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_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", b->written);
529 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
534 static int __btree_err(int ret,
541 const char *fmt, ...)
543 struct printbuf out = PRINTBUF;
546 btree_err_msg(&out, c, ca, b, i, b->written, write);
549 prt_vprintf(&out, fmt, args);
552 if (write == WRITE) {
553 bch2_print_string_as_lines(KERN_ERR, out.buf);
554 ret = c->opts.errors == BCH_ON_ERROR_continue
556 : -BCH_ERR_fsck_errors_not_fixed;
560 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
561 ret = -BCH_ERR_btree_node_read_err_fixable;
562 if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
563 ret = -BCH_ERR_btree_node_read_err_bad_node;
566 case -BCH_ERR_btree_node_read_err_fixable:
567 mustfix_fsck_err(c, "%s", out.buf);
568 ret = -BCH_ERR_fsck_fix;
570 case -BCH_ERR_btree_node_read_err_want_retry:
571 case -BCH_ERR_btree_node_read_err_must_retry:
572 bch2_print_string_as_lines(KERN_ERR, out.buf);
574 case -BCH_ERR_btree_node_read_err_bad_node:
575 bch2_print_string_as_lines(KERN_ERR, out.buf);
576 bch2_topology_error(c);
577 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology) ?: -EIO;
579 case -BCH_ERR_btree_node_read_err_incompatible:
580 bch2_print_string_as_lines(KERN_ERR, out.buf);
581 ret = -BCH_ERR_fsck_errors_not_fixed;
592 #define btree_err(type, c, ca, b, i, msg, ...) \
594 int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
596 if (_ret != -BCH_ERR_fsck_fix) { \
604 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
607 * When btree topology repair changes the start or end of a node, that might
608 * mean we have to drop keys that are no longer inside the node:
611 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
615 for_each_bset(b, t) {
616 struct bset *i = bset(b, t);
617 struct bkey_packed *k;
619 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
620 if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
624 unsigned shift = (u64 *) k - (u64 *) i->start;
626 memmove_u64s_down(i->start, k,
627 (u64 *) vstruct_end(i) - (u64 *) k);
628 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
629 set_btree_bset_end(b, t);
632 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
633 if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
636 if (k != vstruct_last(i)) {
637 i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
638 set_btree_bset_end(b, t);
643 * Always rebuild search trees: eytzinger search tree nodes directly
644 * depend on the values of min/max key:
646 bch2_bset_set_no_aux_tree(b, b->set);
647 bch2_btree_build_aux_trees(b);
650 struct bkey unpacked;
651 struct btree_node_iter iter;
652 for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
653 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
654 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
658 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
659 struct btree *b, struct bset *i,
660 unsigned offset, unsigned sectors,
661 int write, bool have_retry, bool *saw_error)
663 unsigned version = le16_to_cpu(i->version);
664 struct printbuf buf1 = PRINTBUF;
665 struct printbuf buf2 = PRINTBUF;
668 btree_err_on(!bch2_version_compatible(version),
669 -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
670 "unsupported bset version %u.%u",
671 BCH_VERSION_MAJOR(version),
672 BCH_VERSION_MINOR(version));
674 if (btree_err_on(version < c->sb.version_min,
675 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
676 "bset version %u older than superblock version_min %u",
677 version, c->sb.version_min)) {
678 mutex_lock(&c->sb_lock);
679 c->disk_sb.sb->version_min = cpu_to_le16(version);
681 mutex_unlock(&c->sb_lock);
684 if (btree_err_on(BCH_VERSION_MAJOR(version) >
685 BCH_VERSION_MAJOR(c->sb.version),
686 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
687 "bset version %u newer than superblock version %u",
688 version, c->sb.version)) {
689 mutex_lock(&c->sb_lock);
690 c->disk_sb.sb->version = cpu_to_le16(version);
692 mutex_unlock(&c->sb_lock);
695 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
696 -BCH_ERR_btree_node_read_err_incompatible, c, ca, b, i,
697 "BSET_SEPARATE_WHITEOUTS no longer supported");
699 if (btree_err_on(offset + sectors > btree_sectors(c),
700 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
701 "bset past end of btree node")) {
707 btree_err_on(offset && !i->u64s,
708 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
711 btree_err_on(BSET_OFFSET(i) &&
712 BSET_OFFSET(i) != offset,
713 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
714 "bset at wrong sector offset");
717 struct btree_node *bn =
718 container_of(i, struct btree_node, keys);
719 /* These indicate that we read the wrong btree node: */
721 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
722 struct bch_btree_ptr_v2 *bp =
723 &bkey_i_to_btree_ptr_v2(&b->key)->v;
726 btree_err_on(bp->seq != bn->keys.seq,
727 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
728 "incorrect sequence number (wrong btree node)");
731 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
732 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
