1 // SPDX-License-Identifier: GPL-2.0
5 #include "alloc_foreground.h"
6 #include "bkey_on_stack.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/sched/signal.h>
30 #include <linux/task_io_accounting_ops.h>
31 #include <linux/uio.h>
32 #include <linux/writeback.h>
34 #include <trace/events/bcachefs.h>
35 #include <trace/events/writeback.h>
41 struct bch_writepage_io {
43 struct bch_inode_info *inode;
46 struct bch_write_op op;
50 struct completion done;
56 struct quota_res quota_res;
59 struct iovec inline_vecs[2];
62 struct bch_write_op op;
69 struct bch_read_bio rbio;
72 /* pagecache_block must be held */
73 static int write_invalidate_inode_pages_range(struct address_space *mapping,
74 loff_t start, loff_t end)
79 * XXX: the way this is currently implemented, we can spin if a process
80 * is continually redirtying a specific page
83 if (!mapping->nrpages &&
84 !mapping->nrexceptional)
87 ret = filemap_write_and_wait_range(mapping, start, end);
91 if (!mapping->nrpages)
94 ret = invalidate_inode_pages2_range(mapping,
97 } while (ret == -EBUSY);
104 #ifdef CONFIG_BCACHEFS_QUOTA
106 static void bch2_quota_reservation_put(struct bch_fs *c,
107 struct bch_inode_info *inode,
108 struct quota_res *res)
113 mutex_lock(&inode->ei_quota_lock);
114 BUG_ON(res->sectors > inode->ei_quota_reserved);
116 bch2_quota_acct(c, inode->ei_qid, Q_SPC,
117 -((s64) res->sectors), KEY_TYPE_QUOTA_PREALLOC);
118 inode->ei_quota_reserved -= res->sectors;
119 mutex_unlock(&inode->ei_quota_lock);
124 static int bch2_quota_reservation_add(struct bch_fs *c,
125 struct bch_inode_info *inode,
126 struct quota_res *res,
132 mutex_lock(&inode->ei_quota_lock);
133 ret = bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors,
134 check_enospc ? KEY_TYPE_QUOTA_PREALLOC : KEY_TYPE_QUOTA_NOCHECK);
136 inode->ei_quota_reserved += sectors;
137 res->sectors += sectors;
139 mutex_unlock(&inode->ei_quota_lock);
146 static void bch2_quota_reservation_put(struct bch_fs *c,
147 struct bch_inode_info *inode,
148 struct quota_res *res)
152 static int bch2_quota_reservation_add(struct bch_fs *c,
153 struct bch_inode_info *inode,
154 struct quota_res *res,
163 /* i_size updates: */
165 struct inode_new_size {
171 static int inode_set_size(struct bch_inode_info *inode,
172 struct bch_inode_unpacked *bi,
175 struct inode_new_size *s = p;
177 bi->bi_size = s->new_size;
178 if (s->fields & ATTR_ATIME)
179 bi->bi_atime = s->now;
180 if (s->fields & ATTR_MTIME)
181 bi->bi_mtime = s->now;
182 if (s->fields & ATTR_CTIME)
183 bi->bi_ctime = s->now;
188 int __must_check bch2_write_inode_size(struct bch_fs *c,
189 struct bch_inode_info *inode,
190 loff_t new_size, unsigned fields)
192 struct inode_new_size s = {
193 .new_size = new_size,
194 .now = bch2_current_time(c),
198 return bch2_write_inode(c, inode, inode_set_size, &s, fields);
201 static void i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
202 struct quota_res *quota_res, s64 sectors)
207 mutex_lock(&inode->ei_quota_lock);
208 #ifdef CONFIG_BCACHEFS_QUOTA
209 if (quota_res && sectors > 0) {
210 BUG_ON(sectors > quota_res->sectors);
211 BUG_ON(sectors > inode->ei_quota_reserved);
213 quota_res->sectors -= sectors;
214 inode->ei_quota_reserved -= sectors;
216 bch2_quota_acct(c, inode->ei_qid, Q_SPC, sectors, KEY_TYPE_QUOTA_WARN);
219 inode->v.i_blocks += sectors;
220 mutex_unlock(&inode->ei_quota_lock);
225 /* stored in page->private: */
227 struct bch_page_sector {
228 /* Uncompressed, fully allocated replicas: */
229 unsigned nr_replicas:3;
231 /* Owns PAGE_SECTORS * replicas_reserved sized reservation: */
232 unsigned replicas_reserved:3;
243 struct bch_page_state {
245 atomic_t write_count;
246 struct bch_page_sector s[PAGE_SECTORS];
249 static inline struct bch_page_state *__bch2_page_state(struct page *page)
251 return page_has_private(page)
252 ? (struct bch_page_state *) page_private(page)
256 static inline struct bch_page_state *bch2_page_state(struct page *page)
258 EBUG_ON(!PageLocked(page));
260 return __bch2_page_state(page);
263 /* for newly allocated pages: */
264 static void __bch2_page_state_release(struct page *page)
266 struct bch_page_state *s = __bch2_page_state(page);
271 ClearPagePrivate(page);
272 set_page_private(page, 0);
277 static void bch2_page_state_release(struct page *page)
279 struct bch_page_state *s = bch2_page_state(page);
284 ClearPagePrivate(page);
285 set_page_private(page, 0);
290 /* for newly allocated pages: */
291 static struct bch_page_state *__bch2_page_state_create(struct page *page,
294 struct bch_page_state *s;
296 s = kzalloc(sizeof(*s), GFP_NOFS|gfp);
300 spin_lock_init(&s->lock);
302 * migrate_page_move_mapping() assumes that pages with private data
303 * have their count elevated by 1.
306 set_page_private(page, (unsigned long) s);
307 SetPagePrivate(page);
311 static struct bch_page_state *bch2_page_state_create(struct page *page,
314 return bch2_page_state(page) ?: __bch2_page_state_create(page, gfp);
317 static inline unsigned inode_nr_replicas(struct bch_fs *c, struct bch_inode_info *inode)
319 /* XXX: this should not be open coded */
320 return inode->ei_inode.bi_data_replicas
321 ? inode->ei_inode.bi_data_replicas - 1
322 : c->opts.data_replicas;
325 static inline unsigned sectors_to_reserve(struct bch_page_sector *s,
326 unsigned nr_replicas)
328 return max(0, (int) nr_replicas -
330 s->replicas_reserved);
333 static int bch2_get_page_disk_reservation(struct bch_fs *c,
334 struct bch_inode_info *inode,
335 struct page *page, bool check_enospc)
337 struct bch_page_state *s = bch2_page_state_create(page, 0);
338 unsigned nr_replicas = inode_nr_replicas(c, inode);
339 struct disk_reservation disk_res = { 0 };
340 unsigned i, disk_res_sectors = 0;
346 for (i = 0; i < ARRAY_SIZE(s->s); i++)
347 disk_res_sectors += sectors_to_reserve(&s->s[i], nr_replicas);
349 if (!disk_res_sectors)
352 ret = bch2_disk_reservation_get(c, &disk_res,
355 ? BCH_DISK_RESERVATION_NOFAIL
360 for (i = 0; i < ARRAY_SIZE(s->s); i++)
361 s->s[i].replicas_reserved +=
362 sectors_to_reserve(&s->s[i], nr_replicas);
367 struct bch2_page_reservation {
368 struct disk_reservation disk;
369 struct quota_res quota;
372 static void bch2_page_reservation_init(struct bch_fs *c,
373 struct bch_inode_info *inode,
374 struct bch2_page_reservation *res)
376 memset(res, 0, sizeof(*res));
378 res->disk.nr_replicas = inode_nr_replicas(c, inode);
381 static void bch2_page_reservation_put(struct bch_fs *c,
382 struct bch_inode_info *inode,
383 struct bch2_page_reservation *res)
385 bch2_disk_reservation_put(c, &res->disk);
386 bch2_quota_reservation_put(c, inode, &res->quota);
389 static int bch2_page_reservation_get(struct bch_fs *c,
390 struct bch_inode_info *inode, struct page *page,
391 struct bch2_page_reservation *res,
392 unsigned offset, unsigned len, bool check_enospc)
394 struct bch_page_state *s = bch2_page_state_create(page, 0);
395 unsigned i, disk_sectors = 0, quota_sectors = 0;
401 for (i = round_down(offset, block_bytes(c)) >> 9;
402 i < round_up(offset + len, block_bytes(c)) >> 9;
404 disk_sectors += sectors_to_reserve(&s->s[i],
405 res->disk.nr_replicas);
406 quota_sectors += s->s[i].state == SECTOR_UNALLOCATED;
410 ret = bch2_disk_reservation_add(c, &res->disk,
413 ? BCH_DISK_RESERVATION_NOFAIL
420 ret = bch2_quota_reservation_add(c, inode, &res->quota,
424 struct disk_reservation tmp = {
425 .sectors = disk_sectors
428 bch2_disk_reservation_put(c, &tmp);
429 res->disk.sectors -= disk_sectors;
437 static void bch2_clear_page_bits(struct page *page)
439 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
440 struct bch_fs *c = inode->v.i_sb->s_fs_info;
441 struct bch_page_state *s = bch2_page_state(page);
442 struct disk_reservation disk_res = { 0 };
443 int i, dirty_sectors = 0;
448 EBUG_ON(!PageLocked(page));
449 EBUG_ON(PageWriteback(page));
451 for (i = 0; i < ARRAY_SIZE(s->s); i++) {
452 disk_res.sectors += s->s[i].replicas_reserved;
453 s->s[i].replicas_reserved = 0;
