1 // SPDX-License-Identifier: GPL-2.0
4 #include "bkey_methods.h"
6 #include "btree_cache.h"
8 #include "btree_iter.h"
9 #include "btree_locking.h"
10 #include "btree_update.h"
11 #include "btree_update_interior.h"
18 #include "journal_reclaim.h"
19 #include "journal_seq_blacklist.h"
24 #include <linux/sched/mm.h>
26 void bch2_btree_node_io_unlock(struct btree *b)
28 EBUG_ON(!btree_node_write_in_flight(b));
30 clear_btree_node_write_in_flight_inner(b);
31 clear_btree_node_write_in_flight(b);
32 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
35 void bch2_btree_node_io_lock(struct btree *b)
37 bch2_assert_btree_nodes_not_locked();
39 wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
40 TASK_UNINTERRUPTIBLE);
43 void __bch2_btree_node_wait_on_read(struct btree *b)
45 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
46 TASK_UNINTERRUPTIBLE);
49 void __bch2_btree_node_wait_on_write(struct btree *b)
51 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
52 TASK_UNINTERRUPTIBLE);
55 void bch2_btree_node_wait_on_read(struct btree *b)
57 bch2_assert_btree_nodes_not_locked();
59 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
60 TASK_UNINTERRUPTIBLE);
63 void bch2_btree_node_wait_on_write(struct btree *b)
65 bch2_assert_btree_nodes_not_locked();
67 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
68 TASK_UNINTERRUPTIBLE);
71 static void verify_no_dups(struct btree *b,
72 struct bkey_packed *start,
73 struct bkey_packed *end)
75 #ifdef CONFIG_BCACHEFS_DEBUG
76 struct bkey_packed *k, *p;
81 for (p = start, k = bkey_p_next(start);
83 p = k, k = bkey_p_next(k)) {
84 struct bkey l = bkey_unpack_key(b, p);
85 struct bkey r = bkey_unpack_key(b, k);
87 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
92 static void set_needs_whiteout(struct bset *i, int v)
94 struct bkey_packed *k;
96 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
97 k->needs_whiteout = v;
100 static void btree_bounce_free(struct bch_fs *c, size_t size,
101 bool used_mempool, void *p)
104 mempool_free(p, &c->btree_bounce_pool);
109 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
112 unsigned flags = memalloc_nofs_save();
115 BUG_ON(size > c->opts.btree_node_size);
117 *used_mempool = false;
118 p = kvmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
120 *used_mempool = true;
121 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
123 memalloc_nofs_restore(flags);
127 static void sort_bkey_ptrs(const struct btree *bt,
128 struct bkey_packed **ptrs, unsigned nr)
130 unsigned n = nr, a = nr / 2, b, c, d;
135 /* Heap sort: see lib/sort.c: */
140 swap(ptrs[0], ptrs[n]);
144 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
145 b = bch2_bkey_cmp_packed(bt,
147 ptrs[d]) >= 0 ? c : d;
152 bch2_bkey_cmp_packed(bt,
159 swap(ptrs[b], ptrs[c]);
164 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
166 struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
167 bool used_mempool = false;
168 size_t bytes = b->whiteout_u64s * sizeof(u64);
170 if (!b->whiteout_u64s)
173 new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
175 ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
177 for (k = unwritten_whiteouts_start(b);
178 k != unwritten_whiteouts_end(b);
182 sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
186 while (ptrs != ptrs_end) {
187 bkey_p_copy(k, *ptrs);
192 verify_no_dups(b, new_whiteouts,
193 (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
195 memcpy_u64s(unwritten_whiteouts_start(b),
196 new_whiteouts, b->whiteout_u64s);
198 btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
201 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
202 bool compacting, enum compact_mode mode)
204 if (!bset_dead_u64s(b, t))
209 return should_compact_bset_lazy(b, t) ||
210 (compacting && !bset_written(b, bset(b, t)));
218 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
223 for_each_bset(b, t) {
224 struct bset *i = bset(b, t);
225 struct bkey_packed *k, *n, *out, *start, *end;
226 struct btree_node_entry *src = NULL, *dst = NULL;
228 if (t != b->set && !bset_written(b, i)) {
229 src = container_of(i, struct btree_node_entry, keys);
230 dst = max(write_block(b),
231 (void *) btree_bkey_last(b, t - 1));
237 if (!should_compact_bset(b, t, ret, mode)) {
239 memmove(dst, src, sizeof(*src) +
240 le16_to_cpu(src->keys.u64s) *
243 set_btree_bset(b, t, i);
248 start = btree_bkey_first(b, t);
249 end = btree_bkey_last(b, t);
252 memmove(dst, src, sizeof(*src));
254 set_btree_bset(b, t, i);
259 for (k = start; k != end; k = n) {
262 if (!bkey_deleted(k)) {
264 out = bkey_p_next(out);
266 BUG_ON(k->needs_whiteout);
270 i->u64s = cpu_to_le16((u64 *) out - i->_data);
271 set_btree_bset_end(b, t);
272 bch2_bset_set_no_aux_tree(b, t);
276 bch2_verify_btree_nr_keys(b);
278 bch2_btree_build_aux_trees(b);
283 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
284 enum compact_mode mode)
286 return bch2_drop_whiteouts(b, mode);
289 static void btree_node_sort(struct bch_fs *c, struct btree *b,
292 bool filter_whiteouts)
294 struct btree_node *out;
295 struct sort_iter_stack sort_iter;
297 struct bset *start_bset = bset(b, &b->set[start_idx]);
298 bool used_mempool = false;
299 u64 start_time, seq = 0;
300 unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
301 bool sorting_entire_node = start_idx == 0 &&
304 sort_iter_stack_init(&sort_iter, b);
306 for (t = b->set + start_idx;
307 t < b->set + end_idx;
309 u64s += le16_to_cpu(bset(b, t)->u64s);
310 sort_iter_add(&sort_iter.iter,
311 btree_bkey_first(b, t),
312 btree_bkey_last(b, t));
315 bytes = sorting_entire_node
317 : __vstruct_bytes(struct btree_node, u64s);
319 out = btree_bounce_alloc(c, bytes, &used_mempool);
321 start_time = local_clock();
323 u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter, filter_whiteouts);
325 out->keys.u64s = cpu_to_le16(u64s);
327 BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
329 if (sorting_entire_node)
330 time_stats_update(&c->times[BCH_TIME_btree_node_sort],
333 /* Make sure we preserve bset journal_seq: */
334 for (t = b->set + start_idx; t < b->set + end_idx; t++)
335 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
336 start_bset->journal_seq = cpu_to_le64(seq);
338 if (sorting_entire_node) {
339 u64s = le16_to_cpu(out->keys.u64s);
341 BUG_ON(bytes != btree_buf_bytes(b));
344 * Our temporary buffer is the same size as the btree node's
345 * buffer, we can just swap buffers instead of doing a big
349 out->keys.u64s = cpu_to_le16(u64s);
351 set_btree_bset(b, b->set, &b->data->keys);
353 start_bset->u64s = out->keys.u64s;
354 memcpy_u64s(start_bset->start,
356 le16_to_cpu(out->keys.u64s));
359 for (i = start_idx + 1; i < end_idx; i++)