733 "incorrect btree id");
735 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
736 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
740 compat_btree_node(b->c.level, b->c.btree_id, version,
741 BSET_BIG_ENDIAN(i), write, bn);
743 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
744 struct bch_btree_ptr_v2 *bp =
745 &bkey_i_to_btree_ptr_v2(&b->key)->v;
747 if (BTREE_PTR_RANGE_UPDATED(bp)) {
748 b->data->min_key = bp->min_key;
749 b->data->max_key = b->key.k.p;
752 btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
753 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
754 "incorrect min_key: got %s should be %s",
755 (printbuf_reset(&buf1),
756 bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
757 (printbuf_reset(&buf2),
758 bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
761 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
762 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, i,
763 "incorrect max key %s",
764 (printbuf_reset(&buf1),
765 bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
768 compat_btree_node(b->c.level, b->c.btree_id, version,
769 BSET_BIG_ENDIAN(i), write, bn);
771 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
772 -BCH_ERR_btree_node_read_err_bad_node, c, ca, b, i,
773 "invalid bkey format: %s\n %s", buf1.buf,
774 (printbuf_reset(&buf2),
775 bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
776 printbuf_reset(&buf1);
778 compat_bformat(b->c.level, b->c.btree_id, version,
779 BSET_BIG_ENDIAN(i), write,
784 printbuf_exit(&buf2);
785 printbuf_exit(&buf1);
789 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
791 bool updated_range, int rw,
792 struct printbuf *err)
794 return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
795 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
796 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
799 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
800 struct bset *i, int write,
801 bool have_retry, bool *saw_error)
803 unsigned version = le16_to_cpu(i->version);
804 struct bkey_packed *k, *prev = NULL;
805 struct printbuf buf = PRINTBUF;
806 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
807 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
811 k != vstruct_last(i);) {
815 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
816 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
817 "key extends past end of bset")) {
818 i->u64s = cpu_to_le16((u64 *) k - i->_data);
822 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
823 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i,
824 "invalid bkey format %u", k->format)) {
825 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
826 memmove_u64s_down(k, bkey_p_next(k),
827 (u64 *) vstruct_end(i) - (u64 *) k);
831 /* XXX: validate k->u64s */
833 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
834 BSET_BIG_ENDIAN(i), write,
837 u = __bkey_disassemble(b, k, &tmp);
839 printbuf_reset(&buf);
840 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
841 printbuf_reset(&buf);
842 prt_printf(&buf, "invalid bkey: ");
843 bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
844 prt_printf(&buf, "\n ");
845 bch2_bkey_val_to_text(&buf, c, u.s_c);
847 btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
849 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
850 memmove_u64s_down(k, bkey_p_next(k),
851 (u64 *) vstruct_end(i) - (u64 *) k);
856 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
857 BSET_BIG_ENDIAN(i), write,
860 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
861 struct bkey up = bkey_unpack_key(b, prev);
863 printbuf_reset(&buf);
864 prt_printf(&buf, "keys out of order: ");
865 bch2_bkey_to_text(&buf, &up);
866 prt_printf(&buf, " > ");
867 bch2_bkey_to_text(&buf, u.k);
869 bch2_dump_bset(c, b, i, 0);
871 if (btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf)) {
872 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
873 memmove_u64s_down(k, bkey_p_next(k),
874 (u64 *) vstruct_end(i) - (u64 *) k);
887 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
888 struct btree *b, bool have_retry, bool *saw_error)
890 struct btree_node_entry *bne;
891 struct sort_iter *iter;
892 struct btree_node *sorted;
893 struct bkey_packed *k;
894 struct bch_extent_ptr *ptr;
896 bool used_mempool, blacklisted;
897 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
898 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
900 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
901 struct printbuf buf = PRINTBUF;
902 int ret = 0, retry_read = 0, write = READ;
904 b->version_ondisk = U16_MAX;
905 /* We might get called multiple times on read retry: */
908 iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
909 sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
911 if (bch2_meta_read_fault("btree"))
912 btree_err(-BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
915 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
916 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
917 "bad magic: want %llx, got %llx",
918 bset_magic(c), le64_to_cpu(b->data->magic));
920 btree_err_on(!b->data->keys.seq,
921 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
