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 struct address_space *faults_disabled_mapping(void)
40 return (void *) (((unsigned long) current->faults_disabled_mapping) & ~1UL);
43 static inline void set_fdm_dropped_locks(void)
45 current->faults_disabled_mapping =
46 (void *) (((unsigned long) current->faults_disabled_mapping)|1);
49 static inline bool fdm_dropped_locks(void)
51 return ((unsigned long) current->faults_disabled_mapping) & 1;
58 struct bch_writepage_io {
60 struct bch_inode_info *inode;
63 struct bch_write_op op;
67 struct completion done;
73 struct quota_res quota_res;
77 struct iovec inline_vecs[2];
80 struct bch_write_op op;
88 struct bch_read_bio rbio;
91 /* pagecache_block must be held */
92 static int write_invalidate_inode_pages_range(struct address_space *mapping,
93 loff_t start, loff_t end)
98 * XXX: the way this is currently implemented, we can spin if a process
99 * is continually redirtying a specific page
102 if (!mapping->nrpages &&
103 !mapping->nrexceptional)
106 ret = filemap_write_and_wait_range(mapping, start, end);
110 if (!mapping->nrpages)
113 ret = invalidate_inode_pages2_range(mapping,
116 } while (ret == -EBUSY);
123 #ifdef CONFIG_BCACHEFS_QUOTA
125 static void bch2_quota_reservation_put(struct bch_fs *c,
126 struct bch_inode_info *inode,
127 struct quota_res *res)
132 mutex_lock(&inode->ei_quota_lock);
133 BUG_ON(res->sectors > inode->ei_quota_reserved);
135 bch2_quota_acct(c, inode->ei_qid, Q_SPC,
136 -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
137 inode->ei_quota_reserved -= res->sectors;
138 mutex_unlock(&inode->ei_quota_lock);
143 static int bch2_quota_reservation_add(struct bch_fs *c,
144 struct bch_inode_info *inode,
145 struct quota_res *res,
151 mutex_lock(&inode->ei_quota_lock);
152 ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
153 check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
155 inode->ei_quota_reserved += sectors;
156 res->sectors += sectors;
158 mutex_unlock(&inode->ei_quota_lock);
165 static void bch2_quota_reservation_put(struct bch_fs *c,
166 struct bch_inode_info *inode,
167 struct quota_res *res)
171 static int bch2_quota_reservation_add(struct bch_fs *c,
172 struct bch_inode_info *inode,
173 struct quota_res *res,
182 /* i_size updates: */
184 struct inode_new_size {
190 static int inode_set_size(struct bch_inode_info *inode,
191 struct bch_inode_unpacked *bi,
194 struct inode_new_size *s = p;
196 bi->bi_size = s->new_size;
197 if (s->fields & ATTR_ATIME)
198 bi->bi_atime = s->now;
199 if (s->fields & ATTR_MTIME)
200 bi->bi_mtime = s->now;
201 if (s->fields & ATTR_CTIME)
202 bi->bi_ctime = s->now;
207 int __must_check bch2_write_inode_size(struct bch_fs *c,
208 struct bch_inode_info *inode,
209 loff_t new_size, unsigned fields)
211 struct inode_new_size s = {
212 .new_size = new_size,
213 .now = bch2_current_time(c),
217 return bch2_write_inode(c, inode, inode_set_size, &s, fields);
220 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
221 struct quota_res *quota_res, s64 sectors)
226 mutex_lock(&inode->ei_quota_lock);
227 #ifdef CONFIG_BCACHEFS_QUOTA
228 if (quota_res && sectors > 0) {
229 BUG_ON(sectors > quota_res->sectors);
230 BUG_ON(sectors > inode->ei_quota_reserved);
232 quota_res->sectors -= sectors;
233 inode->ei_quota_reserved -= sectors;
235 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
238 inode->v.i_blocks += sectors;
239 mutex_unlock(&inode->ei_quota_lock);
244 /* stored in page->private: */
246 struct bch_page_sector {
247 /* Uncompressed, fully allocated replicas: */
248 unsigned nr_replicas:3;
250 /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
251 unsigned replicas_reserved:3;
262 struct bch_page_state {
264 atomic_t write_count;
265 struct bch_page_sector s[PAGE_SECTORS];
268 static inline struct bch_page_state *__bch2_page_state(struct page *page)
270 return page_has_private(page)
271 ? (struct bch_page_state *) page_private(page)
275 static inline struct bch_page_state *bch2_page_state(struct page *page)
277 EBUG_ON(!PageLocked(page));
279 return __bch2_page_state(page);
282 /* for newly allocated pages: */
283 static void __bch2_page_state_release(struct page *page)
285 kfree(detach_page_private(page));
288 static void bch2_page_state_release(struct page *page)
290 EBUG_ON(!PageLocked(page));
291 __bch2_page_state_release(page);
294 /* for newly allocated pages: */
295 static struct bch_page_state *__bch2_page_state_create(struct page *page,
298 struct bch_page_state *s;
300 s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
304 spin_lock_init(&s->lock);
305 attach_page_private(page, s);
309 static struct bch_page_state *bch2_page_state_create(struct page *page,
312 return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
315 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
317 /* XXX: this should not be open coded */
318 return inode->ei_inode.bi_data_replicas
319 ? inode->ei_inode.bi_data_replicas - 1
320 : c->opts.data_replicas;
323 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
324 unsigned nr_replicas)
326 return max(0, (int) nr_replicas -
328 s->replicas_reserved);
331 static int bch2_get_page_disk_reservation(struct bch_fs *c,
332 struct bch_inode_info *inode,
333 struct page *page, bool check_enospc)
335 struct bch_page_state *s = bch2_page_state_create(page, 0);
336 unsigned nr_replicas = inode_nr_replicas(c, inode);
337 struct disk_reservation disk_res = { 0 };
338 unsigned i, disk_res_sectors = 0;
344 for (i = 0; i < ARRAY_SIZE(s->s); i++)
345 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
347 if (!disk_res_sectors)
350 ret = bch2_disk_reservation_get(c, &disk_res,
353 ? BCH_DISK_RESERVATION_NOFAIL
358 for (i = 0; i < ARRAY_SIZE(s->s); i++)
359 s->s[i].replicas_reserved +=
360 sectors_to_reserve(&s->s[i], nr_replicas);
365 struct bch2_page_reservation {
366 struct disk_reservation disk;
367 struct quota_res quota;
370 static void bch2_page_reservation_init(struct bch_fs *c,
371 struct bch_inode_info *inode,
372 struct bch2_page_reservation *res)
374 memset(res, 0, sizeof(*res));
376 res->disk.nr_replicas = inode_nr_replicas(c, inode);
379 static void bch2_page_reservation_put(struct bch_fs *c,
380 struct bch_inode_info *inode,
381 struct bch2_page_reservation *res)
383 bch2_disk_reservation_put(c, &res->disk);
384 bch2_quota_reservation_put(c, inode, &res->quota);
387 static int bch2_page_reservation_get(struct bch_fs *c,
388 struct bch_inode_info *inode, struct page *page,
389 struct bch2_page_reservation *res,
390 unsigned offset, unsigned len, bool check_enospc)
392 struct bch_page_state *s = bch2_page_state_create(page, 0);
393 unsigned i, disk_sectors = 0, quota_sectors = 0;
399 for (i = round_down(offset, block_bytes(c)) >> 9;
400 i < round_up(offset + len, block_bytes(c)) >> 9;
402 disk_sectors += sectors_to_reserve(&s->s[i],
403 res->disk.nr_replicas);
404 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
408 ret = bch2_disk_reservation_add(c, &res->disk,
411 ? BCH_DISK_RESERVATION_NOFAIL
418 ret = bch2_quota_reservation_add(c, inode, &res->quota,
422 struct disk_reservation tmp = {
423 .sectors = disk_sectors
426 bch2_disk_reservation_put(c, &tmp);
427 res->disk.sectors -= disk_sectors;
435 static void bch2_clear_page_bits(struct page *page)
437 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
438 struct bch_fs *c = inode->v.i_sb->s_fs_info;
439 struct bch_page_state *s = bch2_page_state(page);
440 struct disk_reservation disk_res = { 0 };
441 int i, dirty_sectors = 0;
446 EBUG_ON(!PageLocked(page));
447 EBUG_ON(PageWriteback(page));
449 for (i = 0; i < ARRAY_SIZE(s->s); i++) {
450 disk_res.sectors += s->s[i].replicas_reserved;
451 s->s[i].replicas_reserved = 0;
453 if (s->s[i].state == SECTOR_DIRTY) {
455 s->s[i].state = SECTOR_UNALLOCATED;
459 bch2_disk_reservation_put(c, &disk_res);
462 i_sectors_acct(c, inode, NULL, -dirty_sectors);
