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 > btree_bytes(c));
117 *used_mempool = false;
118 p = vpmalloc(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(c, b);
178 k != unwritten_whiteouts_end(c, b);
182 sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
186 while (ptrs != ptrs_end) {
192 verify_no_dups(b, new_whiteouts,
193 (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
195 memcpy_u64s(unwritten_whiteouts_start(c, 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 bch2_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_bytes(c));
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 bch2_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(c, b, bne);
507 bch2_btree_build_aux_trees(b);
510 bch2_trans_node_reinit_iter(trans, b);
513 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
516 prt_printf(out, "%s level %u/%u\n ",
517 bch2_btree_ids[b->c.btree_id],
519 bch2_btree_id_root(c, b->c.btree_id)->level);
520 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
523 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
525 struct btree *b, struct bset *i,
526 unsigned offset, int write)
528 prt_printf(out, bch2_log_msg(c, "%s"),
530 ? "error validating btree node "
531 : "corrupt btree node before write ");
533 prt_printf(out, "on %s ", ca->name);
534 prt_printf(out, "at btree ");
535 btree_pos_to_text(out, c, b);
537 prt_printf(out, "\n node offset %u", b->written);
539 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
544 static int __btree_err(int ret,
551 const char *fmt, ...)
553 struct printbuf out = PRINTBUF;
556 btree_err_msg(&out, c, ca, b, i, b->written, write);
559 prt_vprintf(&out, fmt, args);
562 if (write == WRITE) {
563 bch2_print_string_as_lines(KERN_ERR, out.buf);
564 ret = c->opts.errors == BCH_ON_ERROR_continue
566 : -BCH_ERR_fsck_errors_not_fixed;
570 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
571 ret = -BCH_ERR_btree_node_read_err_fixable;
572 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
573 ret = -BCH_ERR_btree_node_read_err_bad_node;
576 case -BCH_ERR_btree_node_read_err_fixable:
577 mustfix_fsck_err(c, "%s", out.buf);
578 ret = -BCH_ERR_fsck_fix;
580 case -BCH_ERR_btree_node_read_err_want_retry:
581 case -BCH_ERR_btree_node_read_err_must_retry:
582 bch2_print_string_as_lines(KERN_ERR, out.buf);
584 case -BCH_ERR_btree_node_read_err_bad_node:
585 bch2_print_string_as_lines(KERN_ERR, out.buf);
586 bch2_topology_error(c);
587 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
589 case -BCH_ERR_btree_node_read_err_incompatible:
590 bch2_print_string_as_lines(KERN_ERR, out.buf);
591 ret = -BCH_ERR_fsck_errors_not_fixed;
602 #define btree_err(type, c, ca, b, i, msg, ...) \
604 int _ret = __btree_err(type, c, ca, b, i, write, have_retry, 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, c, ca, b, i,
680 "unsupported bset version %u.%u",
681 BCH_VERSION_MAJOR(version),
682 BCH_VERSION_MINOR(version));
684 if (btree_err_on(version < c->sb.version_min,
685 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
686 "bset version %u older than superblock version_min %u",
687 version, c->sb.version_min)) {
688 mutex_lock(&c->sb_lock);
689 c->disk_sb.sb->version_min = cpu_to_le16(version);
691 mutex_unlock(&c->sb_lock);
694 if (btree_err_on(BCH_VERSION_MAJOR(version) >
695 BCH_VERSION_MAJOR(c->sb.version),
696 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
697 "bset version %u newer than superblock version %u",
698 version, c->sb.version)) {
699 mutex_lock(&c->sb_lock);
700 c->disk_sb.sb->version = cpu_to_le16(version);
702 mutex_unlock(&c->sb_lock);
705 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
706 -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
707 "BSET_SEPARATE_WHITEOUTS no longer supported");
709 if (btree_err_on(offset + sectors > btree_sectors(c),
710 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
711 "bset past end of btree node")) {
717 btree_err_on(offset && !i->u64s,
718 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
721 btree_err_on(BSET_OFFSET(i) &&
722 BSET_OFFSET(i) != offset,
723 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
724 "bset at wrong sector offset");
727 struct btree_node *bn =
728 container_of(i, struct btree_node, keys);
729 /* These indicate that we read the wrong btree node: */
731 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
732 struct bch_btree_ptr_v2 *bp =
733 &bkey_i_to_btree_ptr_v2(&b->key)->v;
736 btree_err_on(bp->seq != bn->keys.seq,
737 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
738 "incorrect sequence number (wrong btree node)");
741 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
742 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
743 "incorrect btree id");
745 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
746 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
750 compat_btree_node(b->c.level, b->c.btree_id, version,
751 BSET_BIG_ENDIAN(i), write, bn);
753 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
754 struct bch_btree_ptr_v2 *bp =
755 &bkey_i_to_btree_ptr_v2(&b->key)->v;