922 "bad btree header: seq 0");
924 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
925 struct bch_btree_ptr_v2 *bp =
926 &bkey_i_to_btree_ptr_v2(&b->key)->v;
928 btree_err_on(b->data->keys.seq != bp->seq,
929 -BCH_ERR_btree_node_read_err_must_retry, c, ca, b, NULL,
930 "got wrong btree node (seq %llx want %llx)",
931 b->data->keys.seq, bp->seq);
934 while (b->written < (ptr_written ?: btree_sectors(c))) {
937 struct bch_csum csum;
938 bool first = !b->written;
943 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
944 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
945 "unknown checksum type %llu",
948 nonce = btree_nonce(i, b->written << 9);
949 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
951 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
952 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
955 ret = bset_encrypt(c, i, b->written << 9);
956 if (bch2_fs_fatal_err_on(ret, c,
957 "error decrypting btree node: %i", ret))
960 btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
961 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
962 -BCH_ERR_btree_node_read_err_incompatible, c, NULL, b, NULL,
963 "btree node does not have NEW_EXTENT_OVERWRITE set");
965 sectors = vstruct_sectors(b->data, c->block_bits);
967 bne = write_block(b);
970 if (i->seq != b->data->keys.seq)
973 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
974 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
975 "unknown checksum type %llu",
978 nonce = btree_nonce(i, b->written << 9);
979 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
981 btree_err_on(bch2_crc_cmp(csum, bne->csum),
982 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, i,
985 ret = bset_encrypt(c, i, b->written << 9);
986 if (bch2_fs_fatal_err_on(ret, c,
987 "error decrypting btree node: %i\n", ret))
990 sectors = vstruct_sectors(bne, c->block_bits);
993 b->version_ondisk = min(b->version_ondisk,
994 le16_to_cpu(i->version));
996 ret = validate_bset(c, ca, b, i, b->written, sectors,
997 READ, have_retry, saw_error);
1002 btree_node_set_format(b, b->data->format);
1004 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1008 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1010 blacklisted = bch2_journal_seq_is_blacklisted(c,
1011 le64_to_cpu(i->journal_seq),
1014 btree_err_on(blacklisted && first,
1015 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1016 "first btree node bset has blacklisted journal seq (%llu)",
1017 le64_to_cpu(i->journal_seq));
1019 btree_err_on(blacklisted && ptr_written,
1020 -BCH_ERR_btree_node_read_err_fixable, c, ca, b, i,
1021 "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1022 le64_to_cpu(i->journal_seq),
1023 b->written, b->written + sectors, ptr_written);
1025 b->written += sectors;
1027 if (blacklisted && !first)
1036 btree_err_on(b->written < ptr_written,
1037 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1038 "btree node data missing: expected %u sectors, found %u",
1039 ptr_written, b->written);
1041 for (bne = write_block(b);
1042 bset_byte_offset(b, bne) < btree_bytes(c);
1043 bne = (void *) bne + block_bytes(c))
1044 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1045 !bch2_journal_seq_is_blacklisted(c,
1046 le64_to_cpu(bne->keys.journal_seq),
1048 -BCH_ERR_btree_node_read_err_want_retry, c, ca, b, NULL,
1049 "found bset signature after last bset");
1052 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1053 sorted->keys.u64s = 0;
1055 set_btree_bset(b, b->set, &b->data->keys);
1057 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1059 u64s = le16_to_cpu(sorted->keys.u64s);
1061 sorted->keys.u64s = cpu_to_le16(u64s);
1062 swap(sorted, b->data);
1063 set_btree_bset(b, b->set, &b->data->keys);
1066 BUG_ON(b->nr.live_u64s != u64s);
1068 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1071 bch2_btree_node_drop_keys_outside_node(b);
1074 for (k = i->start; k != vstruct_last(i);) {
1076 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1078 printbuf_reset(&buf);
1080 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1081 (bch2_inject_invalid_keys &&
1082 !bversion_cmp(u.k->version, MAX_VERSION))) {
1083 printbuf_reset(&buf);
1085 prt_printf(&buf, "invalid bkey: ");
1086 bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1087 prt_printf(&buf, "\n ");
1088 bch2_bkey_val_to_text(&buf, c, u.s_c);
1090 btree_err(-BCH_ERR_btree_node_read_err_fixable, c, NULL, b, i, "%s", buf.buf);
1092 btree_keys_account_key_drop(&b->nr, 0, k);
1094 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1095 memmove_u64s_down(k, bkey_p_next(k),
1096 (u64 *) vstruct_end(i) - (u64 *) k);
1097 set_btree_bset_end(b, b->set);
1101 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1102 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1110 bch2_bset_build_aux_tree(b, b->set, false);
1112 set_needs_whiteout(btree_bset_first(b), true);
1114 btree_node_reset_sib_u64s(b);
1116 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1117 struct bch_dev *ca2 = bch_dev_bkey_exists(c, ptr->dev);
1119 if (ca2->mi.state != BCH_MEMBER_STATE_rw)
1120 set_btree_node_need_rewrite(b);
1124 set_btree_node_need_rewrite(b);
1126 mempool_free(iter, &c->fill_iter);