455 if (s->s[i].state == SECTOR_DIRTY) {
457 s->s[i].state = SECTOR_UNALLOCATED;
461 bch2_disk_reservation_put(c, &disk_res);
464 i_sectors_acct(c, inode, NULL, -dirty_sectors);
466 bch2_page_state_release(page);
469 static void bch2_set_page_dirty(struct bch_fs *c,
470 struct bch_inode_info *inode, struct page *page,
471 struct bch2_page_reservation *res,
472 unsigned offset, unsigned len)
474 struct bch_page_state *s = bch2_page_state(page);
475 unsigned i, dirty_sectors = 0;
477 WARN_ON((u64) page_offset(page) + offset + len >
478 round_up((u64) i_size_read(&inode->v), block_bytes(c)));
482 for (i = round_down(offset, block_bytes(c)) >> 9;
483 i < round_up(offset + len, block_bytes(c)) >> 9;
485 unsigned sectors = sectors_to_reserve(&s->s[i],
486 res->disk.nr_replicas);
489 * This can happen if we race with the error path in
490 * bch2_writepage_io_done():
492 sectors = min_t(unsigned, sectors, res->disk.sectors);
494 s->s[i].replicas_reserved += sectors;
495 res->disk.sectors -= sectors;
497 if (s->s[i].state == SECTOR_UNALLOCATED)
500 s->s[i].state = max_t(unsigned, s->s[i].state, SECTOR_DIRTY);
503 spin_unlock(&s->lock);
506 i_sectors_acct(c, inode, &res->quota, dirty_sectors);
508 if (!PageDirty(page))
509 __set_page_dirty_nobuffers(page);
512 vm_fault_t bch2_page_fault(struct vm_fault *vmf)
514 struct file *file = vmf->vma->vm_file;
515 struct bch_inode_info *inode = file_bch_inode(file);
518 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
519 ret = filemap_fault(vmf);
520 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
525 vm_fault_t bch2_page_mkwrite(struct vm_fault *vmf)
527 struct page *page = vmf->page;
528 struct file *file = vmf->vma->vm_file;
529 struct bch_inode_info *inode = file_bch_inode(file);
530 struct address_space *mapping = file->f_mapping;
531 struct bch_fs *c = inode->v.i_sb->s_fs_info;
532 struct bch2_page_reservation res;
535 int ret = VM_FAULT_LOCKED;
537 bch2_page_reservation_init(c, inode, &res);
539 sb_start_pagefault(inode->v.i_sb);
540 file_update_time(file);
543 * Not strictly necessary, but helps avoid dio writes livelocking in
544 * write_invalidate_inode_pages_range() - can drop this if/when we get
545 * a write_invalidate_inode_pages_range() that works without dropping
546 * page lock before invalidating page
548 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
551 isize = i_size_read(&inode->v);
553 if (page->mapping != mapping || page_offset(page) >= isize) {
555 ret = VM_FAULT_NOPAGE;
559 len = min_t(loff_t, PAGE_SIZE, isize - page_offset(page));
561 if (bch2_page_reservation_get(c, inode, page, &res, 0, len, true)) {
563 ret = VM_FAULT_SIGBUS;
567 bch2_set_page_dirty(c, inode, page, &res, 0, len);
568 bch2_page_reservation_put(c, inode, &res);
570 wait_for_stable_page(page);
572 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
573 sb_end_pagefault(inode->v.i_sb);
578 void bch2_invalidatepage(struct page *page, unsigned int offset,
581 if (offset || length < PAGE_SIZE)
584 bch2_clear_page_bits(page);
587 int bch2_releasepage(struct page *page, gfp_t gfp_mask)
592 bch2_clear_page_bits(page);
596 #ifdef CONFIG_MIGRATION
597 int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
598 struct page *page, enum migrate_mode mode)
602 EBUG_ON(!PageLocked(page));
603 EBUG_ON(!PageLocked(newpage));
605 ret = migrate_page_move_mapping(mapping, newpage, page, 0);
606 if (ret != MIGRATEPAGE_SUCCESS)
609 if (PagePrivate(page)) {
610 ClearPagePrivate(page);
612 set_page_private(newpage, page_private(page));
613 set_page_private(page, 0);
615 SetPagePrivate(newpage);
618 if (mode != MIGRATE_SYNC_NO_COPY)
619 migrate_page_copy(newpage, page);
621 migrate_page_states(newpage, page);
622 return MIGRATEPAGE_SUCCESS;
628 static void bch2_readpages_end_io(struct bio *bio)
630 struct bvec_iter_all iter;
633 bio_for_each_segment_all(bv, bio, iter) {
634 struct page *page = bv->bv_page;
636 if (!bio->bi_status) {
637 SetPageUptodate(page);
639 ClearPageUptodate(page);
648 static inline void page_state_init_for_read(struct page *page)
650 SetPagePrivate(page);
654 struct readpages_iter {
655 struct address_space *mapping;
663 static int readpages_iter_init(struct readpages_iter *iter,
664 struct address_space *mapping,
665 struct list_head *pages, unsigned nr_pages)
667 memset(iter, 0, sizeof(*iter));
669 iter->mapping = mapping;
670 iter->offset = list_last_entry(pages, struct page, lru)->index;
672 iter->pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
676 while (!list_empty(pages)) {
677 struct page *page = list_last_entry(pages, struct page, lru);
679 __bch2_page_state_create(page, __GFP_NOFAIL);
681 iter->pages[iter->nr_pages++] = page;
682 list_del(&page->lru);
688 static inline struct page *readpage_iter_next(struct readpages_iter *iter)
694 BUG_ON(iter->idx > iter->nr_added);
695 BUG_ON(iter->nr_added > iter->nr_pages);
697 if (iter->idx < iter->nr_added)
701 if (iter->idx == iter->nr_pages)
704 ret = add_to_page_cache_lru_vec(iter->mapping,
705 iter->pages + iter->nr_added,
706 iter->nr_pages - iter->nr_added,
707 iter->offset + iter->nr_added,
712 page = iter->pages[iter->nr_added];
716 __bch2_page_state_release(page);
720 iter->nr_added += ret;
722 for (i = iter->idx; i < iter->nr_added; i++)
723 put_page(iter->pages[i]);
725 EBUG_ON(iter->pages[iter->idx]->index != iter->offset + iter->idx);
727 return iter->pages[iter->idx];
730 static void bch2_add_page_sectors(struct bio *bio, struct bkey_s_c k)
732 struct bvec_iter iter;
734 unsigned nr_ptrs = k.k->type == KEY_TYPE_reflink_v
735 ? 0 : bch2_bkey_nr_ptrs_fully_allocated(k);
736 unsigned state = k.k->type == KEY_TYPE_reservation
740 bio_for_each_segment(bv, bio, iter) {
741 struct bch_page_state *s = bch2_page_state(bv.bv_page);
744 for (i = bv.bv_offset >> 9;
745 i < (bv.bv_offset + bv.bv_len) >> 9;
747 s->s[i].nr_replicas = nr_ptrs;
748 s->s[i].state = state;
753 static bool extent_partial_reads_expensive(struct bkey_s_c k)
755 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
756 struct bch_extent_crc_unpacked crc;
757 const union bch_extent_entry *i;
759 bkey_for_each_crc(k.k, ptrs, crc, i)
760 if (crc.csum_type || crc.compression_type)
765 static void readpage_bio_extend(struct readpages_iter *iter,
767 unsigned sectors_this_extent,
770 while (bio_sectors(bio) < sectors_this_extent &&
771 bio->bi_vcnt < bio->bi_max_vecs) {
772 pgoff_t page_offset = bio_end_sector(bio) >> PAGE_SECTOR_SHIFT;
773 struct page *page = readpage_iter_next(iter);
777 if (iter->offset + iter->idx != page_offset)
785 page = xa_load(&iter->mapping->i_pages, page_offset);
786 if (page && !xa_is_value(page))
789 page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
793 if (!__bch2_page_state_create(page, 0)) {
798 ret = add_to_page_cache_lru(page, iter->mapping,
799 page_offset, GFP_NOFS);
801 __bch2_page_state_release(page);
809 BUG_ON(!bio_add_page(bio, page, PAGE_SIZE, 0));
813 static void bchfs_read(struct btree_trans *trans, struct btree_iter *iter,
814 struct bch_read_bio *rbio, u64 inum,
815 struct readpages_iter *readpages_iter)
817 struct bch_fs *c = trans->c;
818 struct bkey_on_stack sk;
819 int flags = BCH_READ_RETRY_IF_STALE|
820 BCH_READ_MAY_PROMOTE;
824 rbio->start_time = local_clock();
826 bkey_on_stack_init(&sk);
830 unsigned bytes, sectors, offset_into_extent;
832 bch2_btree_iter_set_pos(iter,
833 POS(inum, rbio->bio.bi_iter.bi_sector));
835 k = bch2_btree_iter_peek_slot(iter);
840 bkey_on_stack_reassemble(&sk, c, k);
841 k = bkey_i_to_s_c(sk.k);
843 offset_into_extent = iter->pos.offset -
844 bkey_start_offset(k.k);
845 sectors = k.k->size - offset_into_extent;
847 ret = bch2_read_indirect_extent(trans,
848 &offset_into_extent, sk.k);
852 sectors = min(sectors, k.k->size - offset_into_extent);
854 bch2_trans_unlock(trans);
857 readpage_bio_extend(readpages_iter, &rbio->bio, sectors,
858 extent_partial_reads_expensive(k));
860 bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
861 swap(rbio->bio.bi_iter.bi_size, bytes);