757 if (BTREE_PTR_RANGE_UPDATED(bp)) {
758 b->data->min_key = bp->min_key;
759 b->data->max_key = b->key.k.p;
762 btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
763 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
764 "incorrect min_key: got %s should be %s",
765 (printbuf_reset(&buf1),
766 bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
767 (printbuf_reset(&buf2),
768 bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
771 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
772 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
773 "incorrect max key %s",
774 (printbuf_reset(&buf1),
775 bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
778 compat_btree_node(b->c.level, b->c.btree_id, version,
779 BSET_BIG_ENDIAN(i), write, bn);
781 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
782 -BCH_ERR_btree_node_read_err_bad_node, c, ca, b, i,
783 "invalid bkey format: %s\n %s", buf1.buf,
784 (printbuf_reset(&buf2),
785 bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
786 printbuf_reset(&buf1);
788 compat_bformat(b->c.level, b->c.btree_id, version,
789 BSET_BIG_ENDIAN(i), write,
794 printbuf_exit(&buf2);
795 printbuf_exit(&buf1);
799 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
801 bool updated_range, int rw,
802 struct printbuf *err)
804 return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
805 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
806 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
809 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
810 struct bset *i, int write,
811 bool have_retry, bool *saw_error)
813 unsigned version = le16_to_cpu(i->version);
814 struct bkey_packed *k, *prev = NULL;
815 struct printbuf buf = PRINTBUF;
816 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
817 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
821 k != vstruct_last(i);) {
825 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
826 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
827 "key extends past end of bset")) {
828 i->u64s = cpu_to_le16((u64 *) k - i->_data);
832 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
833 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
834 "invalid bkey format %u", k->format)) {
835 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
836 memmove_u64s_down(k, bkey_p_next(k),
837 (u64 *) vstruct_end(i) - (u64 *) k);
841 /* XXX: validate k->u64s */
843 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
844 BSET_BIG_ENDIAN(i), write,
847 u = __bkey_disassemble(b, k, &tmp);
849 printbuf_reset(&buf);
850 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
851 printbuf_reset(&buf);
852 prt_printf(&buf, "invalid bkey: ");
853 bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
854 prt_printf(&buf, "\n ");
855 bch2_bkey_val_to_text(&buf, c, u.s_c);
857 btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
859 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
860 memmove_u64s_down(k, bkey_p_next(k),
861 (u64 *) vstruct_end(i) - (u64 *) k);
866 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
867 BSET_BIG_ENDIAN(i), write,
870 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
871 struct bkey up = bkey_unpack_key(b, prev);
873 printbuf_reset(&buf);
874 prt_printf(&buf, "keys out of order: ");
875 bch2_bkey_to_text(&buf, &up);
876 prt_printf(&buf, " > ");
877 bch2_bkey_to_text(&buf, u.k);
879 bch2_dump_bset(c, b, i, 0);
881 if (btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf)) {
882 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
883 memmove_u64s_down(k, bkey_p_next(k),
884 (u64 *) vstruct_end(i) - (u64 *) k);
897 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
898 struct btree *b, bool have_retry, bool *saw_error)
900 struct btree_node_entry *bne;
901 struct sort_iter *iter;
902 struct btree_node *sorted;
903 struct bkey_packed *k;
904 struct bch_extent_ptr *ptr;
906 bool used_mempool, blacklisted;
907 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
908 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
910 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
911 struct printbuf buf = PRINTBUF;
912 int ret = 0, retry_read = 0, write = READ;
914 b->version_ondisk = U16_MAX;
915 /* We might get called multiple times on read retry: */
918 iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
919 sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
921 if (bch2_meta_read_fault("btree"))
922 btree_err(-BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
925 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
926 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
927 "bad magic: want %llx, got %llx",
928 bset_magic(c), le64_to_cpu(b->data->magic));
930 btree_err_on(!b->data->keys.seq,
931 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
932 "bad btree header: seq 0");
934 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
935 struct bch_btree_ptr_v2 *bp =
936 &bkey_i_to_btree_ptr_v2(&b->key)->v;
938 btree_err_on(b->data->keys.seq != bp->seq,
939 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
940 "got wrong btree node (seq %llx want %llx)",