1127 printbuf_exit(&buf);
1130 if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1131 ret == -BCH_ERR_btree_node_read_err_must_retry)
1134 set_btree_node_read_error(b);
1138 static void btree_node_read_work(struct work_struct *work)
1140 struct btree_read_bio *rb =
1141 container_of(work, struct btree_read_bio, work);
1142 struct bch_fs *c = rb->c;
1143 struct btree *b = rb->b;
1144 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1145 struct bio *bio = &rb->bio;
1146 struct bch_io_failures failed = { .nr = 0 };
1147 struct printbuf buf = PRINTBUF;
1148 bool saw_error = false;
1155 bch_info(c, "retrying read");
1156 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1157 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1158 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1159 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1160 bio->bi_iter.bi_size = btree_bytes(c);
1162 if (rb->have_ioref) {
1163 bio_set_dev(bio, ca->disk_sb.bdev);
1164 submit_bio_wait(bio);
1166 bio->bi_status = BLK_STS_REMOVED;
1169 printbuf_reset(&buf);
1170 bch2_btree_pos_to_text(&buf, c, b);
1171 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1172 bch2_blk_status_to_str(bio->bi_status), buf.buf);
1174 percpu_ref_put(&ca->io_ref);
1175 rb->have_ioref = false;
1177 bch2_mark_io_failure(&failed, &rb->pick);
1179 can_retry = bch2_bkey_pick_read_device(c,
1180 bkey_i_to_s_c(&b->key),
1181 &failed, &rb->pick) > 0;
1183 if (!bio->bi_status &&
1184 !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1186 bch_info(c, "retry success");
1193 set_btree_node_read_error(b);
1198 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1202 if (saw_error && !btree_node_read_error(b)) {
1203 printbuf_reset(&buf);
1204 bch2_bpos_to_text(&buf, b->key.k.p);
1205 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1206 __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1208 bch2_btree_node_rewrite_async(c, b);
1211 printbuf_exit(&buf);
1212 clear_btree_node_read_in_flight(b);
1213 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1216 static void btree_node_read_endio(struct bio *bio)
1218 struct btree_read_bio *rb =
1219 container_of(bio, struct btree_read_bio, bio);
1220 struct bch_fs *c = rb->c;
1222 if (rb->have_ioref) {
1223 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1225 bch2_latency_acct(ca, rb->start_time, READ);
1228 queue_work(c->io_complete_wq, &rb->work);
1231 struct btree_node_read_all {
1236 void *buf[BCH_REPLICAS_MAX];
1237 struct bio *bio[BCH_REPLICAS_MAX];
1238 blk_status_t err[BCH_REPLICAS_MAX];
1241 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1243 struct btree_node *bn = data;
1244 struct btree_node_entry *bne;
1245 unsigned offset = 0;
1247 if (le64_to_cpu(bn->magic) != bset_magic(c))
1250 while (offset < btree_sectors(c)) {
1252 offset += vstruct_sectors(bn, c->block_bits);
1254 bne = data + (offset << 9);
1255 if (bne->keys.seq != bn->keys.seq)
1257 offset += vstruct_sectors(bne, c->block_bits);
1264 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1266 struct btree_node *bn = data;
1267 struct btree_node_entry *bne;
1272 while (offset < btree_sectors(c)) {
1273 bne = data + (offset << 9);
1274 if (bne->keys.seq == bn->keys.seq)
1283 static void btree_node_read_all_replicas_done(struct closure *cl)
1285 struct btree_node_read_all *ra =
1286 container_of(cl, struct btree_node_read_all, cl);
1287 struct bch_fs *c = ra->c;
1288 struct btree *b = ra->b;
1289 struct printbuf buf = PRINTBUF;
1290 bool dump_bset_maps = false;
1291 bool have_retry = false;
1292 int ret = 0, best = -1, write = READ;
1293 unsigned i, written = 0, written2 = 0;
1294 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1295 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1296 bool _saw_error = false, *saw_error = &_saw_error;
1298 for (i = 0; i < ra->nr; i++) {
1299 struct btree_node *bn = ra->buf[i];
1304 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1305 (seq && seq != bn->keys.seq))
1310 written = btree_node_sectors_written(c, bn);
1314 written2 = btree_node_sectors_written(c, ra->buf[i]);
1315 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1316 "btree node sectors written mismatch: %u != %u",
1317 written, written2) ||
1318 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1319 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1320 "found bset signature after last bset") ||
1321 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1322 -BCH_ERR_btree_node_read_err_fixable, c, NULL, b, NULL,
1323 "btree node replicas content mismatch"))
1324 dump_bset_maps = true;
1326 if (written2 > written) {
1332 if (dump_bset_maps) {
1333 for (i = 0; i < ra->nr; i++) {
1334 struct btree_node *bn = ra->buf[i];
1335 struct btree_node_entry *bne = NULL;
1336 unsigned offset = 0, sectors;
1342 printbuf_reset(&buf);
1344 while (offset < btree_sectors(c)) {
1346 sectors = vstruct_sectors(bn, c->block_bits);
1348 bne = ra->buf[i] + (offset << 9);
1349 if (bne->keys.seq != bn->keys.seq)
1351 sectors = vstruct_sectors(bne, c->block_bits);
1354 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1355 if (bne && bch2_journal_seq_is_blacklisted(c,