360 b->nr.bset_u64s[start_idx] +=
365 for (i = start_idx + 1; i < b->nsets; i++) {
366 b->nr.bset_u64s[i] = b->nr.bset_u64s[i + shift];
367 b->set[i] = b->set[i + shift];
370 for (i = b->nsets; i < MAX_BSETS; i++)
371 b->nr.bset_u64s[i] = 0;
373 set_btree_bset_end(b, &b->set[start_idx]);
374 bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
376 btree_bounce_free(c, bytes, used_mempool, out);
378 bch2_verify_btree_nr_keys(b);
381 void bch2_btree_sort_into(struct bch_fs *c,
385 struct btree_nr_keys nr;
386 struct btree_node_iter src_iter;
387 u64 start_time = local_clock();
389 BUG_ON(dst->nsets != 1);
391 bch2_bset_set_no_aux_tree(dst, dst->set);
393 bch2_btree_node_iter_init_from_start(&src_iter, src);
395 nr = bch2_sort_repack(btree_bset_first(dst),
400 time_stats_update(&c->times[BCH_TIME_btree_node_sort],
403 set_btree_bset_end(dst, dst->set);
405 dst->nr.live_u64s += nr.live_u64s;
406 dst->nr.bset_u64s[0] += nr.bset_u64s[0];
407 dst->nr.packed_keys += nr.packed_keys;
408 dst->nr.unpacked_keys += nr.unpacked_keys;
410 bch2_verify_btree_nr_keys(dst);
414 * We're about to add another bset to the btree node, so if there's currently
415 * too many bsets - sort some of them together:
417 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
419 unsigned unwritten_idx;
422 for (unwritten_idx = 0;
423 unwritten_idx < b->nsets;
425 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
428 if (b->nsets - unwritten_idx > 1) {
429 btree_node_sort(c, b, unwritten_idx,
434 if (unwritten_idx > 1) {
435 btree_node_sort(c, b, 0, unwritten_idx, false);
442 void bch2_btree_build_aux_trees(struct btree *b)
447 bch2_bset_build_aux_tree(b, t,
448 !bset_written(b, bset(b, t)) &&
449 t == bset_tree_last(b));
453 * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
455 * The first bset is going to be of similar order to the size of the node, the
456 * last bset is bounded by btree_write_set_buffer(), which is set to keep the
457 * memmove on insert from being too expensive: the middle bset should, ideally,
458 * be the geometric mean of the first and the last.
460 * Returns true if the middle bset is greater than that geometric mean:
462 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
464 unsigned mid_u64s_bits =
465 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
467 return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
471 * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
474 * Safe to call if there already is an unwritten bset - will only add a new bset
475 * if @b doesn't already have one.
477 * Returns true if we sorted (i.e. invalidated iterators
479 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
481 struct bch_fs *c = trans->c;
482 struct btree_node_entry *bne;
483 bool reinit_iter = false;
485 EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
486 BUG_ON(bset_written(b, bset(b, &b->set[1])));
487 BUG_ON(btree_node_just_written(b));
489 if (b->nsets == MAX_BSETS &&
490 !btree_node_write_in_flight(b) &&
491 should_compact_all(c, b)) {
492 bch2_btree_node_write(c, b, SIX_LOCK_write,
493 BTREE_WRITE_init_next_bset);
497 if (b->nsets == MAX_BSETS &&
498 btree_node_compact(c, b))
501 BUG_ON(b->nsets >= MAX_BSETS);
503 bne = want_new_bset(c, b);
505 bch2_bset_init_next(b, bne);
507 bch2_btree_build_aux_trees(b);
510 bch2_trans_node_reinit_iter(trans, b);
513 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
515 struct btree *b, struct bset *i,
516 unsigned offset, int write)
518 prt_printf(out, bch2_log_msg(c, "%s"),
520 ? "error validating btree node "
521 : "corrupt btree node before write ");
523 prt_printf(out, "on %s ", ca->name);
524 prt_printf(out, "at btree ");
525 bch2_btree_pos_to_text(out, c, b);
527 prt_printf(out, "\n node offset %u/%u",
528 b->written, btree_ptr_sectors_written(&b->key));
530 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
535 static int __btree_err(int ret,
542 enum bch_sb_error_id err_type,
543 const char *fmt, ...)
545 struct printbuf out = PRINTBUF;
548 btree_err_msg(&out, c, ca, b, i, b->written, write);
551 prt_vprintf(&out, fmt, args);
554 if (write == WRITE) {
555 bch2_print_string_as_lines(KERN_ERR, out.buf);
556 ret = c->opts.errors == BCH_ON_ERROR_continue
558 : -BCH_ERR_fsck_errors_not_fixed;
562 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
563 ret = -BCH_ERR_btree_node_read_err_fixable;
564 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
565 ret = -BCH_ERR_btree_node_read_err_bad_node;
567 if (ret != -BCH_ERR_btree_node_read_err_fixable)
568 bch2_sb_error_count(c, err_type);
571 case -BCH_ERR_btree_node_read_err_fixable:
572 ret = bch2_fsck_err(c, FSCK_CAN_FIX, err_type, "%s", out.buf);
573 if (ret != -BCH_ERR_fsck_fix &&
574 ret != -BCH_ERR_fsck_ignore)
576 ret = -BCH_ERR_fsck_fix;
578 case -BCH_ERR_btree_node_read_err_want_retry:
579 case -BCH_ERR_btree_node_read_err_must_retry:
580 bch2_print_string_as_lines(KERN_ERR, out.buf);
582 case -BCH_ERR_btree_node_read_err_bad_node:
583 bch2_print_string_as_lines(KERN_ERR, out.buf);
584 bch2_topology_error(c);
585 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
587 case -BCH_ERR_btree_node_read_err_incompatible:
588 bch2_print_string_as_lines(KERN_ERR, out.buf);
589 ret = -BCH_ERR_fsck_errors_not_fixed;
600 #define btree_err(type, c, ca, b, i, _err_type, msg, ...) \
602 int _ret = __btree_err(type, c, ca, b, i, write, have_retry, \
603 BCH_FSCK_ERR_##_err_type, \
604 msg, ##__VA_ARGS__); \
606 if (_ret != -BCH_ERR_fsck_fix) { \
614 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
617 * When btree topology repair changes the start or end of a node, that might
618 * mean we have to drop keys that are no longer inside the node:
621 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
625 for_each_bset(b, t) {
626 struct bset *i = bset(b, t);
627 struct bkey_packed *k;
629 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
630 if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
634 unsigned shift = (u64 *) k - (u64 *) i->start;
636 memmove_u64s_down(i->start, k,
637 (u64 *) vstruct_end(i) - (u64 *) k);
638 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
639 set_btree_bset_end(b, t);
642 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
643 if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
646 if (k != vstruct_last(i)) {
647 i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
648 set_btree_bset_end(b, t);
653 * Always rebuild search trees: eytzinger search tree nodes directly
654 * depend on the values of min/max key:
656 bch2_bset_set_no_aux_tree(b, b->set);
657 bch2_btree_build_aux_trees(b);
660 struct bkey unpacked;
661 struct btree_node_iter iter;
662 for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