941 b->data->keys.seq, bp->seq);
944 while (b->written < (ptr_written ?: btree_sectors(c))) {
947 struct bch_csum csum;
948 bool first = !b->written;
953 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
954 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
955 "unknown checksum type %llu",
958 nonce = btree_nonce(i, b->written << 9);
959 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
961 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
962 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
965 ret = bset_encrypt(c, i, b->written << 9);
966 if (bch2_fs_fatal_err_on(ret, c,
967 "error decrypting btree node: %i", ret))
970 btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
971 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
972 -BCH_ERR_btree_node_read_err_incompatible, c, NULL, b, NULL,
973 "btree node does not have NEW_EXTENT_OVERWRITE set");
975 sectors = vstruct_sectors(b->data, c->block_bits);
977 bne = write_block(b);
980 if (i->seq != b->data->keys.seq)
983 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
984 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
985 "unknown checksum type %llu",
988 nonce = btree_nonce(i, b->written << 9);
989 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
991 btree_err_on(bch2_crc_cmp(csum, bne->csum),
992 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
995 ret = bset_encrypt(c, i, b->written << 9);
996 if (bch2_fs_fatal_err_on(ret, c,
997 "error decrypting btree node: %i\n", ret))
1000 sectors = vstruct_sectors(bne, c->block_bits);
1003 b->version_ondisk = min(b->version_ondisk,
1004 le16_to_cpu(i->version));
1006 ret = validate_bset(c, ca, b, i, b->written, sectors,
1007 READ, have_retry, saw_error);
1012 btree_node_set_format(b, b->data->format);
1014 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1018 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1020 blacklisted = bch2_journal_seq_is_blacklisted(c,
1021 le64_to_cpu(i->journal_seq),
1024 btree_err_on(blacklisted && first,
1025 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1026 "first btree node bset has blacklisted journal seq (%llu)",
1027 le64_to_cpu(i->journal_seq));
1029 btree_err_on(blacklisted && ptr_written,
1030 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1031 "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1032 le64_to_cpu(i->journal_seq),
1033 b->written, b->written + sectors, ptr_written);
1035 b->written += sectors;
1037 if (blacklisted && !first)
1046 btree_err_on(b->written < ptr_written,
1047 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1048 "btree node data missing: expected %u sectors, found %u",
1049 ptr_written, b->written);
1051 for (bne = write_block(b);
1052 bset_byte_offset(b, bne) < btree_bytes(c);
1053 bne = (void *) bne + block_bytes(c))
1054 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1055 !bch2_journal_seq_is_blacklisted(c,
1056 le64_to_cpu(bne->keys.journal_seq),
1058 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1059 "found bset signature after last bset");
1062 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1063 sorted->keys.u64s = 0;
1065 set_btree_bset(b, b->set, &b->data->keys);
1067 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1069 u64s = le16_to_cpu(sorted->keys.u64s);
1071 sorted->keys.u64s = cpu_to_le16(u64s);
1072 swap(sorted, b->data);
1073 set_btree_bset(b, b->set, &b->data->keys);
1076 BUG_ON(b->nr.live_u64s != u64s);
1078 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1081 bch2_btree_node_drop_keys_outside_node(b);
1084 for (k = i->start; k != vstruct_last(i);) {
1086 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1088 printbuf_reset(&buf);
1090 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1091 (bch2_inject_invalid_keys &&
1092 !bversion_cmp(u.k->version, MAX_VERSION))) {
1093 printbuf_reset(&buf);
1095 prt_printf(&buf, "invalid bkey: ");
1096 bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1097 prt_printf(&buf, "\n ");
1098 bch2_bkey_val_to_text(&buf, c, u.s_c);
1100 btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
1102 btree_keys_account_key_drop(&b->nr, 0, k);
1104 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1105 memmove_u64s_down(k, bkey_p_next(k),
1106 (u64 *) vstruct_end(i) - (u64 *) k);
1107 set_btree_bset_end(b, b->set);
1111 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1112 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1120 bch2_bset_build_aux_tree(b, b->set, false);
1122 set_needs_whiteout(btree_bset_first(b), true);
1124 btree_node_reset_sib_u64s(b);
1126 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1127 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1129 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1130 set_btree_node_need_rewrite(b);
1134 set_btree_node_need_rewrite(b);
1136 mempool_free(iter, &c->fill_iter);
1137 printbuf_exit(&buf);
1140 if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1141 ret == -BCH_ERR_btree_node_read_err_must_retry)
1144 set_btree_node_read_error(b);
1148 static void btree_node_read_work(struct work_struct *work)