1356 le64_to_cpu(bne->keys.journal_seq), false))
1357 prt_printf(&buf, "*");
1361 while (offset < btree_sectors(c)) {
1362 bne = ra->buf[i] + (offset << 9);
1363 if (bne->keys.seq == bn->keys.seq) {
1365 prt_printf(&buf, " GAP");
1368 sectors = vstruct_sectors(bne, c->block_bits);
1369 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1370 if (bch2_journal_seq_is_blacklisted(c,
1371 le64_to_cpu(bne->keys.journal_seq), false))
1372 prt_printf(&buf, "*");
1377 bch_err(c, "replica %u:%s", i, buf.buf);
1382 memcpy(b->data, ra->buf[best], btree_bytes(c));
1383 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1389 set_btree_node_read_error(b);
1390 else if (*saw_error)
1391 bch2_btree_node_rewrite_async(c, b);
1393 for (i = 0; i < ra->nr; i++) {
1394 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1395 bio_put(ra->bio[i]);
1398 closure_debug_destroy(&ra->cl);
1400 printbuf_exit(&buf);
1402 clear_btree_node_read_in_flight(b);
1403 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1406 static void btree_node_read_all_replicas_endio(struct bio *bio)
1408 struct btree_read_bio *rb =
1409 container_of(bio, struct btree_read_bio, bio);
1410 struct bch_fs *c = rb->c;
1411 struct btree_node_read_all *ra = rb->ra;
1413 if (rb->have_ioref) {
1414 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1416 bch2_latency_acct(ca, rb->start_time, READ);
1419 ra->err[rb->idx] = bio->bi_status;
1420 closure_put(&ra->cl);
1424 * XXX This allocates multiple times from the same mempools, and can deadlock
1425 * under sufficient memory pressure (but is only a debug path)
1427 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1429 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1430 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1431 const union bch_extent_entry *entry;
1432 struct extent_ptr_decoded pick;
1433 struct btree_node_read_all *ra;
1436 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1438 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1440 closure_init(&ra->cl, NULL);
1443 ra->nr = bch2_bkey_nr_ptrs(k);
1445 for (i = 0; i < ra->nr; i++) {
1446 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1447 ra->bio[i] = bio_alloc_bioset(NULL,
1448 buf_pages(ra->buf[i], btree_bytes(c)),
1449 REQ_OP_READ|REQ_SYNC|REQ_META,
1455 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1456 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1457 struct btree_read_bio *rb =
1458 container_of(ra->bio[i], struct btree_read_bio, bio);
1462 rb->start_time = local_clock();
1463 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1466 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1467 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1468 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1470 if (rb->have_ioref) {
1471 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1472 bio_sectors(&rb->bio));
1473 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1475 closure_get(&ra->cl);
1476 submit_bio(&rb->bio);
1478 ra->err[i] = BLK_STS_REMOVED;
1485 closure_sync(&ra->cl);
1486 btree_node_read_all_replicas_done(&ra->cl);
1488 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1495 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1498 struct extent_ptr_decoded pick;
1499 struct btree_read_bio *rb;
1504 trace_and_count(c, btree_node_read, c, b);
1506 if (bch2_verify_all_btree_replicas &&
1507 !btree_node_read_all_replicas(c, b, sync))
1510 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1514 struct printbuf buf = PRINTBUF;
1516 prt_str(&buf, "btree node read error: no device to read from\n at ");
1517 bch2_btree_pos_to_text(&buf, c, b);
1518 bch_err(c, "%s", buf.buf);
1520 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1521 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1522 bch2_fatal_error(c);
1524 set_btree_node_read_error(b);
1525 clear_btree_node_read_in_flight(b);
1526 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1527 printbuf_exit(&buf);
1531 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1533 bio = bio_alloc_bioset(NULL,
1534 buf_pages(b->data, btree_bytes(c)),
1535 REQ_OP_READ|REQ_SYNC|REQ_META,
1538 rb = container_of(bio, struct btree_read_bio, bio);
1542 rb->start_time = local_clock();
1543 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1545 INIT_WORK(&rb->work, btree_node_read_work);
1546 bio->bi_iter.bi_sector = pick.ptr.offset;
1547 bio->bi_end_io = btree_node_read_endio;
1548 bch2_bio_map(bio, b->data, btree_bytes(c));
1550 if (rb->have_ioref) {
1551 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1553 bio_set_dev(bio, ca->disk_sb.bdev);
1556 submit_bio_wait(bio);
1558 btree_node_read_work(&rb->work);
1563 bio->bi_status = BLK_STS_REMOVED;
1566 btree_node_read_work(&rb->work);
1568 queue_work(c->io_complete_wq, &rb->work);
1572 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1573 const struct bkey_i *k, unsigned level)
1575 struct bch_fs *c = trans->c;
1580 closure_init_stack(&cl);
1583 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1587 b = bch2_btree_node_mem_alloc(trans, level != 0);
1588 bch2_btree_cache_cannibalize_unlock(c);
1592 bkey_copy(&b->key, k);
