1 // SPDX-License-Identifier: GPL-2.0
3 * Some low level IO code, and hacks for various block layer limitations
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
10 #include "alloc_background.h"
11 #include "alloc_foreground.h"
14 #include "btree_update.h"
19 #include "data_update.h"
21 #include "disk_groups.h"
24 #include "extent_update.h"
30 #include "nocow_locking.h"
31 #include "rebalance.h"
32 #include "subvolume.h"
36 #include <linux/blkdev.h>
37 #include <linux/prefetch.h>
38 #include <linux/random.h>
39 #include <linux/sched/mm.h>
41 #include <trace/events/bcachefs.h>
43 const char *bch2_blk_status_to_str(blk_status_t status)
45 if (status == BLK_STS_REMOVED)
46 return "device removed";
47 return blk_status_to_str(status);
50 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
52 static bool bch2_target_congested(struct bch_fs *c, u16 target)
54 const struct bch_devs_mask *devs;
55 unsigned d, nr = 0, total = 0;
56 u64 now = local_clock(), last;
64 devs = bch2_target_to_mask(c, target) ?:
65 &c->rw_devs[BCH_DATA_user];
67 for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
68 ca = rcu_dereference(c->devs[d]);
72 congested = atomic_read(&ca->congested);
73 last = READ_ONCE(ca->congested_last);
74 if (time_after64(now, last))
75 congested -= (now - last) >> 12;
77 total += max(congested, 0LL);
82 return bch2_rand_range(nr * CONGESTED_MAX) < total;
85 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
89 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
90 /* ideally we'd be taking into account the device's variance here: */
91 u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
92 s64 latency_over = io_latency - latency_threshold;
94 if (latency_threshold && latency_over > 0) {
96 * bump up congested by approximately latency_over * 4 /
97 * latency_threshold - we don't need much accuracy here so don't
98 * bother with the divide:
100 if (atomic_read(&ca->congested) < CONGESTED_MAX)
101 atomic_add(latency_over >>
102 max_t(int, ilog2(latency_threshold) - 2, 0),
105 ca->congested_last = now;
106 } else if (atomic_read(&ca->congested) > 0) {
107 atomic_dec(&ca->congested);
111 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
113 atomic64_t *latency = &ca->cur_latency[rw];
114 u64 now = local_clock();
115 u64 io_latency = time_after64(now, submit_time)
118 u64 old, new, v = atomic64_read(latency);
124 * If the io latency was reasonably close to the current
125 * latency, skip doing the update and atomic operation - most of
128 if (abs((int) (old - io_latency)) < (old >> 1) &&
132 new = ewma_add(old, io_latency, 5);
133 } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
135 bch2_congested_acct(ca, io_latency, now, rw);
137 __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
142 static bool bch2_target_congested(struct bch_fs *c, u16 target)
149 /* Allocate, free from mempool: */
151 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
153 struct bvec_iter_all iter;
156 bio_for_each_segment_all(bv, bio, iter)
157 if (bv->bv_page != ZERO_PAGE(0))
158 mempool_free(bv->bv_page, &c->bio_bounce_pages);
162 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
166 if (likely(!*using_mempool)) {
167 page = alloc_page(GFP_NOIO);
168 if (unlikely(!page)) {
169 mutex_lock(&c->bio_bounce_pages_lock);
170 *using_mempool = true;
176 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
182 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
185 bool using_mempool = false;
188 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
189 unsigned len = min_t(size_t, PAGE_SIZE, size);
191 BUG_ON(!bio_add_page(bio, page, len, 0));
196 mutex_unlock(&c->bio_bounce_pages_lock);
199 /* Extent update path: */
201 int bch2_sum_sector_overwrites(struct btree_trans *trans,
202 struct btree_iter *extent_iter,
204 bool *usage_increasing,
205 s64 *i_sectors_delta,
206 s64 *disk_sectors_delta)
208 struct bch_fs *c = trans->c;
209 struct btree_iter iter;
211 unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
212 bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
215 *usage_increasing = false;
216 *i_sectors_delta = 0;
217 *disk_sectors_delta = 0;
219 bch2_trans_copy_iter(&iter, extent_iter);
221 for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) {
222 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
223 max(bkey_start_offset(&new->k),
224 bkey_start_offset(old.k));
226 *i_sectors_delta += sectors *
227 (bkey_extent_is_allocation(&new->k) -
228 bkey_extent_is_allocation(old.k));
230 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
231 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
232 ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
235 if (!*usage_increasing &&
236 (new->k.p.snapshot != old.k->p.snapshot ||
237 new_replicas > bch2_bkey_replicas(c, old) ||
238 (!new_compressed && bch2_bkey_sectors_compressed(old))))
239 *usage_increasing = true;
241 if (bkey_ge(old.k->p, new->k.p))
245 bch2_trans_iter_exit(trans, &iter);
249 static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
250 struct btree_iter *extent_iter,
254 struct btree_iter iter;
256 struct bkey_i_inode_v3 *inode;
257 unsigned inode_update_flags = BTREE_UPDATE_NOJOURNAL;
260 bch2_trans_iter_init(trans, &iter, BTREE_ID_inodes,
262 extent_iter->pos.inode,
263 extent_iter->snapshot),
264 BTREE_ITER_INTENT|BTREE_ITER_CACHED);
265 k = bch2_bkey_get_mut(trans, &iter);
266 ret = PTR_ERR_OR_ZERO(k);
270 if (unlikely(k->k.type != KEY_TYPE_inode_v3)) {
271 k = bch2_inode_to_v3(trans, k);
272 ret = PTR_ERR_OR_ZERO(k);
277 inode = bkey_i_to_inode_v3(k);
279 if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_I_SIZE_DIRTY) &&
280 new_i_size > le64_to_cpu(inode->v.bi_size)) {
281 inode->v.bi_size = cpu_to_le64(new_i_size);
282 inode_update_flags = 0;
285 if (i_sectors_delta) {
286 le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
287 inode_update_flags = 0;
290 if (inode->k.p.snapshot != iter.snapshot) {
291 inode->k.p.snapshot = iter.snapshot;
292 inode_update_flags = 0;
295 ret = bch2_trans_update(trans, &iter, &inode->k_i,
296 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
299 bch2_trans_iter_exit(trans, &iter);
303 int bch2_extent_update(struct btree_trans *trans,
305 struct btree_iter *iter,
307 struct disk_reservation *disk_res,
309 s64 *i_sectors_delta_total,
312 struct bpos next_pos;
313 bool usage_increasing;
314 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
318 * This traverses us the iterator without changing iter->path->pos to
319 * search_key() (which is pos + 1 for extents): we want there to be a
320 * path already traversed at iter->pos because
321 * bch2_trans_extent_update() will use it to attempt extent merging
323 ret = __bch2_btree_iter_traverse(iter);
327 ret = bch2_extent_trim_atomic(trans, iter, k);
333 ret = bch2_sum_sector_overwrites(trans, iter, k,
336 &disk_sectors_delta);
341 disk_sectors_delta > (s64) disk_res->sectors) {
342 ret = bch2_disk_reservation_add(trans->c, disk_res,
343 disk_sectors_delta - disk_res->sectors,
344 !check_enospc || !usage_increasing
345 ? BCH_DISK_RESERVATION_NOFAIL : 0);
352 * We always have to do an inode update - even when i_size/i_sectors
353 * aren't changing - for fsync to work properly; fsync relies on
354 * inode->bi_journal_seq which is updated by the trigger code:
356 ret = bch2_extent_update_i_size_sectors(trans, iter,
357 min(k->k.p.offset << 9, new_i_size),
359 bch2_trans_update(trans, iter, k, 0) ?:
360 bch2_trans_commit(trans, disk_res, NULL,
361 BTREE_INSERT_NOCHECK_RW|
362 BTREE_INSERT_NOFAIL);
366 if (i_sectors_delta_total)
367 *i_sectors_delta_total += i_sectors_delta;
368 bch2_btree_iter_set_pos(iter, next_pos);
372 /* Overwrites whatever was present with zeroes: */
373 int bch2_extent_fallocate(struct btree_trans *trans,
375 struct btree_iter *iter,
377 struct bch_io_opts opts,
378 s64 *i_sectors_delta,
379 struct write_point_specifier write_point)
381 struct bch_fs *c = trans->c;
382 struct disk_reservation disk_res = { 0 };
384 struct open_buckets open_buckets;
386 struct bkey_buf old, new;
387 bool have_reservation = false;
388 bool unwritten = opts.nocow &&
389 c->sb.version >= bcachefs_metadata_version_unwritten_extents;
392 bch2_bkey_buf_init(&old);
393 bch2_bkey_buf_init(&new);
394 closure_init_stack(&cl);
397 k = bch2_btree_iter_peek_slot(iter);
402 sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset);
404 if (!have_reservation) {
405 unsigned new_replicas =
406 max(0, (int) opts.data_replicas -
407 (int) bch2_bkey_nr_ptrs_fully_allocated(k));
409 * Get a disk reservation before (in the nocow case) calling
410 * into the allocator:
412 ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0);
416 bch2_bkey_buf_reassemble(&old, c, k);
419 if (have_reservation) {
420 if (!bch2_extents_match(k, bkey_i_to_s_c(old.k)))
423 bch2_key_resize(&new.k->k, sectors);
424 } else if (!unwritten) {
425 struct bkey_i_reservation *reservation;
427 bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64));
428 reservation = bkey_reservation_init(new.k);
429 reservation->k.p = iter->pos;
430 bch2_key_resize(&reservation->k, sectors);