464 bch2_page_state_release(page);
467 static void bch2_set_page_dirty(struct bch_fs *c,
468 struct bch_inode_info *inode, struct page *page,
469 struct bch2_page_reservation *res,
470 unsigned offset, unsigned len)
472 struct bch_page_state *s = bch2_page_state(page);
473 unsigned i, dirty_sectors = 0;
475 WARN_ON((u64) page_offset(page) + offset + len >
476 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
480 for (i = round_down(offset, block_bytes(c)) >> 9;
481 i < round_up(offset + len, block_bytes(c)) >> 9;
483 unsigned sectors = sectors_to_reserve(&s->s[i],
484 res->disk.nr_replicas);
487 * This can happen if we race with the error path in
488 * bch2_writepage_io_done():
490 sectors = min_t(unsigned, sectors, res->disk.sectors);
492 s->s[i].replicas_reserved += sectors;
493 res->disk.sectors -= sectors;
495 if (s->s[i].state == SECTOR_UNALLOCATED)
498 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
501 spin_unlock(&s->lock);
504 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
506 if (!PageDirty(page))
507 __set_page_dirty_nobuffers(page);
510 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
512 struct file *file = vmf->vma->vm_file;
513 struct address_space *mapping = file->f_mapping;
514 struct address_space *fdm = faults_disabled_mapping();
515 struct bch_inode_info *inode = file_bch_inode(file);
519 return VM_FAULT_SIGBUS;
523 struct bch_inode_info *fdm_host = to_bch_ei(fdm->host);
525 if (bch2_pagecache_add_tryget(&inode->ei_pagecache_lock))
528 bch2_pagecache_block_put(&fdm_host->ei_pagecache_lock);
530 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
531 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
533 bch2_pagecache_block_get(&fdm_host->ei_pagecache_lock);
535 /* Signal that lock has been dropped: */
536 set_fdm_dropped_locks();
537 return VM_FAULT_SIGBUS;
540 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
542 ret = filemap_fault(vmf);
543 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
548 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
550 struct page *page = vmf->page;
551 struct file *file = vmf->vma->vm_file;
552 struct bch_inode_info *inode = file_bch_inode(file);
553 struct address_space *mapping = file->f_mapping;
554 struct bch_fs *c = inode->v.i_sb->s_fs_info;
555 struct bch2_page_reservation res;
558 int ret = VM_FAULT_LOCKED;
560 bch2_page_reservation_init(c, inode, &res);
562 sb_start_pagefault(inode->v.i_sb);
563 file_update_time(file);
566 * Not strictly necessary, but helps avoid dio writes livelocking in
567 * write_invalidate_inode_pages_range() - can drop this if/when we get
568 * a write_invalidate_inode_pages_range() that works without dropping
569 * page lock before invalidating page
571 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
574 isize = i_size_read(&inode->v);
576 if (page->mapping != mapping || page_offset(page) >= isize) {
578 ret = VM_FAULT_NOPAGE;
582 len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
584 if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
586 ret = VM_FAULT_SIGBUS;
590 bch2_set_page_dirty(c, inode, page, &res, 0, len);
591 bch2_page_reservation_put(c, inode, &res);
593 wait_for_stable_page(page);
595 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
596 sb_end_pagefault(inode->v.i_sb);
601 void bch2_invalidatepage(struct page *page, unsigned int offset,
604 if (offset || length < PAGE_SIZE)
607 bch2_clear_page_bits(page);
610 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
615 bch2_clear_page_bits(page);
619 #ifdef CONFIG_MIGRATION
620 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
621 struct page *page, enum migrate_mode mode)
625 EBUG_ON(!PageLocked(page));
626 EBUG_ON(!PageLocked(newpage));
628 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
629 if (ret != MIGRATEPAGE_SUCCESS)
632 if (PagePrivate(page))
633 attach_page_private(newpage, detach_page_private(page));
635 if (mode != MIGRATE_SYNC_NO_COPY)
636 migrate_page_copy(newpage, page);
638 migrate_page_states(newpage, page);
639 return MIGRATEPAGE_SUCCESS;
645 static void bch2_readpages_end_io(struct bio *bio)
647 struct bvec_iter_all iter;
650 bio_for_each_segment_all(bv, bio, iter) {
651 struct page *page = bv->bv_page;
653 if (!bio->bi_status) {
654 SetPageUptodate(page);
656 ClearPageUptodate(page);
665 struct readpages_iter {
666 struct address_space *mapping;
673 static int readpages_iter_init(struct readpages_iter *iter,
674 struct readahead_control *ractl)
676 unsigned i, nr_pages = readahead_count(ractl);
678 memset(iter, 0, sizeof(*iter));
680 iter->mapping = ractl->mapping;
681 iter->offset = readahead_index(ractl);
682 iter->nr_pages = nr_pages;
684 iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
688 nr_pages = __readahead_batch(ractl, iter->pages, nr_pages);
689 for (i = 0; i < nr_pages; i++) {
690 __bch2_page_state_create(iter->pages[i], __GFP_NOFAIL);
691 put_page(iter->pages[i]);
697 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
699 if (iter->idx >= iter->nr_pages)
702 EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
704 return iter->pages[iter->idx];
707 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
709 struct bvec_iter iter;
711 unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
712 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
713 unsigned state = k.k->type == KEY_TYPE_reservation
717 bio_for_each_segment(bv, bio, iter) {
718 struct bch_page_state *s = bch2_page_state(bv.bv_page);
721 for (i = bv.bv_offset >> 9;
722 i < (bv.bv_offset + bv.bv_len) >> 9;
724 s->s[i].nr_replicas = nr_ptrs;
725 s->s[i].state = state;
730 static bool extent_partial_reads_expensive(struct bkey_s_c k)
732 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
733 struct bch_extent_crc_unpacked crc;
734 const union bch_extent_entry *i;
736 bkey_for_each_crc(k.k, ptrs, crc, i)
737 if (crc.csum_type || crc.compression_type)
742 static void readpage_bio_extend(struct readpages_iter *iter,
744 unsigned sectors_this_extent,
747 while (bio_sectors(bio) < sectors_this_extent &&
748 bio->bi_vcnt < bio->bi_max_vecs) {
749 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
750 struct page *page = readpage_iter_next(iter);
754 if (iter->offset + iter->idx != page_offset)
762 page = xa_load(&iter->mapping->i_pages, page_offset);
763 if (page && !xa_is_value(page))
766 page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
770 if (!__bch2_page_state_create(page, 0)) {
775 ret = add_to_page_cache_lru(page, iter->mapping,
776 page_offset, GFP_NOFS);
778 __bch2_page_state_release(page);
786 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
790 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
791 struct bch_read_bio *rbio, u64 inum,
792 struct readpages_iter *readpages_iter)
794 struct bch_fs *c = trans->c;
796 int flags = BCH_READ_RETRY_IF_STALE|
797 BCH_READ_MAY_PROMOTE;
801 rbio->start_time = local_clock();
803 bch2_bkey_buf_init(&sk);
807 unsigned bytes, sectors, offset_into_extent;
808 enum btree_id data_btree = BTREE_ID_extents;
810 bch2_btree_iter_set_pos(iter,
811 POS(inum, rbio->bio.bi_iter.bi_sector));
813 k = bch2_btree_iter_peek_slot(iter);
818 offset_into_extent = iter->pos.offset -
819 bkey_start_offset(k.k);
820 sectors = k.k->size - offset_into_extent;
822 bch2_bkey_buf_reassemble(&sk, c, k);
824 ret = bch2_read_indirect_extent(trans, &data_btree,
825 &offset_into_extent, &sk);
829 k = bkey_i_to_s_c(sk.k);
831 sectors = min(sectors, k.k->size - offset_into_extent);
833 bch2_trans_unlock(trans);
836 readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
837 extent_partial_reads_expensive(k));
839 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
840 swap(rbio->bio.bi_iter.bi_size, bytes);
842 if (rbio->bio.bi_iter.bi_size == bytes)
843 flags |= BCH_READ_LAST_FRAGMENT;
845 if (bkey_extent_is_allocation(k.k))
846 bch2_add_page_sectors(&rbio->bio, k);
848 bch2_read_extent(trans, rbio, iter->pos,