1593 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1595 set_btree_node_read_in_flight(b);
1597 bch2_btree_node_read(c, b, true);
1599 if (btree_node_read_error(b)) {
1600 bch2_btree_node_hash_remove(&c->btree_cache, b);
1602 mutex_lock(&c->btree_cache.lock);
1603 list_move(&b->list, &c->btree_cache.freeable);
1604 mutex_unlock(&c->btree_cache.lock);
1610 bch2_btree_set_root_for_read(c, b);
1612 six_unlock_write(&b->c.lock);
1613 six_unlock_intent(&b->c.lock);
1618 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1619 const struct bkey_i *k, unsigned level)
1621 return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1624 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1625 struct btree_write *w)
1627 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1635 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1638 closure_put(&((struct btree_update *) new)->cl);
1640 bch2_journal_pin_drop(&c->journal, &w->journal);
1643 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1645 struct btree_write *w = btree_prev_write(b);
1646 unsigned long old, new, v;
1649 bch2_btree_complete_write(c, b, w);
1651 v = READ_ONCE(b->flags);
1655 if ((old & (1U << BTREE_NODE_dirty)) &&
1656 (old & (1U << BTREE_NODE_need_write)) &&
1657 !(old & (1U << BTREE_NODE_never_write)) &&
1658 !(old & (1U << BTREE_NODE_write_blocked)) &&
1659 !(old & (1U << BTREE_NODE_will_make_reachable))) {
1660 new &= ~(1U << BTREE_NODE_dirty);
1661 new &= ~(1U << BTREE_NODE_need_write);
1662 new |= (1U << BTREE_NODE_write_in_flight);
1663 new |= (1U << BTREE_NODE_write_in_flight_inner);
1664 new |= (1U << BTREE_NODE_just_written);
1665 new ^= (1U << BTREE_NODE_write_idx);
1667 type = new & BTREE_WRITE_TYPE_MASK;
1668 new &= ~BTREE_WRITE_TYPE_MASK;
1670 new &= ~(1U << BTREE_NODE_write_in_flight);
1671 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1673 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1675 if (new & (1U << BTREE_NODE_write_in_flight))
1676 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1678 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1681 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1683 struct btree_trans *trans = bch2_trans_get(c);
1685 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1686 __btree_node_write_done(c, b);
1687 six_unlock_read(&b->c.lock);
1689 bch2_trans_put(trans);
1692 static void btree_node_write_work(struct work_struct *work)
1694 struct btree_write_bio *wbio =
1695 container_of(work, struct btree_write_bio, work);
1696 struct bch_fs *c = wbio->wbio.c;
1697 struct btree *b = wbio->wbio.bio.bi_private;
1698 struct bch_extent_ptr *ptr;
1701 btree_bounce_free(c,
1703 wbio->wbio.used_mempool,
1706 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1707 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1709 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1712 if (wbio->wbio.first_btree_write) {
1713 if (wbio->wbio.failed.nr) {
1717 ret = bch2_trans_do(c, NULL, NULL, 0,
1718 bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1719 BCH_WATERMARK_reclaim|
1720 BTREE_INSERT_JOURNAL_RECLAIM|
1721 BTREE_INSERT_NOFAIL|
1722 BTREE_INSERT_NOCHECK_RW,
1723 !wbio->wbio.failed.nr));
1728 bio_put(&wbio->wbio.bio);
1729 btree_node_write_done(c, b);
1732 set_btree_node_noevict(b);
1733 if (!bch2_err_matches(ret, EROFS))
1734 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1738 static void btree_node_write_endio(struct bio *bio)
1740 struct bch_write_bio *wbio = to_wbio(bio);
1741 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1742 struct bch_write_bio *orig = parent ?: wbio;
1743 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1744 struct bch_fs *c = wbio->c;
1745 struct btree *b = wbio->bio.bi_private;
1746 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1747 unsigned long flags;
1749 if (wbio->have_ioref)
1750 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1752 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1753 bch2_blk_status_to_str(bio->bi_status)) ||
1754 bch2_meta_write_fault("btree")) {
1755 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1756 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1757 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1760 if (wbio->have_ioref)
1761 percpu_ref_put(&ca->io_ref);
1765 bio_endio(&parent->bio);
1769 clear_btree_node_write_in_flight_inner(b);
1770 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1771 INIT_WORK(&wb->work, btree_node_write_work);
1772 queue_work(c->btree_io_complete_wq, &wb->work);
1775 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1776 struct bset *i, unsigned sectors)
1778 struct printbuf buf = PRINTBUF;
1782 ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1783 BKEY_TYPE_btree, WRITE, &buf);
1786 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1787 printbuf_exit(&buf);
1791 ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1792 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1794 bch2_inconsistent_error(c);
1801 static void btree_write_submit(struct work_struct *work)
1803 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1804 struct bch_extent_ptr *ptr;
1805 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1807 bkey_copy(&tmp.k, &wbio->key);