431 reservation->v.nr_replicas = opts.data_replicas;
433 struct bkey_i_extent *e;
434 struct bch_devs_list devs_have;
435 struct write_point *wp;
436 struct bch_extent_ptr *ptr;
440 bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX);
442 e = bkey_extent_init(new.k);
445 ret = bch2_alloc_sectors_start_trans(trans,
446 opts.foreground_target,
452 RESERVE_none, 0, &cl, &wp);
453 if (bch2_err_matches(ret, BCH_ERR_operation_blocked)) {
454 bch2_trans_unlock(trans);
461 sectors = min(sectors, wp->sectors_free);
463 bch2_key_resize(&e->k, sectors);
465 bch2_open_bucket_get(c, wp, &open_buckets);
466 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
467 bch2_alloc_sectors_done(c, wp);
469 extent_for_each_ptr(extent_i_to_s(e), ptr)
470 ptr->unwritten = true;
473 have_reservation = true;
475 ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res,
476 0, i_sectors_delta, true);
478 if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
479 bch2_trans_unlock(trans);
483 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
484 bch2_trans_begin(trans);
488 bch2_open_buckets_put(c, &open_buckets);
489 bch2_disk_reservation_put(c, &disk_res);
490 bch2_bkey_buf_exit(&new, c);
491 bch2_bkey_buf_exit(&old, c);
497 * Returns -BCH_ERR_transacton_restart if we had to drop locks:
499 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
500 subvol_inum inum, u64 end,
501 s64 *i_sectors_delta)
503 struct bch_fs *c = trans->c;
504 unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
505 struct bpos end_pos = POS(inum.inum, end);
507 int ret = 0, ret2 = 0;
511 bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
512 struct disk_reservation disk_res =
513 bch2_disk_reservation_init(c, 0);
514 struct bkey_i delete;
519 bch2_trans_begin(trans);
521 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
525 bch2_btree_iter_set_snapshot(iter, snapshot);
528 * peek_upto() doesn't have ideal semantics for extents:
530 k = bch2_btree_iter_peek_upto(iter, end_pos);
538 bkey_init(&delete.k);
539 delete.k.p = iter->pos;
541 /* create the biggest key we can */
542 bch2_key_resize(&delete.k, max_sectors);
543 bch2_cut_back(end_pos, &delete);
545 ret = bch2_extent_update(trans, inum, iter, &delete,
546 &disk_res, 0, i_sectors_delta, false);
547 bch2_disk_reservation_put(c, &disk_res);
553 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
554 s64 *i_sectors_delta)
556 struct btree_trans trans;
557 struct btree_iter iter;
560 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
561 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
562 POS(inum.inum, start),
565 ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
567 bch2_trans_iter_exit(&trans, &iter);
568 bch2_trans_exit(&trans);
570 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
576 static int bch2_write_index_default(struct bch_write_op *op)
578 struct bch_fs *c = op->c;
580 struct keylist *keys = &op->insert_keys;
581 struct bkey_i *k = bch2_keylist_front(keys);
582 struct btree_trans trans;
583 struct btree_iter iter;
585 .subvol = op->subvol,
586 .inum = k->k.p.inode,
590 BUG_ON(!inum.subvol);
592 bch2_bkey_buf_init(&sk);
593 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
596 bch2_trans_begin(&trans);
598 k = bch2_keylist_front(keys);
599 bch2_bkey_buf_copy(&sk, c, k);
601 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
602 &sk.k->k.p.snapshot);
603 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
608 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
609 bkey_start_pos(&sk.k->k),
610 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
612 ret = bch2_extent_update(&trans, inum, &iter, sk.k,
614 op->new_i_size, &op->i_sectors_delta,
615 op->flags & BCH_WRITE_CHECK_ENOSPC);
616 bch2_trans_iter_exit(&trans, &iter);
618 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
623 if (bkey_ge(iter.pos, k->k.p))
624 bch2_keylist_pop_front(&op->insert_keys);
626 bch2_cut_front(iter.pos, k);
627 } while (!bch2_keylist_empty(keys));
629 bch2_trans_exit(&trans);
630 bch2_bkey_buf_exit(&sk, c);
637 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
638 enum bch_data_type type,
639 const struct bkey_i *k,
642 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
643 const struct bch_extent_ptr *ptr;
644 struct bch_write_bio *n;
647 BUG_ON(c->opts.nochanges);
649 bkey_for_each_ptr(ptrs, ptr) {
650 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
653 ca = bch_dev_bkey_exists(c, ptr->dev);
655 if (to_entry(ptr + 1) < ptrs.end) {
656 n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
657 GFP_NOIO, &ca->replica_set));
659 n->bio.bi_end_io = wbio->bio.bi_end_io;
660 n->bio.bi_private = wbio->bio.bi_private;
665 n->bio.bi_opf = wbio->bio.bi_opf;
666 bio_inc_remaining(&wbio->bio);
674 n->have_ioref = nocow || bch2_dev_get_ioref(ca,
675 type == BCH_DATA_btree ? READ : WRITE);
677 n->submit_time = local_clock();
678 n->inode_offset = bkey_start_offset(&k->k);
679 n->bio.bi_iter.bi_sector = ptr->offset;
681 if (likely(n->have_ioref)) {
682 this_cpu_add(ca->io_done->sectors[WRITE][type],
683 bio_sectors(&n->bio));
685 bio_set_dev(&n->bio, ca->disk_sb.bdev);
687 if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
694 n->bio.bi_status = BLK_STS_REMOVED;
700 static void __bch2_write(struct bch_write_op *);
702 static void bch2_write_done(struct closure *cl)
704 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
705 struct bch_fs *c = op->c;
707 bch2_disk_reservation_put(c, &op->res);
708 bch2_write_ref_put(c, BCH_WRITE_REF_write);
709 bch2_keylist_free(&op->insert_keys, op->inline_keys);
711 bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
714 closure_debug_destroy(cl);
719 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
721 struct keylist *keys = &op->insert_keys;
722 struct bch_extent_ptr *ptr;
723 struct bkey_i *src, *dst = keys->keys, *n;
725 for (src = keys->keys; src != keys->top; src = n) {
728 if (bkey_extent_is_direct_data(&src->k)) {
729 bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
730 test_bit(ptr->dev, op->failed.d));
732 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
737 memmove_u64s_down(dst, src, src->k.u64s);
738 dst = bkey_next(dst);
746 * bch_write_index - after a write, update index to point to new data
748 static void __bch2_write_index(struct bch_write_op *op)
750 struct bch_fs *c = op->c;
751 struct keylist *keys = &op->insert_keys;
756 if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
757 ret = bch2_write_drop_io_error_ptrs(op);
763 * probably not the ideal place to hook this in, but I don't
764 * particularly want to plumb io_opts all the way through the btree
765 * update stack right now
767 for_each_keylist_key(keys, k)
768 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
770 if (!bch2_keylist_empty(keys)) {
771 u64 sectors_start = keylist_sectors(keys);
773 ret = !(op->flags & BCH_WRITE_MOVE)
774 ? bch2_write_index_default(op)
775 : bch2_data_update_index_update(op);
777 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
778 BUG_ON(keylist_sectors(keys) && !ret);
780 op->written += sectors_start - keylist_sectors(keys);
782 if (ret && !bch2_err_matches(ret, EROFS)) {
783 struct bkey_i *k = bch2_keylist_front(&op->insert_keys);
785 bch_err_inum_offset_ratelimited(c,
786 k->k.p.inode, k->k.p.offset << 9,
787 "write error while doing btree update: %s",
795 /* If some a bucket wasn't written, we can't erasure code it: */
796 for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
797 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
799 bch2_open_buckets_put(c, &op->open_buckets);
802 keys->top = keys->keys;
804 op->flags |= BCH_WRITE_DONE;
808 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
810 if (state != wp->state) {
811 u64 now = ktime_get_ns();
813 if (wp->last_state_change &&
814 time_after64(now, wp->last_state_change))
815 wp->time[wp->state] += now - wp->last_state_change;
817 wp->last_state_change = now;
821 static inline void wp_update_state(struct write_point *wp, bool running)
823 enum write_point_state state;
825 state = running ? WRITE_POINT_running :
826 !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
827 : WRITE_POINT_stopped;
829 __wp_update_state(wp, state);
832 static void bch2_write_index(struct closure *cl)
834 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
835 struct write_point *wp = op->wp;
836 struct workqueue_struct *wq = index_update_wq(op);
841 * We're not using wp->writes_lock here, so this is racey: that's ok,
842 * because this is just for diagnostic purposes, and we're running out
843 * of interrupt context here so if we were to take the log we'd have to
844 * switch to spin_lock_irq()/irqsave(), which is not free:
846 if (wp->state == WRITE_POINT_waiting_io)
847 __wp_update_state(wp, WRITE_POINT_waiting_work);
849 op->btree_update_ready = true;
850 queue_work(wq, &wp->index_update_work);
853 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
855 op->btree_update_ready = false;
858 spin_lock(&wp->writes_lock);
859 list_add_tail(&op->wp_list, &wp->writes);
860 if (wp->state == WRITE_POINT_stopped)
861 __wp_update_state(wp, WRITE_POINT_waiting_io);