663 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
664 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
668 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
669 struct btree *b, struct bset *i,
670 unsigned offset, unsigned sectors,
671 int write, bool have_retry, bool *saw_error)
673 unsigned version = le16_to_cpu(i->version);
674 struct printbuf buf1 = PRINTBUF;
675 struct printbuf buf2 = PRINTBUF;
678 btree_err_on(!bch2_version_compatible(version),
679 -BCH_ERR_btree_node_read_err_incompatible,
681 btree_node_unsupported_version,
682 "unsupported bset version %u.%u",
683 BCH_VERSION_MAJOR(version),
684 BCH_VERSION_MINOR(version));
686 if (btree_err_on(version < c->sb.version_min,
687 -BCH_ERR_btree_node_read_err_fixable,
689 btree_node_bset_older_than_sb_min,
690 "bset version %u older than superblock version_min %u",
691 version, c->sb.version_min)) {
692 mutex_lock(&c->sb_lock);
693 c->disk_sb.sb->version_min = cpu_to_le16(version);
695 mutex_unlock(&c->sb_lock);
698 if (btree_err_on(BCH_VERSION_MAJOR(version) >
699 BCH_VERSION_MAJOR(c->sb.version),
700 -BCH_ERR_btree_node_read_err_fixable,
702 btree_node_bset_newer_than_sb,
703 "bset version %u newer than superblock version %u",
704 version, c->sb.version)) {
705 mutex_lock(&c->sb_lock);
706 c->disk_sb.sb->version = cpu_to_le16(version);
708 mutex_unlock(&c->sb_lock);
711 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
712 -BCH_ERR_btree_node_read_err_incompatible,
714 btree_node_unsupported_version,
715 "BSET_SEPARATE_WHITEOUTS no longer supported");
717 if (btree_err_on(offset + sectors > btree_sectors(c),
718 -BCH_ERR_btree_node_read_err_fixable,
720 bset_past_end_of_btree_node,
721 "bset past end of btree node")) {
727 btree_err_on(offset && !i->u64s,
728 -BCH_ERR_btree_node_read_err_fixable,
733 btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset,
734 -BCH_ERR_btree_node_read_err_want_retry,
736 bset_wrong_sector_offset,
737 "bset at wrong sector offset");
740 struct btree_node *bn =
741 container_of(i, struct btree_node, keys);
742 /* These indicate that we read the wrong btree node: */
744 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
745 struct bch_btree_ptr_v2 *bp =
746 &bkey_i_to_btree_ptr_v2(&b->key)->v;
749 btree_err_on(bp->seq != bn->keys.seq,
750 -BCH_ERR_btree_node_read_err_must_retry,
753 "incorrect sequence number (wrong btree node)");
756 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
757 -BCH_ERR_btree_node_read_err_must_retry,
759 btree_node_bad_btree,
760 "incorrect btree id");
762 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
763 -BCH_ERR_btree_node_read_err_must_retry,
765 btree_node_bad_level,
769 compat_btree_node(b->c.level, b->c.btree_id, version,
770 BSET_BIG_ENDIAN(i), write, bn);
772 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
773 struct bch_btree_ptr_v2 *bp =
774 &bkey_i_to_btree_ptr_v2(&b->key)->v;
776 if (BTREE_PTR_RANGE_UPDATED(bp)) {
777 b->data->min_key = bp->min_key;
778 b->data->max_key = b->key.k.p;
781 btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
782 -BCH_ERR_btree_node_read_err_must_retry,
784 btree_node_bad_min_key,
785 "incorrect min_key: got %s should be %s",
786 (printbuf_reset(&buf1),
787 bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
788 (printbuf_reset(&buf2),
789 bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
792 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
793 -BCH_ERR_btree_node_read_err_must_retry,
795 btree_node_bad_max_key,
796 "incorrect max key %s",
797 (printbuf_reset(&buf1),
798 bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
801 compat_btree_node(b->c.level, b->c.btree_id, version,
802 BSET_BIG_ENDIAN(i), write, bn);
804 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
805 -BCH_ERR_btree_node_read_err_bad_node,
807 btree_node_bad_format,
808 "invalid bkey format: %s\n %s", buf1.buf,
809 (printbuf_reset(&buf2),
810 bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
811 printbuf_reset(&buf1);
813 compat_bformat(b->c.level, b->c.btree_id, version,
814 BSET_BIG_ENDIAN(i), write,
819 printbuf_exit(&buf2);
820 printbuf_exit(&buf1);
824 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
826 bool updated_range, int rw,
827 struct printbuf *err)
829 return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
830 (!updated_range ? bch2_bkey_in_btree_node(c, b, k, err) : 0) ?:
831 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
834 static bool __bkey_valid(struct bch_fs *c, struct btree *b,
835 struct bset *i, struct bkey_packed *k)
837 if (bkey_p_next(k) > vstruct_last(i))
840 if (k->format > KEY_FORMAT_CURRENT)
843 struct printbuf buf = PRINTBUF;
845 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
846 bool ret = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b), READ, &buf);
851 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
852 struct bset *i, int write,
853 bool have_retry, bool *saw_error)
855 unsigned version = le16_to_cpu(i->version);
856 struct bkey_packed *k, *prev = NULL;
857 struct printbuf buf = PRINTBUF;
858 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
859 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
863 k != vstruct_last(i);) {
866 unsigned next_good_key;
868 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
869 -BCH_ERR_btree_node_read_err_fixable,
871 btree_node_bkey_past_bset_end,
872 "key extends past end of bset")) {
873 i->u64s = cpu_to_le16((u64 *) k - i->_data);
877 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
878 -BCH_ERR_btree_node_read_err_fixable,
880 btree_node_bkey_bad_format,
881 "invalid bkey format %u", k->format))
884 /* XXX: validate k->u64s */
886 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
887 BSET_BIG_ENDIAN(i), write,
890 u = __bkey_disassemble(b, k, &tmp);
892 printbuf_reset(&buf);
893 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
894 printbuf_reset(&buf);
895 bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
896 prt_printf(&buf, "\n ");
897 bch2_bkey_val_to_text(&buf, c, u.s_c);
899 btree_err(-BCH_ERR_btree_node_read_err_fixable,
902 "invalid bkey: %s", buf.buf);
907 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
908 BSET_BIG_ENDIAN(i), write,
911 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
912 struct bkey up = bkey_unpack_key(b, prev);
914 printbuf_reset(&buf);
915 prt_printf(&buf, "keys out of order: ");
916 bch2_bkey_to_text(&buf, &up);
917 prt_printf(&buf, " > ");
918 bch2_bkey_to_text(&buf, u.k);
920 if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
922 btree_node_bkey_out_of_order,
931 next_good_key = k->u64s;
933 if (!next_good_key ||
934 (BSET_BIG_ENDIAN(i) == CPU_BIG_ENDIAN &&