849 data_btree, k, offset_into_extent, flags);
851 if (flags & BCH_READ_LAST_FRAGMENT)
854 swap(rbio->bio.bi_iter.bi_size, bytes);
855 bio_advance(&rbio->bio, bytes);
862 bch_err_inum_ratelimited(c, inum,
863 "read error %i from btree lookup", ret);
864 rbio->bio.bi_status = BLK_STS_IOERR;
865 bio_endio(&rbio->bio);
868 bch2_bkey_buf_exit(&sk, c);
871 void bch2_readahead(struct readahead_control *ractl)
873 struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
874 struct bch_fs *c = inode->v.i_sb->s_fs_info;
875 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
876 struct btree_trans trans;
877 struct btree_iter *iter;
879 struct readpages_iter readpages_iter;
882 ret = readpages_iter_init(&readpages_iter, ractl);
885 bch2_trans_init(&trans, c, 0, 0);
886 iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
889 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
891 while ((page = readpage_iter_next(&readpages_iter))) {
892 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
893 unsigned n = min_t(unsigned,
894 readpages_iter.nr_pages -
897 struct bch_read_bio *rbio =
898 rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
901 readpages_iter.idx++;
903 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
904 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
905 rbio->bio.bi_end_io = bch2_readpages_end_io;
906 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
908 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
912 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
914 bch2_trans_iter_put(&trans, iter);
915 bch2_trans_exit(&trans);
916 kfree(readpages_iter.pages);
919 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
920 u64 inum, struct page *page)
922 struct btree_trans trans;
923 struct btree_iter *iter;
925 bch2_page_state_create(page, __GFP_NOFAIL);
927 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
928 rbio->bio.bi_iter.bi_sector =
929 (sector_t) page->index << PAGE_SECTOR_SHIFT;
930 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
932 bch2_trans_init(&trans, c, 0, 0);
933 iter = bch2_trans_get_iter(&trans, BTREE_ID_extents, POS_MIN,
936 bchfs_read(&trans, iter, rbio, inum, NULL);
938 bch2_trans_iter_put(&trans, iter);
939 bch2_trans_exit(&trans);
942 int bch2_readpage(struct file *file, struct page *page)
944 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
945 struct bch_fs *c = inode->v.i_sb->s_fs_info;
946 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
947 struct bch_read_bio *rbio;
949 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
950 rbio->bio.bi_end_io = bch2_readpages_end_io;
952 __bchfs_readpage(c, rbio, inode->v.i_ino, page);
956 static void bch2_read_single_page_end_io(struct bio *bio)
958 complete(bio->bi_private);
961 static int bch2_read_single_page(struct page *page,
962 struct address_space *mapping)
964 struct bch_inode_info *inode = to_bch_ei(mapping->host);
965 struct bch_fs *c = inode->v.i_sb->s_fs_info;
966 struct bch_read_bio *rbio;
968 DECLARE_COMPLETION_ONSTACK(done);
970 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
971 io_opts(c, &inode->ei_inode));
972 rbio->bio.bi_private = &done;
973 rbio->bio.bi_end_io = bch2_read_single_page_end_io;
975 __bchfs_readpage(c, rbio, inode->v.i_ino, page);
976 wait_for_completion(&done);
978 ret = blk_status_to_errno(rbio->bio.bi_status);
984 SetPageUptodate(page);
990 struct bch_writepage_state {
991 struct bch_writepage_io *io;
992 struct bch_io_opts opts;
995 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
996 struct bch_inode_info *inode)
998 return (struct bch_writepage_state) {
999 .opts = io_opts(c, &inode->ei_inode)
1003 static void bch2_writepage_io_free(struct closure *cl)
1005 struct bch_writepage_io *io = container_of(cl,
1006 struct bch_writepage_io, cl);
1008 bio_put(&io->op.wbio.bio);
1011 static void bch2_writepage_io_done(struct closure *cl)
1013 struct bch_writepage_io *io = container_of(cl,
1014 struct bch_writepage_io, cl);
1015 struct bch_fs *c = io->op.c;
1016 struct bio *bio = &io->op.wbio.bio;
1017 struct bvec_iter_all iter;
1018 struct bio_vec *bvec;
1021 up(&io->op.c->io_in_flight);
1024 set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
1026 bio_for_each_segment_all(bvec, bio, iter) {
1027 struct bch_page_state *s;
1029 SetPageError(bvec->bv_page);
1030 mapping_set_error(bvec->bv_page->mapping, -EIO);
1032 s = __bch2_page_state(bvec->bv_page);
1033 spin_lock(&s->lock);
1034 for (i = 0; i < PAGE_SECTORS; i++)
1035 s->s[i].nr_replicas = 0;
1036 spin_unlock(&s->lock);
1040 if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1041 bio_for_each_segment_all(bvec, bio, iter) {
1042 struct bch_page_state *s;
1044 s = __bch2_page_state(bvec->bv_page);
1045 spin_lock(&s->lock);
1046 for (i = 0; i < PAGE_SECTORS; i++)
1047 s->s[i].nr_replicas = 0;
1048 spin_unlock(&s->lock);
1053 * racing with fallocate can cause us to add fewer sectors than
1054 * expected - but we shouldn't add more sectors than expected:
1056 BUG_ON(io->op.i_sectors_delta > 0);
1059 * (error (due to going RO) halfway through a page can screw that up
1062 BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1066 * PageWriteback is effectively our ref on the inode - fixup i_blocks
1067 * before calling end_page_writeback:
1069 i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1071 bio_for_each_segment_all(bvec, bio, iter) {
1072 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1074 if (atomic_dec_and_test(&s->write_count))
1075 end_page_writeback(bvec->bv_page);
1078 closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1081 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1083 struct bch_writepage_io *io = w->io;
1085 down(&io->op.c->io_in_flight);
1088 closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1089 continue_at(&io->cl, bch2_writepage_io_done, NULL);
1093 * Get a bch_writepage_io and add @page to it - appending to an existing one if
1094 * possible, else allocating a new one:
1096 static void bch2_writepage_io_alloc(struct bch_fs *c,
1097 struct writeback_control *wbc,
1098 struct bch_writepage_state *w,
1099 struct bch_inode_info *inode,
1101 unsigned nr_replicas)
1103 struct bch_write_op *op;
1105 w->io = container_of(bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS,
1106 &c->writepage_bioset),
1107 struct bch_writepage_io, op.wbio.bio);
1109 closure_init(&w->io->cl, NULL);
1110 w->io->inode = inode;
1113 bch2_write_op_init(op, c, w->opts);
1114 op->target = w->opts.foreground_target;
1115 op_journal_seq_set(op, &inode->ei_journal_seq);
1116 op->nr_replicas = nr_replicas;
1117 op->res.nr_replicas = nr_replicas;
1118 op->write_point = writepoint_hashed(inode->ei_last_dirtied);
1119 op->pos = POS(inode->v.i_ino, sector);
1120 op->wbio.bio.bi_iter.bi_sector = sector;
1121 op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
1124 static int __bch2_writepage(struct page *page,
1125 struct writeback_control *wbc,
1128 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1129 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1130 struct bch_writepage_state *w = data;
1131 struct bch_page_state *s, orig;
1132 unsigned i, offset, nr_replicas_this_write = U32_MAX;
1133 loff_t i_size = i_size_read(&inode->v);
1134 pgoff_t end_index = i_size >> PAGE_SHIFT;
1137 EBUG_ON(!PageUptodate(page));
1139 /* Is the page fully inside i_size? */
1140 if (page->index < end_index)
1143 /* Is the page fully outside i_size? (truncate in progress) */
1144 offset = i_size & (PAGE_SIZE - 1);
1145 if (page->index > end_index || !offset) {
1151 * The page straddles i_size. It must be zeroed out on each and every
1152 * writepage invocation because it may be mmapped. "A file is mapped
1153 * in multiples of the page size. For a file that is not a multiple of
1154 * the page size, the remaining memory is zeroed when mapped, and
1155 * writes to that region are not written out to the file."
