 1);
2698 mutex_lock(&inode->ei_update_lock);
2699 ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2701 mutex_unlock(&inode->ei_update_lock);
2706 truncate_setsize(&inode->v, iattr->ia_size);
2708 ret = bch2_fpunch(c, inode_inum(inode),
2709 round_up(iattr->ia_size, block_bytes(c)) >> 9,
2710 U64_MAX, &i_sectors_delta);
2711 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2713 WARN_ON(!inode->v.i_size && inode->v.i_blocks &&
2714 !bch2_journal_error(&c->journal));
2719 mutex_lock(&inode->ei_update_lock);
2720 ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2721 mutex_unlock(&inode->ei_update_lock);
2723 ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2725 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2731 static int inode_update_times_fn(struct bch_inode_info *inode,
2732 struct bch_inode_unpacked *bi, void *p)
2734 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2736 bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2740 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2742 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2743 u64 end = offset + len;
2744 u64 block_start = round_up(offset, block_bytes(c));
2745 u64 block_end = round_down(end, block_bytes(c));
2746 bool truncated_last_page;
2749 ret = bch2_truncate_pages(inode, offset, end);
2750 if (unlikely(ret < 0))
2753 truncated_last_page = ret;
2755 truncate_pagecache_range(&inode->v, offset, end - 1);
2757 if (block_start < block_end ) {
2758 s64 i_sectors_delta = 0;
2760 ret = bch2_fpunch(c, inode_inum(inode),
2761 block_start >> 9, block_end >> 9,
2763 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2766 mutex_lock(&inode->ei_update_lock);
2767 if (end >= inode->v.i_size && !truncated_last_page) {
2768 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2769 ATTR_MTIME|ATTR_CTIME);
2771 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2772 ATTR_MTIME|ATTR_CTIME);
2774 mutex_unlock(&inode->ei_update_lock);
2779 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2780 loff_t offset, loff_t len,
2783 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2784 struct address_space *mapping = inode->v.i_mapping;
2785 struct bkey_buf copy;
2786 struct btree_trans trans;
2787 struct btree_iter src, dst, del;
2788 loff_t shift, new_size;
2792 if ((offset | len) & (block_bytes(c) - 1))
2796 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2799 if (offset >= inode->v.i_size)
2802 src_start = U64_MAX;
2805 if (offset + len >= inode->v.i_size)
2808 src_start = offset + len;
2812 new_size = inode->v.i_size + shift;
2814 ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2819 i_size_write(&inode->v, new_size);
2820 mutex_lock(&inode->ei_update_lock);
2821 ret = bch2_write_inode_size(c, inode, new_size,
2822 ATTR_MTIME|ATTR_CTIME);
2823 mutex_unlock(&inode->ei_update_lock);
2825 s64 i_sectors_delta = 0;
2827 ret = bch2_fpunch(c, inode_inum(inode),
2828 offset >> 9, (offset + len) >> 9,
2830 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2836 bch2_bkey_buf_init(©);
2837 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2838 bch2_trans_iter_init(&trans, &src, BTREE_ID_extents,
2839 POS(inode->v.i_ino, src_start >> 9),
2841 bch2_trans_copy_iter(&dst, &src);
2842 bch2_trans_copy_iter(&del, &src);
2844 while (ret == 0 || ret == -EINTR) {
2845 struct disk_reservation disk_res =
2846 bch2_disk_reservation_init(c, 0);
2847 struct bkey_i delete;
2849 struct bpos next_pos;
2850 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2851 struct bpos atomic_end;
2852 unsigned trigger_flags = 0;