1150 struct btree_read_bio *rb =
1151 container_of(work, struct btree_read_bio, work);
1152 struct bch_fs *c = rb->c;
1153 struct btree *b = rb->b;
1154 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1155 struct bio *bio = &rb->bio;
1156 struct bch_io_failures failed = { .nr = 0 };
1157 struct printbuf buf = PRINTBUF;
1158 bool saw_error = false;
1165 bch_info(c, "retrying read");
1166 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1167 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1168 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1169 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1170 bio->bi_iter.bi_size = btree_bytes(c);
1172 if (rb->have_ioref) {
1173 bio_set_dev(bio, ca->disk_sb.bdev);
1174 submit_bio_wait(bio);
1176 bio->bi_status = BLK_STS_REMOVED;
1179 printbuf_reset(&buf);
1180 btree_pos_to_text(&buf, c, b);
1181 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1182 bch2_blk_status_to_str(bio->bi_status), buf.buf);
1184 percpu_ref_put(&ca->io_ref);
1185 rb->have_ioref = false;
1187 bch2_mark_io_failure(&failed, &rb->pick);
1189 can_retry = bch2_bkey_pick_read_device(c,
1190 bkey_i_to_s_c(&b->key),
1191 &failed, &rb->pick) > 0;
1193 if (!bio->bi_status &&
1194 !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1196 bch_info(c, "retry success");
1203 set_btree_node_read_error(b);
1208 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1212 if (saw_error && !btree_node_read_error(b)) {
1213 printbuf_reset(&buf);
1214 bch2_bpos_to_text(&buf, b->key.k.p);
1215 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1216 __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1218 bch2_btree_node_rewrite_async(c, b);
1221 printbuf_exit(&buf);
1222 clear_btree_node_read_in_flight(b);
1223 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1226 static void btree_node_read_endio(struct bio *bio)
1228 struct btree_read_bio *rb =
1229 container_of(bio, struct btree_read_bio, bio);
1230 struct bch_fs *c = rb->c;
1232 if (rb->have_ioref) {
1233 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1235 bch2_latency_acct(ca, rb->start_time, READ);
1238 queue_work(c->io_complete_wq, &rb->work);
1241 struct btree_node_read_all {
1246 void *buf[BCH_REPLICAS_MAX];
1247 struct bio *bio[BCH_REPLICAS_MAX];
1248 blk_status_t err[BCH_REPLICAS_MAX];
1251 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1253 struct btree_node *bn = data;
1254 struct btree_node_entry *bne;
1255 unsigned offset = 0;
1257 if (le64_to_cpu(bn->magic) != bset_magic(c))
1260 while (offset < btree_sectors(c)) {
1262 offset += vstruct_sectors(bn, c->block_bits);
1264 bne = data + (offset << 9);
1265 if (bne->keys.seq != bn->keys.seq)
1267 offset += vstruct_sectors(bne, c->block_bits);
1274 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1276 struct btree_node *bn = data;
1277 struct btree_node_entry *bne;
1282 while (offset < btree_sectors(c)) {
1283 bne = data + (offset << 9);
1284 if (bne->keys.seq == bn->keys.seq)
1293 static void btree_node_read_all_replicas_done(struct closure *cl)
1295 struct btree_node_read_all *ra =
1296 container_of(cl, struct btree_node_read_all, cl);
1297 struct bch_fs *c = ra->c;
1298 struct btree *b = ra->b;
1299 struct printbuf buf = PRINTBUF;
1300 bool dump_bset_maps = false;
1301 bool have_retry = false;
1302 int ret = 0, best = -1, write = READ;
1303 unsigned i, written = 0, written2 = 0;
1304 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1305 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1306 bool _saw_error = false, *saw_error = &_saw_error;
1308 for (i = 0; i < ra->nr; i++) {
1309 struct btree_node *bn = ra->buf[i];
1314 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1315 (seq && seq != bn->keys.seq))
1320 written = btree_node_sectors_written(c, bn);
1324 written2 = btree_node_sectors_written(c, ra->buf[i]);
1325 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1326 "btree node sectors written mismatch: %u != %u",
1327 written, written2) ||
1328 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1329 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1330 "found bset signature after last bset") ||
1331 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1332 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1333 "btree node replicas content mismatch"))
1334 dump_bset_maps = true;
1336 if (written2 > written) {
1342 if (dump_bset_maps) {
1343 for (i = 0; i < ra->nr; i++) {
1344 struct btree_node *bn = ra->buf[i];
1345 struct btree_node_entry *bne = NULL;
1346 unsigned offset = 0, sectors;
1352 printbuf_reset(&buf);
1354 while (offset < btree_sectors(c)) {
1356 sectors = vstruct_sectors(bn, c->block_bits);
1358 bne = ra->buf[i] + (offset << 9);
1359 if (bne->keys.seq != bn->keys.seq)
1361 sectors = vstruct_sectors(bne, c->block_bits);
1364 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1365 if (bne && bch2_journal_seq_is_blacklisted(c,
1366 le64_to_cpu(bne->keys.journal_seq), false))
1367 prt_printf(&buf, "*");
1371 while (offset < btree_sectors(c)) {
1372 bne = ra->buf[i] + (offset << 9);
1373 if (bne->keys.seq == bn->keys.seq) {