862 spin_unlock(&wp->writes_lock);
865 void bch2_write_point_do_index_updates(struct work_struct *work)
867 struct write_point *wp =
868 container_of(work, struct write_point, index_update_work);
869 struct bch_write_op *op;
872 spin_lock(&wp->writes_lock);
873 list_for_each_entry(op, &wp->writes, wp_list)
874 if (op->btree_update_ready) {
875 list_del(&op->wp_list);
880 wp_update_state(wp, op != NULL);
881 spin_unlock(&wp->writes_lock);
886 op->flags |= BCH_WRITE_IN_WORKER;
888 __bch2_write_index(op);
890 if (!(op->flags & BCH_WRITE_DONE))
893 bch2_write_done(&op->cl);
897 static void bch2_write_endio(struct bio *bio)
899 struct closure *cl = bio->bi_private;
900 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
901 struct bch_write_bio *wbio = to_wbio(bio);
902 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
903 struct bch_fs *c = wbio->c;
904 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
906 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
908 wbio->inode_offset << 9,
909 "data write error: %s",
910 bch2_blk_status_to_str(bio->bi_status))) {
911 set_bit(wbio->dev, op->failed.d);
912 op->flags |= BCH_WRITE_IO_ERROR;
916 set_bit(wbio->dev, op->devs_need_flush->d);
918 if (wbio->have_ioref) {
919 bch2_latency_acct(ca, wbio->submit_time, WRITE);
920 percpu_ref_put(&ca->io_ref);
924 bch2_bio_free_pages_pool(c, bio);
930 bio_endio(&parent->bio);
935 static void init_append_extent(struct bch_write_op *op,
936 struct write_point *wp,
937 struct bversion version,
938 struct bch_extent_crc_unpacked crc)
940 struct bkey_i_extent *e;
942 op->pos.offset += crc.uncompressed_size;
944 e = bkey_extent_init(op->insert_keys.top);
946 e->k.size = crc.uncompressed_size;
947 e->k.version = version;
950 crc.compression_type ||
952 bch2_extent_crc_append(&e->k_i, crc);
954 bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
955 op->flags & BCH_WRITE_CACHED);
957 bch2_keylist_push(&op->insert_keys);
960 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
961 struct write_point *wp,
963 bool *page_alloc_failed,
966 struct bch_write_bio *wbio;
968 unsigned output_available =
969 min(wp->sectors_free << 9, src->bi_iter.bi_size);
970 unsigned pages = DIV_ROUND_UP(output_available +
972 ? ((unsigned long) buf & (PAGE_SIZE - 1))
975 pages = min(pages, BIO_MAX_VECS);
977 bio = bio_alloc_bioset(NULL, pages, 0,
978 GFP_NOIO, &c->bio_write);
979 wbio = wbio_init(bio);
980 wbio->put_bio = true;
981 /* copy WRITE_SYNC flag */
982 wbio->bio.bi_opf = src->bi_opf;
985 bch2_bio_map(bio, buf, output_available);
992 * We can't use mempool for more than c->sb.encoded_extent_max
993 * worth of pages, but we'd like to allocate more if we can:
995 bch2_bio_alloc_pages_pool(c, bio,
996 min_t(unsigned, output_available,
997 c->opts.encoded_extent_max));
999 if (bio->bi_iter.bi_size < output_available)
1000 *page_alloc_failed =
1001 bch2_bio_alloc_pages(bio,
1003 bio->bi_iter.bi_size,
1009 static int bch2_write_rechecksum(struct bch_fs *c,
1010 struct bch_write_op *op,
1011 unsigned new_csum_type)
1013 struct bio *bio = &op->wbio.bio;
1014 struct bch_extent_crc_unpacked new_crc;
1017 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
1019 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
1020 bch2_csum_type_is_encryption(new_csum_type))
1021 new_csum_type = op->crc.csum_type;
1023 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
1025 op->crc.offset, op->crc.live_size,
1030 bio_advance(bio, op->crc.offset << 9);
1031 bio->bi_iter.bi_size = op->crc.live_size << 9;
1036 static int bch2_write_decrypt(struct bch_write_op *op)
1038 struct bch_fs *c = op->c;
1039 struct nonce nonce = extent_nonce(op->version, op->crc);
1040 struct bch_csum csum;
1043 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
1047 * If we need to decrypt data in the write path, we'll no longer be able
1048 * to verify the existing checksum (poly1305 mac, in this case) after
1049 * it's decrypted - this is the last point we'll be able to reverify the
1052 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
1053 if (bch2_crc_cmp(op->crc.csum, csum))
1056 ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
1057 op->crc.csum_type = 0;
1058 op->crc.csum = (struct bch_csum) { 0, 0 };
1062 static enum prep_encoded_ret {
1065 PREP_ENCODED_CHECKSUM_ERR,
1066 PREP_ENCODED_DO_WRITE,
1067 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
1069 struct bch_fs *c = op->c;
1070 struct bio *bio = &op->wbio.bio;
1072 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
1073 return PREP_ENCODED_OK;
1075 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
1077 /* Can we just write the entire extent as is? */
1078 if (op->crc.uncompressed_size == op->crc.live_size &&
1079 op->crc.compressed_size <= wp->sectors_free &&
1080 (op->crc.compression_type == op->compression_type ||
1081 op->incompressible)) {
1082 if (!crc_is_compressed(op->crc) &&
1083 op->csum_type != op->crc.csum_type &&
1084 bch2_write_rechecksum(c, op, op->csum_type))
1085 return PREP_ENCODED_CHECKSUM_ERR;
1087 return PREP_ENCODED_DO_WRITE;
1091 * If the data is compressed and we couldn't write the entire extent as
1092 * is, we have to decompress it:
1094 if (crc_is_compressed(op->crc)) {
1095 struct bch_csum csum;
1097 if (bch2_write_decrypt(op))
1098 return PREP_ENCODED_CHECKSUM_ERR;
1100 /* Last point we can still verify checksum: */
1101 csum = bch2_checksum_bio(c, op->crc.csum_type,
1102 extent_nonce(op->version, op->crc),
1104 if (bch2_crc_cmp(op->crc.csum, csum))
1105 return PREP_ENCODED_CHECKSUM_ERR;
1107 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
1108 return PREP_ENCODED_ERR;
1112 * No longer have compressed data after this point - data might be
1117 * If the data is checksummed and we're only writing a subset,
1118 * rechecksum and adjust bio to point to currently live data:
1120 if ((op->crc.live_size != op->crc.uncompressed_size ||
1121 op->crc.csum_type != op->csum_type) &&
1122 bch2_write_rechecksum(c, op, op->csum_type))
1123 return PREP_ENCODED_CHECKSUM_ERR;
1126 * If we want to compress the data, it has to be decrypted:
1128 if ((op->compression_type ||
1129 bch2_csum_type_is_encryption(op->crc.csum_type) !=
1130 bch2_csum_type_is_encryption(op->csum_type)) &&
1131 bch2_write_decrypt(op))
1132 return PREP_ENCODED_CHECKSUM_ERR;
1134 return PREP_ENCODED_OK;
1137 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
1140 struct bch_fs *c = op->c;
1141 struct bio *src = &op->wbio.bio, *dst = src;
1142 struct bvec_iter saved_iter;
1144 unsigned total_output = 0, total_input = 0;
1145 bool bounce = false;
1146 bool page_alloc_failed = false;
1149 BUG_ON(!bio_sectors(src));
1151 ec_buf = bch2_writepoint_ec_buf(c, wp);
1153 switch (bch2_write_prep_encoded_data(op, wp)) {
1154 case PREP_ENCODED_OK:
1156 case PREP_ENCODED_ERR:
1159 case PREP_ENCODED_CHECKSUM_ERR:
1161 case PREP_ENCODED_DO_WRITE:
1162 /* XXX look for bug here */
1164 dst = bch2_write_bio_alloc(c, wp, src,
1167 bio_copy_data(dst, src);
1170 init_append_extent(op, wp, op->version, op->crc);
1175 op->compression_type ||
1177 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
1178 (bch2_csum_type_is_encryption(op->csum_type) &&
1179 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
1180 dst = bch2_write_bio_alloc(c, wp, src,
1186 saved_iter = dst->bi_iter;
1189 struct bch_extent_crc_unpacked crc = { 0 };
1190 struct bversion version = op->version;
1191 size_t dst_len, src_len;
1193 if (page_alloc_failed &&
1194 dst->bi_iter.bi_size < (wp->sectors_free << 9) &&
1195 dst->bi_iter.bi_size < c->opts.encoded_extent_max)
1198 BUG_ON(op->compression_type &&
1199 (op->flags & BCH_WRITE_DATA_ENCODED) &&
1200 bch2_csum_type_is_encryption(op->crc.csum_type));
1201 BUG_ON(op->compression_type && !bounce);
1203 crc.compression_type = op->incompressible
1204 ? BCH_COMPRESSION_TYPE_incompressible
1205 : op->compression_type
1206 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
1207 op->compression_type)
1209 if (!crc_is_compressed(crc)) {
1210 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
1211 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
1214 dst_len = min_t(unsigned, dst_len,
1215 c->opts.encoded_extent_max);
1218 swap(dst->bi_iter.bi_size, dst_len);
1219 bio_copy_data(dst, src);
1220 swap(dst->bi_iter.bi_size, dst_len);
1226 BUG_ON(!src_len || !dst_len);
1228 if (bch2_csum_type_is_encryption(op->csum_type)) {
1229 if (bversion_zero(version)) {
1230 version.lo = atomic64_inc_return(&c->key_version);
1232 crc.nonce = op->nonce;
1233 op->nonce += src_len >> 9;
1237 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1238 !crc_is_compressed(crc) &&
1239 bch2_csum_type_is_encryption(op->crc.csum_type) ==
1240 bch2_csum_type_is_encryption(op->csum_type)) {
1241 u8 compression_type = crc.compression_type;
1242 u16 nonce = crc.nonce;
1244 * Note: when we're using rechecksum(), we need to be
1245 * checksumming @src because it has all the data our
1246 * existing checksum covers - if we bounced (because we
1247 * were trying to compress), @dst will only have the
1248 * part of the data the new checksum will cover.