935 version >= bcachefs_metadata_version_snapshot)) {
937 * only do scanning if bch2_bkey_compat() has nothing to
941 if (!__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key))) {
942 for (next_good_key = 1;
943 next_good_key < (u64 *) vstruct_last(i) - (u64 *) k;
945 if (__bkey_valid(c, b, i, (void *) ((u64 *) k + next_good_key)))
951 * didn't find a good key, have to truncate the rest of
954 next_good_key = (u64 *) vstruct_last(i) - (u64 *) k;
957 le16_add_cpu(&i->u64s, -next_good_key);
958 memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
965 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
966 struct btree *b, bool have_retry, bool *saw_error)
968 struct btree_node_entry *bne;
969 struct sort_iter *iter;
970 struct btree_node *sorted;
971 struct bkey_packed *k;
973 bool used_mempool, blacklisted;
974 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
975 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
977 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
978 struct printbuf buf = PRINTBUF;
979 int ret = 0, retry_read = 0, write = READ;
980 u64 start_time = local_clock();
982 b->version_ondisk = U16_MAX;
983 /* We might get called multiple times on read retry: */
986 iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
987 sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
989 if (bch2_meta_read_fault("btree"))
990 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
992 btree_node_fault_injected,
995 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
996 -BCH_ERR_btree_node_read_err_must_retry,
998 btree_node_bad_magic,
999 "bad magic: want %llx, got %llx",
1000 bset_magic(c), le64_to_cpu(b->data->magic));
1002 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
1003 struct bch_btree_ptr_v2 *bp =
1004 &bkey_i_to_btree_ptr_v2(&b->key)->v;
1006 bch2_bpos_to_text(&buf, b->data->min_key);
1008 bch2_bpos_to_text(&buf, b->data->max_key);
1010 btree_err_on(b->data->keys.seq != bp->seq,
1011 -BCH_ERR_btree_node_read_err_must_retry,
1014 "got wrong btree node (want %llx got %llx)\n"
1015 "got btree %s level %llu pos %s",
1016 bp->seq, b->data->keys.seq,
1017 bch2_btree_id_str(BTREE_NODE_ID(b->data)),
1018 BTREE_NODE_LEVEL(b->data),
1021 btree_err_on(!b->data->keys.seq,
1022 -BCH_ERR_btree_node_read_err_must_retry,
1025 "bad btree header: seq 0");
1028 while (b->written < (ptr_written ?: btree_sectors(c))) {
1031 bool first = !b->written;
1037 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1038 -BCH_ERR_btree_node_read_err_want_retry,
1041 "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1043 nonce = btree_nonce(i, b->written << 9);
1045 struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1046 csum_bad = bch2_crc_cmp(b->data->csum, csum);
1048 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1050 btree_err_on(csum_bad,
1051 -BCH_ERR_btree_node_read_err_want_retry,
1055 (printbuf_reset(&buf),
1056 bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
1059 ret = bset_encrypt(c, i, b->written << 9);
1060 if (bch2_fs_fatal_err_on(ret, c,
1061 "error decrypting btree node: %i", ret))
1064 btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1065 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1066 -BCH_ERR_btree_node_read_err_incompatible,
1068 btree_node_unsupported_version,
1069 "btree node does not have NEW_EXTENT_OVERWRITE set");
1071 sectors = vstruct_sectors(b->data, c->block_bits);
1073 bne = write_block(b);
1076 if (i->seq != b->data->keys.seq)
1079 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1080 -BCH_ERR_btree_node_read_err_want_retry,
1083 "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1085 nonce = btree_nonce(i, b->written << 9);
1086 struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1087 csum_bad = bch2_crc_cmp(bne->csum, csum);
1089 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1091 btree_err_on(csum_bad,
1092 -BCH_ERR_btree_node_read_err_want_retry,
1096 (printbuf_reset(&buf),
1097 bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
1100 ret = bset_encrypt(c, i, b->written << 9);
1101 if (bch2_fs_fatal_err_on(ret, c,
1102 "error decrypting btree node: %i\n", ret))
1105 sectors = vstruct_sectors(bne, c->block_bits);
1108 b->version_ondisk = min(b->version_ondisk,
1109 le16_to_cpu(i->version));
1111 ret = validate_bset(c, ca, b, i, b->written, sectors,
1112 READ, have_retry, saw_error);
1117 btree_node_set_format(b, b->data->format);
1119 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1123 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1125 blacklisted = bch2_journal_seq_is_blacklisted(c,
1126 le64_to_cpu(i->journal_seq),
1129 btree_err_on(blacklisted && first,
1130 -BCH_ERR_btree_node_read_err_fixable,
1132 bset_blacklisted_journal_seq,
1133 "first btree node bset has blacklisted journal seq (%llu)",
1134 le64_to_cpu(i->journal_seq));
1136 btree_err_on(blacklisted && ptr_written,
1137 -BCH_ERR_btree_node_read_err_fixable,
1139 first_bset_blacklisted_journal_seq,
1140 "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1141 le64_to_cpu(i->journal_seq),
1142 b->written, b->written + sectors, ptr_written);
1144 b->written += sectors;
1146 if (blacklisted && !first)
1155 btree_err_on(b->written < ptr_written,
1156 -BCH_ERR_btree_node_read_err_want_retry,
1158 btree_node_data_missing,
1159 "btree node data missing: expected %u sectors, found %u",
1160 ptr_written, b->written);
1162 for (bne = write_block(b);
1163 bset_byte_offset(b, bne) < btree_buf_bytes(b);
1164 bne = (void *) bne + block_bytes(c))
1165 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1166 !bch2_journal_seq_is_blacklisted(c,
1167 le64_to_cpu(bne->keys.journal_seq),
1169 -BCH_ERR_btree_node_read_err_want_retry,
1171 btree_node_bset_after_end,
1172 "found bset signature after last bset");
1175 sorted = btree_bounce_alloc(c, btree_buf_bytes(b), &used_mempool);
1176 sorted->keys.u64s = 0;
1178 set_btree_bset(b, b->set, &b->data->keys);
1180 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1182 u64s = le16_to_cpu(sorted->keys.u64s);
1184 sorted->keys.u64s = cpu_to_le16(u64s);
1185 swap(sorted, b->data);
1186 set_btree_bset(b, b->set, &b->data->keys);
1189 BUG_ON(b->nr.live_u64s != u64s);
1191 btree_bounce_free(c, btree_buf_bytes(b), used_mempool, sorted);
1194 bch2_btree_node_drop_keys_outside_node(b);
1197 for (k = i->start; k != vstruct_last(i);) {
1199 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1201 printbuf_reset(&buf);
1203 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1204 (bch2_inject_invalid_keys &&
1205 !bversion_cmp(u.k->version, MAX_VERSION))) {
1206 printbuf_reset(&buf);
1208 prt_printf(&buf, "invalid bkey: ");
1209 bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1210 prt_printf(&buf, "\n ");