1157 zero_user_segment(page, offset, PAGE_SIZE);
1159 s = bch2_page_state_create(page, __GFP_NOFAIL);
1161 ret = bch2_get_page_disk_reservation(c, inode, page, true);
1164 mapping_set_error(page->mapping, ret);
1169 /* Before unlocking the page, get copy of reservations: */
1172 for (i = 0; i < PAGE_SECTORS; i++) {
1173 if (s->s[i].state < SECTOR_DIRTY)
1176 nr_replicas_this_write =
1177 min_t(unsigned, nr_replicas_this_write,
1178 s->s[i].nr_replicas +
1179 s->s[i].replicas_reserved);
1182 for (i = 0; i < PAGE_SECTORS; i++) {
1183 if (s->s[i].state < SECTOR_DIRTY)
1186 s->s[i].nr_replicas = w->opts.compression
1187 ? 0 : nr_replicas_this_write;
1189 s->s[i].replicas_reserved = 0;
1190 s->s[i].state = SECTOR_ALLOCATED;
1193 BUG_ON(atomic_read(&s->write_count));
1194 atomic_set(&s->write_count, 1);
1196 BUG_ON(PageWriteback(page));
1197 set_page_writeback(page);
1203 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1206 while (offset < PAGE_SECTORS &&
1207 orig.s[offset].state < SECTOR_DIRTY)
1210 if (offset == PAGE_SECTORS)
1213 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1215 while (offset + sectors < PAGE_SECTORS &&
1216 orig.s[offset + sectors].state >= SECTOR_DIRTY)
1219 for (i = offset; i < offset + sectors; i++) {
1220 reserved_sectors += orig.s[i].replicas_reserved;
1221 dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1225 (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1226 bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1227 w->io->op.wbio.bio.bi_iter.bi_size + (sectors << 9) >=
1228 (BIO_MAX_VECS * PAGE_SIZE) ||
1229 bio_end_sector(&w->io->op.wbio.bio) != sector))
1230 bch2_writepage_do_io(w);
1233 bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1234 nr_replicas_this_write);
1236 atomic_inc(&s->write_count);
1238 BUG_ON(inode != w->io->inode);
1239 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1240 sectors << 9, offset << 9));
1242 /* Check for writing past i_size: */
1243 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1244 round_up(i_size, block_bytes(c)));
1246 w->io->op.res.sectors += reserved_sectors;
1247 w->io->op.i_sectors_delta -= dirty_sectors;
1248 w->io->op.new_i_size = i_size;
1253 if (atomic_dec_and_test(&s->write_count))
1254 end_page_writeback(page);
1259 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1261 struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1262 struct bch_writepage_state w =
1263 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1264 struct blk_plug plug;
1267 blk_start_plug(&plug);
1268 ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1270 bch2_writepage_do_io(&w);
1271 blk_finish_plug(&plug);
1275 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1277 struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1278 struct bch_writepage_state w =
1279 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1282 ret = __bch2_writepage(page, wbc, &w);
1284 bch2_writepage_do_io(&w);
1289 /* buffered writes: */
1291 int bch2_write_begin(struct file *file, struct address_space *mapping,
1292 loff_t pos, unsigned len, unsigned flags,
1293 struct page **pagep, void **fsdata)
1295 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1296 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1297 struct bch2_page_reservation *res;
1298 pgoff_t index = pos >> PAGE_SHIFT;
1299 unsigned offset = pos & (PAGE_SIZE - 1);
1303 res = kmalloc(sizeof(*res), GFP_KERNEL);
1307 bch2_page_reservation_init(c, inode, res);
1310 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1312 page = grab_cache_page_write_begin(mapping, index, flags);
1316 if (PageUptodate(page))
1319 /* If we're writing entire page, don't need to read it in first: */
1320 if (len == PAGE_SIZE)
1323 if (!offset && pos + len >= inode->v.i_size) {
1324 zero_user_segment(page, len, PAGE_SIZE);
1325 flush_dcache_page(page);
1329 if (index > inode->v.i_size >> PAGE_SHIFT) {
1330 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1331 flush_dcache_page(page);
1335 ret = bch2_read_single_page(page, mapping);
1339 ret = bch2_page_reservation_get(c, inode, page, res,
1342 if (!PageUptodate(page)) {
1344 * If the page hasn't been read in, we won't know if we
1345 * actually need a reservation - we don't actually need
1346 * to read here, we just need to check if the page is
1347 * fully backed by uncompressed data:
1362 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1368 int bch2_write_end(struct file *file, struct address_space *mapping,
1369 loff_t pos, unsigned len, unsigned copied,
1370 struct page *page, void *fsdata)
1372 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1373 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1374 struct bch2_page_reservation *res = fsdata;
1375 unsigned offset = pos & (PAGE_SIZE - 1);
1377 lockdep_assert_held(&inode->v.i_rwsem);
1379 if (unlikely(copied < len && !PageUptodate(page))) {
1381 * The page needs to be read in, but that would destroy
1382 * our partial write - simplest thing is to just force
1383 * userspace to redo the write:
1385 zero_user(page, 0, PAGE_SIZE);
1386 flush_dcache_page(page);
1390 spin_lock(&inode->v.i_lock);
1391 if (pos + copied > inode->v.i_size)
1392 i_size_write(&inode->v, pos + copied);
1393 spin_unlock(&inode->v.i_lock);
1396 if (!PageUptodate(page))
1397 SetPageUptodate(page);
1399 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1401 inode->ei_last_dirtied = (unsigned long) current;
1406 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1408 bch2_page_reservation_put(c, inode, res);
1414 #define WRITE_BATCH_PAGES 32
1416 static int __bch2_buffered_write(struct bch_inode_info *inode,
1417 struct address_space *mapping,
1418 struct iov_iter *iter,
1419 loff_t pos, unsigned len)
1421 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1422 struct page *pages[WRITE_BATCH_PAGES];
1423 struct bch2_page_reservation res;
1424 unsigned long index = pos >> PAGE_SHIFT;
1425 unsigned offset = pos & (PAGE_SIZE - 1);
1426 unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1427 unsigned i, reserved = 0, set_dirty = 0;
1428 unsigned copied = 0, nr_pages_copied = 0;
1432 BUG_ON(nr_pages > ARRAY_SIZE(pages));
1434 bch2_page_reservation_init(c, inode, &res);
1436 for (i = 0; i < nr_pages; i++) {
1437 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1444 len = min_t(unsigned, len,
1445 nr_pages * PAGE_SIZE - offset);
1450 if (offset && !PageUptodate(pages[0])) {
1451 ret = bch2_read_single_page(pages[0], mapping);
1456 if ((pos + len) & (PAGE_SIZE - 1) &&
1457 !PageUptodate(pages[nr_pages - 1])) {
1458 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1459 zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1461 ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1467 while (reserved < len) {
1468 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1469 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1470 unsigned pg_len = min_t(unsigned, len - reserved,
1471 PAGE_SIZE - pg_offset);
1473 ret = bch2_page_reservation_get(c, inode, page, &res,
1474 pg_offset, pg_len, true);
1476 if (ret && !PageUptodate(page)) {
1477 ret = bch2_read_single_page(page, mapping);
1479 goto retry_reservation;
1488 if (mapping_writably_mapped(mapping))
1489 for (i = 0; i < nr_pages; i++)
1490 flush_dcache_page(pages[i]);
1492 while (copied < len) {
1493 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1494 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1495 unsigned pg_len = min_t(unsigned, len - copied,
1496 PAGE_SIZE - pg_offset);
1497 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1498 iter, pg_offset, pg_len);
1503 if (!PageUptodate(page) &&
1504 pg_copied != PAGE_SIZE &&
1505 pos + copied + pg_copied < inode->v.i_size) {
1506 zero_user(page, 0, PAGE_SIZE);
1510 flush_dcache_page(page);
1511 iov_iter_advance(iter, pg_copied);
1512 copied += pg_copied;
1514 if (pg_copied != pg_len)
1521 spin_lock(&inode->v.i_lock);
1522 if (pos + copied > inode->v.i_size)
1523 i_size_write(&inode->v, pos + copied);
1524 spin_unlock(&inode->v.i_lock);
1526 while (set_dirty < copied) {
1527 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1528 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1529 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1530 PAGE_SIZE - pg_offset);
1532 if (!PageUptodate(page))
1533 SetPageUptodate(page);
1535 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1539 set_dirty += pg_len;
1542 nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1543 inode->ei_last_dirtied = (unsigned long) current;
1545 for (i = nr_pages_copied; i < nr_pages; i++) {
1546 unlock_page(pages[i]);
1550 bch2_page_reservation_put(c, inode, &res);
1552 return copied ?: ret;
1555 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1557 struct file *file = iocb->ki_filp;
1558 struct address_space *mapping = file->f_mapping;
1559 struct bch_inode_info *inode = file_bch_inode(file);
1560 loff_t pos = iocb->ki_pos;
1561 ssize_t written = 0;
1564 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1567 unsigned offset = pos & (PAGE_SIZE - 1);
1568 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1569 PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1572 * Bring in the user page that we will copy from _first_.
1573 * Otherwise there's a nasty deadlock on copying from the
1574 * same page as we're writing to, without it being marked
1577 * Not only is this an optimisation, but it is also required
1578 * to check that the address is actually valid, when atomic
1579 * usercopies are used, below.
1581 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1582 bytes = min_t(unsigned long, iov_iter_count(iter),
1583 PAGE_SIZE - offset);
1585 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1591 if (unlikely(fatal_signal_pending(current))) {
1596 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1597 if (unlikely(ret < 0))
1602 if (unlikely(ret == 0)) {
1604 * If we were unable to copy any data at all, we must
1605 * fall back to a single segment length write.
1607 * If we didn't fallback here, we could livelock
1608 * because not all segments in the iov can be copied at
1609 * once without a pagefault.