1250 * But normally we want to be checksumming post bounce,
1251 * because part of the reason for bouncing is so the
1252 * data can't be modified (by userspace) while it's in
1255 if (bch2_rechecksum_bio(c, src, version, op->crc,
1258 bio_sectors(src) - (src_len >> 9),
1262 * rchecksum_bio sets compression_type on crc from op->crc,
1263 * this isn't always correct as sometimes we're changing
1264 * an extent from uncompressed to incompressible.
1266 crc.compression_type = compression_type;
1269 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1270 bch2_rechecksum_bio(c, src, version, op->crc,
1273 bio_sectors(src) - (src_len >> 9),
1277 crc.compressed_size = dst_len >> 9;
1278 crc.uncompressed_size = src_len >> 9;
1279 crc.live_size = src_len >> 9;
1281 swap(dst->bi_iter.bi_size, dst_len);
1282 ret = bch2_encrypt_bio(c, op->csum_type,
1283 extent_nonce(version, crc), dst);
1287 crc.csum = bch2_checksum_bio(c, op->csum_type,
1288 extent_nonce(version, crc), dst);
1289 crc.csum_type = op->csum_type;
1290 swap(dst->bi_iter.bi_size, dst_len);
1293 init_append_extent(op, wp, version, crc);
1296 bio_advance(dst, dst_len);
1297 bio_advance(src, src_len);
1298 total_output += dst_len;
1299 total_input += src_len;
1300 } while (dst->bi_iter.bi_size &&
1301 src->bi_iter.bi_size &&
1303 !bch2_keylist_realloc(&op->insert_keys,
1305 ARRAY_SIZE(op->inline_keys),
1306 BKEY_EXTENT_U64s_MAX));
1308 more = src->bi_iter.bi_size != 0;
1310 dst->bi_iter = saved_iter;
1312 if (dst == src && more) {
1313 BUG_ON(total_output != total_input);
1315 dst = bio_split(src, total_input >> 9,
1316 GFP_NOIO, &c->bio_write);
1317 wbio_init(dst)->put_bio = true;
1318 /* copy WRITE_SYNC flag */
1319 dst->bi_opf = src->bi_opf;
1322 dst->bi_iter.bi_size = total_output;
1327 bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1330 if (to_wbio(dst)->bounce)
1331 bch2_bio_free_pages_pool(c, dst);
1332 if (to_wbio(dst)->put_bio)
1338 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1341 struct bch_fs *c = op->c;
1342 struct bkey_s_c_extent e;
1343 struct extent_ptr_decoded p;
1344 const union bch_extent_entry *entry;
1345 unsigned replicas = 0;
1347 if (k.k->type != KEY_TYPE_extent)
1350 e = bkey_s_c_to_extent(k);
1351 extent_for_each_ptr_decode(e, p, entry) {
1352 if (p.crc.csum_type ||
1353 crc_is_compressed(p.crc) ||
1357 replicas += bch2_extent_ptr_durability(c, &p);
1360 return replicas >= op->opts.data_replicas;
1363 static inline void bch2_nocow_write_unlock(struct bch_write_op *op)
1365 struct bch_fs *c = op->c;
1366 const struct bch_extent_ptr *ptr;
1369 for_each_keylist_key(&op->insert_keys, k) {
1370 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
1372 bkey_for_each_ptr(ptrs, ptr)
1373 bch2_bucket_nocow_unlock(&c->nocow_locks,
1374 PTR_BUCKET_POS(c, ptr),
1375 BUCKET_NOCOW_LOCK_UPDATE);
1379 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1380 struct btree_iter *iter,
1381 struct bkey_i *orig,
1386 struct bkey_ptrs ptrs;
1387 struct bch_extent_ptr *ptr;
1390 if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1395 new = bch2_bkey_make_mut(trans, k);
1396 ret = PTR_ERR_OR_ZERO(new);
1400 bch2_cut_front(bkey_start_pos(&orig->k), new);
1401 bch2_cut_back(orig->k.p, new);
1403 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1404 bkey_for_each_ptr(ptrs, ptr)
1408 * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1409 * that was done when we kicked off the write, and here it's important
1410 * that we update the extent that we wrote to - even if a snapshot has
1411 * since been created. The write is still outstanding, so we're ok
1412 * w.r.t. snapshot atomicity:
1414 return bch2_extent_update_i_size_sectors(trans, iter,
1415 min(new->k.p.offset << 9, new_i_size), 0) ?:
1416 bch2_trans_update(trans, iter, new,
1417 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1420 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1422 struct bch_fs *c = op->c;
1423 struct btree_trans trans;
1424 struct btree_iter iter;
1425 struct bkey_i *orig;
1429 bch2_trans_init(&trans, c, 0, 0);
1431 for_each_keylist_key(&op->insert_keys, orig) {
1432 ret = for_each_btree_key_upto_commit(&trans, iter, BTREE_ID_extents,
1433 bkey_start_pos(&orig->k), orig->k.p,
1434 BTREE_ITER_INTENT, k,
1435 NULL, NULL, BTREE_INSERT_NOFAIL, ({
1436 bch2_nocow_write_convert_one_unwritten(&trans, &iter, orig, k, op->new_i_size);
1439 if (ret && !bch2_err_matches(ret, EROFS)) {
1440 struct bkey_i *k = bch2_keylist_front(&op->insert_keys);
1442 bch_err_inum_offset_ratelimited(c,
1443 k->k.p.inode, k->k.p.offset << 9,
1444 "write error while doing btree update: %s",
1454 bch2_trans_exit(&trans);
1457 static void __bch2_nocow_write_done(struct bch_write_op *op)
1459 bch2_nocow_write_unlock(op);
1461 if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1463 } else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1464 bch2_nocow_write_convert_unwritten(op);
1467 static void bch2_nocow_write_done(struct closure *cl)
1469 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1471 __bch2_nocow_write_done(op);
1472 bch2_write_done(cl);
1475 static void bch2_nocow_write(struct bch_write_op *op)
1477 struct bch_fs *c = op->c;
1478 struct btree_trans trans;
1479 struct btree_iter iter;
1481 struct bkey_ptrs_c ptrs;
1482 const struct bch_extent_ptr *ptr, *ptr2;
1486 struct nocow_lock_bucket *l;
1487 } buckets[BCH_REPLICAS_MAX];
1488 unsigned nr_buckets = 0;
1492 if (op->flags & BCH_WRITE_MOVE)
1495 bch2_trans_init(&trans, c, 0, 0);
1497 bch2_trans_begin(&trans);
1499 ret = bch2_subvolume_get_snapshot(&trans, op->subvol, &snapshot);
1503 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
1504 SPOS(op->pos.inode, op->pos.offset, snapshot),
1507 struct bio *bio = &op->wbio.bio;
1511 k = bch2_btree_iter_peek_slot(&iter);
1516 /* fall back to normal cow write path? */
1517 if (unlikely(k.k->p.snapshot != snapshot ||
1518 !bch2_extent_is_writeable(op, k)))
1521 if (bch2_keylist_realloc(&op->insert_keys,
1523 ARRAY_SIZE(op->inline_keys),
1527 /* Get iorefs before dropping btree locks: */
1528 ptrs = bch2_bkey_ptrs_c(k);
1529 bkey_for_each_ptr(ptrs, ptr) {
1530 buckets[nr_buckets].b = PTR_BUCKET_POS(c, ptr);
1531 buckets[nr_buckets].gen = ptr->gen;
1532 buckets[nr_buckets].l =
1533 bucket_nocow_lock(&c->nocow_locks,
1534 bucket_to_u64(buckets[nr_buckets].b));
1536 prefetch(buckets[nr_buckets].l);
1539 if (unlikely(!bch2_dev_get_ioref(bch_dev_bkey_exists(c, ptr->dev), WRITE)))
1543 op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1546 /* Unlock before taking nocow locks, doing IO: */
1547 bkey_reassemble(op->insert_keys.top, k);
1548 bch2_trans_unlock(&trans);