1211 bch2_bkey_val_to_text(&buf, c, u.s_c);
1213 btree_err(-BCH_ERR_btree_node_read_err_fixable,
1215 btree_node_bad_bkey,
1218 btree_keys_account_key_drop(&b->nr, 0, k);
1220 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1221 memmove_u64s_down(k, bkey_p_next(k),
1222 (u64 *) vstruct_end(i) - (u64 *) k);
1223 set_btree_bset_end(b, b->set);
1227 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1228 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1236 bch2_bset_build_aux_tree(b, b->set, false);
1238 set_needs_whiteout(btree_bset_first(b), true);
1240 btree_node_reset_sib_u64s(b);
1242 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1243 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1245 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1246 set_btree_node_need_rewrite(b);
1250 set_btree_node_need_rewrite(b);
1252 mempool_free(iter, &c->fill_iter);
1253 printbuf_exit(&buf);
1254 time_stats_update(&c->times[BCH_TIME_btree_node_read_done], start_time);
1257 if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1258 ret == -BCH_ERR_btree_node_read_err_must_retry)
1261 set_btree_node_read_error(b);
1265 static void btree_node_read_work(struct work_struct *work)
1267 struct btree_read_bio *rb =
1268 container_of(work, struct btree_read_bio, work);
1269 struct bch_fs *c = rb->c;
1270 struct btree *b = rb->b;
1271 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1272 struct bio *bio = &rb->bio;
1273 struct bch_io_failures failed = { .nr = 0 };
1274 struct printbuf buf = PRINTBUF;
1275 bool saw_error = false;
1282 bch_info(c, "retrying read");
1283 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1284 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1285 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1286 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1287 bio->bi_iter.bi_size = btree_buf_bytes(b);
1289 if (rb->have_ioref) {
1290 bio_set_dev(bio, ca->disk_sb.bdev);
1291 submit_bio_wait(bio);
1293 bio->bi_status = BLK_STS_REMOVED;
1296 printbuf_reset(&buf);
1297 bch2_btree_pos_to_text(&buf, c, b);
1298 bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1299 "btree read error %s for %s",
1300 bch2_blk_status_to_str(bio->bi_status), buf.buf);
1302 percpu_ref_put(&ca->io_ref);
1303 rb->have_ioref = false;
1305 bch2_mark_io_failure(&failed, &rb->pick);
1307 can_retry = bch2_bkey_pick_read_device(c,
1308 bkey_i_to_s_c(&b->key),
1309 &failed, &rb->pick) > 0;
1311 if (!bio->bi_status &&
1312 !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1314 bch_info(c, "retry success");
1321 set_btree_node_read_error(b);
1326 time_stats_update(&c->times[BCH_TIME_btree_node_read],
1330 if (saw_error && !btree_node_read_error(b)) {
1331 printbuf_reset(&buf);
1332 bch2_bpos_to_text(&buf, b->key.k.p);
1333 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1334 __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1336 bch2_btree_node_rewrite_async(c, b);
1339 printbuf_exit(&buf);
1340 clear_btree_node_read_in_flight(b);
1341 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1344 static void btree_node_read_endio(struct bio *bio)
1346 struct btree_read_bio *rb =
1347 container_of(bio, struct btree_read_bio, bio);
1348 struct bch_fs *c = rb->c;
1350 if (rb->have_ioref) {
1351 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1353 bch2_latency_acct(ca, rb->start_time, READ);
1356 queue_work(c->io_complete_wq, &rb->work);
1359 struct btree_node_read_all {
1364 void *buf[BCH_REPLICAS_MAX];
1365 struct bio *bio[BCH_REPLICAS_MAX];
1366 blk_status_t err[BCH_REPLICAS_MAX];
1369 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1371 struct btree_node *bn = data;
1372 struct btree_node_entry *bne;
1373 unsigned offset = 0;
1375 if (le64_to_cpu(bn->magic) != bset_magic(c))
1378 while (offset < btree_sectors(c)) {
1380 offset += vstruct_sectors(bn, c->block_bits);
1382 bne = data + (offset << 9);
1383 if (bne->keys.seq != bn->keys.seq)
1385 offset += vstruct_sectors(bne, c->block_bits);
1392 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1394 struct btree_node *bn = data;
1395 struct btree_node_entry *bne;
1400 while (offset < btree_sectors(c)) {
1401 bne = data + (offset << 9);
1402 if (bne->keys.seq == bn->keys.seq)
1411 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1413 closure_type(ra, struct btree_node_read_all, cl);
1414 struct bch_fs *c = ra->c;
1415 struct btree *b = ra->b;
1416 struct printbuf buf = PRINTBUF;
1417 bool dump_bset_maps = false;
1418 bool have_retry = false;
1419 int ret = 0, best = -1, write = READ;
1420 unsigned i, written = 0, written2 = 0;
1421 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1422 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1423 bool _saw_error = false, *saw_error = &_saw_error;
1425 for (i = 0; i < ra->nr; i++) {
1426 struct btree_node *bn = ra->buf[i];
1431 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1432 (seq && seq != bn->keys.seq))
1437 written = btree_node_sectors_written(c, bn);
1441 written2 = btree_node_sectors_written(c, ra->buf[i]);
1442 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1444 btree_node_replicas_sectors_written_mismatch,
1445 "btree node sectors written mismatch: %u != %u",
1446 written, written2) ||
1447 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1448 -BCH_ERR_btree_node_read_err_fixable,
1450 btree_node_bset_after_end,
1451 "found bset signature after last bset") ||
1452 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1453 -BCH_ERR_btree_node_read_err_fixable,
1455 btree_node_replicas_data_mismatch,
1456 "btree node replicas content mismatch"))
1457 dump_bset_maps = true;
1459 if (written2 > written) {
1465 if (dump_bset_maps) {
1466 for (i = 0; i < ra->nr; i++) {
1467 struct btree_node *bn = ra->buf[i];
1468 struct btree_node_entry *bne = NULL;
1469 unsigned offset = 0, sectors;
1475 printbuf_reset(&buf);
1477 while (offset < btree_sectors(c)) {
1479 sectors = vstruct_sectors(bn, c->block_bits);
1481 bne = ra->buf[i] + (offset << 9);
1482 if (bne->keys.seq != bn->keys.seq)
1484 sectors = vstruct_sectors(bne, c->block_bits);
1487 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1488 if (bne && bch2_journal_seq_is_blacklisted(c,
1489 le64_to_cpu(bne->keys.journal_seq), false))
1490 prt_printf(&buf, "*");
1494 while (offset < btree_sectors(c)) {
1495 bne = ra->buf[i] + (offset << 9);
1496 if (bne->keys.seq == bn->keys.seq) {
1498 prt_printf(&buf, " GAP");
1501 sectors = vstruct_sectors(bne, c->block_bits);
1502 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1503 if (bch2_journal_seq_is_blacklisted(c,
1504 le64_to_cpu(bne->keys.journal_seq), false))
1505 prt_printf(&buf, "*");