1375 prt_printf(&buf, " GAP");
1378 sectors = vstruct_sectors(bne, c->block_bits);
1379 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1380 if (bch2_journal_seq_is_blacklisted(c,
1381 le64_to_cpu(bne->keys.journal_seq), false))
1382 prt_printf(&buf, "*");
1387 bch_err(c, "replica %u:%s", i, buf.buf);
1392 memcpy(b->data, ra->buf[best], btree_bytes(c));
1393 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1399 set_btree_node_read_error(b);
1400 else if (*saw_error)
1401 bch2_btree_node_rewrite_async(c, b);
1403 for (i = 0; i < ra->nr; i++) {
1404 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1405 bio_put(ra->bio[i]);
1408 closure_debug_destroy(&ra->cl);
1410 printbuf_exit(&buf);
1412 clear_btree_node_read_in_flight(b);
1413 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1416 static void btree_node_read_all_replicas_endio(struct bio *bio)
1418 struct btree_read_bio *rb =
1419 container_of(bio, struct btree_read_bio, bio);
1420 struct bch_fs *c = rb->c;
1421 struct btree_node_read_all *ra = rb->ra;
1423 if (rb->have_ioref) {
1424 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1426 bch2_latency_acct(ca, rb->start_time, READ);
1429 ra->err[rb->idx] = bio->bi_status;
1430 closure_put(&ra->cl);
1434 * XXX This allocates multiple times from the same mempools, and can deadlock
1435 * under sufficient memory pressure (but is only a debug path)
1437 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1439 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1440 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1441 const union bch_extent_entry *entry;
1442 struct extent_ptr_decoded pick;
1443 struct btree_node_read_all *ra;
1446 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1448 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1450 closure_init(&ra->cl, NULL);
1453 ra->nr = bch2_bkey_nr_ptrs(k);
1455 for (i = 0; i < ra->nr; i++) {
1456 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1457 ra->bio[i] = bio_alloc_bioset(NULL,
1458 buf_pages(ra->buf[i], btree_bytes(c)),
1459 REQ_OP_READ|REQ_SYNC|REQ_META,
1465 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1466 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1467 struct btree_read_bio *rb =
1468 container_of(ra->bio[i], struct btree_read_bio, bio);
1472 rb->start_time = local_clock();
1473 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1476 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1477 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1478 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1480 if (rb->have_ioref) {
1481 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1482 bio_sectors(&rb->bio));
1483 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1485 closure_get(&ra->cl);
1486 submit_bio(&rb->bio);
1488 ra->err[i] = BLK_STS_REMOVED;
1495 closure_sync(&ra->cl);
1496 btree_node_read_all_replicas_done(&ra->cl);
1498 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1505 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1508 struct extent_ptr_decoded pick;
1509 struct btree_read_bio *rb;
1514 trace_and_count(c, btree_node_read, c, b);
1516 if (bch2_verify_all_btree_replicas &&
1517 !btree_node_read_all_replicas(c, b, sync))
1520 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1524 struct printbuf buf = PRINTBUF;
1526 prt_str(&buf, "btree node read error: no device to read from\n at ");
1527 btree_pos_to_text(&buf, c, b);
1528 bch_err(c, "%s", buf.buf);
1530 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1531 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1532 bch2_fatal_error(c);
1534 set_btree_node_read_error(b);
1535 clear_btree_node_read_in_flight(b);
1536 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1537 printbuf_exit(&buf);
1541 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1543 bio = bio_alloc_bioset(NULL,
1544 buf_pages(b->data, btree_bytes(c)),
1545 REQ_OP_READ|REQ_SYNC|REQ_META,
1548 rb = container_of(bio, struct btree_read_bio, bio);
1552 rb->start_time = local_clock();
1553 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1555 INIT_WORK(&rb->work, btree_node_read_work);
1556 bio->bi_iter.bi_sector = pick.ptr.offset;
1557 bio->bi_end_io = btree_node_read_endio;
1558 bch2_bio_map(bio, b->data, btree_bytes(c));
1560 if (rb->have_ioref) {
1561 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1563 bio_set_dev(bio, ca->disk_sb.bdev);
1566 submit_bio_wait(bio);
1568 btree_node_read_work(&rb->work);
1573 bio->bi_status = BLK_STS_REMOVED;
1576 btree_node_read_work(&rb->work);
1578 queue_work(c->io_complete_wq, &rb->work);
1582 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1583 const struct bkey_i *k, unsigned level)
1585 struct bch_fs *c = trans->c;
1590 closure_init_stack(&cl);
1593 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1597 b = bch2_btree_node_mem_alloc(trans, level != 0);
1598 bch2_btree_cache_cannibalize_unlock(c);
1602 bkey_copy(&b->key, k);
1603 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1605 set_btree_node_read_in_flight(b);
1607 bch2_btree_node_read(c, b, true);
1609 if (btree_node_read_error(b)) {