863 if (rbio->bio.bi_iter.bi_size == bytes)
864 flags |= BCH_READ_LAST_FRAGMENT;
866 if (bkey_extent_is_allocation(k.k))
867 bch2_add_page_sectors(&rbio->bio, k);
869 bch2_read_extent(c, rbio, k, offset_into_extent, flags);
871 if (flags & BCH_READ_LAST_FRAGMENT)
874 swap(rbio->bio.bi_iter.bi_size, bytes);
875 bio_advance(&rbio->bio, bytes);
882 bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
883 bio_endio(&rbio->bio);
886 bkey_on_stack_exit(&sk, c);
889 int bch2_readpages(struct file *file, struct address_space *mapping,
890 struct list_head *pages, unsigned nr_pages)
892 struct bch_inode_info *inode = to_bch_ei(mapping->host);
893 struct bch_fs *c = inode->v.i_sb->s_fs_info;
894 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
895 struct btree_trans trans;
896 struct btree_iter *iter;
898 struct readpages_iter readpages_iter;
901 ret = readpages_iter_init(&readpages_iter, mapping, pages, nr_pages);
904 bch2_trans_init(&trans, c, 0, 0);
906 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
909 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
911 while ((page = readpage_iter_next(&readpages_iter))) {
912 pgoff_t index = readpages_iter.offset + readpages_iter.idx;
913 unsigned n = min_t(unsigned,
914 readpages_iter.nr_pages -
917 struct bch_read_bio *rbio =
918 rbio_init(bio_alloc_bioset(GFP_NOFS, n, &c->bio_read),
921 readpages_iter.idx++;
923 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, 0);
924 rbio->bio.bi_iter.bi_sector = (sector_t) index << PAGE_SECTOR_SHIFT;
925 rbio->bio.bi_end_io = bch2_readpages_end_io;
926 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
928 bchfs_read(&trans, iter, rbio, inode->v.i_ino,
932 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
934 bch2_trans_exit(&trans);
935 kfree(readpages_iter.pages);
940 static void __bchfs_readpage(struct bch_fs *c, struct bch_read_bio *rbio,
941 u64 inum, struct page *page)
943 struct btree_trans trans;
944 struct btree_iter *iter;
946 bch2_page_state_create(page, __GFP_NOFAIL);
948 bio_set_op_attrs(&rbio->bio, REQ_OP_READ, REQ_SYNC);
949 rbio->bio.bi_iter.bi_sector =
950 (sector_t) page->index << PAGE_SECTOR_SHIFT;
951 BUG_ON(!bio_add_page(&rbio->bio, page, PAGE_SIZE, 0));
953 bch2_trans_init(&trans, c, 0, 0);
954 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, POS_MIN,
957 bchfs_read(&trans, iter, rbio, inum, NULL);
959 bch2_trans_exit(&trans);
962 int bch2_readpage(struct file *file, struct page *page)
964 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
965 struct bch_fs *c = inode->v.i_sb->s_fs_info;
966 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
967 struct bch_read_bio *rbio;
969 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read), opts);
970 rbio->bio.bi_end_io = bch2_readpages_end_io;
972 __bchfs_readpage(c, rbio, inode->v.i_ino, page);
976 static void bch2_read_single_page_end_io(struct bio *bio)
978 complete(bio->bi_private);
981 static int bch2_read_single_page(struct page *page,
982 struct address_space *mapping)
984 struct bch_inode_info *inode = to_bch_ei(mapping->host);
985 struct bch_fs *c = inode->v.i_sb->s_fs_info;
986 struct bch_read_bio *rbio;
988 DECLARE_COMPLETION_ONSTACK(done);
990 rbio = rbio_init(bio_alloc_bioset(GFP_NOFS, 1, &c->bio_read),
991 io_opts(c, &inode->ei_inode));
992 rbio->bio.bi_private = &done;
993 rbio->bio.bi_end_io = bch2_read_single_page_end_io;
995 __bchfs_readpage(c, rbio, inode->v.i_ino, page);
996 wait_for_completion(&done);
998 ret = blk_status_to_errno(rbio->bio.bi_status);
1004 SetPageUptodate(page);
1010 struct bch_writepage_state {
1011 struct bch_writepage_io *io;
1012 struct bch_io_opts opts;
1015 static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
1016 struct bch_inode_info *inode)
1018 return (struct bch_writepage_state) {
1019 .opts = io_opts(c, &inode->ei_inode)
1023 static void bch2_writepage_io_free(struct closure *cl)
1025 struct bch_writepage_io *io = container_of(cl,
1026 struct bch_writepage_io, cl);
1028 bio_put(&io->op.wbio.bio);
1031 static void bch2_writepage_io_done(struct closure *cl)
1033 struct bch_writepage_io *io = container_of(cl,
1034 struct bch_writepage_io, cl);
1035 struct bch_fs *c = io->op.c;
1036 struct bio *bio = &io->op.wbio.bio;
1037 struct bvec_iter_all iter;
1038 struct bio_vec *bvec;
1042 bio_for_each_segment_all(bvec, bio, iter) {
1043 struct bch_page_state *s;
1045 SetPageError(bvec->bv_page);
1046 mapping_set_error(bvec->bv_page->mapping, -EIO);
1048 s = __bch2_page_state(bvec->bv_page);
1049 spin_lock(&s->lock);
1050 for (i = 0; i < PAGE_SECTORS; i++)
1051 s->s[i].nr_replicas = 0;
1052 spin_unlock(&s->lock);
1056 if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
1057 bio_for_each_segment_all(bvec, bio, iter) {
1058 struct bch_page_state *s;
1060 s = __bch2_page_state(bvec->bv_page);
1061 spin_lock(&s->lock);
1062 for (i = 0; i < PAGE_SECTORS; i++)
1063 s->s[i].nr_replicas = 0;
1064 spin_unlock(&s->lock);
1069 * racing with fallocate can cause us to add fewer sectors than
1070 * expected - but we shouldn't add more sectors than expected:
1072 BUG_ON(io->op.i_sectors_delta > 0);
1075 * (error (due to going RO) halfway through a page can screw that up
1078 BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
1082 * PageWriteback is effectively our ref on the inode - fixup i_blocks
1083 * before calling end_page_writeback:
1085 i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
1087 bio_for_each_segment_all(bvec, bio, iter) {
1088 struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
1090 if (atomic_dec_and_test(&s->write_count))
1091 end_page_writeback(bvec->bv_page);
1094 closure_return_with_destructor(&io->cl, bch2_writepage_io_free);
1097 static void bch2_writepage_do_io(struct bch_writepage_state *w)
1099 struct bch_writepage_io *io = w->io;
1102 closure_call(&io->op.cl, bch2_write, NULL, &io->cl);
1103 continue_at(&io->cl, bch2_writepage_io_done, NULL);
1107 * Get a bch_writepage_io and add @page to it - appending to an existing one if
1108 * possible, else allocating a new one:
1110 static void bch2_writepage_io_alloc(struct bch_fs *c,
1111 struct writeback_control *wbc,
1112 struct bch_writepage_state *w,
1113 struct bch_inode_info *inode,
1115 unsigned nr_replicas)
1117 struct bch_write_op *op;
1119 w->io = container_of(bio_alloc_bioset(GFP_NOFS,
1121 &c->writepage_bioset),
1122 struct bch_writepage_io, op.wbio.bio);
1124 closure_init(&w->io->cl, NULL);
1125 w->io->inode = inode;
1128 bch2_write_op_init(op, c, w->opts);
1129 op->target = w->opts.foreground_target;
1130 op_journal_seq_set(op, &inode->ei_journal_seq);
1131 op->nr_replicas = nr_replicas;
1132 op->res.nr_replicas = nr_replicas;
1133 op->write_point = writepoint_hashed(inode->ei_last_dirtied);
1134 op->pos = POS(inode->v.i_ino, sector);
1135 op->wbio.bio.bi_iter.bi_sector = sector;
1136 op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
1139 static int __bch2_writepage(struct page *page,
1140 struct writeback_control *wbc,
1143 struct bch_inode_info *inode = to_bch_ei(page->mapping->host);
1144 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1145 struct bch_writepage_state *w = data;
1146 struct bch_page_state *s, orig;
1147 unsigned i, offset, nr_replicas_this_write = U32_MAX;
1148 loff_t i_size = i_size_read(&inode->v);
1149 pgoff_t end_index = i_size >> PAGE_SHIFT;
1152 EBUG_ON(!PageUptodate(page));
1154 /* Is the page fully inside i_size? */
1155 if (page->index < end_index)
1158 /* Is the page fully outside i_size? (truncate in progress) */
1159 offset = i_size & (PAGE_SIZE - 1);
1160 if (page->index > end_index || !offset) {
1166 * The page straddles i_size. It must be zeroed out on each and every
1167 * writepage invocation because it may be mmapped. "A file is mapped
1168 * in multiples of the page size. For a file that is not a multiple of
1169 * the page size, the remaining memory is zeroed when mapped, and
1170 * writes to that region are not written out to the file."