1550 bch2_cut_front(op->pos, op->insert_keys.top);
1551 if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1552 bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1554 for (i = 0; i < nr_buckets; i++) {
1555 struct bch_dev *ca = bch_dev_bkey_exists(c, buckets[i].b.inode);
1556 struct nocow_lock_bucket *l = buckets[i].l;
1559 __bch2_bucket_nocow_lock(&c->nocow_locks, l,
1560 bucket_to_u64(buckets[i].b),
1561 BUCKET_NOCOW_LOCK_UPDATE);
1564 stale = gen_after(*bucket_gen(ca, buckets[i].b.offset), buckets[i].gen);
1567 if (unlikely(stale))
1568 goto err_bucket_stale;
1571 bio = &op->wbio.bio;
1572 if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1573 bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1574 GFP_KERNEL, &c->bio_write);
1575 wbio_init(bio)->put_bio = true;
1576 bio->bi_opf = op->wbio.bio.bi_opf;
1578 op->flags |= BCH_WRITE_DONE;
1581 op->pos.offset += bio_sectors(bio);
1582 op->written += bio_sectors(bio);
1584 bio->bi_end_io = bch2_write_endio;
1585 bio->bi_private = &op->cl;
1586 bio->bi_opf |= REQ_OP_WRITE;
1587 closure_get(&op->cl);
1588 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1589 op->insert_keys.top, true);
1591 bch2_keylist_push(&op->insert_keys);
1592 if (op->flags & BCH_WRITE_DONE)
1594 bch2_btree_iter_advance(&iter);
1597 bch2_trans_iter_exit(&trans, &iter);
1599 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1603 bch_err_inum_offset_ratelimited(c,
1605 op->pos.offset << 9,
1606 "%s: btree lookup error %s",
1607 __func__, bch2_err_str(ret));
1609 op->flags |= BCH_WRITE_DONE;
1612 bch2_trans_exit(&trans);
1614 /* fallback to cow write path? */
1615 if (!(op->flags & BCH_WRITE_DONE)) {
1616 closure_sync(&op->cl);
1617 __bch2_nocow_write_done(op);
1618 op->insert_keys.top = op->insert_keys.keys;
1619 } else if (op->flags & BCH_WRITE_SYNC) {
1620 closure_sync(&op->cl);
1621 bch2_nocow_write_done(&op->cl);
1625 * needs to run out of process context because ei_quota_lock is
1628 continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1632 bkey_for_each_ptr(ptrs, ptr2) {
1636 percpu_ref_put(&bch_dev_bkey_exists(c, ptr2->dev)->io_ref);
1639 /* Fall back to COW path: */
1643 bch2_bucket_nocow_unlock(&c->nocow_locks,
1645 BUCKET_NOCOW_LOCK_UPDATE);
1647 bkey_for_each_ptr(ptrs, ptr2)
1648 percpu_ref_put(&bch_dev_bkey_exists(c, ptr2->dev)->io_ref);
1650 /* We can retry this: */
1651 ret = BCH_ERR_transaction_restart;
1655 static void __bch2_write(struct bch_write_op *op)
1657 struct bch_fs *c = op->c;
1658 struct write_point *wp = NULL;
1659 struct bio *bio = NULL;
1660 unsigned nofs_flags;
1663 nofs_flags = memalloc_nofs_save();
1665 if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1666 bch2_nocow_write(op);
1667 if (op->flags & BCH_WRITE_DONE)
1668 goto out_nofs_restore;
1671 memset(&op->failed, 0, sizeof(op->failed));
1672 op->btree_update_ready = false;
1675 struct bkey_i *key_to_write;
1676 unsigned key_to_write_offset = op->insert_keys.top_p -
1677 op->insert_keys.keys_p;
1679 /* +1 for possible cache device: */
1680 if (op->open_buckets.nr + op->nr_replicas + 1 >
1681 ARRAY_SIZE(op->open_buckets.v))
1684 if (bch2_keylist_realloc(&op->insert_keys,
1686 ARRAY_SIZE(op->inline_keys),
1687 BKEY_EXTENT_U64s_MAX))
1691 * The copygc thread is now global, which means it's no longer
1692 * freeing up space on specific disks, which means that
1693 * allocations for specific disks may hang arbitrarily long:
1695 ret = bch2_trans_do(c, NULL, NULL, 0,
1696 bch2_alloc_sectors_start_trans(&trans,
1698 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1702 op->nr_replicas_required,
1705 (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1706 BCH_WRITE_ONLY_SPECIFIED_DEVS))
1707 ? NULL : &op->cl, &wp));
1708 if (unlikely(ret)) {
1709 if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1717 bch2_open_bucket_get(c, wp, &op->open_buckets);
1718 ret = bch2_write_extent(op, wp, &bio);
1720 bch2_alloc_sectors_done_inlined(c, wp);
1723 op->flags |= BCH_WRITE_DONE;
1731 bio->bi_end_io = bch2_write_endio;
1732 bio->bi_private = &op->cl;
1733 bio->bi_opf |= REQ_OP_WRITE;
1735 closure_get(bio->bi_private);
1737 key_to_write = (void *) (op->insert_keys.keys_p +
1738 key_to_write_offset);
1740 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1741 key_to_write, false);
1747 * If we're running asynchronously, wne may still want to block
1748 * synchronously here if we weren't able to submit all of the IO at
1749 * once, as that signals backpressure to the caller.
1751 if ((op->flags & BCH_WRITE_SYNC) ||
1752 (!(op->flags & BCH_WRITE_DONE) &&
1753 !(op->flags & BCH_WRITE_IN_WORKER))) {
1754 closure_sync(&op->cl);
1755 __bch2_write_index(op);
1757 if (!(op->flags & BCH_WRITE_DONE))
1759 bch2_write_done(&op->cl);
1761 bch2_write_queue(op, wp);
1762 continue_at(&op->cl, bch2_write_index, NULL);
1765 memalloc_nofs_restore(nofs_flags);
1768 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1770 struct bio *bio = &op->wbio.bio;
1771 struct bvec_iter iter;
1772 struct bkey_i_inline_data *id;
1776 op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1777 op->flags |= BCH_WRITE_DONE;
1779 bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1781 ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1782 ARRAY_SIZE(op->inline_keys),
1783 BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1789 sectors = bio_sectors(bio);
1790 op->pos.offset += sectors;
1792 id = bkey_inline_data_init(op->insert_keys.top);
1794 id->k.version = op->version;
1795 id->k.size = sectors;
1797 iter = bio->bi_iter;
1798 iter.bi_size = data_len;
1799 memcpy_from_bio(id->v.data, bio, iter);
1801 while (data_len & 7)
1802 id->v.data[data_len++] = '\0';
1803 set_bkey_val_bytes(&id->k, data_len);
1804 bch2_keylist_push(&op->insert_keys);
1806 __bch2_write_index(op);
1808 bch2_write_done(&op->cl);
1812 * bch_write - handle a write to a cache device or flash only volume
1814 * This is the starting point for any data to end up in a cache device; it could
1815 * be from a normal write, or a writeback write, or a write to a flash only
1816 * volume - it's also used by the moving garbage collector to compact data in
1817 * mostly empty buckets.
1819 * It first writes the data to the cache, creating a list of keys to be inserted
1820 * (if the data won't fit in a single open bucket, there will be multiple keys);
1821 * after the data is written it calls bch_journal, and after the keys have been
1822 * added to the next journal write they're inserted into the btree.
1824 * If op->discard is true, instead of inserting the data it invalidates the
1825 * region of the cache represented by op->bio and op->inode.