1611 bytes = min_t(unsigned long, PAGE_SIZE - offset,
1612 iov_iter_single_seg_count(iter));
1619 balance_dirty_pages_ratelimited(mapping);
1620 } while (iov_iter_count(iter));
1622 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1624 return written ? written : ret;
1627 /* O_DIRECT reads */
1629 static void bio_check_or_release(struct bio *bio, bool check_dirty)
1632 bio_check_pages_dirty(bio);
1634 bio_release_pages(bio, false);
1639 static void bch2_dio_read_complete(struct closure *cl)
1641 struct dio_read *dio = container_of(cl, struct dio_read, cl);
1643 dio->req->ki_complete(dio->req, dio->ret, 0);
1644 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1647 static void bch2_direct_IO_read_endio(struct bio *bio)
1649 struct dio_read *dio = bio->bi_private;
1652 dio->ret = blk_status_to_errno(bio->bi_status);
1654 closure_put(&dio->cl);
1657 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1659 struct dio_read *dio = bio->bi_private;
1660 bool should_dirty = dio->should_dirty;
1662 bch2_direct_IO_read_endio(bio);
1663 bio_check_or_release(bio, should_dirty);
1666 static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
1668 struct file *file = req->ki_filp;
1669 struct bch_inode_info *inode = file_bch_inode(file);
1670 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1671 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1672 struct dio_read *dio;
1674 loff_t offset = req->ki_pos;
1675 bool sync = is_sync_kiocb(req);
1679 if ((offset|iter->count) & (block_bytes(c) - 1))
1682 ret = min_t(loff_t, iter->count,
1683 max_t(loff_t, 0, i_size_read(&inode->v) - offset));
1688 shorten = iov_iter_count(iter) - round_up(ret, block_bytes(c));
1689 iter->count -= shorten;
1691 bio = bio_alloc_bioset(GFP_KERNEL,
1692 iov_iter_npages(iter, BIO_MAX_VECS),
1693 &c->dio_read_bioset);
1695 bio->bi_end_io = bch2_direct_IO_read_endio;
1697 dio = container_of(bio, struct dio_read, rbio.bio);
1698 closure_init(&dio->cl, NULL);
1701 * this is a _really_ horrible hack just to avoid an atomic sub at the
1705 set_closure_fn(&dio->cl, bch2_dio_read_complete, NULL);
1706 atomic_set(&dio->cl.remaining,
1707 CLOSURE_REMAINING_INITIALIZER -
1709 CLOSURE_DESTRUCTOR);
1711 atomic_set(&dio->cl.remaining,
1712 CLOSURE_REMAINING_INITIALIZER + 1);
1718 * This is one of the sketchier things I've encountered: we have to skip
1719 * the dirtying of requests that are internal from the kernel (i.e. from
1720 * loopback), because we'll deadlock on page_lock.
1722 dio->should_dirty = iter_is_iovec(iter);
1725 while (iter->count) {
1726 bio = bio_alloc_bioset(GFP_KERNEL,
1727 iov_iter_npages(iter, BIO_MAX_VECS),
1729 bio->bi_end_io = bch2_direct_IO_read_split_endio;
1731 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1732 bio->bi_iter.bi_sector = offset >> 9;
1733 bio->bi_private = dio;
1735 ret = bio_iov_iter_get_pages(bio, iter);
1737 /* XXX: fault inject this path */
1738 bio->bi_status = BLK_STS_RESOURCE;
1743 offset += bio->bi_iter.bi_size;
1745 if (dio->should_dirty)
1746 bio_set_pages_dirty(bio);
1749 closure_get(&dio->cl);
1751 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1754 iter->count += shorten;
1757 closure_sync(&dio->cl);
1758 closure_debug_destroy(&dio->cl);
1760 bio_check_or_release(&dio->rbio.bio, dio->should_dirty);
1763 return -EIOCBQUEUED;
1767 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1769 struct file *file = iocb->ki_filp;
1770 struct bch_inode_info *inode = file_bch_inode(file);
1771 struct address_space *mapping = file->f_mapping;
1772 size_t count = iov_iter_count(iter);
1776 return 0; /* skip atime */
1778 if (iocb->ki_flags & IOCB_DIRECT) {
1779 struct blk_plug plug;
1781 ret = filemap_write_and_wait_range(mapping,
1783 iocb->ki_pos + count - 1);
1787 file_accessed(file);
1789 blk_start_plug(&plug);
1790 ret = bch2_direct_IO_read(iocb, iter);
1791 blk_finish_plug(&plug);
1794 iocb->ki_pos += ret;
1796 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1797 ret = generic_file_read_iter(iocb, iter);
1798 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1804 /* O_DIRECT writes */
1806 static void bch2_dio_write_loop_async(struct bch_write_op *);
1808 static long bch2_dio_write_loop(struct dio_write *dio)
1810 bool kthread = (current->flags & PF_KTHREAD) != 0;
1811 struct kiocb *req = dio->req;
1812 struct address_space *mapping = req->ki_filp->f_mapping;
1813 struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1814 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1815 struct bio *bio = &dio->op.wbio.bio;
1816 struct bvec_iter_all iter;
1818 unsigned unaligned, iter_count;
1819 bool sync = dio->sync, dropped_locks;
1825 down(&c->io_in_flight);
1828 iter_count = dio->iter.count;
1831 kthread_use_mm(dio->mm);
1832 BUG_ON(current->faults_disabled_mapping);
1833 current->faults_disabled_mapping = mapping;
1835 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1837 dropped_locks = fdm_dropped_locks();
1839 current->faults_disabled_mapping = NULL;
1841 kthread_unuse_mm(dio->mm);
1844 * If the fault handler returned an error but also signalled
1845 * that it dropped & retook ei_pagecache_lock, we just need to
1846 * re-shoot down the page cache and retry:
1848 if (dropped_locks && ret)
1851 if (unlikely(ret < 0))
1854 if (unlikely(dropped_locks)) {
1855 ret = write_invalidate_inode_pages_range(mapping,
1857 req->ki_pos + iter_count - 1);
1861 if (!bio->bi_iter.bi_size)
1865 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1866 bio->bi_iter.bi_size -= unaligned;
1867 iov_iter_revert(&dio->iter, unaligned);
1869 if (!bio->bi_iter.bi_size) {
1871 * bio_iov_iter_get_pages was only able to get <
1872 * blocksize worth of pages:
1874 bio_for_each_segment_all(bv, bio, iter)
1875 put_page(bv->bv_page);
1880 bch2_write_op_init(&dio->op, c, io_opts(c, &inode->ei_inode));
1881 dio->op.end_io = bch2_dio_write_loop_async;
1882 dio->op.target = dio->op.opts.foreground_target;
1883 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1884 dio->op.write_point = writepoint_hashed((unsigned long) current);
1885 dio->op.nr_replicas = dio->op.opts.data_replicas;
1886 dio->op.pos = POS(inode->v.i_ino, (u64) req->ki_pos >> 9);
1888 if ((req->ki_flags & IOCB_DSYNC) &&
1889 !c->opts.journal_flush_disabled)
1890 dio->op.flags |= BCH_WRITE_FLUSH;
1891 dio->op.flags |= BCH_WRITE_CHECK_ENOSPC;
1893 ret = bch2_disk_reservation_get(c, &dio->op.res, bio_sectors(bio),
1894 dio->op.opts.data_replicas, 0);
1895 if (unlikely(ret) &&
1896 !bch2_check_range_allocated(c, dio->op.pos,
1898 dio->op.opts.data_replicas,
1899 dio->op.opts.compression != 0))
1902 task_io_account_write(bio->bi_iter.bi_size);
1904 if (!dio->sync && !dio->loop && dio->iter.count) {
1905 struct iovec *iov = dio->inline_vecs;
1907 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1908 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1910 if (unlikely(!iov)) {
1911 dio->sync = sync = true;
1915 dio->free_iov = true;
1918 memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1919 dio->iter.iov = iov;
1923 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1926 wait_for_completion(&dio->done);
1928 return -EIOCBQUEUED;
1930 i_sectors_acct(c, inode, &dio->quota_res,
1931 dio->op.i_sectors_delta);
1932 req->ki_pos += (u64) dio->op.written << 9;
1933 dio->written += dio->op.written;
1935 spin_lock(&inode->v.i_lock);
1936 if (req->ki_pos > inode->v.i_size)
1937 i_size_write(&inode->v, req->ki_pos);
1938 spin_unlock(&inode->v.i_lock);
1940 if (likely(!bio_flagged(bio, BIO_NO_PAGE_REF)))
1941 bio_for_each_segment_all(bv, bio, iter)
1942 put_page(bv->bv_page);
1944 if (dio->op.error) {
1945 set_bit(EI_INODE_ERROR, &inode->ei_flags);
1949 if (!dio->iter.count)
1953 reinit_completion(&dio->done);
1956 ret = dio->op.error ?: ((long) dio->written << 9);
1958 up(&c->io_in_flight);
1959 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1960 bch2_quota_reservation_put(c, inode, &dio->quota_res);
1963 kfree(dio->iter.iov);
1967 /* inode->i_dio_count is our ref on inode and thus bch_fs */
1968 inode_dio_end(&inode->v);
1971 req->ki_complete(req, ret, 0);
1977 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1979 struct dio_write *dio = container_of(op, struct dio_write, op);
1982 complete(&dio->done);
1984 bch2_dio_write_loop(dio);