1172 zero_user_segment(page, offset, PAGE_SIZE);
1174 s = bch2_page_state_create(page, __GFP_NOFAIL);
1176 ret = bch2_get_page_disk_reservation(c, inode, page, true);
1179 mapping_set_error(page->mapping, ret);
1184 /* Before unlocking the page, get copy of reservations: */
1187 for (i = 0; i < PAGE_SECTORS; i++) {
1188 if (s->s[i].state < SECTOR_DIRTY)
1191 nr_replicas_this_write =
1192 min_t(unsigned, nr_replicas_this_write,
1193 s->s[i].nr_replicas +
1194 s->s[i].replicas_reserved);
1197 for (i = 0; i < PAGE_SECTORS; i++) {
1198 if (s->s[i].state < SECTOR_DIRTY)
1201 s->s[i].nr_replicas = w->opts.compression
1202 ? 0 : nr_replicas_this_write;
1204 s->s[i].replicas_reserved = 0;
1205 s->s[i].state = SECTOR_ALLOCATED;
1208 BUG_ON(atomic_read(&s->write_count));
1209 atomic_set(&s->write_count, 1);
1211 BUG_ON(PageWriteback(page));
1212 set_page_writeback(page);
1218 unsigned sectors = 1, dirty_sectors = 0, reserved_sectors = 0;
1221 while (offset < PAGE_SECTORS &&
1222 orig.s[offset].state < SECTOR_DIRTY)
1225 if (offset == PAGE_SECTORS)
1228 sector = ((u64) page->index << PAGE_SECTOR_SHIFT) + offset;
1230 while (offset + sectors < PAGE_SECTORS &&
1231 orig.s[offset + sectors].state >= SECTOR_DIRTY)
1234 for (i = offset; i < offset + sectors; i++) {
1235 reserved_sectors += orig.s[i].replicas_reserved;
1236 dirty_sectors += orig.s[i].state == SECTOR_DIRTY;
1240 (w->io->op.res.nr_replicas != nr_replicas_this_write ||
1241 bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
1242 w->io->op.wbio.bio.bi_iter.bi_size >= (256U << 20) ||
1243 bio_end_sector(&w->io->op.wbio.bio) != sector))
1244 bch2_writepage_do_io(w);
1247 bch2_writepage_io_alloc(c, wbc, w, inode, sector,
1248 nr_replicas_this_write);
1250 atomic_inc(&s->write_count);
1252 BUG_ON(inode != w->io->inode);
1253 BUG_ON(!bio_add_page(&w->io->op.wbio.bio, page,
1254 sectors << 9, offset << 9));
1256 /* Check for writing past i_size: */
1257 WARN_ON((bio_end_sector(&w->io->op.wbio.bio) << 9) >
1258 round_up(i_size, block_bytes(c)));
1260 w->io->op.res.sectors += reserved_sectors;
1261 w->io->op.i_sectors_delta -= dirty_sectors;
1262 w->io->op.new_i_size = i_size;
1267 if (atomic_dec_and_test(&s->write_count))
1268 end_page_writeback(page);
1273 int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
1275 struct bch_fs *c = mapping->host->i_sb->s_fs_info;
1276 struct bch_writepage_state w =
1277 bch_writepage_state_init(c, to_bch_ei(mapping->host));
1278 struct blk_plug plug;
1281 blk_start_plug(&plug);
1282 ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
1284 bch2_writepage_do_io(&w);
1285 blk_finish_plug(&plug);
1289 int bch2_writepage(struct page *page, struct writeback_control *wbc)
1291 struct bch_fs *c = page->mapping->host->i_sb->s_fs_info;
1292 struct bch_writepage_state w =
1293 bch_writepage_state_init(c, to_bch_ei(page->mapping->host));
1296 ret = __bch2_writepage(page, wbc, &w);
1298 bch2_writepage_do_io(&w);
1303 /* buffered writes: */
1305 int bch2_write_begin(struct file *file, struct address_space *mapping,
1306 loff_t pos, unsigned len, unsigned flags,
1307 struct page **pagep, void **fsdata)
1309 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1310 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1311 struct bch2_page_reservation *res;
1312 pgoff_t index = pos >> PAGE_SHIFT;
1313 unsigned offset = pos & (PAGE_SIZE - 1);
1317 res = kmalloc(sizeof(*res), GFP_KERNEL);
1321 bch2_page_reservation_init(c, inode, res);
1324 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1326 page = grab_cache_page_write_begin(mapping, index, flags);
1330 if (PageUptodate(page))
1333 /* If we're writing entire page, don't need to read it in first: */
1334 if (len == PAGE_SIZE)
1337 if (!offset && pos + len >= inode->v.i_size) {
1338 zero_user_segment(page, len, PAGE_SIZE);
1339 flush_dcache_page(page);
1343 if (index > inode->v.i_size >> PAGE_SHIFT) {
1344 zero_user_segments(page, 0, offset, offset + len, PAGE_SIZE);
1345 flush_dcache_page(page);
1349 ret = bch2_read_single_page(page, mapping);
1353 ret = bch2_page_reservation_get(c, inode, page, res,
1356 if (!PageUptodate(page)) {
1358 * If the page hasn't been read in, we won't know if we
1359 * actually need a reservation - we don't actually need
1360 * to read here, we just need to check if the page is
1361 * fully backed by uncompressed data:
1376 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1382 int bch2_write_end(struct file *file, struct address_space *mapping,
1383 loff_t pos, unsigned len, unsigned copied,
1384 struct page *page, void *fsdata)
1386 struct bch_inode_info *inode = to_bch_ei(mapping->host);
1387 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1388 struct bch2_page_reservation *res = fsdata;
1389 unsigned offset = pos & (PAGE_SIZE - 1);
1391 lockdep_assert_held(&inode->v.i_rwsem);
1393 if (unlikely(copied < len && !PageUptodate(page))) {
1395 * The page needs to be read in, but that would destroy
1396 * our partial write - simplest thing is to just force
1397 * userspace to redo the write:
1399 zero_user(page, 0, PAGE_SIZE);
1400 flush_dcache_page(page);
1404 spin_lock(&inode->v.i_lock);
1405 if (pos + copied > inode->v.i_size)
1406 i_size_write(&inode->v, pos + copied);
1407 spin_unlock(&inode->v.i_lock);
1410 if (!PageUptodate(page))
1411 SetPageUptodate(page);
1413 bch2_set_page_dirty(c, inode, page, res, offset, copied);
1415 inode->ei_last_dirtied = (unsigned long) current;
1420 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1422 bch2_page_reservation_put(c, inode, res);
1428 #define WRITE_BATCH_PAGES 32
1430 static int __bch2_buffered_write(struct bch_inode_info *inode,
1431 struct address_space *mapping,
1432 struct iov_iter *iter,
1433 loff_t pos, unsigned len)
1435 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1436 struct page *pages[WRITE_BATCH_PAGES];
1437 struct bch2_page_reservation res;
1438 unsigned long index = pos >> PAGE_SHIFT;
1439 unsigned offset = pos & (PAGE_SIZE - 1);
1440 unsigned nr_pages = DIV_ROUND_UP(offset + len, PAGE_SIZE);
1441 unsigned i, reserved = 0, set_dirty = 0;
1442 unsigned copied = 0, nr_pages_copied = 0;
1446 BUG_ON(nr_pages > ARRAY_SIZE(pages));
1448 bch2_page_reservation_init(c, inode, &res);
1450 for (i = 0; i < nr_pages; i++) {
1451 pages[i] = grab_cache_page_write_begin(mapping, index + i, 0);
1458 len = min_t(unsigned, len,
1459 nr_pages * PAGE_SIZE - offset);
1464 if (offset && !PageUptodate(pages[0])) {
1465 ret = bch2_read_single_page(pages[0], mapping);
1470 if ((pos + len) & (PAGE_SIZE - 1) &&
1471 !PageUptodate(pages[nr_pages - 1])) {
1472 if ((index + nr_pages - 1) << PAGE_SHIFT >= inode->v.i_size) {
1473 zero_user(pages[nr_pages - 1], 0, PAGE_SIZE);
1475 ret = bch2_read_single_page(pages[nr_pages - 1], mapping);
1481 while (reserved < len) {
1482 struct page *page = pages[(offset + reserved) >> PAGE_SHIFT];
1483 unsigned pg_offset = (offset + reserved) & (PAGE_SIZE - 1);
1484 unsigned pg_len = min_t(unsigned, len - reserved,
1485 PAGE_SIZE - pg_offset);
1487 ret = bch2_page_reservation_get(c, inode, page, &res,
1488 pg_offset, pg_len, true);
1490 if (ret && !PageUptodate(page)) {
1491 ret = bch2_read_single_page(page, mapping);
1493 goto retry_reservation;
1502 if (mapping_writably_mapped(mapping))
1503 for (i = 0; i < nr_pages; i++)
1504 flush_dcache_page(pages[i]);
1506 while (copied < len) {
1507 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1508 unsigned pg_offset = (offset + copied) & (PAGE_SIZE - 1);
1509 unsigned pg_len = min_t(unsigned, len - copied,
1510 PAGE_SIZE - pg_offset);
1511 unsigned pg_copied = iov_iter_copy_from_user_atomic(page,
1512 iter, pg_offset, pg_len);
1517 flush_dcache_page(page);
1518 iov_iter_advance(iter, pg_copied);
1519 copied += pg_copied;
1526 ((offset + copied) & (PAGE_SIZE - 1))) {
1527 struct page *page = pages[(offset + copied) >> PAGE_SHIFT];
1529 if (!PageUptodate(page)) {
1530 zero_user(page, 0, PAGE_SIZE);
1531 copied -= (offset + copied) & (PAGE_SIZE - 1);
1535 spin_lock(&inode->v.i_lock);
1536 if (pos + copied > inode->v.i_size)
1537 i_size_write(&inode->v, pos + copied);
1538 spin_unlock(&inode->v.i_lock);
1540 while (set_dirty < copied) {
1541 struct page *page = pages[(offset + set_dirty) >> PAGE_SHIFT];
1542 unsigned pg_offset = (offset + set_dirty) & (PAGE_SIZE - 1);
1543 unsigned pg_len = min_t(unsigned, copied - set_dirty,
1544 PAGE_SIZE - pg_offset);
1546 if (!PageUptodate(page))
1547 SetPageUptodate(page);
1549 bch2_set_page_dirty(c, inode, page, &res, pg_offset, pg_len);
1553 set_dirty += pg_len;
1556 nr_pages_copied = DIV_ROUND_UP(offset + copied, PAGE_SIZE);
1557 inode->ei_last_dirtied = (unsigned long) current;
1559 for (i = nr_pages_copied; i < nr_pages; i++) {
1560 unlock_page(pages[i]);
1564 bch2_page_reservation_put(c, inode, &res);
1566 return copied ?: ret;
1569 static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
1571 struct file *file = iocb->ki_filp;
1572 struct address_space *mapping = file->f_mapping;
1573 struct bch_inode_info *inode = file_bch_inode(file);
1574 loff_t pos = iocb->ki_pos;
1575 ssize_t written = 0;
1578 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1581 unsigned offset = pos & (PAGE_SIZE - 1);
1582 unsigned bytes = min_t(unsigned long, iov_iter_count(iter),
1583 PAGE_SIZE * WRITE_BATCH_PAGES - offset);
1586 * Bring in the user page that we will copy from _first_.