1610 bch2_btree_node_hash_remove(&c->btree_cache, b);
1612 mutex_lock(&c->btree_cache.lock);
1613 list_move(&b->list, &c->btree_cache.freeable);
1614 mutex_unlock(&c->btree_cache.lock);
1620 bch2_btree_set_root_for_read(c, b);
1622 six_unlock_write(&b->c.lock);
1623 six_unlock_intent(&b->c.lock);
1628 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1629 const struct bkey_i *k, unsigned level)
1631 return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1634 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1635 struct btree_write *w)
1637 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1645 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1648 closure_put(&((struct btree_update *) new)->cl);
1650 bch2_journal_pin_drop(&c->journal, &w->journal);
1653 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1655 struct btree_write *w = btree_prev_write(b);
1656 unsigned long old, new, v;
1659 bch2_btree_complete_write(c, b, w);
1661 v = READ_ONCE(b->flags);
1665 if ((old & (1U << BTREE_NODE_dirty)) &&
1666 (old & (1U << BTREE_NODE_need_write)) &&
1667 !(old & (1U << BTREE_NODE_never_write)) &&
1668 !(old & (1U << BTREE_NODE_write_blocked)) &&
1669 !(old & (1U << BTREE_NODE_will_make_reachable))) {
1670 new &= ~(1U << BTREE_NODE_dirty);
1671 new &= ~(1U << BTREE_NODE_need_write);
1672 new |= (1U << BTREE_NODE_write_in_flight);
1673 new |= (1U << BTREE_NODE_write_in_flight_inner);
1674 new |= (1U << BTREE_NODE_just_written);
1675 new ^= (1U << BTREE_NODE_write_idx);
1677 type = new & BTREE_WRITE_TYPE_MASK;
1678 new &= ~BTREE_WRITE_TYPE_MASK;
1680 new &= ~(1U << BTREE_NODE_write_in_flight);
1681 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1683 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1685 if (new & (1U << BTREE_NODE_write_in_flight))
1686 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1688 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1691 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1693 struct btree_trans *trans = bch2_trans_get(c);
1695 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1696 __btree_node_write_done(c, b);
1697 six_unlock_read(&b->c.lock);
1699 bch2_trans_put(trans);
1702 static void btree_node_write_work(struct work_struct *work)
1704 struct btree_write_bio *wbio =
1705 container_of(work, struct btree_write_bio, work);
1706 struct bch_fs *c = wbio->wbio.c;
1707 struct btree *b = wbio->wbio.bio.bi_private;
1708 struct bch_extent_ptr *ptr;
1711 btree_bounce_free(c,
1713 wbio->wbio.used_mempool,
1716 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1717 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1719 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1722 if (wbio->wbio.first_btree_write) {
1723 if (wbio->wbio.failed.nr) {
1727 ret = bch2_trans_do(c, NULL, NULL, 0,
1728 bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1729 BCH_WATERMARK_reclaim|
1730 BTREE_INSERT_JOURNAL_RECLAIM|
1731 BTREE_INSERT_NOFAIL|
1732 BTREE_INSERT_NOCHECK_RW,
1733 !wbio->wbio.failed.nr));
1738 bio_put(&wbio->wbio.bio);
1739 btree_node_write_done(c, b);
1742 set_btree_node_noevict(b);
1743 if (!bch2_err_matches(ret, EROFS))
1744 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1748 static void btree_node_write_endio(struct bio *bio)
1750 struct bch_write_bio *wbio = to_wbio(bio);
1751 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1752 struct bch_write_bio *orig = parent ?: wbio;
1753 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1754 struct bch_fs *c = wbio->c;
1755 struct btree *b = wbio->bio.bi_private;
1756 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1757 unsigned long flags;
1759 if (wbio->have_ioref)
1760 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1762 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1763 bch2_blk_status_to_str(bio->bi_status)) ||
1764 bch2_meta_write_fault("btree")) {
1765 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1766 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1767 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1770 if (wbio->have_ioref)
1771 percpu_ref_put(&ca->io_ref);
1775 bio_endio(&parent->bio);
1779 clear_btree_node_write_in_flight_inner(b);
1780 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1781 INIT_WORK(&wb->work, btree_node_write_work);
1782 queue_work(c->btree_io_complete_wq, &wb->work);
1785 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1786 struct bset *i, unsigned sectors)
1788 struct printbuf buf = PRINTBUF;
1792 ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1793 BKEY_TYPE_btree, WRITE, &buf);
1796 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1797 printbuf_exit(&buf);
1801 ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1802 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1804 bch2_inconsistent_error(c);
1811 static void btree_write_submit(struct work_struct *work)
1813 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1814 struct bch_extent_ptr *ptr;
1815 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1817 bkey_copy(&tmp.k, &wbio->key);
1819 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1820 ptr->offset += wbio->sector_offset;