1988 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1990 struct file *file = req->ki_filp;
1991 struct address_space *mapping = file->f_mapping;
1992 struct bch_inode_info *inode = file_bch_inode(file);
1993 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1994 struct dio_write *dio;
1996 bool locked = true, extending;
2000 prefetch((void *) &c->opts + 64);
2001 prefetch(&inode->ei_inode);
2002 prefetch((void *) &inode->ei_inode + 64);
2004 inode_lock(&inode->v);
2006 ret = generic_write_checks(req, iter);
2007 if (unlikely(ret <= 0))
2010 ret = file_remove_privs(file);
2014 ret = file_update_time(file);
2018 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
2021 inode_dio_begin(&inode->v);
2022 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2024 extending = req->ki_pos + iter->count > inode->v.i_size;
2026 inode_unlock(&inode->v);
2030 bio = bio_alloc_bioset(GFP_KERNEL,
2031 iov_iter_is_bvec(iter)
2033 : iov_iter_npages(iter, BIO_MAX_VECS),
2034 &c->dio_write_bioset);
2035 dio = container_of(bio, struct dio_write, op.wbio.bio);
2036 init_completion(&dio->done);
2038 dio->mm = current->mm;
2040 dio->sync = is_sync_kiocb(req) || extending;
2041 dio->free_iov = false;
2042 dio->quota_res.sectors = 0;
2046 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
2047 iter->count >> 9, true);
2051 ret = write_invalidate_inode_pages_range(mapping,
2053 req->ki_pos + iter->count - 1);
2057 ret = bch2_dio_write_loop(dio);
2060 inode_unlock(&inode->v);
2063 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2064 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2066 inode_dio_end(&inode->v);
2070 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2072 struct file *file = iocb->ki_filp;
2073 struct bch_inode_info *inode = file_bch_inode(file);
2076 if (iocb->ki_flags & IOCB_DIRECT)
2077 return bch2_direct_write(iocb, from);
2079 /* We can write back this queue in page reclaim */
2080 current->backing_dev_info = inode_to_bdi(&inode->v);
2081 inode_lock(&inode->v);
2083 ret = generic_write_checks(iocb, from);
2087 ret = file_remove_privs(file);
2091 ret = file_update_time(file);
2095 ret = bch2_buffered_write(iocb, from);
2096 if (likely(ret > 0))
2097 iocb->ki_pos += ret;
2099 inode_unlock(&inode->v);
2100 current->backing_dev_info = NULL;
2103 ret = generic_write_sync(iocb, ret);
2110 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2112 struct bch_inode_info *inode = file_bch_inode(file);
2113 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2116 ret = file_write_and_wait_range(file, start, end);
2120 if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2123 ret = sync_inode_metadata(&inode->v, 1);
2127 if (!c->opts.journal_flush_disabled)
2128 ret = bch2_journal_flush_seq(&c->journal,
2129 inode->ei_journal_seq);
2130 ret2 = file_check_and_advance_wb_err(file);
2137 static inline int range_has_data(struct bch_fs *c,
2141 struct btree_trans trans;
2142 struct btree_iter *iter;
2146 bch2_trans_init(&trans, c, 0, 0);
2148 for_each_btree_key(&trans, iter, BTREE_ID_extents, start, 0, k, ret) {
2149 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2152 if (bkey_extent_is_data(k.k)) {
2157 bch2_trans_iter_put(&trans, iter);
2159 return bch2_trans_exit(&trans) ?: ret;
2162 static int __bch2_truncate_page(struct bch_inode_info *inode,
2163 pgoff_t index, loff_t start, loff_t end)
2165 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2166 struct address_space *mapping = inode->v.i_mapping;
2167 struct bch_page_state *s;
2168 unsigned start_offset = start & (PAGE_SIZE - 1);
2169 unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2174 /* Page boundary? Nothing to do */
2175 if (!((index == start >> PAGE_SHIFT && start_offset) ||
2176 (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2180 if (index << PAGE_SHIFT >= inode->v.i_size)
2183 page = find_lock_page(mapping, index);
2186 * XXX: we're doing two index lookups when we end up reading the
2189 ret = range_has_data(c,
2190 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2191 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2195 page = find_or_create_page(mapping, index, GFP_KERNEL);
2196 if (unlikely(!page)) {
2202 s = bch2_page_state_create(page, 0);
2208 if (!PageUptodate(page)) {
2209 ret = bch2_read_single_page(page, mapping);
2214 if (index != start >> PAGE_SHIFT)
2216 if (index != end >> PAGE_SHIFT)
2217 end_offset = PAGE_SIZE;
2219 for (i = round_up(start_offset, block_bytes(c)) >> 9;
2220 i < round_down(end_offset, block_bytes(c)) >> 9;
2222 s->s[i].nr_replicas = 0;
2223 s->s[i].state = SECTOR_UNALLOCATED;
2226 zero_user_segment(page, start_offset, end_offset);
2229 * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2231 * XXX: because we aren't currently tracking whether the page has actual
2232 * data in it (vs. just 0s, or only partially written) this wrong. ick.
2234 ret = bch2_get_page_disk_reservation(c, inode, page, false);
2238 * This removes any writeable userspace mappings; we need to force
2239 * .page_mkwrite to be called again before any mmapped writes, to
2240 * redirty the full page:
2243 __set_page_dirty_nobuffers(page);
2251 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2253 return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2254 from, round_up(from, PAGE_SIZE));
2257 static int bch2_extend(struct user_namespace *mnt_userns,
2258 struct bch_inode_info *inode,
2259 struct bch_inode_unpacked *inode_u,
2260 struct iattr *iattr)
2262 struct address_space *mapping = inode->v.i_mapping;
2268 * this has to be done _before_ extending i_size:
2270 ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2274 truncate_setsize(&inode->v, iattr->ia_size);
2276 return bch2_setattr_nonsize(mnt_userns, inode, iattr);
2279 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2280 struct bch_inode_unpacked *bi,
2283 bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2287 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2288 struct bch_inode_unpacked *bi, void *p)
2290 u64 *new_i_size = p;
2292 bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2293 bi->bi_size = *new_i_size;
2297 int bch2_truncate(struct user_namespace *mnt_userns,
2298 struct bch_inode_info *inode, struct iattr *iattr)
2300 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2301 struct address_space *mapping = inode->v.i_mapping;
2302 struct bch_inode_unpacked inode_u;
2303 struct btree_trans trans;
2304 struct btree_iter *iter;
2305 u64 new_i_size = iattr->ia_size;
2306 s64 i_sectors_delta = 0;
2310 * If the truncate call with change the size of the file, the
2311 * cmtimes should be updated. If the size will not change, we
2312 * do not need to update the cmtimes.
2314 if (iattr->ia_size != inode->v.i_size) {
2315 if (!(iattr->ia_valid & ATTR_MTIME))
2316 ktime_get_coarse_real_ts64(&iattr->ia_mtime);
2317 if (!(iattr->ia_valid & ATTR_CTIME))
2318 ktime_get_coarse_real_ts64(&iattr->ia_ctime);
2319 iattr->ia_valid |= ATTR_MTIME|ATTR_CTIME;
2322 inode_dio_wait(&inode->v);
2323 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2326 * fetch current on disk i_size: inode is locked, i_size can only
2327 * increase underneath us:
2329 bch2_trans_init(&trans, c, 0, 0);
2330 iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2331 ret = PTR_ERR_OR_ZERO(iter);
2332 bch2_trans_iter_put(&trans, iter);
2333 bch2_trans_exit(&trans);
2339 * check this before next assertion; on filesystem error our normal
2340 * invariants are a bit broken (truncate has to truncate the page cache
2341 * before the inode).
2343 ret = bch2_journal_error(&c->journal);
2347 WARN_ON(!test_bit(EI_INODE_ERROR, &inode->ei_flags) &&
2348 inode->v.i_size < inode_u.bi_size);
2350 if (iattr->ia_size > inode->v.i_size) {
2351 ret = bch2_extend(mnt_userns, inode, &inode_u, iattr);
2355 iattr->ia_valid &= ~ATTR_SIZE;
2357 ret = bch2_truncate_page(inode, iattr->ia_size);
2362 * When extending, we're going to write the new i_size to disk
2363 * immediately so we need to flush anything above the current on disk
2366 * Also, when extending we need to flush the page that i_size currently
2367 * straddles - if it's mapped to userspace, we need to ensure that
2368 * userspace has to redirty it and call .mkwrite -> set_page_dirty
2369 * again to allocate the part of the page that was extended.