1587 * Otherwise there's a nasty deadlock on copying from the
1588 * same page as we're writing to, without it being marked
1591 * Not only is this an optimisation, but it is also required
1592 * to check that the address is actually valid, when atomic
1593 * usercopies are used, below.
1595 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1596 bytes = min_t(unsigned long, iov_iter_count(iter),
1597 PAGE_SIZE - offset);
1599 if (unlikely(iov_iter_fault_in_readable(iter, bytes))) {
1605 if (unlikely(fatal_signal_pending(current))) {
1610 ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
1611 if (unlikely(ret < 0))
1616 if (unlikely(ret == 0)) {
1618 * If we were unable to copy any data at all, we must
1619 * fall back to a single segment length write.
1621 * If we didn't fallback here, we could livelock
1622 * because not all segments in the iov can be copied at
1623 * once without a pagefault.
1625 bytes = min_t(unsigned long, PAGE_SIZE - offset,
1626 iov_iter_single_seg_count(iter));
1632 balance_dirty_pages_ratelimited(mapping);
1633 } while (iov_iter_count(iter));
1635 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1637 return written ? written : ret;
1640 /* O_DIRECT reads */
1642 static void bch2_dio_read_complete(struct closure *cl)
1644 struct dio_read *dio = container_of(cl, struct dio_read, cl);
1646 dio->req->ki_complete(dio->req, dio->ret, 0);
1647 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1650 static void bch2_direct_IO_read_endio(struct bio *bio)
1652 struct dio_read *dio = bio->bi_private;
1655 dio->ret = blk_status_to_errno(bio->bi_status);
1657 closure_put(&dio->cl);
1660 static void bch2_direct_IO_read_split_endio(struct bio *bio)
1662 bch2_direct_IO_read_endio(bio);
1663 bio_check_pages_dirty(bio); /* transfers ownership */
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_PAGES),
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);
1719 while (iter->count) {
1720 bio = bio_alloc_bioset(GFP_KERNEL,
1721 iov_iter_npages(iter, BIO_MAX_PAGES),
1723 bio->bi_end_io = bch2_direct_IO_read_split_endio;
1725 bio_set_op_attrs(bio, REQ_OP_READ, REQ_SYNC);
1726 bio->bi_iter.bi_sector = offset >> 9;
1727 bio->bi_private = dio;
1729 ret = bio_iov_iter_get_pages(bio, iter);
1731 /* XXX: fault inject this path */
1732 bio->bi_status = BLK_STS_RESOURCE;
1737 offset += bio->bi_iter.bi_size;
1738 bio_set_pages_dirty(bio);
1741 closure_get(&dio->cl);
1743 bch2_read(c, rbio_init(bio, opts), inode->v.i_ino);
1746 iter->count += shorten;
1749 closure_sync(&dio->cl);
1750 closure_debug_destroy(&dio->cl);
1752 bio_check_pages_dirty(&dio->rbio.bio); /* transfers ownership */
1755 return -EIOCBQUEUED;
1759 ssize_t bch2_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1761 struct file *file = iocb->ki_filp;
1762 struct bch_inode_info *inode = file_bch_inode(file);
1763 struct address_space *mapping = file->f_mapping;
1764 size_t count = iov_iter_count(iter);
1768 return 0; /* skip atime */
1770 if (iocb->ki_flags & IOCB_DIRECT) {
1771 struct blk_plug plug;
1773 ret = filemap_write_and_wait_range(mapping,
1775 iocb->ki_pos + count - 1);
1779 file_accessed(file);
1781 blk_start_plug(&plug);
1782 ret = bch2_direct_IO_read(iocb, iter);
1783 blk_finish_plug(&plug);
1786 iocb->ki_pos += ret;
1788 bch2_pagecache_add_get(&inode->ei_pagecache_lock);
1789 ret = generic_file_read_iter(iocb, iter);
1790 bch2_pagecache_add_put(&inode->ei_pagecache_lock);
1796 /* O_DIRECT writes */
1798 static long bch2_dio_write_loop(struct dio_write *dio)
1800 bool kthread = (current->flags & PF_KTHREAD) != 0;
1801 struct bch_fs *c = dio->op.c;
1802 struct kiocb *req = dio->req;
1803 struct address_space *mapping = req->ki_filp->f_mapping;
1804 struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
1805 struct bio *bio = &dio->op.wbio.bio;
1806 struct bvec_iter_all iter;
1810 bool sync = dio->sync;
1819 BUG_ON(current->faults_disabled_mapping);
1820 current->faults_disabled_mapping = mapping;
1822 ret = bio_iov_iter_get_pages(bio, &dio->iter);
1824 current->faults_disabled_mapping = NULL;
1828 if (unlikely(ret < 0))
1831 unaligned = bio->bi_iter.bi_size & (block_bytes(c) - 1);
1832 bio->bi_iter.bi_size -= unaligned;
1833 iov_iter_revert(&dio->iter, unaligned);
1835 if (!bio->bi_iter.bi_size) {
1837 * bio_iov_iter_get_pages was only able to get <
1838 * blocksize worth of pages:
1840 bio_for_each_segment_all(bv, bio, iter)
1841 put_page(bv->bv_page);
1846 dio->op.pos = POS(inode->v.i_ino,
1847 (req->ki_pos >> 9) + dio->op.written);
1849 task_io_account_write(bio->bi_iter.bi_size);
1851 if (!dio->sync && !dio->loop && dio->iter.count) {
1852 struct iovec *iov = dio->inline_vecs;
1854 if (dio->iter.nr_segs > ARRAY_SIZE(dio->inline_vecs)) {
1855 iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
1857 if (unlikely(!iov)) {
1858 dio->sync = sync = true;
1862 dio->free_iov = true;
1865 memcpy(iov, dio->iter.iov, dio->iter.nr_segs * sizeof(*iov));
1866 dio->iter.iov = iov;
1870 closure_call(&dio->op.cl, bch2_write, NULL, NULL);
1873 wait_for_completion(&dio->done);
1875 return -EIOCBQUEUED;
1877 i_sectors_acct(c, inode, &dio->quota_res,
1878 dio->op.i_sectors_delta);
1879 dio->op.i_sectors_delta = 0;
1881 new_i_size = req->ki_pos + ((u64) dio->op.written << 9);
1883 spin_lock(&inode->v.i_lock);
1884 if (new_i_size > inode->v.i_size)
1885 i_size_write(&inode->v, new_i_size);
1886 spin_unlock(&inode->v.i_lock);
1888 bio_for_each_segment_all(bv, bio, iter)
1889 put_page(bv->bv_page);
1890 if (!dio->iter.count || dio->op.error)
1894 reinit_completion(&dio->done);
1897 ret = dio->op.error ?: ((long) dio->op.written << 9);
1899 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
1900 bch2_disk_reservation_put(c, &dio->op.res);
1901 bch2_quota_reservation_put(c, inode, &dio->quota_res);
1904 kfree(dio->iter.iov);
1908 /* inode->i_dio_count is our ref on inode and thus bch_fs */
1909 inode_dio_end(&inode->v);
1912 req->ki_complete(req, ret, 0);
1918 static void bch2_dio_write_loop_async(struct bch_write_op *op)
1920 struct dio_write *dio = container_of(op, struct dio_write, op);
1923 complete(&dio->done);
1925 bch2_dio_write_loop(dio);
1929 ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
1931 struct file *file = req->ki_filp;
1932 struct address_space *mapping = file->f_mapping;
1933 struct bch_inode_info *inode = file_bch_inode(file);
1934 struct bch_fs *c = inode->v.i_sb->s_fs_info;
1935 struct bch_io_opts opts = io_opts(c, &inode->ei_inode);
1936 struct dio_write *dio;
1938 bool locked = true, extending;
1942 prefetch((void *) &c->opts + 64);
1943 prefetch(&inode->ei_inode);
1944 prefetch((void *) &inode->ei_inode + 64);
1946 inode_lock(&inode->v);
1948 ret = generic_write_checks(req, iter);
1949 if (unlikely(ret <= 0))
1952 ret = file_remove_privs(file);
1956 ret = file_update_time(file);
1960 if (unlikely((req->ki_pos|iter->count) & (block_bytes(c) - 1)))
1963 inode_dio_begin(&inode->v);
1964 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
1966 extending = req->ki_pos + iter->count > inode->v.i_size;
1968 inode_unlock(&inode->v);
1972 bio = bio_alloc_bioset(GFP_KERNEL,
1973 iov_iter_npages(iter, BIO_MAX_PAGES),
1974 &c->dio_write_bioset);
1975 dio = container_of(bio, struct dio_write, op.wbio.bio);
1976 init_completion(&dio->done);
1978 dio->mm = current->mm;
1980 dio->sync = is_sync_kiocb(req) || extending;
1981 dio->free_iov = false;