2855 bch2_trans_begin(&trans);
2857 ret = bch2_subvolume_get_snapshot(&trans,
2858 inode->ei_subvol, &snapshot);
2862 bch2_btree_iter_set_snapshot(&src, snapshot);
2863 bch2_btree_iter_set_snapshot(&dst, snapshot);
2864 bch2_btree_iter_set_snapshot(&del, snapshot);
2866 bch2_trans_begin(&trans);
2869 ? bch2_btree_iter_peek_prev(&src)
2870 : bch2_btree_iter_peek(&src);
2871 if ((ret = bkey_err(k)))
2874 if (!k.k || k.k->p.inode != inode->v.i_ino)
2878 bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2881 bch2_bkey_buf_reassemble(©, c, k);
2884 bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2885 bch2_cut_front(move_pos, copy.k);
2887 copy.k->k.p.offset += shift >> 9;
2888 bch2_btree_iter_set_pos(&dst, bkey_start_pos(©.k->k));
2890 ret = bch2_extent_atomic_end(&trans, &dst, copy.k, &atomic_end);
2894 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2896 move_pos = atomic_end;
2897 move_pos.offset -= shift >> 9;
2900 bch2_cut_back(atomic_end, copy.k);
2904 bkey_init(&delete.k);
2905 delete.k.p = copy.k->k.p;
2906 delete.k.size = copy.k->k.size;
2907 delete.k.p.offset -= shift >> 9;
2908 bch2_btree_iter_set_pos(&del, bkey_start_pos(&delete.k));
2910 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2912 if (copy.k->k.size == k.k->size) {
2914 * If we're moving the entire extent, we can skip
2917 trigger_flags |= BTREE_TRIGGER_NORUN;
2919 /* We might end up splitting compressed extents: */
2921 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2923 ret = bch2_disk_reservation_get(c, &disk_res,
2924 copy.k->k.size, nr_ptrs,
2925 BCH_DISK_RESERVATION_NOFAIL);
2929 ret = bch2_btree_iter_traverse(&del) ?:
2930 bch2_trans_update(&trans, &del, &delete, trigger_flags) ?:
2931 bch2_trans_update(&trans, &dst, copy.k, trigger_flags) ?:
2932 bch2_trans_commit(&trans, &disk_res, NULL,
2933 BTREE_INSERT_NOFAIL);
2934 bch2_disk_reservation_put(c, &disk_res);
2937 bch2_btree_iter_set_pos(&src, next_pos);
2939 bch2_trans_iter_exit(&trans, &del);
2940 bch2_trans_iter_exit(&trans, &dst);
2941 bch2_trans_iter_exit(&trans, &src);
2942 bch2_trans_exit(&trans);
2943 bch2_bkey_buf_exit(©, c);
2948 mutex_lock(&inode->ei_update_lock);
2950 i_size_write(&inode->v, new_size);
2951 ret = bch2_write_inode_size(c, inode, new_size,
2952 ATTR_MTIME|ATTR_CTIME);
2954 /* We need an inode update to update bi_journal_seq for fsync: */
2955 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2956 ATTR_MTIME|ATTR_CTIME);
2958 mutex_unlock(&inode->ei_update_lock);
2962 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2963 u64 start_sector, u64 end_sector)
2965 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2966 struct btree_trans trans;
2967 struct btree_iter iter;
2968 struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2969 unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2972 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2974 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2975 POS(inode->v.i_ino, start_sector),
2976 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2978 while (!ret && bkey_cmp(iter.pos, end_pos) < 0) {
2979 s64 i_sectors_delta = 0;
2980 struct disk_reservation disk_res = { 0 };
2981 struct quota_res quota_res = { 0 };
2982 struct bkey_i_reservation reservation;
2987 bch2_trans_begin(&trans);
2989 ret = bch2_subvolume_get_snapshot(&trans,
2990 inode->ei_subvol, &snapshot);
2994 bch2_btree_iter_set_snapshot(&iter, snapshot);
2996 k = bch2_btree_iter_peek_slot(&iter);
2997 if ((ret = bkey_err(k)))
3000 /* already reserved */
3001 if (k.k->type == KEY_TYPE_reservation &&