1510 bch_err(c, "replica %u:%s", i, buf.buf);
1515 memcpy(b->data, ra->buf[best], btree_buf_bytes(b));
1516 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1522 set_btree_node_read_error(b);
1523 else if (*saw_error)
1524 bch2_btree_node_rewrite_async(c, b);
1526 for (i = 0; i < ra->nr; i++) {
1527 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1528 bio_put(ra->bio[i]);
1531 closure_debug_destroy(&ra->cl);
1533 printbuf_exit(&buf);
1535 clear_btree_node_read_in_flight(b);
1536 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1539 static void btree_node_read_all_replicas_endio(struct bio *bio)
1541 struct btree_read_bio *rb =
1542 container_of(bio, struct btree_read_bio, bio);
1543 struct bch_fs *c = rb->c;
1544 struct btree_node_read_all *ra = rb->ra;
1546 if (rb->have_ioref) {
1547 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1549 bch2_latency_acct(ca, rb->start_time, READ);
1552 ra->err[rb->idx] = bio->bi_status;
1553 closure_put(&ra->cl);
1557 * XXX This allocates multiple times from the same mempools, and can deadlock
1558 * under sufficient memory pressure (but is only a debug path)
1560 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1562 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1563 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1564 const union bch_extent_entry *entry;
1565 struct extent_ptr_decoded pick;
1566 struct btree_node_read_all *ra;
1569 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1571 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1573 closure_init(&ra->cl, NULL);
1576 ra->nr = bch2_bkey_nr_ptrs(k);
1578 for (i = 0; i < ra->nr; i++) {
1579 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1580 ra->bio[i] = bio_alloc_bioset(NULL,
1581 buf_pages(ra->buf[i], btree_buf_bytes(b)),
1582 REQ_OP_READ|REQ_SYNC|REQ_META,
1588 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1589 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1590 struct btree_read_bio *rb =
1591 container_of(ra->bio[i], struct btree_read_bio, bio);
1595 rb->start_time = local_clock();
1596 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1599 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1600 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1601 bch2_bio_map(&rb->bio, ra->buf[i], btree_buf_bytes(b));
1603 if (rb->have_ioref) {
1604 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1605 bio_sectors(&rb->bio));
1606 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1608 closure_get(&ra->cl);
1609 submit_bio(&rb->bio);
1611 ra->err[i] = BLK_STS_REMOVED;
1618 closure_sync(&ra->cl);
1619 btree_node_read_all_replicas_done(&ra->cl.work);
1621 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1628 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1631 struct bch_fs *c = trans->c;
1632 struct extent_ptr_decoded pick;
1633 struct btree_read_bio *rb;
1638 trace_and_count(c, btree_node_read, trans, b);
1640 if (bch2_verify_all_btree_replicas &&
1641 !btree_node_read_all_replicas(c, b, sync))
1644 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1648 struct printbuf buf = PRINTBUF;
1650 prt_str(&buf, "btree node read error: no device to read from\n at ");
1651 bch2_btree_pos_to_text(&buf, c, b);
1652 bch_err(c, "%s", buf.buf);
1654 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1655 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1656 bch2_fatal_error(c);
1658 set_btree_node_read_error(b);
1659 clear_btree_node_read_in_flight(b);
1660 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1661 printbuf_exit(&buf);
1665 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1667 bio = bio_alloc_bioset(NULL,
1668 buf_pages(b->data, btree_buf_bytes(b)),
1669 REQ_OP_READ|REQ_SYNC|REQ_META,
1672 rb = container_of(bio, struct btree_read_bio, bio);
1676 rb->start_time = local_clock();
1677 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1679 INIT_WORK(&rb->work, btree_node_read_work);
1680 bio->bi_iter.bi_sector = pick.ptr.offset;
1681 bio->bi_end_io = btree_node_read_endio;
1682 bch2_bio_map(bio, b->data, btree_buf_bytes(b));
1684 if (rb->have_ioref) {
1685 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1687 bio_set_dev(bio, ca->disk_sb.bdev);
1690 submit_bio_wait(bio);
1691 bch2_latency_acct(ca, rb->start_time, READ);
1692 btree_node_read_work(&rb->work);
1697 bio->bi_status = BLK_STS_REMOVED;
1700 btree_node_read_work(&rb->work);
1702 queue_work(c->io_complete_wq, &rb->work);
1706 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1707 const struct bkey_i *k, unsigned level)
1709 struct bch_fs *c = trans->c;
1714 closure_init_stack(&cl);
1717 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1721 b = bch2_btree_node_mem_alloc(trans, level != 0);
1722 bch2_btree_cache_cannibalize_unlock(trans);
1726 bkey_copy(&b->key, k);
1727 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1729 set_btree_node_read_in_flight(b);
1731 bch2_btree_node_read(trans, b, true);
1733 if (btree_node_read_error(b)) {
1734 bch2_btree_node_hash_remove(&c->btree_cache, b);
1736 mutex_lock(&c->btree_cache.lock);
1737 list_move(&b->list, &c->btree_cache.freeable);
1738 mutex_unlock(&c->btree_cache.lock);
1744 bch2_btree_set_root_for_read(c, b);
1746 six_unlock_write(&b->c.lock);
1747 six_unlock_intent(&b->c.lock);
1752 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1753 const struct bkey_i *k, unsigned level)
1755 return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1758 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1759 struct btree_write *w)
1761 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1769 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1772 closure_put(&((struct btree_update *) new)->cl);
1774 bch2_journal_pin_drop(&c->journal, &w->journal);
1777 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1779 struct btree_write *w = btree_prev_write(b);
1780 unsigned long old, new, v;
1783 bch2_btree_complete_write(c, b, w);
1785 v = READ_ONCE(b->flags);
1789 if ((old & (1U << BTREE_NODE_dirty)) &&
1790 (old & (1U << BTREE_NODE_need_write)) &&
1791 !(old & (1U << BTREE_NODE_never_write)) &&
1792 !(old & (1U << BTREE_NODE_write_blocked)) &&
1793 !(old & (1U << BTREE_NODE_will_make_reachable))) {
1794 new &= ~(1U << BTREE_NODE_dirty);
1795 new &= ~(1U << BTREE_NODE_need_write);
1796 new |= (1U << BTREE_NODE_write_in_flight);
1797 new |= (1U << BTREE_NODE_write_in_flight_inner);
1798 new |= (1U << BTREE_NODE_just_written);
1799 new ^= (1U << BTREE_NODE_write_idx);
1801 type = new & BTREE_WRITE_TYPE_MASK;
1802 new &= ~BTREE_WRITE_TYPE_MASK;
1804 new &= ~(1U << BTREE_NODE_write_in_flight);