2371 if (iattr->ia_size > inode_u.bi_size)
2372 ret = filemap_write_and_wait_range(mapping,
2374 iattr->ia_size - 1);
2375 else if (iattr->ia_size & (PAGE_SIZE - 1))
2376 ret = filemap_write_and_wait_range(mapping,
2377 round_down(iattr->ia_size, PAGE_SIZE),
2378 iattr->ia_size - 1);
2382 mutex_lock(&inode->ei_update_lock);
2383 ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2385 mutex_unlock(&inode->ei_update_lock);
2390 truncate_setsize(&inode->v, iattr->ia_size);
2392 ret = bch2_fpunch(c, inode->v.i_ino,
2393 round_up(iattr->ia_size, block_bytes(c)) >> 9,
2394 U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2395 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2400 mutex_lock(&inode->ei_update_lock);
2401 ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL, 0);
2402 mutex_unlock(&inode->ei_update_lock);
2404 ret = bch2_setattr_nonsize(mnt_userns, inode, iattr);
2406 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2412 static int inode_update_times_fn(struct bch_inode_info *inode,
2413 struct bch_inode_unpacked *bi, void *p)
2415 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2417 bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2421 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2423 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2424 u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2425 u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2428 inode_lock(&inode->v);
2429 inode_dio_wait(&inode->v);
2430 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2432 ret = __bch2_truncate_page(inode,
2433 offset >> PAGE_SHIFT,
2434 offset, offset + len);
2438 if (offset >> PAGE_SHIFT !=
2439 (offset + len) >> PAGE_SHIFT) {
2440 ret = __bch2_truncate_page(inode,
2441 (offset + len) >> PAGE_SHIFT,
2442 offset, offset + len);
2447 truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2449 if (discard_start < discard_end) {
2450 s64 i_sectors_delta = 0;
2452 ret = bch2_fpunch(c, inode->v.i_ino,
2453 discard_start, discard_end,
2454 &inode->ei_journal_seq,
2456 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2459 mutex_lock(&inode->ei_update_lock);
2460 ret = bch2_write_inode(c, inode, inode_update_times_fn, NULL,
2461 ATTR_MTIME|ATTR_CTIME) ?: ret;
2462 mutex_unlock(&inode->ei_update_lock);
2464 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2465 inode_unlock(&inode->v);
2470 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2471 loff_t offset, loff_t len,
2474 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2475 struct address_space *mapping = inode->v.i_mapping;
2476 struct bkey_buf copy;
2477 struct btree_trans trans;
2478 struct btree_iter *src, *dst, *del;
2479 loff_t shift, new_size;
2483 if ((offset | len) & (block_bytes(c) - 1))
2487 * We need i_mutex to keep the page cache consistent with the extents
2488 * btree, and the btree consistent with i_size - we don't need outside
2489 * locking for the extents btree itself, because we're using linked
2492 inode_lock(&inode->v);
2493 inode_dio_wait(&inode->v);
2494 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2498 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2502 if (offset >= inode->v.i_size)
2505 src_start = U64_MAX;
2509 if (offset + len >= inode->v.i_size)
2512 src_start = offset + len;
2516 new_size = inode->v.i_size + shift;
2518 ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2523 i_size_write(&inode->v, new_size);
2524 mutex_lock(&inode->ei_update_lock);
2525 ret = bch2_write_inode_size(c, inode, new_size,
2526 ATTR_MTIME|ATTR_CTIME);
2527 mutex_unlock(&inode->ei_update_lock);
2529 s64 i_sectors_delta = 0;
2531 ret = bch2_fpunch(c, inode->v.i_ino,
2532 offset >> 9, (offset + len) >> 9,
2533 &inode->ei_journal_seq,
2535 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2541 bch2_bkey_buf_init(©);
2542 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
2543 src = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2544 POS(inode->v.i_ino, src_start >> 9),
2546 dst = bch2_trans_copy_iter(&trans, src);
2547 del = bch2_trans_copy_iter(&trans, src);
2549 while (ret == 0 || ret == -EINTR) {
2550 struct disk_reservation disk_res =
2551 bch2_disk_reservation_init(c, 0);
2552 struct bkey_i delete;
2554 struct bpos next_pos;
2555 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2556 struct bpos atomic_end;
2557 unsigned trigger_flags = 0;
2560 ? bch2_btree_iter_peek_prev(src)
2561 : bch2_btree_iter_peek(src);
2562 if ((ret = bkey_err(k)))
2565 if (!k.k || k.k->p.inode != inode->v.i_ino)
2569 bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2572 bch2_bkey_buf_reassemble(©, c, k);
2575 bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
2576 bch2_cut_front(move_pos, copy.k);
2578 copy.k->k.p.offset += shift >> 9;
2579 bch2_btree_iter_set_pos(dst, bkey_start_pos(©.k->k));
2581 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2585 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2587 move_pos = atomic_end;
2588 move_pos.offset -= shift >> 9;
2591 bch2_cut_back(atomic_end, copy.k);
2595 bkey_init(&delete.k);
2596 delete.k.p = copy.k->k.p;
2597 delete.k.size = copy.k->k.size;
2598 delete.k.p.offset -= shift >> 9;
2599 bch2_btree_iter_set_pos(del, bkey_start_pos(&delete.k));
2601 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2603 if (copy.k->k.size == k.k->size) {
2605 * If we're moving the entire extent, we can skip
2608 trigger_flags |= BTREE_TRIGGER_NORUN;
2610 /* We might end up splitting compressed extents: */
2612 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2614 ret = bch2_disk_reservation_get(c, &disk_res,
2615 copy.k->k.size, nr_ptrs,
2616 BCH_DISK_RESERVATION_NOFAIL);
2620 ret = bch2_btree_iter_traverse(del) ?:
2621 bch2_trans_update(&trans, del, &delete, trigger_flags) ?:
2622 bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2623 bch2_trans_commit(&trans, &disk_res,
2624 &inode->ei_journal_seq,
2625 BTREE_INSERT_NOFAIL);
2626 bch2_disk_reservation_put(c, &disk_res);
2629 bch2_btree_iter_set_pos(src, next_pos);
2631 bch2_trans_iter_put(&trans, del);
2632 bch2_trans_iter_put(&trans, dst);
2633 bch2_trans_iter_put(&trans, src);
2634 bch2_trans_exit(&trans);
2635 bch2_bkey_buf_exit(©, c);
2641 i_size_write(&inode->v, new_size);
2642 mutex_lock(&inode->ei_update_lock);
2643 ret = bch2_write_inode_size(c, inode, new_size,
2644 ATTR_MTIME|ATTR_CTIME);
2645 mutex_unlock(&inode->ei_update_lock);
2648 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2649 inode_unlock(&inode->v);
2653 static int __bchfs_fallocate(struct bch_inode_info *inode, int mode,
2654 u64 start_sector, u64 end_sector)
2656 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2657 struct btree_trans trans;
2658 struct btree_iter *iter;
2659 struct bpos end_pos = POS(inode->v.i_ino, end_sector);
2660 unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2663 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
2665 iter = bch2_trans_get_iter(&trans, BTREE_ID_extents,
2666 POS(inode->v.i_ino, start_sector),
2667 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2669 while (!ret && bkey_cmp(iter->pos, end_pos) < 0) {
2670 s64 i_sectors_delta = 0;
2671 struct disk_reservation disk_res = { 0 };
2672 struct quota_res quota_res = { 0 };
2673 struct bkey_i_reservation reservation;
2677 bch2_trans_begin(&trans);
2679 k = bch2_btree_iter_peek_slot(iter);
2680 if ((ret = bkey_err(k)))
2683 /* already reserved */
2684 if (k.k->type == KEY_TYPE_reservation &&
2685 bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2686 bch2_btree_iter_next_slot(iter);
2690 if (bkey_extent_is_data(k.k) &&
2691 !(mode & FALLOC_FL_ZERO_RANGE)) {
2692 bch2_btree_iter_next_slot(iter);
2696 bkey_reservation_init(&reservation.k_i);
2697 reservation.k.type = KEY_TYPE_reservation;
2698 reservation.k.p = k.k->p;
2699 reservation.k.size = k.k->size;
2701 bch2_cut_front(iter->pos, &reservation.k_i);
2702 bch2_cut_back(end_pos, &reservation.k_i);
2704 sectors = reservation.k.size;
2705 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2707 if (!bkey_extent_is_allocation(k.k)) {
2708 ret = bch2_quota_reservation_add(c, inode,
2715 if (reservation.v.nr_replicas < replicas ||
2716 bch2_bkey_sectors_compressed(k)) {
2717 ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2722 reservation.v.nr_replicas = disk_res.nr_replicas;
2725 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2726 &disk_res, &inode->ei_journal_seq,
2727 0, &i_sectors_delta, true);
2728 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
2730 bch2_quota_reservation_put(c, inode, "a_res);
2731 bch2_disk_reservation_put(c, &disk_res);
2735 bch2_trans_iter_put(&trans, iter);
2736 bch2_trans_exit(&trans);
2740 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2741 loff_t offset, loff_t len)
2743 struct address_space *mapping = inode->v.i_mapping;
2744 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2745 loff_t end = offset + len;
2746 loff_t block_start = round_down(offset, block_bytes(c));
2747 loff_t block_end = round_up(end, block_bytes(c));
2750 inode_lock(&inode->v);
2751 inode_dio_wait(&inode->v);
2752 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2754 if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2755 ret = inode_newsize_ok(&inode->v, end);
2760 if (mode & FALLOC_FL_ZERO_RANGE) {
2761 ret = __bch2_truncate_page(inode,
2762 offset >> PAGE_SHIFT,
2766 offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2767 ret = __bch2_truncate_page(inode,
2774 truncate_pagecache_range(&inode->v, offset, end - 1);
2777 ret = __bchfs_fallocate(inode, mode, block_start >> 9, block_end >> 9);
2782 * Do we need to extend the file?