1982 dio->quota_res.sectors = 0;
1985 bch2_write_op_init(&dio->op, c, opts);
1986 dio->op.end_io = bch2_dio_write_loop_async;
1987 dio->op.target = opts.foreground_target;
1988 op_journal_seq_set(&dio->op, &inode->ei_journal_seq);
1989 dio->op.write_point = writepoint_hashed((unsigned long) current);
1990 dio->op.flags |= BCH_WRITE_NOPUT_RESERVATION;
1992 if ((req->ki_flags & IOCB_DSYNC) &&
1993 !c->opts.journal_flush_disabled)
1994 dio->op.flags |= BCH_WRITE_FLUSH;
1996 ret = bch2_quota_reservation_add(c, inode, &dio->quota_res,
1997 iter->count >> 9, true);
2001 dio->op.nr_replicas = dio->op.opts.data_replicas;
2003 ret = bch2_disk_reservation_get(c, &dio->op.res, iter->count >> 9,
2004 dio->op.opts.data_replicas, 0);
2005 if (unlikely(ret) &&
2006 !bch2_check_range_allocated(c, POS(inode->v.i_ino,
2009 dio->op.opts.data_replicas))
2012 ret = write_invalidate_inode_pages_range(mapping,
2014 req->ki_pos + iter->count - 1);
2018 ret = bch2_dio_write_loop(dio);
2021 inode_unlock(&inode->v);
2026 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2027 bch2_disk_reservation_put(c, &dio->op.res);
2028 bch2_quota_reservation_put(c, inode, &dio->quota_res);
2030 inode_dio_end(&inode->v);
2034 ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
2036 struct file *file = iocb->ki_filp;
2037 struct bch_inode_info *inode = file_bch_inode(file);
2040 if (iocb->ki_flags & IOCB_DIRECT)
2041 return bch2_direct_write(iocb, from);
2043 /* We can write back this queue in page reclaim */
2044 current->backing_dev_info = inode_to_bdi(&inode->v);
2045 inode_lock(&inode->v);
2047 ret = generic_write_checks(iocb, from);
2051 ret = file_remove_privs(file);
2055 ret = file_update_time(file);
2059 ret = bch2_buffered_write(iocb, from);
2060 if (likely(ret > 0))
2061 iocb->ki_pos += ret;
2063 inode_unlock(&inode->v);
2064 current->backing_dev_info = NULL;
2067 ret = generic_write_sync(iocb, ret);
2074 int bch2_fsync(struct file *file, loff_t start, loff_t end, int datasync)
2076 struct bch_inode_info *inode = file_bch_inode(file);
2077 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2080 ret = file_write_and_wait_range(file, start, end);
2084 if (datasync && !(inode->v.i_state & I_DIRTY_DATASYNC))
2087 ret = sync_inode_metadata(&inode->v, 1);
2091 if (!c->opts.journal_flush_disabled)
2092 ret = bch2_journal_flush_seq(&c->journal,
2093 inode->ei_journal_seq);
2094 ret2 = file_check_and_advance_wb_err(file);
2101 static inline int range_has_data(struct bch_fs *c,
2105 struct btree_trans trans;
2106 struct btree_iter *iter;
2110 bch2_trans_init(&trans, c, 0, 0);
2112 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS, start, 0, k, ret) {
2113 if (bkey_cmp(bkey_start_pos(k.k), end) >= 0)
2116 if (bkey_extent_is_data(k.k)) {
2122 return bch2_trans_exit(&trans) ?: ret;
2125 static int __bch2_truncate_page(struct bch_inode_info *inode,
2126 pgoff_t index, loff_t start, loff_t end)
2128 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2129 struct address_space *mapping = inode->v.i_mapping;
2130 struct bch_page_state *s;
2131 unsigned start_offset = start & (PAGE_SIZE - 1);
2132 unsigned end_offset = ((end - 1) & (PAGE_SIZE - 1)) + 1;
2137 /* Page boundary? Nothing to do */
2138 if (!((index == start >> PAGE_SHIFT && start_offset) ||
2139 (index == end >> PAGE_SHIFT && end_offset != PAGE_SIZE)))
2143 if (index << PAGE_SHIFT >= inode->v.i_size)
2146 page = find_lock_page(mapping, index);
2149 * XXX: we're doing two index lookups when we end up reading the
2152 ret = range_has_data(c,
2153 POS(inode->v.i_ino, index << PAGE_SECTOR_SHIFT),
2154 POS(inode->v.i_ino, (index + 1) << PAGE_SECTOR_SHIFT));
2158 page = find_or_create_page(mapping, index, GFP_KERNEL);
2159 if (unlikely(!page)) {
2165 s = bch2_page_state_create(page, 0);
2171 if (!PageUptodate(page)) {
2172 ret = bch2_read_single_page(page, mapping);
2177 if (index != start >> PAGE_SHIFT)
2179 if (index != end >> PAGE_SHIFT)
2180 end_offset = PAGE_SIZE;
2182 for (i = round_up(start_offset, block_bytes(c)) >> 9;
2183 i < round_down(end_offset, block_bytes(c)) >> 9;
2185 s->s[i].nr_replicas = 0;
2186 s->s[i].state = SECTOR_UNALLOCATED;
2189 zero_user_segment(page, start_offset, end_offset);
2192 * Bit of a hack - we don't want truncate to fail due to -ENOSPC.
2194 * XXX: because we aren't currently tracking whether the page has actual
2195 * data in it (vs. just 0s, or only partially written) this wrong. ick.
2197 ret = bch2_get_page_disk_reservation(c, inode, page, false);
2200 __set_page_dirty_nobuffers(page);
2208 static int bch2_truncate_page(struct bch_inode_info *inode, loff_t from)
2210 return __bch2_truncate_page(inode, from >> PAGE_SHIFT,
2211 from, round_up(from, PAGE_SIZE));
2214 static int bch2_extend(struct bch_inode_info *inode,
2215 struct bch_inode_unpacked *inode_u,
2216 struct iattr *iattr)
2218 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2219 struct address_space *mapping = inode->v.i_mapping;
2225 * this has to be done _before_ extending i_size:
2227 ret = filemap_write_and_wait_range(mapping, inode_u->bi_size, S64_MAX);
2231 truncate_setsize(&inode->v, iattr->ia_size);
2232 setattr_copy(&inode->v, iattr);
2234 mutex_lock(&inode->ei_update_lock);
2235 ret = bch2_write_inode_size(c, inode, inode->v.i_size,
2236 ATTR_MTIME|ATTR_CTIME);
2237 mutex_unlock(&inode->ei_update_lock);
2242 static int bch2_truncate_finish_fn(struct bch_inode_info *inode,
2243 struct bch_inode_unpacked *bi,
2246 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2248 bi->bi_flags &= ~BCH_INODE_I_SIZE_DIRTY;
2249 bi->bi_mtime = bi->bi_ctime = bch2_current_time(c);
2253 static int bch2_truncate_start_fn(struct bch_inode_info *inode,
2254 struct bch_inode_unpacked *bi, void *p)
2256 u64 *new_i_size = p;
2258 bi->bi_flags |= BCH_INODE_I_SIZE_DIRTY;
2259 bi->bi_size = *new_i_size;
2263 int bch2_truncate(struct bch_inode_info *inode, struct iattr *iattr)
2265 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2266 struct address_space *mapping = inode->v.i_mapping;
2267 struct bch_inode_unpacked inode_u;
2268 struct btree_trans trans;
2269 struct btree_iter *iter;
2270 u64 new_i_size = iattr->ia_size;
2271 s64 i_sectors_delta = 0;
2274 inode_dio_wait(&inode->v);
2275 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2278 * fetch current on disk i_size: inode is locked, i_size can only
2279 * increase underneath us:
2281 bch2_trans_init(&trans, c, 0, 0);
2282 iter = bch2_inode_peek(&trans, &inode_u, inode->v.i_ino, 0);
2283 ret = PTR_ERR_OR_ZERO(iter);
2284 bch2_trans_exit(&trans);
2290 * check this before next assertion; on filesystem error our normal
2291 * invariants are a bit broken (truncate has to truncate the page cache
2292 * before the inode).
2294 ret = bch2_journal_error(&c->journal);
2298 BUG_ON(inode->v.i_size < inode_u.bi_size);
2300 if (iattr->ia_size > inode->v.i_size) {
2301 ret = bch2_extend(inode, &inode_u, iattr);
2305 ret = bch2_truncate_page(inode, iattr->ia_size);
2310 * When extending, we're going to write the new i_size to disk
2311 * immediately so we need to flush anything above the current on disk
2314 * Also, when extending we need to flush the page that i_size currently
2315 * straddles - if it's mapped to userspace, we need to ensure that
2316 * userspace has to redirty it and call .mkwrite -> set_page_dirty
2317 * again to allocate the part of the page that was extended.