1805 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1807 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1809 if (new & (1U << BTREE_NODE_write_in_flight))
1810 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1812 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1815 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1817 struct btree_trans *trans = bch2_trans_get(c);
1819 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1820 __btree_node_write_done(c, b);
1821 six_unlock_read(&b->c.lock);
1823 bch2_trans_put(trans);
1826 static void btree_node_write_work(struct work_struct *work)
1828 struct btree_write_bio *wbio =
1829 container_of(work, struct btree_write_bio, work);
1830 struct bch_fs *c = wbio->wbio.c;
1831 struct btree *b = wbio->wbio.bio.bi_private;
1832 struct bch_extent_ptr *ptr;
1835 btree_bounce_free(c,
1837 wbio->wbio.used_mempool,
1840 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1841 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1843 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) {
1844 ret = -BCH_ERR_btree_write_all_failed;
1848 if (wbio->wbio.first_btree_write) {
1849 if (wbio->wbio.failed.nr) {
1853 ret = bch2_trans_do(c, NULL, NULL, 0,
1854 bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1855 BCH_WATERMARK_reclaim|
1856 BCH_TRANS_COMMIT_journal_reclaim|
1857 BCH_TRANS_COMMIT_no_enospc|
1858 BCH_TRANS_COMMIT_no_check_rw,
1859 !wbio->wbio.failed.nr));
1864 bio_put(&wbio->wbio.bio);
1865 btree_node_write_done(c, b);
1868 set_btree_node_noevict(b);
1869 if (!bch2_err_matches(ret, EROFS))
1870 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1874 static void btree_node_write_endio(struct bio *bio)
1876 struct bch_write_bio *wbio = to_wbio(bio);
1877 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1878 struct bch_write_bio *orig = parent ?: wbio;
1879 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1880 struct bch_fs *c = wbio->c;
1881 struct btree *b = wbio->bio.bi_private;
1882 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1883 unsigned long flags;
1885 if (wbio->have_ioref)
1886 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1888 if (bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1889 "btree write error: %s",
1890 bch2_blk_status_to_str(bio->bi_status)) ||
1891 bch2_meta_write_fault("btree")) {
1892 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1893 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1894 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1897 if (wbio->have_ioref)
1898 percpu_ref_put(&ca->io_ref);
1902 bio_endio(&parent->bio);
1906 clear_btree_node_write_in_flight_inner(b);
1907 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1908 INIT_WORK(&wb->work, btree_node_write_work);
1909 queue_work(c->btree_io_complete_wq, &wb->work);
1912 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1913 struct bset *i, unsigned sectors)
1915 struct printbuf buf = PRINTBUF;
1919 ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1920 BKEY_TYPE_btree, WRITE, &buf);
1923 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1924 printbuf_exit(&buf);
1928 ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1929 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1931 bch2_inconsistent_error(c);
1938 static void btree_write_submit(struct work_struct *work)
1940 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1941 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1943 bkey_copy(&tmp.k, &wbio->key);
1945 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1946 ptr->offset += wbio->sector_offset;
1948 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1952 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1954 struct btree_write_bio *wbio;
1955 struct bset_tree *t;
1957 struct btree_node *bn = NULL;
1958 struct btree_node_entry *bne = NULL;
1959 struct sort_iter_stack sort_iter;
1961 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1964 unsigned long old, new;
1965 bool validate_before_checksum = false;
1966 enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1970 if (flags & BTREE_WRITE_ALREADY_STARTED)
1974 * We may only have a read lock on the btree node - the dirty bit is our
1975 * "lock" against racing with other threads that may be trying to start
1976 * a write, we do a write iff we clear the dirty bit. Since setting the
1977 * dirty bit requires a write lock, we can't race with other threads
1981 old = new = READ_ONCE(b->flags);
1983 if (!(old & (1 << BTREE_NODE_dirty)))
1986 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1987 !(old & (1 << BTREE_NODE_need_write)))
1991 ((1 << BTREE_NODE_never_write)|
1992 (1 << BTREE_NODE_write_blocked)))
1996 (old & (1 << BTREE_NODE_will_make_reachable)))
1999 if (old & (1 << BTREE_NODE_write_in_flight))
2002 if (flags & BTREE_WRITE_ONLY_IF_NEED)
2003 type = new & BTREE_WRITE_TYPE_MASK;
2004 new &= ~BTREE_WRITE_TYPE_MASK;
2006 new &= ~(1 << BTREE_NODE_dirty);
2007 new &= ~(1 << BTREE_NODE_need_write);
2008 new |= (1 << BTREE_NODE_write_in_flight);
2009 new |= (1 << BTREE_NODE_write_in_flight_inner);
2010 new |= (1 << BTREE_NODE_just_written);
2011 new ^= (1 << BTREE_NODE_write_idx);
2012 } while (cmpxchg_acquire(&b->flags, old, new) != old);
2014 if (new & (1U << BTREE_NODE_need_write))
2017 BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
2019 atomic_dec(&c->btree_cache.dirty);
2021 BUG_ON(btree_node_fake(b));
2022 BUG_ON((b->will_make_reachable != 0) != !b->written);
2024 BUG_ON(b->written >= btree_sectors(c));
2025 BUG_ON(b->written & (block_sectors(c) - 1));
2026 BUG_ON(bset_written(b, btree_bset_last(b)));
2027 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
2028 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
2030 bch2_sort_whiteouts(c, b);
2032 sort_iter_stack_init(&sort_iter, b);
2035 ? sizeof(struct btree_node)
2036 : sizeof(struct btree_node_entry);
2038 bytes += b->whiteout_u64s * sizeof(u64);
2040 for_each_bset(b, t) {
2043 if (bset_written(b, i))
2046 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2047 sort_iter_add(&sort_iter.iter,
2048 btree_bkey_first(b, t),
2049 btree_bkey_last(b, t));
2050 seq = max(seq, le64_to_cpu(i->journal_seq));
2053 BUG_ON(b->written && !seq);
2055 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2058 /* buffer must be a multiple of the block size */
2059 bytes = round_up(bytes, block_bytes(c));
2061 data = btree_bounce_alloc(c, bytes, &used_mempool);
2069 bne->keys = b->data->keys;
2073 i->journal_seq = cpu_to_le64(seq);
2076 sort_iter_add(&sort_iter.iter,
2077 unwritten_whiteouts_start(b),
2078 unwritten_whiteouts_end(b));
2079 SET_BSET_SEPARATE_WHITEOUTS(i, false);
2081 b->whiteout_u64s = 0;
2083 u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
2084 le16_add_cpu(&i->u64s, u64s);