2784 * If we zeroed up to the end of the file, we dropped whatever writes
2785 * were going to write out the current i_size, so we have to extend
2786 * manually even if FL_KEEP_SIZE was set:
2788 if (end >= inode->v.i_size &&
2789 (!(mode & FALLOC_FL_KEEP_SIZE) ||
2790 (mode & FALLOC_FL_ZERO_RANGE))) {
2793 * Sync existing appends before extending i_size,
2794 * as in bch2_extend():
2796 ret = filemap_write_and_wait_range(mapping,
2797 inode->ei_inode.bi_size, S64_MAX);
2801 if (mode & FALLOC_FL_KEEP_SIZE)
2802 end = inode->v.i_size;
2804 i_size_write(&inode->v, end);
2806 mutex_lock(&inode->ei_update_lock);
2807 ret = bch2_write_inode_size(c, inode, end, 0);
2808 mutex_unlock(&inode->ei_update_lock);
2811 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2812 inode_unlock(&inode->v);
2816 long bch2_fallocate_dispatch(struct file *file, int mode,
2817 loff_t offset, loff_t len)
2819 struct bch_inode_info *inode = file_bch_inode(file);
2820 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2823 if (!percpu_ref_tryget(&c->writes))
2826 if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2827 ret = bchfs_fallocate(inode, mode, offset, len);
2828 else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2829 ret = bchfs_fpunch(inode, offset, len);
2830 else if (mode == FALLOC_FL_INSERT_RANGE)
2831 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2832 else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2833 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2837 percpu_ref_put(&c->writes);
2842 static void mark_range_unallocated(struct bch_inode_info *inode,
2843 loff_t start, loff_t end)
2845 pgoff_t index = start >> PAGE_SHIFT;
2846 pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2847 struct pagevec pvec;
2849 pagevec_init(&pvec);
2852 unsigned nr_pages, i, j;
2854 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2859 for (i = 0; i < nr_pages; i++) {
2860 struct page *page = pvec.pages[i];
2861 struct bch_page_state *s;
2864 s = bch2_page_state(page);
2867 spin_lock(&s->lock);
2868 for (j = 0; j < PAGE_SECTORS; j++)
2869 s->s[j].nr_replicas = 0;
2870 spin_unlock(&s->lock);
2875 pagevec_release(&pvec);
2876 } while (index <= end_index);
2879 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2880 struct file *file_dst, loff_t pos_dst,
2881 loff_t len, unsigned remap_flags)
2883 struct bch_inode_info *src = file_bch_inode(file_src);
2884 struct bch_inode_info *dst = file_bch_inode(file_dst);
2885 struct bch_fs *c = src->v.i_sb->s_fs_info;
2886 s64 i_sectors_delta = 0;
2890 if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2893 if (remap_flags & REMAP_FILE_DEDUP)
2896 if ((pos_src & (block_bytes(c) - 1)) ||
2897 (pos_dst & (block_bytes(c) - 1)))
2901 abs(pos_src - pos_dst) < len)
2904 bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2906 file_update_time(file_dst);
2908 inode_dio_wait(&src->v);
2909 inode_dio_wait(&dst->v);
2911 ret = generic_remap_file_range_prep(file_src, pos_src,
2914 if (ret < 0 || len == 0)
2917 aligned_len = round_up((u64) len, block_bytes(c));
2919 ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2920 pos_dst, pos_dst + len - 1);
2924 mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2926 ret = bch2_remap_range(c,
2927 POS(dst->v.i_ino, pos_dst >> 9),
2928 POS(src->v.i_ino, pos_src >> 9),
2930 &dst->ei_journal_seq,
2931 pos_dst + len, &i_sectors_delta);
2936 * due to alignment, we might have remapped slightly more than requsted
2938 ret = min((u64) ret << 9, (u64) len);
2940 /* XXX get a quota reservation */
2941 i_sectors_acct(c, dst, NULL, i_sectors_delta);
2943 spin_lock(&dst->v.i_lock);
2944 if (pos_dst + ret > dst->v.i_size)
2945 i_size_write(&dst->v, pos_dst + ret);
2946 spin_unlock(&dst->v.i_lock);
2948 if (((file_dst->f_flags & (__O_SYNC | O_DSYNC)) ||
2949 IS_SYNC(file_inode(file_dst))) &&
2950 !c->opts.journal_flush_disabled)
2951 ret = bch2_journal_flush_seq(&c->journal, dst->ei_journal_seq);
2953 bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2960 static int page_data_offset(struct page *page, unsigned offset)
2962 struct bch_page_state *s = bch2_page_state(page);
2966 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2967 if (s->s[i].state >= SECTOR_DIRTY)
2973 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2974 loff_t start_offset,
2977 struct address_space *mapping = vinode->i_mapping;
2979 pgoff_t start_index = start_offset >> PAGE_SHIFT;
2980 pgoff_t end_index = end_offset >> PAGE_SHIFT;
2981 pgoff_t index = start_index;
2985 while (index <= end_index) {
2986 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2989 offset = page_data_offset(page,
2990 page->index == start_index
2991 ? start_offset & (PAGE_SIZE - 1)
2994 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2996 start_offset, end_offset);
3012 static loff_t bch2_seek_data(struct file *file, u64 offset)
3014 struct bch_inode_info *inode = file_bch_inode(file);
3015 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3016 struct btree_trans trans;
3017 struct btree_iter *iter;
3019 u64 isize, next_data = MAX_LFS_FILESIZE;
3022 isize = i_size_read(&inode->v);
3023 if (offset >= isize)
3026 bch2_trans_init(&trans, c, 0, 0);
3028 for_each_btree_key(&trans, iter, BTREE_ID_extents,
3029 POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
3030 if (k.k->p.inode != inode->v.i_ino) {
3032 } else if (bkey_extent_is_data(k.k)) {
3033 next_data = max(offset, bkey_start_offset(k.k) << 9);
3035 } else if (k.k->p.offset >> 9 > isize)
3038 bch2_trans_iter_put(&trans, iter);
3040 ret = bch2_trans_exit(&trans) ?: ret;
3044 if (next_data > offset)
3045 next_data = bch2_seek_pagecache_data(&inode->v,
3048 if (next_data >= isize)
3051 return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
3054 static int __page_hole_offset(struct page *page, unsigned offset)
3056 struct bch_page_state *s = bch2_page_state(page);
3062 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3063 if (s->s[i].state < SECTOR_DIRTY)
3069 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3071 pgoff_t index = offset >> PAGE_SHIFT;
3076 page = find_lock_page(mapping, index);
3080 pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3082 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3089 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3090 loff_t start_offset,
3093 struct address_space *mapping = vinode->i_mapping;
3094 loff_t offset = start_offset, hole;
3096 while (offset < end_offset) {
3097 hole = page_hole_offset(mapping, offset);
3098 if (hole >= 0 && hole <= end_offset)
3099 return max(start_offset, hole);
3101 offset += PAGE_SIZE;
3102 offset &= PAGE_MASK;
3108 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3110 struct bch_inode_info *inode = file_bch_inode(file);
3111 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3112 struct btree_trans trans;
3113 struct btree_iter *iter;
3115 u64 isize, next_hole = MAX_LFS_FILESIZE;
3118 isize = i_size_read(&inode->v);
3119 if (offset >= isize)
3122 bch2_trans_init(&trans, c, 0, 0);
3124 for_each_btree_key(&trans, iter, BTREE_ID_extents,
3125 POS(inode->v.i_ino, offset >> 9),
3126 BTREE_ITER_SLOTS, k, ret) {
3127 if (k.k->p.inode != inode->v.i_ino) {
3128 next_hole = bch2_seek_pagecache_hole(&inode->v,
3129 offset, MAX_LFS_FILESIZE);
3131 } else if (!bkey_extent_is_data(k.k)) {
3132 next_hole = bch2_seek_pagecache_hole(&inode->v,
3133 max(offset, bkey_start_offset(k.k) << 9),
3134 k.k->p.offset << 9);
3136 if (next_hole < k.k->p.offset << 9)
3139 offset = max(offset, bkey_start_offset(k.k) << 9);
3142 bch2_trans_iter_put(&trans, iter);
3144 ret = bch2_trans_exit(&trans) ?: ret;
3148 if (next_hole > isize)
3151 return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3154 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3160 return generic_file_llseek(file, offset, whence);
3162 return bch2_seek_data(file, offset);
3164 return bch2_seek_hole(file, offset);
3170 void bch2_fs_fsio_exit(struct bch_fs *c)
3172 bioset_exit(&c->dio_write_bioset);
3173 bioset_exit(&c->dio_read_bioset);
3174 bioset_exit(&c->writepage_bioset);
3177 int bch2_fs_fsio_init(struct bch_fs *c)
3181 pr_verbose_init(c->opts, "");
3183 if (bioset_init(&c->writepage_bioset,
3184 4, offsetof(struct bch_writepage_io, op.wbio.bio),
3185 BIOSET_NEED_BVECS) ||
3186 bioset_init(&c->dio_read_bioset,
3187 4, offsetof(struct dio_read, rbio.bio),
3188 BIOSET_NEED_BVECS) ||
3189 bioset_init(&c->dio_write_bioset,
3190 4, offsetof(struct dio_write, op.wbio.bio),
3194 pr_verbose_init(c->opts, "ret %i", ret);
3198 #endif /* NO_BCACHEFS_FS */