2319 if (iattr->ia_size > inode_u.bi_size)
2320 ret = filemap_write_and_wait_range(mapping,
2322 iattr->ia_size - 1);
2323 else if (iattr->ia_size & (PAGE_SIZE - 1))
2324 ret = filemap_write_and_wait_range(mapping,
2325 round_down(iattr->ia_size, PAGE_SIZE),
2326 iattr->ia_size - 1);
2330 mutex_lock(&inode->ei_update_lock);
2331 ret = bch2_write_inode(c, inode, bch2_truncate_start_fn,
2333 mutex_unlock(&inode->ei_update_lock);
2338 truncate_setsize(&inode->v, iattr->ia_size);
2340 ret = bch2_fpunch(c, inode->v.i_ino,
2341 round_up(iattr->ia_size, block_bytes(c)) >> 9,
2342 U64_MAX, &inode->ei_journal_seq, &i_sectors_delta);
2343 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2348 setattr_copy(&inode->v, iattr);
2350 mutex_lock(&inode->ei_update_lock);
2351 ret = bch2_write_inode(c, inode, bch2_truncate_finish_fn, NULL,
2352 ATTR_MTIME|ATTR_CTIME);
2353 mutex_unlock(&inode->ei_update_lock);
2355 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2361 static long bchfs_fpunch(struct bch_inode_info *inode, loff_t offset, loff_t len)
2363 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2364 u64 discard_start = round_up(offset, block_bytes(c)) >> 9;
2365 u64 discard_end = round_down(offset + len, block_bytes(c)) >> 9;
2368 inode_lock(&inode->v);
2369 inode_dio_wait(&inode->v);
2370 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2372 ret = __bch2_truncate_page(inode,
2373 offset >> PAGE_SHIFT,
2374 offset, offset + len);
2378 if (offset >> PAGE_SHIFT !=
2379 (offset + len) >> PAGE_SHIFT) {
2380 ret = __bch2_truncate_page(inode,
2381 (offset + len) >> PAGE_SHIFT,
2382 offset, offset + len);
2387 truncate_pagecache_range(&inode->v, offset, offset + len - 1);
2389 if (discard_start < discard_end) {
2390 s64 i_sectors_delta = 0;
2392 ret = bch2_fpunch(c, inode->v.i_ino,
2393 discard_start, discard_end,
2394 &inode->ei_journal_seq,
2396 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2399 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2400 inode_unlock(&inode->v);
2405 static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
2406 loff_t offset, loff_t len,
2409 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2410 struct address_space *mapping = inode->v.i_mapping;
2411 struct bkey_on_stack copy;
2412 struct btree_trans trans;
2413 struct btree_iter *src, *dst;
2414 loff_t shift, new_size;
2418 if ((offset | len) & (block_bytes(c) - 1))
2421 bkey_on_stack_init(©);
2422 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 256);
2425 * We need i_mutex to keep the page cache consistent with the extents
2426 * btree, and the btree consistent with i_size - we don't need outside
2427 * locking for the extents btree itself, because we're using linked
2430 inode_lock(&inode->v);
2431 inode_dio_wait(&inode->v);
2432 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2436 if (inode->v.i_sb->s_maxbytes - inode->v.i_size < len)
2440 if (offset >= inode->v.i_size)
2443 src_start = U64_MAX;
2447 if (offset + len >= inode->v.i_size)
2450 src_start = offset + len;
2454 new_size = inode->v.i_size + shift;
2456 ret = write_invalidate_inode_pages_range(mapping, offset, LLONG_MAX);
2461 i_size_write(&inode->v, new_size);
2462 mutex_lock(&inode->ei_update_lock);
2463 ret = bch2_write_inode_size(c, inode, new_size,
2464 ATTR_MTIME|ATTR_CTIME);
2465 mutex_unlock(&inode->ei_update_lock);
2467 s64 i_sectors_delta = 0;
2469 ret = bch2_fpunch(c, inode->v.i_ino,
2470 offset >> 9, (offset + len) >> 9,
2471 &inode->ei_journal_seq,
2473 i_sectors_acct(c, inode, NULL, i_sectors_delta);
2479 src = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2480 POS(inode->v.i_ino, src_start >> 9),
2482 BUG_ON(IS_ERR_OR_NULL(src));
2484 dst = bch2_trans_copy_iter(&trans, src);
2485 BUG_ON(IS_ERR_OR_NULL(dst));
2488 struct disk_reservation disk_res =
2489 bch2_disk_reservation_init(c, 0);
2490 struct bkey_i delete;
2492 struct bpos next_pos;
2493 struct bpos move_pos = POS(inode->v.i_ino, offset >> 9);
2494 struct bpos atomic_end;
2495 unsigned trigger_flags = 0;
2498 ? bch2_btree_iter_peek_prev(src)
2499 : bch2_btree_iter_peek(src);
2500 if ((ret = bkey_err(k)))
2503 if (!k.k || k.k->p.inode != inode->v.i_ino)
2506 BUG_ON(bkey_cmp(src->pos, bkey_start_pos(k.k)));
2509 bkey_cmp(k.k->p, POS(inode->v.i_ino, offset >> 9)) <= 0)
2512 bkey_on_stack_reassemble(©, c, k);
2515 bkey_cmp(bkey_start_pos(k.k), move_pos) < 0) {
2516 bch2_cut_front(move_pos, copy.k);
2517 bch2_btree_iter_set_pos(src, bkey_start_pos(©.k->k));
2520 copy.k->k.p.offset += shift >> 9;
2521 bch2_btree_iter_set_pos(dst, bkey_start_pos(©.k->k));
2523 ret = bch2_extent_atomic_end(dst, copy.k, &atomic_end);
2527 if (bkey_cmp(atomic_end, copy.k->k.p)) {
2529 move_pos = atomic_end;
2530 move_pos.offset -= shift >> 9;
2533 bch2_cut_back(atomic_end, copy.k);
2537 bkey_init(&delete.k);
2538 delete.k.p = src->pos;
2539 bch2_key_resize(&delete.k, copy.k->k.size);
2541 next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
2543 if (copy.k->k.size == k.k->size) {
2545 * If we're moving the entire extent, we can skip
2548 trigger_flags |= BTREE_TRIGGER_NORUN;
2550 /* We might end up splitting compressed extents: */
2552 bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy.k));
2554 ret = bch2_disk_reservation_get(c, &disk_res,
2555 copy.k->k.size, nr_ptrs,
2556 BCH_DISK_RESERVATION_NOFAIL);
2560 ret = bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
2561 bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
2562 bch2_trans_commit(&trans, &disk_res,
2563 &inode->ei_journal_seq,
2564 BTREE_INSERT_NOFAIL);
2565 bch2_disk_reservation_put(c, &disk_res);
2568 bch2_btree_iter_set_pos(src, next_pos);
2575 bch2_trans_cond_resched(&trans);
2577 bch2_trans_unlock(&trans);
2580 i_size_write(&inode->v, new_size);
2581 mutex_lock(&inode->ei_update_lock);
2582 ret = bch2_write_inode_size(c, inode, new_size,
2583 ATTR_MTIME|ATTR_CTIME);
2584 mutex_unlock(&inode->ei_update_lock);
2587 bch2_trans_exit(&trans);
2588 bkey_on_stack_exit(©, c);
2589 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2590 inode_unlock(&inode->v);
2594 static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
2595 loff_t offset, loff_t len)
2597 struct address_space *mapping = inode->v.i_mapping;
2598 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2599 struct btree_trans trans;
2600 struct btree_iter *iter;
2601 struct bpos end_pos;
2602 loff_t end = offset + len;
2603 loff_t block_start = round_down(offset, block_bytes(c));
2604 loff_t block_end = round_up(end, block_bytes(c));
2606 unsigned replicas = io_opts(c, &inode->ei_inode).data_replicas;
2609 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
2611 inode_lock(&inode->v);
2612 inode_dio_wait(&inode->v);
2613 bch2_pagecache_block_get(&inode->ei_pagecache_lock);
2615 if (!(mode & FALLOC_FL_KEEP_SIZE) && end > inode->v.i_size) {
2616 ret = inode_newsize_ok(&inode->v, end);
2621 if (mode & FALLOC_FL_ZERO_RANGE) {
2622 ret = __bch2_truncate_page(inode,
2623 offset >> PAGE_SHIFT,
2627 offset >> PAGE_SHIFT != end >> PAGE_SHIFT)
2628 ret = __bch2_truncate_page(inode,
2635 truncate_pagecache_range(&inode->v, offset, end - 1);
2638 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
2639 POS(inode->v.i_ino, block_start >> 9),
2640 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2641 end_pos = POS(inode->v.i_ino, block_end >> 9);
2643 while (bkey_cmp(iter->pos, end_pos) < 0) {
2644 s64 i_sectors_delta = 0;
2645 struct disk_reservation disk_res = { 0 };
2646 struct quota_res quota_res = { 0 };
2647 struct bkey_i_reservation reservation;
2650 bch2_trans_begin(&trans);
2652 k = bch2_btree_iter_peek_slot(iter);
2653 if ((ret = bkey_err(k)))
2656 /* already reserved */
2657 if (k.k->type == KEY_TYPE_reservation &&
2658 bkey_s_c_to_reservation(k).v->nr_replicas >= replicas) {
2659 bch2_btree_iter_next_slot(iter);
2663 if (bkey_extent_is_data(k.k) &&
2664 !(mode & FALLOC_FL_ZERO_RANGE)) {
2665 bch2_btree_iter_next_slot(iter);
2669 bkey_reservation_init(&reservation.k_i);
2670 reservation.k.type = KEY_TYPE_reservation;
2671 reservation.k.p = k.k->p;
2672 reservation.k.size = k.k->size;
2674 bch2_cut_front(iter->pos, &reservation.k_i);
2675 bch2_cut_back(end_pos, &reservation.k_i);
2677 sectors = reservation.k.size;
2678 reservation.v.nr_replicas = bch2_bkey_nr_ptrs_allocated(k);
2680 if (!bkey_extent_is_allocation(k.k)) {
2681 ret = bch2_quota_reservation_add(c, inode,
2688 if (reservation.v.nr_replicas < replicas ||
2689 bch2_bkey_sectors_compressed(k)) {
2690 ret = bch2_disk_reservation_get(c, &disk_res, sectors,
2695 reservation.v.nr_replicas = disk_res.nr_replicas;
2698 ret = bch2_extent_update(&trans, iter, &reservation.k_i,
2699 &disk_res, &inode->ei_journal_seq,
2700 0, &i_sectors_delta);
2701 i_sectors_acct(c, inode, "a_res, i_sectors_delta);
2703 bch2_quota_reservation_put(c, inode, "a_res);
2704 bch2_disk_reservation_put(c, &disk_res);
2712 * Do we need to extend the file?