1809 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1810 ptr->offset += wbio->sector_offset;
1812 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1816 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1818 struct btree_write_bio *wbio;
1819 struct bset_tree *t;
1821 struct btree_node *bn = NULL;
1822 struct btree_node_entry *bne = NULL;
1823 struct sort_iter_stack sort_iter;
1825 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1828 unsigned long old, new;
1829 bool validate_before_checksum = false;
1830 enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1834 if (flags & BTREE_WRITE_ALREADY_STARTED)
1838 * We may only have a read lock on the btree node - the dirty bit is our
1839 * "lock" against racing with other threads that may be trying to start
1840 * a write, we do a write iff we clear the dirty bit. Since setting the
1841 * dirty bit requires a write lock, we can't race with other threads
1845 old = new = READ_ONCE(b->flags);
1847 if (!(old & (1 << BTREE_NODE_dirty)))
1850 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1851 !(old & (1 << BTREE_NODE_need_write)))
1855 ((1 << BTREE_NODE_never_write)|
1856 (1 << BTREE_NODE_write_blocked)))
1860 (old & (1 << BTREE_NODE_will_make_reachable)))
1863 if (old & (1 << BTREE_NODE_write_in_flight))
1866 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1867 type = new & BTREE_WRITE_TYPE_MASK;
1868 new &= ~BTREE_WRITE_TYPE_MASK;
1870 new &= ~(1 << BTREE_NODE_dirty);
1871 new &= ~(1 << BTREE_NODE_need_write);
1872 new |= (1 << BTREE_NODE_write_in_flight);
1873 new |= (1 << BTREE_NODE_write_in_flight_inner);
1874 new |= (1 << BTREE_NODE_just_written);
1875 new ^= (1 << BTREE_NODE_write_idx);
1876 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1878 if (new & (1U << BTREE_NODE_need_write))
1881 BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1883 atomic_dec(&c->btree_cache.dirty);
1885 BUG_ON(btree_node_fake(b));
1886 BUG_ON((b->will_make_reachable != 0) != !b->written);
1888 BUG_ON(b->written >= btree_sectors(c));
1889 BUG_ON(b->written & (block_sectors(c) - 1));
1890 BUG_ON(bset_written(b, btree_bset_last(b)));
1891 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1892 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1894 bch2_sort_whiteouts(c, b);
1896 sort_iter_stack_init(&sort_iter, b);
1899 ? sizeof(struct btree_node)
1900 : sizeof(struct btree_node_entry);
1902 bytes += b->whiteout_u64s * sizeof(u64);
1904 for_each_bset(b, t) {
1907 if (bset_written(b, i))
1910 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1911 sort_iter_add(&sort_iter.iter,
1912 btree_bkey_first(b, t),
1913 btree_bkey_last(b, t));
1914 seq = max(seq, le64_to_cpu(i->journal_seq));
1917 BUG_ON(b->written && !seq);
1919 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1922 /* buffer must be a multiple of the block size */
1923 bytes = round_up(bytes, block_bytes(c));
1925 data = btree_bounce_alloc(c, bytes, &used_mempool);
1933 bne->keys = b->data->keys;
1937 i->journal_seq = cpu_to_le64(seq);
1940 sort_iter_add(&sort_iter.iter,
1941 unwritten_whiteouts_start(c, b),
1942 unwritten_whiteouts_end(c, b));
1943 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1945 b->whiteout_u64s = 0;
1947 u64s = bch2_sort_keys(i->start, &sort_iter.iter, false);
1948 le16_add_cpu(&i->u64s, u64s);
1950 BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1952 set_needs_whiteout(i, false);
1954 /* do we have data to write? */
1955 if (b->written && !i->u64s)
1958 bytes_to_write = vstruct_end(i) - data;
1959 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1962 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1963 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1965 memset(data + bytes_to_write, 0,
1966 (sectors_to_write << 9) - bytes_to_write);
1968 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1969 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1970 BUG_ON(i->seq != b->data->keys.seq);
1972 i->version = cpu_to_le16(c->sb.version);
1973 SET_BSET_OFFSET(i, b->written);
1974 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1976 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1977 validate_before_checksum = true;
1979 /* validate_bset will be modifying: */
1980 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1981 validate_before_checksum = true;
1983 /* if we're going to be encrypting, check metadata validity first: */
1984 if (validate_before_checksum &&
1985 validate_bset_for_write(c, b, i, sectors_to_write))
1988 ret = bset_encrypt(c, i, b->written << 9);
1989 if (bch2_fs_fatal_err_on(ret, c,
1990 "error encrypting btree node: %i\n", ret))
1993 nonce = btree_nonce(i, b->written << 9);
1996 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1998 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2000 /* if we're not encrypting, check metadata after checksumming: */
2001 if (!validate_before_checksum &&
2002 validate_bset_for_write(c, b, i, sectors_to_write))
2006 * We handle btree write errors by immediately halting the journal -
2007 * after we've done that, we can't issue any subsequent btree writes
2008 * because they might have pointers to new nodes that failed to write.