3002 bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
3003 bch2_btree_iter_advance(&iter);
3007 if (bkey_extent_is_data(k.k) &&
3008 !(mode & FALLOC_FL_ZERO_RANGE)) {
3009 bch2_btree_iter_advance(&iter);
3013 bkey_reservation_init(&reservation.k_i);
3014 reservation.k.type = KEY_TYPE_reservation;
3015 reservation.k.p = k.k->p;
3016 reservation.k.size = k.k->size;
3018 bch2_cut_front(iter.pos, &reservation.k_i);
3019 bch2_cut_back(end_pos, &reservation.k_i);
3021 sectors = reservation.k.size;
3022 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
3024 if (!bkey_extent_is_allocation(k.k)) {
3025 ret = bch2_quota_reservation_add(c, inode,
3032 if (reservation.v.nr_replicas < replicas ||
3033 bch2_bkey_sectors_compressed(k)) {
3034 ret = bch2_disk_reservation_get(c, &disk_res, sectors,
3039 reservation.v.nr_replicas = disk_res.nr_replicas;
3042 ret = bch2_extent_update(&trans, inode_inum(inode), &iter,
3045 0, &i_sectors_delta, true);
3048 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3050 bch2_quota_reservation_put(c, inode, "a_res);
3051 bch2_disk_reservation_put(c, &disk_res);
3056 bch2_trans_unlock(&trans); /* lock ordering, before taking pagecache locks: */
3057 mark_pagecache_reserved(inode, start_sector, iter.pos.offset);
3059 if (ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)) {
3060 struct quota_res quota_res = { 0 };
3061 s64 i_sectors_delta = 0;
3063 bch2_fpunch_at(&trans, &iter, inode_inum(inode),
3064 end_sector, &i_sectors_delta);
3065 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
3066 bch2_quota_reservation_put(c, inode, "a_res);
3069 bch2_trans_iter_exit(&trans, &iter);
3070 bch2_trans_exit(&trans);
3074 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
3075 loff_t offset, loff_t len)
3077 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3078 u64 end = offset + len;
3079 u64 block_start = round_down(offset, block_bytes(c));
3080 u64 block_end = round_up(end, block_bytes(c));
3081 bool truncated_last_page = false;
3084 if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
3085 ret = inode_newsize_ok(&inode->v, end);
3090 if (mode & FALLOC_FL_ZERO_RANGE) {
3091 ret = bch2_truncate_pages(inode, offset, end);
3092 if (unlikely(ret < 0))
3095 truncated_last_page = ret;
3097 truncate_pagecache_range(&inode->v, offset, end - 1);
3099 block_start = round_up(offset, block_bytes(c));
3100 block_end = round_down(end, block_bytes(c));
3103 ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
3106 * On -ENOSPC in ZERO_RANGE mode, we still want to do the inode update,
3107 * so that the VFS cache i_size is consistent with the btree i_size:
3110 !(ret == -ENOSPC && (mode & FALLOC_FL_ZERO_RANGE)))
3113 if (mode & FALLOC_FL_KEEP_SIZE && end > inode->v.i_size)
3114 end = inode->v.i_size;
3116 if (end >= inode->v.i_size &&
3117 (((mode & FALLOC_FL_ZERO_RANGE) && !truncated_last_page) ||
3118 !(mode & FALLOC_FL_KEEP_SIZE))) {
3119 spin_lock(&inode->v.i_lock);
3120 i_size_write(&inode->v, end);
3121 spin_unlock(&inode->v.i_lock);
3123 mutex_lock(&inode->ei_update_lock);
3124 ret2 = bch2_write_inode_size(c, inode, end, 0);
3125 mutex_unlock(&inode->ei_update_lock);
3131 long bch2_fallocate_dispatch(struct file *file, int mode,
3132 loff_t offset, loff_t len)
3134 struct bch_inode_info *inode = file_bch_inode(file);
3135 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3138 if (!percpu_ref_tryget(&c->writes))
3141 inode_lock(&inode->v);
3142 inode_dio_wait(&inode->v);
3143 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