2714 * If we zeroed up to the end of the file, we dropped whatever writes
2715 * were going to write out the current i_size, so we have to extend
2716 * manually even if FL_KEEP_SIZE was set:
2718 if (end >= inode->v.i_size &&
2719 (!(mode & FALLOC_FL_KEEP_SIZE) ||
2720 (mode & FALLOC_FL_ZERO_RANGE))) {
2721 struct btree_iter *inode_iter;
2722 struct bch_inode_unpacked inode_u;
2725 bch2_trans_begin(&trans);
2726 inode_iter = bch2_inode_peek(&trans, &inode_u,
2728 ret = PTR_ERR_OR_ZERO(inode_iter);
2729 } while (ret == -EINTR);
2731 bch2_trans_unlock(&trans);
2737 * Sync existing appends before extending i_size,
2738 * as in bch2_extend():
2740 ret = filemap_write_and_wait_range(mapping,
2741 inode_u.bi_size, S64_MAX);
2745 if (mode & FALLOC_FL_KEEP_SIZE)
2746 end = inode->v.i_size;
2748 i_size_write(&inode->v, end);
2750 mutex_lock(&inode->ei_update_lock);
2751 ret = bch2_write_inode_size(c, inode, end, 0);
2752 mutex_unlock(&inode->ei_update_lock);
2755 bch2_trans_exit(&trans);
2756 bch2_pagecache_block_put(&inode->ei_pagecache_lock);
2757 inode_unlock(&inode->v);
2761 long bch2_fallocate_dispatch(struct file *file, int mode,
2762 loff_t offset, loff_t len)
2764 struct bch_inode_info *inode = file_bch_inode(file);
2765 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2768 if (!percpu_ref_tryget(&c->writes))
2771 if (!(mode & ~(FALLOC_FL_KEEP_SIZE|FALLOC_FL_ZERO_RANGE)))
2772 ret = bchfs_fallocate(inode, mode, offset, len);
2773 else if (mode == (FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE))
2774 ret = bchfs_fpunch(inode, offset, len);
2775 else if (mode == FALLOC_FL_INSERT_RANGE)
2776 ret = bchfs_fcollapse_finsert(inode, offset, len, true);
2777 else if (mode == FALLOC_FL_COLLAPSE_RANGE)
2778 ret = bchfs_fcollapse_finsert(inode, offset, len, false);
2782 percpu_ref_put(&c->writes);
2787 static void mark_range_unallocated(struct bch_inode_info *inode,
2788 loff_t start, loff_t end)
2790 pgoff_t index = start >> PAGE_SHIFT;
2791 pgoff_t end_index = (end - 1) >> PAGE_SHIFT;
2792 struct pagevec pvec;
2794 pagevec_init(&pvec);
2797 unsigned nr_pages, i, j;
2799 nr_pages = pagevec_lookup_range(&pvec, inode->v.i_mapping,
2804 for (i = 0; i < nr_pages; i++) {
2805 struct page *page = pvec.pages[i];
2806 struct bch_page_state *s;
2809 s = bch2_page_state(page);
2812 spin_lock(&s->lock);
2813 for (j = 0; j < PAGE_SECTORS; j++)
2814 s->s[j].nr_replicas = 0;
2815 spin_unlock(&s->lock);
2820 pagevec_release(&pvec);
2821 } while (index <= end_index);
2824 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
2825 struct file *file_dst, loff_t pos_dst,
2826 loff_t len, unsigned remap_flags)
2828 struct bch_inode_info *src = file_bch_inode(file_src);
2829 struct bch_inode_info *dst = file_bch_inode(file_dst);
2830 struct bch_fs *c = src->v.i_sb->s_fs_info;
2831 s64 i_sectors_delta = 0;
2835 if (remap_flags & ~(REMAP_FILE_DEDUP|REMAP_FILE_ADVISORY))
2838 if (remap_flags & REMAP_FILE_DEDUP)
2841 if ((pos_src & (block_bytes(c) - 1)) ||
2842 (pos_dst & (block_bytes(c) - 1)))
2846 abs(pos_src - pos_dst) < len)
2849 bch2_lock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2851 file_update_time(file_dst);
2853 inode_dio_wait(&src->v);
2854 inode_dio_wait(&dst->v);
2856 ret = generic_remap_file_range_prep(file_src, pos_src,
2859 if (ret < 0 || len == 0)
2862 aligned_len = round_up((u64) len, block_bytes(c));
2864 ret = write_invalidate_inode_pages_range(dst->v.i_mapping,
2865 pos_dst, pos_dst + len - 1);
2869 mark_range_unallocated(src, pos_src, pos_src + aligned_len);
2871 ret = bch2_remap_range(c,
2872 POS(dst->v.i_ino, pos_dst >> 9),
2873 POS(src->v.i_ino, pos_src >> 9),
2875 &dst->ei_journal_seq,
2876 pos_dst + len, &i_sectors_delta);
2881 * due to alignment, we might have remapped slightly more than requsted
2883 ret = min((u64) ret << 9, (u64) len);
2885 /* XXX get a quota reservation */
2886 i_sectors_acct(c, dst, NULL, i_sectors_delta);
2888 spin_lock(&dst->v.i_lock);
2889 if (pos_dst + ret > dst->v.i_size)
2890 i_size_write(&dst->v, pos_dst + ret);
2891 spin_unlock(&dst->v.i_lock);
2893 bch2_unlock_inodes(INODE_LOCK|INODE_PAGECACHE_BLOCK, src, dst);
2900 static int page_data_offset(struct page *page, unsigned offset)
2902 struct bch_page_state *s = bch2_page_state(page);
2906 for (i = offset >> 9; i < PAGE_SECTORS; i++)
2907 if (s->s[i].state >= SECTOR_DIRTY)
2913 static loff_t bch2_seek_pagecache_data(struct inode *vinode,
2914 loff_t start_offset,
2917 struct address_space *mapping = vinode->i_mapping;
2919 pgoff_t start_index = start_offset >> PAGE_SHIFT;
2920 pgoff_t end_index = end_offset >> PAGE_SHIFT;
2921 pgoff_t index = start_index;
2925 while (index <= end_index) {
2926 if (find_get_pages_range(mapping, &index, end_index, 1, &page)) {
2929 offset = page_data_offset(page,
2930 page->index == start_index
2931 ? start_offset & (PAGE_SIZE - 1)
2934 ret = clamp(((loff_t) page->index << PAGE_SHIFT) +
2936 start_offset, end_offset);
2952 static loff_t bch2_seek_data(struct file *file, u64 offset)
2954 struct bch_inode_info *inode = file_bch_inode(file);
2955 struct bch_fs *c = inode->v.i_sb->s_fs_info;
2956 struct btree_trans trans;
2957 struct btree_iter *iter;
2959 u64 isize, next_data = MAX_LFS_FILESIZE;
2962 isize = i_size_read(&inode->v);
2963 if (offset >= isize)
2966 bch2_trans_init(&trans, c, 0, 0);
2968 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
2969 POS(inode->v.i_ino, offset >> 9), 0, k, ret) {
2970 if (k.k->p.inode != inode->v.i_ino) {
2972 } else if (bkey_extent_is_data(k.k)) {
2973 next_data = max(offset, bkey_start_offset(k.k) << 9);
2975 } else if (k.k->p.offset >> 9 > isize)
2979 ret = bch2_trans_exit(&trans) ?: ret;
2983 if (next_data > offset)
2984 next_data = bch2_seek_pagecache_data(&inode->v,
2987 if (next_data >= isize)
2990 return vfs_setpos(file, next_data, MAX_LFS_FILESIZE);
2993 static int __page_hole_offset(struct page *page, unsigned offset)
2995 struct bch_page_state *s = bch2_page_state(page);
3001 for (i = offset >> 9; i < PAGE_SECTORS; i++)
3002 if (s->s[i].state < SECTOR_DIRTY)
3008 static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
3010 pgoff_t index = offset >> PAGE_SHIFT;
3015 page = find_lock_entry(mapping, index);
3016 if (!page || xa_is_value(page))
3019 pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
3021 ret = ((loff_t) index << PAGE_SHIFT) + pg_offset;
3028 static loff_t bch2_seek_pagecache_hole(struct inode *vinode,
3029 loff_t start_offset,
3032 struct address_space *mapping = vinode->i_mapping;
3033 loff_t offset = start_offset, hole;
3035 while (offset < end_offset) {
3036 hole = page_hole_offset(mapping, offset);
3037 if (hole >= 0 && hole <= end_offset)
3038 return max(start_offset, hole);
3040 offset += PAGE_SIZE;
3041 offset &= PAGE_MASK;
3047 static loff_t bch2_seek_hole(struct file *file, u64 offset)
3049 struct bch_inode_info *inode = file_bch_inode(file);
3050 struct bch_fs *c = inode->v.i_sb->s_fs_info;
3051 struct btree_trans trans;
3052 struct btree_iter *iter;
3054 u64 isize, next_hole = MAX_LFS_FILESIZE;
3057 isize = i_size_read(&inode->v);
3058 if (offset >= isize)
3061 bch2_trans_init(&trans, c, 0, 0);
3063 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
3064 POS(inode->v.i_ino, offset >> 9),
3065 BTREE_ITER_SLOTS, k, ret) {
3066 if (k.k->p.inode != inode->v.i_ino) {
3067 next_hole = bch2_seek_pagecache_hole(&inode->v,
3068 offset, MAX_LFS_FILESIZE);
3070 } else if (!bkey_extent_is_data(k.k)) {
3071 next_hole = bch2_seek_pagecache_hole(&inode->v,
3072 max(offset, bkey_start_offset(k.k) << 9),
3073 k.k->p.offset << 9);
3075 if (next_hole < k.k->p.offset << 9)
3078 offset = max(offset, bkey_start_offset(k.k) << 9);
3082 ret = bch2_trans_exit(&trans) ?: ret;
3086 if (next_hole > isize)
3089 return vfs_setpos(file, next_hole, MAX_LFS_FILESIZE);
3092 loff_t bch2_llseek(struct file *file, loff_t offset, int whence)
3098 return generic_file_llseek(file, offset, whence);
3100 return bch2_seek_data(file, offset);
3102 return bch2_seek_hole(file, offset);
3108 void bch2_fs_fsio_exit(struct bch_fs *c)
3110 bioset_exit(&c->dio_write_bioset);
3111 bioset_exit(&c->dio_read_bioset);
3112 bioset_exit(&c->writepage_bioset);
3115 int bch2_fs_fsio_init(struct bch_fs *c)
3119 pr_verbose_init(c->opts, "");
3121 if (bioset_init(&c->writepage_bioset,
3122 4, offsetof(struct bch_writepage_io, op.wbio.bio),
3123 BIOSET_NEED_BVECS) ||
3124 bioset_init(&c->dio_read_bioset,
3125 4, offsetof(struct dio_read, rbio.bio),
3126 BIOSET_NEED_BVECS) ||
3127 bioset_init(&c->dio_write_bioset,
3128 4, offsetof(struct dio_write, op.wbio.bio),
3132 pr_verbose_init(c->opts, "ret %i", ret);
3136 #endif /* NO_BCACHEFS_FS */