2086 BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2088 set_needs_whiteout(i, false);
2090 /* do we have data to write? */
2091 if (b->written && !i->u64s)
2094 bytes_to_write = vstruct_end(i) - data;
2095 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2098 b->key.k.type == KEY_TYPE_btree_ptr_v2)
2099 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2101 memset(data + bytes_to_write, 0,
2102 (sectors_to_write << 9) - bytes_to_write);
2104 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2105 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2106 BUG_ON(i->seq != b->data->keys.seq);
2108 i->version = cpu_to_le16(c->sb.version);
2109 SET_BSET_OFFSET(i, b->written);
2110 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2112 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2113 validate_before_checksum = true;
2115 /* validate_bset will be modifying: */
2116 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2117 validate_before_checksum = true;
2119 /* if we're going to be encrypting, check metadata validity first: */
2120 if (validate_before_checksum &&
2121 validate_bset_for_write(c, b, i, sectors_to_write))
2124 ret = bset_encrypt(c, i, b->written << 9);
2125 if (bch2_fs_fatal_err_on(ret, c,
2126 "error encrypting btree node: %i\n", ret))
2129 nonce = btree_nonce(i, b->written << 9);
2132 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2134 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2136 /* if we're not encrypting, check metadata after checksumming: */
2137 if (!validate_before_checksum &&
2138 validate_bset_for_write(c, b, i, sectors_to_write))
2142 * We handle btree write errors by immediately halting the journal -
2143 * after we've done that, we can't issue any subsequent btree writes
2144 * because they might have pointers to new nodes that failed to write.
2146 * Furthermore, there's no point in doing any more btree writes because
2147 * with the journal stopped, we're never going to update the journal to
2148 * reflect that those writes were done and the data flushed from the
2151 * Also on journal error, the pending write may have updates that were
2152 * never journalled (interior nodes, see btree_update_nodes_written()) -
2153 * it's critical that we don't do the write in that case otherwise we
2154 * will have updates visible that weren't in the journal:
2156 * Make sure to update b->written so bch2_btree_init_next() doesn't
2159 if (bch2_journal_error(&c->journal) ||
2163 trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2165 wbio = container_of(bio_alloc_bioset(NULL,
2166 buf_pages(data, sectors_to_write << 9),
2167 REQ_OP_WRITE|REQ_META,
2170 struct btree_write_bio, wbio.bio);
2171 wbio_init(&wbio->wbio.bio);
2173 wbio->data_bytes = bytes;
2174 wbio->sector_offset = b->written;
2176 wbio->wbio.used_mempool = used_mempool;
2177 wbio->wbio.first_btree_write = !b->written;
2178 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
2179 wbio->wbio.bio.bi_private = b;
2181 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2183 bkey_copy(&wbio->key, &b->key);
2185 b->written += sectors_to_write;
2187 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2188 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2189 cpu_to_le16(b->written);
2191 atomic64_inc(&c->btree_write_stats[type].nr);
2192 atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2194 INIT_WORK(&wbio->work, btree_write_submit);
2195 queue_work(c->io_complete_wq, &wbio->work);
2198 set_btree_node_noevict(b);
2199 b->written += sectors_to_write;
2201 btree_bounce_free(c, bytes, used_mempool, data);
2202 __btree_node_write_done(c, b);
2206 * Work that must be done with write lock held:
2208 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2210 bool invalidated_iter = false;
2211 struct btree_node_entry *bne;
2212 struct bset_tree *t;
2214 if (!btree_node_just_written(b))
2217 BUG_ON(b->whiteout_u64s);
2219 clear_btree_node_just_written(b);
2222 * Note: immediately after write, bset_written() doesn't work - the
2223 * amount of data we had to write after compaction might have been
2224 * smaller than the offset of the last bset.
2226 * However, we know that all bsets have been written here, as long as
2227 * we're still holding the write lock:
2231 * XXX: decide if we really want to unconditionally sort down to a
2235 btree_node_sort(c, b, 0, b->nsets, true);
2236 invalidated_iter = true;
2238 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2242 set_needs_whiteout(bset(b, t), true);
2244 bch2_btree_verify(c, b);
2247 * If later we don't unconditionally sort down to a single bset, we have
2248 * to ensure this is still true:
2250 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2252 bne = want_new_bset(c, b);
2254 bch2_bset_init_next(b, bne);
2256 bch2_btree_build_aux_trees(b);
2258 return invalidated_iter;
2262 * Use this one if the node is intent locked:
2264 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2265 enum six_lock_type lock_type_held,
2268 if (lock_type_held == SIX_LOCK_intent ||
2269 (lock_type_held == SIX_LOCK_read &&
2270 six_lock_tryupgrade(&b->c.lock))) {
2271 __bch2_btree_node_write(c, b, flags);
2273 /* don't cycle lock unnecessarily: */
2274 if (btree_node_just_written(b) &&
2275 six_trylock_write(&b->c.lock)) {
2276 bch2_btree_post_write_cleanup(c, b);
2277 six_unlock_write(&b->c.lock);
2280 if (lock_type_held == SIX_LOCK_read)
2281 six_lock_downgrade(&b->c.lock);
2283 __bch2_btree_node_write(c, b, flags);
2284 if (lock_type_held == SIX_LOCK_write &&
2285 btree_node_just_written(b))
2286 bch2_btree_post_write_cleanup(c, b);
2290 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2292 struct bucket_table *tbl;
2293 struct rhash_head *pos;
2299 for_each_cached_btree(b, c, tbl, i, pos)
2300 if (test_bit(flag, &b->flags)) {
2302 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2311 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2313 return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2316 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2318 return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2321 static const char * const bch2_btree_write_types[] = {
2322 #define x(t, n) [n] = #t,
2323 BCH_BTREE_WRITE_TYPES()
2327 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2329 printbuf_tabstop_push(out, 20);
2330 printbuf_tabstop_push(out, 10);
2335 prt_str(out, "size");
2338 for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2339 u64 nr = atomic64_read(&c->btree_write_stats[i].nr);
2340 u64 bytes = atomic64_read(&c->btree_write_stats[i].bytes);
2342 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2346 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);