3145 if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
3146 ret = bchfs_fallocate(inode, mode, offset, len);
3147 else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
3148 ret = bchfs_fpunch(inode, offset, len);
3149 else if (mode == FALLOC_FL_INSERT_RANGE)
3150 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
3151 else if (mode == FALLOC_FL_COLLAPSE_RANGE)
3152 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
3157 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
3158 inode_unlock(&inode->v);
3159 percpu_ref_put(&c->writes);
3164 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
3165 struct file *file_dst, loff_t pos_dst,
3166 loff_t len, unsigned remap_flags)
3168 struct bch_inode_info *src = file_bch_inode(file_src);
3169 struct bch_inode_info *dst = file_bch_inode(file_dst);
3170 struct bch_fs *c = src->v.i_sb->s_fs_info;
3171 s64 i_sectors_delta = 0;
3175 if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
3178 if (remap_flags & REMAP_FILE_DEDUP)
3181 if ((pos_src & (block_bytes(c) - 1)) ||
3182 (pos_dst & (block_bytes(c) - 1)))
3186 abs(pos_src - pos_dst) < len)
3189 bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3191 file_update_time(file_dst);
3193 inode_dio_wait(&src->v);
3194 inode_dio_wait(&dst->v);
3196 ret = generic_remap_file_range_prep(file_src, pos_src,
3199 if (ret < 0 || len == 0)
3202 aligned_len = round_up((u64) len, block_bytes(c));
3204 ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
3205 pos_dst, pos_dst + len - 1);
3209 mark_pagecache_unallocated(src, pos_src >> 9,
3210 (pos_src + aligned_len) >> 9);
3212 ret = bch2_remap_range(c,
3213 inode_inum(dst), pos_dst >> 9,
3214 inode_inum(src), pos_src >> 9,
3216 pos_dst + len, &i_sectors_delta);
3221 * due to alignment, we might have remapped slightly more than requsted
3223 ret = min((u64) ret << 9, (u64) len);
3225 /* XXX get a quota reservation */
3226 i_sectors_acct(c, dst, NULL, i_sectors_delta);
3228 spin_lock(&dst->v.i_lock);
3229 if (pos_dst + ret > dst->v.i_size)
3230 i_size_write(&dst->v, pos_dst + ret);
3231 spin_unlock(&dst->v.i_lock);
3233 if ((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
3234 IS_SYNC(file_inode(file_dst)))
3235 ret = bch2_flush_inode(c, inode_inum(dst));
3237 bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
3244 static int page_data_offset(struct page *page, unsigned offset)
3246 struct bch_page_state *s = bch2_page_state(page);
3250 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3251 if (s->s[i].state >= SECTOR_DIRTY)
3257 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
3258 loff_t start_offset,
3261 struct address_space *mapping = vinode->i_mapping;
3263 pgoff_t start_index = start_offset >> PAGE_SHIFT;
3264 pgoff_t end_index = end_offset >> PAGE_SHIFT;
3265 pgoff_t index = start_index;
3269 while (index <= end_index) {
3270 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
3273 offset = page_data_offset(page,
3274 page->index == start_index
3275 ? start_offset & (PAGE_SIZE - 1)
3278 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
3280 start_offset, end_offset);
3296 static loff_t bch2_seek_data(struct file *file, u64 offset)
3298 struct bch_inode_info *inode = file_bch_inode(file);
3299 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3300 struct btree_trans trans;
3301 struct btree_iter iter;
3303 subvol_inum inum = inode_inum(inode);
3304 u64 isize, next_data = MAX_LFS_FILESIZE;
3308 isize = i_size_read(&inode->v);
3309 if (offset >= isize)
3312 bch2_trans_init(&trans, c, 0, 0);
3314 bch2_trans_begin(&trans);