1827 void bch2_write(struct closure *cl)
1829 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1830 struct bio *bio = &op->wbio.bio;
1831 struct bch_fs *c = op->c;
1834 EBUG_ON(op->cl.parent);
1835 BUG_ON(!op->nr_replicas);
1836 BUG_ON(!op->write_point.v);
1837 BUG_ON(bkey_eq(op->pos, POS_MAX));
1839 op->start_time = local_clock();
1840 bch2_keylist_init(&op->insert_keys, op->inline_keys);
1841 wbio_init(bio)->put_bio = false;
1843 if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1844 bch_err_inum_offset_ratelimited(c,
1846 op->pos.offset << 9,
1847 "misaligned write");
1852 if (c->opts.nochanges ||
1853 !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1854 op->error = -BCH_ERR_erofs_no_writes;
1858 this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1859 bch2_increment_clock(c, bio_sectors(bio), WRITE);
1861 data_len = min_t(u64, bio->bi_iter.bi_size,
1862 op->new_i_size - (op->pos.offset << 9));
1864 if (c->opts.inline_data &&
1865 data_len <= min(block_bytes(c) / 2, 1024U)) {
1866 bch2_write_data_inline(op, data_len);
1873 bch2_disk_reservation_put(c, &op->res);
1875 closure_debug_destroy(&op->cl);
1880 /* Cache promotion on read */
1883 struct rcu_head rcu;
1886 struct rhash_head hash;
1889 struct data_update write;
1890 struct bio_vec bi_inline_vecs[0]; /* must be last */
1893 static const struct rhashtable_params bch_promote_params = {
1894 .head_offset = offsetof(struct promote_op, hash),
1895 .key_offset = offsetof(struct promote_op, pos),
1896 .key_len = sizeof(struct bpos),
1899 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1901 struct bch_io_opts opts,
1904 if (!(flags & BCH_READ_MAY_PROMOTE))
1907 if (!opts.promote_target)
1910 if (bch2_bkey_has_target(c, k, opts.promote_target))
1913 if (bkey_extent_is_unwritten(k))
1916 if (bch2_target_congested(c, opts.promote_target)) {
1917 /* XXX trace this */
1921 if (rhashtable_lookup_fast(&c->promote_table, &pos,
1922 bch_promote_params))
1928 static void promote_free(struct bch_fs *c, struct promote_op *op)
1932 bch2_data_update_exit(&op->write);
1934 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1935 bch_promote_params);
1937 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
1941 static void promote_done(struct bch_write_op *wop)
1943 struct promote_op *op =
1944 container_of(wop, struct promote_op, write.op);
1945 struct bch_fs *c = op->write.op.c;
1947 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1949 promote_free(c, op);
1952 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1954 struct bio *bio = &op->write.op.wbio.bio;
1956 trace_and_count(op->write.op.c, read_promote, &rbio->bio);
1958 /* we now own pages: */
1959 BUG_ON(!rbio->bounce);
1960 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1962 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1963 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1964 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1966 bch2_data_update_read_done(&op->write, rbio->pick.crc);
1969 static struct promote_op *__promote_alloc(struct btree_trans *trans,
1970 enum btree_id btree_id,
1973 struct extent_ptr_decoded *pick,
1974 struct bch_io_opts opts,
1976 struct bch_read_bio **rbio)
1978 struct bch_fs *c = trans->c;
1979 struct promote_op *op = NULL;
1981 unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1984 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
1987 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1991 op->start_time = local_clock();
1995 * We don't use the mempool here because extents that aren't
1996 * checksummed or compressed can be too big for the mempool:
1998 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1999 sizeof(struct bio_vec) * pages,
2004 rbio_init(&(*rbio)->bio, opts);
2005 bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
2007 if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
2011 (*rbio)->bounce = true;
2012 (*rbio)->split = true;
2013 (*rbio)->kmalloc = true;
2015 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
2016 bch_promote_params))
2019 bio = &op->write.op.wbio.bio;
2020 bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
2022 ret = bch2_data_update_init(trans, NULL, &op->write,
2023 writepoint_hashed((unsigned long) current),
2025 (struct data_update_opts) {
2026 .target = opts.promote_target,
2027 .extra_replicas = 1,
2028 .write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
2031 if (ret == -BCH_ERR_nocow_lock_blocked) {
2032 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
2033 bch_promote_params);
2039 op->write.op.end_io = promote_done;
2044 bio_free_pages(&(*rbio)->bio);
2048 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
2053 static struct promote_op *promote_alloc(struct btree_trans *trans,
2054 struct bvec_iter iter,
2056 struct extent_ptr_decoded *pick,
2057 struct bch_io_opts opts,
2059 struct bch_read_bio **rbio,
2063 struct bch_fs *c = trans->c;
2064 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
2065 /* data might have to be decompressed in the write path: */
2066 unsigned sectors = promote_full
2067 ? max(pick->crc.compressed_size, pick->crc.live_size)
2068 : bvec_iter_sectors(iter);
2069 struct bpos pos = promote_full
2070 ? bkey_start_pos(k.k)
2071 : POS(k.k->p.inode, iter.bi_sector);
2072 struct promote_op *promote;
2074 if (!should_promote(c, k, pos, opts, flags))
2077 promote = __promote_alloc(trans,
2078 k.k->type == KEY_TYPE_reflink_v
2081 k, pos, pick, opts, sectors, rbio);
2086 *read_full = promote_full;
2092 #define READ_RETRY_AVOID 1
2093 #define READ_RETRY 2
2098 RBIO_CONTEXT_HIGHPRI,
2099 RBIO_CONTEXT_UNBOUND,
2102 static inline struct bch_read_bio *
2103 bch2_rbio_parent(struct bch_read_bio *rbio)
2105 return rbio->split ? rbio->parent : rbio;
2109 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
2110 enum rbio_context context,
2111 struct workqueue_struct *wq)
2113 if (context <= rbio->context) {
2116 rbio->work.func = fn;
2117 rbio->context = context;
2118 queue_work(wq, &rbio->work);
2122 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
2124 BUG_ON(rbio->bounce && !rbio->split);
2127 promote_free(rbio->c, rbio->promote);
2128 rbio->promote = NULL;
2131 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
2134 struct bch_read_bio *parent = rbio->parent;
2139 bio_put(&rbio->bio);
2148 * Only called on a top level bch_read_bio to complete an entire read request,
2151 static void bch2_rbio_done(struct bch_read_bio *rbio)
2153 if (rbio->start_time)
2154 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
2156 bio_endio(&rbio->bio);
2159 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
2160 struct bvec_iter bvec_iter,
2161 struct bch_io_failures *failed,
2164 struct btree_trans trans;
2165 struct btree_iter iter;
2170 flags &= ~BCH_READ_LAST_FRAGMENT;
2171 flags |= BCH_READ_MUST_CLONE;
2173 bch2_bkey_buf_init(&sk);
2174 bch2_trans_init(&trans, c, 0, 0);
2176 bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
2177 rbio->read_pos, BTREE_ITER_SLOTS);
2179 rbio->bio.bi_status = 0;
2181 k = bch2_btree_iter_peek_slot(&iter);
2185 bch2_bkey_buf_reassemble(&sk, c, k);
2186 k = bkey_i_to_s_c(sk.k);
2187 bch2_trans_unlock(&trans);
2189 if (!bch2_bkey_matches_ptr(c, k,
2191 rbio->data_pos.offset -
2192 rbio->pick.crc.offset)) {
2193 /* extent we wanted to read no longer exists: */
2198 ret = __bch2_read_extent(&trans, rbio, bvec_iter,
2201 k, 0, failed, flags);
2202 if (ret == READ_RETRY)
2207 bch2_rbio_done(rbio);
2208 bch2_trans_iter_exit(&trans, &iter);
2209 bch2_trans_exit(&trans);
2210 bch2_bkey_buf_exit(&sk, c);
2213 rbio->bio.bi_status = BLK_STS_IOERR;
2217 static void bch2_rbio_retry(struct work_struct *work)
2219 struct bch_read_bio *rbio =
2220 container_of(work, struct bch_read_bio, work);
2221 struct bch_fs *c = rbio->c;
2222 struct bvec_iter iter = rbio->bvec_iter;
2223 unsigned flags = rbio->flags;
2224 subvol_inum inum = {
2225 .subvol = rbio->subvol,
2226 .inum = rbio->read_pos.inode,
2228 struct bch_io_failures failed = { .nr = 0 };
2230 trace_and_count(c, read_retry, &rbio->bio);
2232 if (rbio->retry == READ_RETRY_AVOID)
2233 bch2_mark_io_failure(&failed, &rbio->pick);
2235 rbio->bio.bi_status = 0;
2237 rbio = bch2_rbio_free(rbio);
2239 flags |= BCH_READ_IN_RETRY;
2240 flags &= ~BCH_READ_MAY_PROMOTE;
2242 if (flags & BCH_READ_NODECODE) {
2243 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
2245 flags &= ~BCH_READ_LAST_FRAGMENT;
2246 flags |= BCH_READ_MUST_CLONE;
2248 __bch2_read(c, rbio, iter, inum, &failed, flags);
2252 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
2255 rbio->retry = retry;
2257 if (rbio->flags & BCH_READ_IN_RETRY)
2260 if (retry == READ_ERR) {
2261 rbio = bch2_rbio_free(rbio);
2263 rbio->bio.bi_status = error;
2264 bch2_rbio_done(rbio);
2266 bch2_rbio_punt(rbio, bch2_rbio_retry,
2267 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
2271 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
2272 struct bch_read_bio *rbio)
2274 struct bch_fs *c = rbio->c;
2275 u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
2276 struct bch_extent_crc_unpacked new_crc;
2277 struct btree_iter iter;
2282 if (crc_is_compressed(rbio->pick.crc))