2010 * Furthermore, there's no point in doing any more btree writes because
2011 * with the journal stopped, we're never going to update the journal to
2012 * reflect that those writes were done and the data flushed from the
2015 * Also on journal error, the pending write may have updates that were
2016 * never journalled (interior nodes, see btree_update_nodes_written()) -
2017 * it's critical that we don't do the write in that case otherwise we
2018 * will have updates visible that weren't in the journal:
2020 * Make sure to update b->written so bch2_btree_init_next() doesn't
2023 if (bch2_journal_error(&c->journal) ||
2027 trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2029 wbio = container_of(bio_alloc_bioset(NULL,
2030 buf_pages(data, sectors_to_write << 9),
2031 REQ_OP_WRITE|REQ_META,
2034 struct btree_write_bio, wbio.bio);
2035 wbio_init(&wbio->wbio.bio);
2037 wbio->data_bytes = bytes;
2038 wbio->sector_offset = b->written;
2040 wbio->wbio.used_mempool = used_mempool;
2041 wbio->wbio.first_btree_write = !b->written;
2042 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
2043 wbio->wbio.bio.bi_private = b;
2045 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2047 bkey_copy(&wbio->key, &b->key);
2049 b->written += sectors_to_write;
2051 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2052 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2053 cpu_to_le16(b->written);
2055 atomic64_inc(&c->btree_write_stats[type].nr);
2056 atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2058 INIT_WORK(&wbio->work, btree_write_submit);
2059 queue_work(c->io_complete_wq, &wbio->work);
2062 set_btree_node_noevict(b);
2063 b->written += sectors_to_write;
2065 btree_bounce_free(c, bytes, used_mempool, data);
2066 __btree_node_write_done(c, b);
2070 * Work that must be done with write lock held:
2072 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2074 bool invalidated_iter = false;
2075 struct btree_node_entry *bne;
2076 struct bset_tree *t;
2078 if (!btree_node_just_written(b))
2081 BUG_ON(b->whiteout_u64s);
2083 clear_btree_node_just_written(b);
2086 * Note: immediately after write, bset_written() doesn't work - the
2087 * amount of data we had to write after compaction might have been
2088 * smaller than the offset of the last bset.
2090 * However, we know that all bsets have been written here, as long as
2091 * we're still holding the write lock:
2095 * XXX: decide if we really want to unconditionally sort down to a
2099 btree_node_sort(c, b, 0, b->nsets, true);
2100 invalidated_iter = true;
2102 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2106 set_needs_whiteout(bset(b, t), true);
2108 bch2_btree_verify(c, b);
2111 * If later we don't unconditionally sort down to a single bset, we have
2112 * to ensure this is still true:
2114 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2116 bne = want_new_bset(c, b);
2118 bch2_bset_init_next(c, b, bne);
2120 bch2_btree_build_aux_trees(b);
2122 return invalidated_iter;
2126 * Use this one if the node is intent locked:
2128 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2129 enum six_lock_type lock_type_held,
2132 if (lock_type_held == SIX_LOCK_intent ||
2133 (lock_type_held == SIX_LOCK_read &&
2134 six_lock_tryupgrade(&b->c.lock))) {
2135 __bch2_btree_node_write(c, b, flags);
2137 /* don't cycle lock unnecessarily: */
2138 if (btree_node_just_written(b) &&
2139 six_trylock_write(&b->c.lock)) {
2140 bch2_btree_post_write_cleanup(c, b);
2141 six_unlock_write(&b->c.lock);
2144 if (lock_type_held == SIX_LOCK_read)
2145 six_lock_downgrade(&b->c.lock);
2147 __bch2_btree_node_write(c, b, flags);
2148 if (lock_type_held == SIX_LOCK_write &&
2149 btree_node_just_written(b))
2150 bch2_btree_post_write_cleanup(c, b);
2154 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2156 struct bucket_table *tbl;
2157 struct rhash_head *pos;
2163 for_each_cached_btree(b, c, tbl, i, pos)
2164 if (test_bit(flag, &b->flags)) {
2166 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2175 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2177 return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2180 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2182 return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2185 static const char * const bch2_btree_write_types[] = {
2186 #define x(t, n) [n] = #t,
2187 BCH_BTREE_WRITE_TYPES()
2191 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2193 printbuf_tabstop_push(out, 20);
2194 printbuf_tabstop_push(out, 10);
2199 prt_str(out, "size");
2202 for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2203 u64 nr = atomic64_read(&c->btree_write_stats[i].nr);
2204 u64 bytes = atomic64_read(&c->btree_write_stats[i].bytes);
2206 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2210 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);