1822 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1826 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1828 struct btree_write_bio *wbio;
1829 struct bset_tree *t;
1831 struct btree_node *bn = NULL;
1832 struct btree_node_entry *bne = NULL;
1833 struct sort_iter_stack sort_iter;
1835 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1838 unsigned long old, new;
1839 bool validate_before_checksum = false;
1840 enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1844 if (flags & BTREE_WRITE_ALREADY_STARTED)
1848 * We may only have a read lock on the btree node - the dirty bit is our
1849 * "lock" against racing with other threads that may be trying to start
1850 * a write, we do a write iff we clear the dirty bit. Since setting the
1851 * dirty bit requires a write lock, we can't race with other threads
1855 old = new = READ_ONCE(b->flags);
1857 if (!(old & (1 << BTREE_NODE_dirty)))
1860 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1861 !(old & (1 << BTREE_NODE_need_write)))
1865 ((1 << BTREE_NODE_never_write)|
1866 (1 << BTREE_NODE_write_blocked)))
1870 (old & (1 << BTREE_NODE_will_make_reachable)))
1873 if (old & (1 << BTREE_NODE_write_in_flight))
1876 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1877 type = new & BTREE_WRITE_TYPE_MASK;
1878 new &= ~BTREE_WRITE_TYPE_MASK;
1880 new &= ~(1 << BTREE_NODE_dirty);
1881 new &= ~(1 << BTREE_NODE_need_write);
1882 new |= (1 << BTREE_NODE_write_in_flight);
1883 new |= (1 << BTREE_NODE_write_in_flight_inner);
1884 new |= (1 << BTREE_NODE_just_written);
1885 new ^= (1 << BTREE_NODE_write_idx);
1886 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1888 if (new & (1U << BTREE_NODE_need_write))
1891 BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1893 atomic_dec(&c->btree_cache.dirty);
1895 BUG_ON(btree_node_fake(b));
1896 BUG_ON((b->will_make_reachable != 0) != !b->written);
1898 BUG_ON(b->written >= btree_sectors(c));
1899 BUG_ON(b->written & (block_sectors(c) - 1));
1900 BUG_ON(bset_written(b, btree_bset_last(b)));
1901 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1902 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1904 bch2_sort_whiteouts(c, b);
1906 sort_iter_stack_init(&sort_iter, b);
1909 ? sizeof(struct btree_node)
1910 : sizeof(struct btree_node_entry);
1912 bytes += b->whiteout_u64s * sizeof(u64);
1914 for_each_bset(b, t) {
1917 if (bset_written(b, i))
1920 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1921 sort_iter_add(&sort_iter.iter,
1922 btree_bkey_first(b, t),
1923 btree_bkey_last(b, t));
1924 seq = max(seq, le64_to_cpu(i->journal_seq));
1927 BUG_ON(b->written && !seq);
1929 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1932 /* buffer must be a multiple of the block size */
1933 bytes = round_up(bytes, block_bytes(c));
1935 data = btree_bounce_alloc(c, bytes, &used_mempool);
1943 bne->keys = b->data->keys;
1947 i->journal_seq = cpu_to_le64(seq);
1950 sort_iter_add(&sort_iter.iter,
1951 unwritten_whiteouts_start(c, b),
1952 unwritten_whiteouts_end(c, b));
1953 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1955 b->whiteout_u64s = 0;
1957 u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
1958 le16_add_cpu(&i->u64s, u64s);
1960 BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1962 set_needs_whiteout(i, false);
1964 /* do we have data to write? */
1965 if (b->written && !i->u64s)
1968 bytes_to_write = vstruct_end(i) - data;
1969 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1972 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1973 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1975 memset(data + bytes_to_write, 0,
1976 (sectors_to_write << 9) - bytes_to_write);
1978 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1979 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1980 BUG_ON(i->seq != b->data->keys.seq);
1982 i->version = cpu_to_le16(c->sb.version);
1983 SET_BSET_OFFSET(i, b->written);
1984 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1986 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1987 validate_before_checksum = true;
1989 /* validate_bset will be modifying: */
1990 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1991 validate_before_checksum = true;
1993 /* if we're going to be encrypting, check metadata validity first: */
1994 if (validate_before_checksum &&
1995 validate_bset_for_write(c, b, i, sectors_to_write))
1998 ret = bset_encrypt(c, i, b->written << 9);
1999 if (bch2_fs_fatal_err_on(ret, c,
2000 "error encrypting btree node: %i\n", ret))
2003 nonce = btree_nonce(i, b->written << 9);
2006 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2008 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2010 /* if we're not encrypting, check metadata after checksumming: */
2011 if (!validate_before_checksum &&
2012 validate_bset_for_write(c, b, i, sectors_to_write))
2016 * We handle btree write errors by immediately halting the journal -
2017 * after we've done that, we can't issue any subsequent btree writes
2018 * because they might have pointers to new nodes that failed to write.