3316 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3320 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3321 SPOS(inode->v.i_ino, offset >> 9, snapshot), 0, k, ret) {
3322 if (k.k->p.inode != inode->v.i_ino) {
3324 } else if (bkey_extent_is_data(k.k)) {
3325 next_data = max(offset, bkey_start_offset(k.k) << 9);
3327 } else if (k.k->p.offset >> 9 > isize)
3330 bch2_trans_iter_exit(&trans, &iter);
3335 bch2_trans_exit(&trans);
3339 if (next_data > offset)
3340 next_data = bch2_seek_pagecache_data(&inode->v,
3343 if (next_data >= isize)
3346 return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3349 static int __page_hole_offset(struct page *page, unsigned offset)
3351 struct bch_page_state *s = bch2_page_state(page);
3357 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3358 if (s->s[i].state < SECTOR_DIRTY)
3364 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3366 pgoff_t index = offset >> PAGE_SHIFT;
3371 page = find_lock_page(mapping, index);
3375 pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3377 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3384 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3385 loff_t start_offset,
3388 struct address_space *mapping = vinode->i_mapping;
3389 loff_t offset = start_offset, hole;
3391 while (offset < end_offset) {
3392 hole = page_hole_offset(mapping, offset);
3393 if (hole >= 0 && hole <= end_offset)
3394 return max(start_offset, hole);
3396 offset += PAGE_SIZE;
3397 offset &= PAGE_MASK;
3403 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3405 struct bch_inode_info *inode = file_bch_inode(file);
3406 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3407 struct btree_trans trans;
3408 struct btree_iter iter;
3410 subvol_inum inum = inode_inum(inode);
3411 u64 isize, next_hole = MAX_LFS_FILESIZE;
3415 isize = i_size_read(&inode->v);
3416 if (offset >= isize)
3419 bch2_trans_init(&trans, c, 0, 0);
3421 bch2_trans_begin(&trans);
3423 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
3427 for_each_btree_key_norestart(&trans, iter, BTREE_ID_extents,
3428 SPOS(inode->v.i_ino, offset >> 9, snapshot),
3429 BTREE_ITER_SLOTS, k, ret) {
3430 if (k.k->p.inode != inode->v.i_ino) {
3431 next_hole = bch2_seek_pagecache_hole(&inode->v,
3432 offset, MAX_LFS_FILESIZE);
3434 } else if (!bkey_extent_is_data(k.k)) {
3435 next_hole = bch2_seek_pagecache_hole(&inode->v,
3436 max(offset, bkey_start_offset(k.k) << 9),
3437 k.k->p.offset << 9);
3439 if (next_hole < k.k->p.offset << 9)
3442 offset = max(offset, bkey_start_offset(k.k) << 9);
3445 bch2_trans_iter_exit(&trans, &iter);
3450 bch2_trans_exit(&trans);
3454 if (next_hole > isize)
3457 return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3460 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3466 return generic_file_llseek(file, offset, whence);
3468 return bch2_seek_data(file, offset);
3470 return bch2_seek_hole(file, offset);
3476 void bch2_fs_fsio_exit(struct bch_fs *c)
3478 bioset_exit(&c->dio_write_bioset);
3479 bioset_exit(&c->dio_read_bioset);
3480 bioset_exit(&c->writepage_bioset);
3483 int bch2_fs_fsio_init(struct bch_fs *c)
3487 pr_verbose_init(c->opts, "");
3489 if (bioset_init(&c->writepage_bioset,
3490 4, offsetof(struct bch_writepage_io, op.wbio.bio),
3491 BIOSET_NEED_BVECS) ||
3492 bioset_init(&c->dio_read_bioset,
3493 4, offsetof(struct dio_read, rbio.bio),
3494 BIOSET_NEED_BVECS) ||
3495 bioset_init(&c->dio_write_bioset,
3496 4, offsetof(struct dio_write, op.wbio.bio),
3500 pr_verbose_init(c->opts, "ret %i", ret);
3504 #endif /* NO_BCACHEFS_FS */