2285 bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
2286 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
2287 k = bch2_btree_iter_peek_slot(&iter);
2288 if ((ret = bkey_err(k)))
2291 if (bversion_cmp(k.k->version, rbio->version) ||
2292 !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
2295 /* Extent was merged? */
2296 if (bkey_start_offset(k.k) < data_offset ||
2297 k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
2300 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
2301 rbio->pick.crc, NULL, &new_crc,
2302 bkey_start_offset(k.k) - data_offset, k.k->size,
2303 rbio->pick.crc.csum_type)) {
2304 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
2310 * going to be temporarily appending another checksum entry:
2312 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
2313 sizeof(struct bch_extent_crc128));
2314 if ((ret = PTR_ERR_OR_ZERO(new)))
2317 bkey_reassemble(new, k);
2319 if (!bch2_bkey_narrow_crcs(new, new_crc))
2322 ret = bch2_trans_update(trans, &iter, new,
2323 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
2325 bch2_trans_iter_exit(trans, &iter);
2329 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
2331 bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
2332 __bch2_rbio_narrow_crcs(&trans, rbio));
2335 /* Inner part that may run in process context */
2336 static void __bch2_read_endio(struct work_struct *work)
2338 struct bch_read_bio *rbio =
2339 container_of(work, struct bch_read_bio, work);
2340 struct bch_fs *c = rbio->c;
2341 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
2342 struct bio *src = &rbio->bio;
2343 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
2344 struct bvec_iter dst_iter = rbio->bvec_iter;
2345 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
2346 struct nonce nonce = extent_nonce(rbio->version, crc);
2347 unsigned nofs_flags;
2348 struct bch_csum csum;
2351 nofs_flags = memalloc_nofs_save();
2353 /* Reset iterator for checksumming and copying bounced data: */
2355 src->bi_iter.bi_size = crc.compressed_size << 9;
2356 src->bi_iter.bi_idx = 0;
2357 src->bi_iter.bi_bvec_done = 0;
2359 src->bi_iter = rbio->bvec_iter;
2362 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
2363 if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
2368 * We need to rework the narrow_crcs path to deliver the read completion
2369 * first, and then punt to a different workqueue, otherwise we're
2370 * holding up reads while doing btree updates which is bad for memory
2373 if (unlikely(rbio->narrow_crcs))
2374 bch2_rbio_narrow_crcs(rbio);
2376 if (rbio->flags & BCH_READ_NODECODE)
2379 /* Adjust crc to point to subset of data we want: */
2380 crc.offset += rbio->offset_into_extent;
2381 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
2383 if (crc_is_compressed(crc)) {
2384 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
2388 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
2389 goto decompression_err;
2391 /* don't need to decrypt the entire bio: */
2392 nonce = nonce_add(nonce, crc.offset << 9);
2393 bio_advance(src, crc.offset << 9);
2395 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
2396 src->bi_iter.bi_size = dst_iter.bi_size;
2398 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
2403 struct bvec_iter src_iter = src->bi_iter;
2404 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
2408 if (rbio->promote) {
2410 * Re encrypt data we decrypted, so it's consistent with
2413 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
2417 promote_start(rbio->promote, rbio);
2418 rbio->promote = NULL;
2421 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
2422 rbio = bch2_rbio_free(rbio);
2423 bch2_rbio_done(rbio);
2426 memalloc_nofs_restore(nofs_flags);
2430 * Checksum error: if the bio wasn't bounced, we may have been
2431 * reading into buffers owned by userspace (that userspace can
2432 * scribble over) - retry the read, bouncing it this time:
2434 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
2435 rbio->flags |= BCH_READ_MUST_BOUNCE;
2436 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
2440 bch_err_inum_offset_ratelimited(ca,
2441 rbio->read_pos.inode,
2442 rbio->read_pos.offset << 9,
2443 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
2444 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
2445 csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
2447 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2450 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
2451 rbio->read_pos.offset << 9,
2452 "decompression error");
2453 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
2456 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
2457 rbio->read_pos.offset << 9,
2459 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
2463 static void bch2_read_endio(struct bio *bio)
2465 struct bch_read_bio *rbio =
2466 container_of(bio, struct bch_read_bio, bio);
2467 struct bch_fs *c = rbio->c;
2468 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
2469 struct workqueue_struct *wq = NULL;
2470 enum rbio_context context = RBIO_CONTEXT_NULL;
2472 if (rbio->have_ioref) {
2473 bch2_latency_acct(ca, rbio->submit_time, READ);
2474 percpu_ref_put(&ca->io_ref);
2478 rbio->bio.bi_end_io = rbio->end_io;
2480 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
2481 rbio->read_pos.inode,
2482 rbio->read_pos.offset,
2483 "data read error: %s",
2484 bch2_blk_status_to_str(bio->bi_status))) {
2485 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
2489 if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
2490 ptr_stale(ca, &rbio->pick.ptr)) {
2491 trace_and_count(c, read_reuse_race, &rbio->bio);
2493 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
2494 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
2496 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
2500 if (rbio->narrow_crcs ||
2502 crc_is_compressed(rbio->pick.crc) ||
2503 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
2504 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
2505 else if (rbio->pick.crc.csum_type)
2506 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
2508 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
2511 int __bch2_read_indirect_extent(struct btree_trans *trans,
2512 unsigned *offset_into_extent,
2513 struct bkey_buf *orig_k)
2515 struct btree_iter iter;
2520 reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
2521 *offset_into_extent;
2523 bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
2524 POS(0, reflink_offset),
2526 k = bch2_btree_iter_peek_slot(&iter);
2531 if (k.k->type != KEY_TYPE_reflink_v &&
2532 k.k->type != KEY_TYPE_indirect_inline_data) {
2533 bch_err_inum_offset_ratelimited(trans->c,
2534 orig_k->k->k.p.inode,
2535 orig_k->k->k.p.offset << 9,
2536 "%llu len %u points to nonexistent indirect extent %llu",
2537 orig_k->k->k.p.offset,
2540 bch2_inconsistent_error(trans->c);
2545 *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
2546 bch2_bkey_buf_reassemble(orig_k, trans->c, k);
2548 bch2_trans_iter_exit(trans, &iter);
2552 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
2554 struct bch_extent_ptr ptr)
2556 struct bch_fs *c = trans->c;
2557 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
2558 struct btree_iter iter;
2559 struct printbuf buf = PRINTBUF;
2562 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
2563 PTR_BUCKET_POS(c, &ptr),
2566 prt_printf(&buf, "Attempting to read from stale dirty pointer:");
2567 printbuf_indent_add(&buf, 2);
2570 bch2_bkey_val_to_text(&buf, c, k);
2573 prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
2575 ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
2578 bch2_bkey_val_to_text(&buf, c, k);
2581 bch2_fs_inconsistent(c, "%s", buf.buf);
2583 bch2_trans_iter_exit(trans, &iter);
2584 printbuf_exit(&buf);
2587 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2588 struct bvec_iter iter, struct bpos read_pos,
2589 enum btree_id data_btree, struct bkey_s_c k,
2590 unsigned offset_into_extent,
2591 struct bch_io_failures *failed, unsigned flags)
2593 struct bch_fs *c = trans->c;
2594 struct extent_ptr_decoded pick;
2595 struct bch_read_bio *rbio = NULL;
2596 struct bch_dev *ca = NULL;
2597 struct promote_op *promote = NULL;
2598 bool bounce = false, read_full = false, narrow_crcs = false;
2599 struct bpos data_pos = bkey_start_pos(k.k);
2602 if (bkey_extent_is_inline_data(k.k)) {
2603 unsigned bytes = min_t(unsigned, iter.bi_size,
2604 bkey_inline_data_bytes(k.k));
2606 swap(iter.bi_size, bytes);
2607 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2608 swap(iter.bi_size, bytes);
2609 bio_advance_iter(&orig->bio, &iter, bytes);
2610 zero_fill_bio_iter(&orig->bio, iter);
2614 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2616 /* hole or reservation - just zero fill: */
2621 bch_err_inum_offset_ratelimited(c,
2622 read_pos.inode, read_pos.offset << 9,
2623 "no device to read from");
2627 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2630 * Stale dirty pointers are treated as IO errors, but @failed isn't
2631 * allocated unless we're in the retry path - so if we're not in the
2632 * retry path, don't check here, it'll be caught in bch2_read_endio()
2633 * and we'll end up in the retry path:
2635 if ((flags & BCH_READ_IN_RETRY) &&
2637 unlikely(ptr_stale(ca, &pick.ptr))) {
2638 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2639 bch2_mark_io_failure(failed, &pick);
2644 * Unlock the iterator while the btree node's lock is still in
2645 * cache, before doing the IO:
2647 bch2_trans_unlock(trans);
2649 if (flags & BCH_READ_NODECODE) {
2651 * can happen if we retry, and the extent we were going to read
2652 * has been merged in the meantime:
2654 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2657 iter.bi_size = pick.crc.compressed_size << 9;
2661 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2662 bio_flagged(&orig->bio, BIO_CHAIN))
2663 flags |= BCH_READ_MUST_CLONE;
2665 narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2666 bch2_can_narrow_extent_crcs(k, pick.crc);
2668 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2669 flags |= BCH_READ_MUST_BOUNCE;
2671 EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2673 if (crc_is_compressed(pick.crc) ||
2674 (pick.crc.csum_type != BCH_CSUM_none &&
2675 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2676 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2677 (flags & BCH_READ_USER_MAPPED)) ||
2678 (flags & BCH_READ_MUST_BOUNCE)))) {
2683 if (orig->opts.promote_target)
2684 promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
2685 &rbio, &bounce, &read_full);
2688 EBUG_ON(crc_is_compressed(pick.crc));
2689 EBUG_ON(pick.crc.csum_type &&
2690 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2691 bvec_iter_sectors(iter) != pick.crc.live_size ||
2693 offset_into_extent));
2695 data_pos.offset += offset_into_extent;
2696 pick.ptr.offset += pick.crc.offset +
2698 offset_into_extent = 0;
2699 pick.crc.compressed_size = bvec_iter_sectors(iter);
2700 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
2701 pick.crc.offset = 0;
2702 pick.crc.live_size = bvec_iter_sectors(iter);
2703 offset_into_extent = 0;
2708 * promote already allocated bounce rbio:
2709 * promote needs to allocate a bio big enough for uncompressing
2710 * data in the write path, but we're not going to use it all
2713 EBUG_ON(rbio->bio.bi_iter.bi_size <
2714 pick.crc.compressed_size << 9);
2715 rbio->bio.bi_iter.bi_size =
2716 pick.crc.compressed_size << 9;
2717 } else if (bounce) {
2718 unsigned sectors = pick.crc.compressed_size;
2720 rbio = rbio_init(bio_alloc_bioset(NULL,
2721 DIV_ROUND_UP(sectors, PAGE_SECTORS),
2724 &c->bio_read_split),
2727 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2728 rbio->bounce = true;
2730 } else if (flags & BCH_READ_MUST_CLONE) {
2732 * Have to clone if there were any splits, due to error
2733 * reporting issues (if a split errored, and retrying didn't
2734 * work, when it reports the error to its parent (us) we don't
2735 * know if the error was from our bio, and we should retry, or
2736 * from the whole bio, in which case we don't want to retry and
2739 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOIO,
2740 &c->bio_read_split),
2742 rbio->bio.bi_iter = iter;
2746 rbio->bio.bi_iter = iter;
2747 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2750 EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2753 rbio->submit_time = local_clock();
2755 rbio->parent = orig;
2757 rbio->end_io = orig->bio.bi_end_io;
2758 rbio->bvec_iter = iter;
2759 rbio->offset_into_extent= offset_into_extent;
2760 rbio->flags = flags;
2761 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2762 rbio->narrow_crcs = narrow_crcs;
2766 /* XXX: only initialize this if needed */
2767 rbio->devs_have = bch2_bkey_devs(k);
2769 rbio->subvol = orig->subvol;
2770 rbio->read_pos = read_pos;
2771 rbio->data_btree = data_btree;
2772 rbio->data_pos = data_pos;
2773 rbio->version = k.k->version;
2774 rbio->promote = promote;
2775 INIT_WORK(&rbio->work, NULL);
2777 rbio->bio.bi_opf = orig->bio.bi_opf;
2778 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2779 rbio->bio.bi_end_io = bch2_read_endio;
2782 trace_and_count(c, read_bounce, &rbio->bio);
2784 this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
2785 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2788 * If it's being moved internally, we don't want to flag it as a cache
2791 if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2792 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2793 PTR_BUCKET_NR(ca, &pick.ptr), READ);
2795 if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2796 bio_inc_remaining(&orig->bio);
2797 trace_and_count(c, read_split, &orig->bio);
2800 if (!rbio->pick.idx) {
2801 if (!rbio->have_ioref) {
2802 bch_err_inum_offset_ratelimited(c,
2804 read_pos.offset << 9,
2805 "no device to read from");
2806 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2810 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2811 bio_sectors(&rbio->bio));
2812 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2814 if (unlikely(c->opts.no_data_io)) {
2815 if (likely(!(flags & BCH_READ_IN_RETRY)))
2816 bio_endio(&rbio->bio);
2818 if (likely(!(flags & BCH_READ_IN_RETRY)))
2819 submit_bio(&rbio->bio);
2821 submit_bio_wait(&rbio->bio);
2825 * We just submitted IO which may block, we expect relock fail
2826 * events and shouldn't count them:
2828 trans->notrace_relock_fail = true;
2830 /* Attempting reconstruct read: */
2831 if (bch2_ec_read_extent(c, rbio)) {
2832 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2836 if (likely(!(flags & BCH_READ_IN_RETRY)))
2837 bio_endio(&rbio->bio);
2840 if (likely(!(flags & BCH_READ_IN_RETRY))) {
2845 rbio->context = RBIO_CONTEXT_UNBOUND;
2846 bch2_read_endio(&rbio->bio);
2849 rbio = bch2_rbio_free(rbio);
2851 if (ret == READ_RETRY_AVOID) {
2852 bch2_mark_io_failure(failed, &pick);
2863 if (flags & BCH_READ_IN_RETRY)
2866 orig->bio.bi_status = BLK_STS_IOERR;
2871 * won't normally happen in the BCH_READ_NODECODE
2872 * (bch2_move_extent()) path, but if we retry and the extent we wanted
2873 * to read no longer exists we have to signal that:
2875 if (flags & BCH_READ_NODECODE)
2878 zero_fill_bio_iter(&orig->bio, iter);
2880 if (flags & BCH_READ_LAST_FRAGMENT)
2881 bch2_rbio_done(orig);
2885 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2886 struct bvec_iter bvec_iter, subvol_inum inum,
2887 struct bch_io_failures *failed, unsigned flags)
2889 struct btree_trans trans;
2890 struct btree_iter iter;
2896 BUG_ON(flags & BCH_READ_NODECODE);
2898 bch2_bkey_buf_init(&sk);
2899 bch2_trans_init(&trans, c, 0, 0);
2901 bch2_trans_begin(&trans);
2902 iter = (struct btree_iter) { NULL };
2904 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2908 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2909 SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2912 unsigned bytes, sectors, offset_into_extent;
2913 enum btree_id data_btree = BTREE_ID_extents;
2916 * read_extent -> io_time_reset may cause a transaction restart
2917 * without returning an error, we need to check for that here:
2919 ret = bch2_trans_relock(&trans);
2923 bch2_btree_iter_set_pos(&iter,
2924 POS(inum.inum, bvec_iter.bi_sector));
2926 k = bch2_btree_iter_peek_slot(&iter);
2931 offset_into_extent = iter.pos.offset -
2932 bkey_start_offset(k.k);
2933 sectors = k.k->size - offset_into_extent;
2935 bch2_bkey_buf_reassemble(&sk, c, k);
2937 ret = bch2_read_indirect_extent(&trans, &data_btree,
2938 &offset_into_extent, &sk);
2942 k = bkey_i_to_s_c(sk.k);
2945 * With indirect extents, the amount of data to read is the min
2946 * of the original extent and the indirect extent:
2948 sectors = min(sectors, k.k->size - offset_into_extent);
2950 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2951 swap(bvec_iter.bi_size, bytes);
2953 if (bvec_iter.bi_size == bytes)
2954 flags |= BCH_READ_LAST_FRAGMENT;
2956 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2958 offset_into_extent, failed, flags);
2962 if (flags & BCH_READ_LAST_FRAGMENT)
2965 swap(bvec_iter.bi_size, bytes);
2966 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2968 ret = btree_trans_too_many_iters(&trans);
2973 bch2_trans_iter_exit(&trans, &iter);
2975 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
2976 ret == READ_RETRY ||
2977 ret == READ_RETRY_AVOID)
2980 bch2_trans_exit(&trans);
2981 bch2_bkey_buf_exit(&sk, c);
2984 bch_err_inum_offset_ratelimited(c, inum.inum,
2985 bvec_iter.bi_sector << 9,
2986 "read error %i from btree lookup", ret);
2987 rbio->bio.bi_status = BLK_STS_IOERR;
2988 bch2_rbio_done(rbio);
2992 void bch2_fs_io_exit(struct bch_fs *c)
2994 if (c->promote_table.tbl)
2995 rhashtable_destroy(&c->promote_table);
2996 mempool_exit(&c->bio_bounce_pages);
2997 bioset_exit(&c->bio_write);
2998 bioset_exit(&c->bio_read_split);
2999 bioset_exit(&c->bio_read);
3002 int bch2_fs_io_init(struct bch_fs *c)
3004 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
3005 BIOSET_NEED_BVECS) ||
3006 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
3007 BIOSET_NEED_BVECS) ||
3008 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
3009 BIOSET_NEED_BVECS) ||
3010 mempool_init_page_pool(&c->bio_bounce_pages,
3012 c->opts.btree_node_size,
3013 c->opts.encoded_extent_max) /
3015 rhashtable_init(&c->promote_table, &bch_promote_params))