2020 * Furthermore, there's no point in doing any more btree writes because
2021 * with the journal stopped, we're never going to update the journal to
2022 * reflect that those writes were done and the data flushed from the
2025 * Also on journal error, the pending write may have updates that were
2026 * never journalled (interior nodes, see btree_update_nodes_written()) -
2027 * it's critical that we don't do the write in that case otherwise we
2028 * will have updates visible that weren't in the journal:
2030 * Make sure to update b->written so bch2_btree_init_next() doesn't
2033 if (bch2_journal_error(&c->journal) ||
2037 trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2039 wbio = container_of(bio_alloc_bioset(NULL,
2040 buf_pages(data, sectors_to_write << 9),
2041 REQ_OP_WRITE|REQ_META,
2044 struct btree_write_bio, wbio.bio);
2045 wbio_init(&wbio->wbio.bio);
2047 wbio->data_bytes = bytes;
2048 wbio->sector_offset = b->written;
2050 wbio->wbio.used_mempool = used_mempool;
2051 wbio->wbio.first_btree_write = !b->written;
2052 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
2053 wbio->wbio.bio.bi_private = b;
2055 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2057 bkey_copy(&wbio->key, &b->key);
2059 b->written += sectors_to_write;
2061 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2062 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2063 cpu_to_le16(b->written);
2065 atomic64_inc(&c->btree_write_stats[type].nr);
2066 atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2068 INIT_WORK(&wbio->work, btree_write_submit);
2069 queue_work(c->io_complete_wq, &wbio->work);
2072 set_btree_node_noevict(b);
2073 b->written += sectors_to_write;
2075 btree_bounce_free(c, bytes, used_mempool, data);
2076 __btree_node_write_done(c, b);
2080 * Work that must be done with write lock held:
2082 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2084 bool invalidated_iter = false;
2085 struct btree_node_entry *bne;
2086 struct bset_tree *t;
2088 if (!btree_node_just_written(b))
2091 BUG_ON(b->whiteout_u64s);
2093 clear_btree_node_just_written(b);
2096 * Note: immediately after write, bset_written() doesn't work - the
2097 * amount of data we had to write after compaction might have been
2098 * smaller than the offset of the last bset.
2100 * However, we know that all bsets have been written here, as long as
2101 * we're still holding the write lock:
2105 * XXX: decide if we really want to unconditionally sort down to a
2109 btree_node_sort(c, b, 0, b->nsets, true);
2110 invalidated_iter = true;
2112 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2116 set_needs_whiteout(bset(b, t), true);
2118 bch2_btree_verify(c, b);
2121 * If later we don't unconditionally sort down to a single bset, we have
2122 * to ensure this is still true:
2124 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2126 bne = want_new_bset(c, b);
2128 bch2_bset_init_next(c, b, bne);
2130 bch2_btree_build_aux_trees(b);
2132 return invalidated_iter;
2136 * Use this one if the node is intent locked:
2138 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2139 enum six_lock_type lock_type_held,
2142 if (lock_type_held == SIX_LOCK_intent ||
2143 (lock_type_held == SIX_LOCK_read &&
2144 six_lock_tryupgrade(&b->c.lock))) {
2145 __bch2_btree_node_write(c, b, flags);
2147 /* don't cycle lock unnecessarily: */
2148 if (btree_node_just_written(b) &&
2149 six_trylock_write(&b->c.lock)) {
2150 bch2_btree_post_write_cleanup(c, b);
2151 six_unlock_write(&b->c.lock);
2154 if (lock_type_held == SIX_LOCK_read)
2155 six_lock_downgrade(&b->c.lock);
2157 __bch2_btree_node_write(c, b, flags);
2158 if (lock_type_held == SIX_LOCK_write &&
2159 btree_node_just_written(b))
2160 bch2_btree_post_write_cleanup(c, b);
2164 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2166 struct bucket_table *tbl;
2167 struct rhash_head *pos;
2173 for_each_cached_btree(b, c, tbl, i, pos)
2174 if (test_bit(flag, &b->flags)) {
2176 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2185 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2187 return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2190 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2192 return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2195 static const char * const bch2_btree_write_types[] = {
2196 #define x(t, n) [n] = #t,
2197 BCH_BTREE_WRITE_TYPES()
2201 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2203 printbuf_tabstop_push(out, 20);
2204 printbuf_tabstop_push(out, 10);
2209 prt_str(out, "size");
2212 for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2213 u64 nr = atomic64_read(&c->btree_write_stats[i].nr);
2214 u64 bytes = atomic64_read(&c->btree_write_stats[i].bytes);
2216 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2220 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);