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"
22 #include <linux/sched/mm.h>
23 #include <trace/events/bcachefs.h>
25 void bch2_btree_node_io_unlock(struct btree *b)
27 EBUG_ON(!btree_node_write_in_flight(b));
29 clear_btree_node_write_in_flight_inner(b);
30 clear_btree_node_write_in_flight(b);
31 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
34 void bch2_btree_node_io_lock(struct btree *b)
36 BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
38 wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
39 TASK_UNINTERRUPTIBLE);
42 void __bch2_btree_node_wait_on_read(struct btree *b)
44 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
45 TASK_UNINTERRUPTIBLE);
48 void __bch2_btree_node_wait_on_write(struct btree *b)
50 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
51 TASK_UNINTERRUPTIBLE);
54 void bch2_btree_node_wait_on_read(struct btree *b)
56 BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
58 wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
59 TASK_UNINTERRUPTIBLE);
62 void bch2_btree_node_wait_on_write(struct btree *b)
64 BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
66 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
67 TASK_UNINTERRUPTIBLE);
70 static void verify_no_dups(struct btree *b,
71 struct bkey_packed *start,
72 struct bkey_packed *end)
74 #ifdef CONFIG_BCACHEFS_DEBUG
75 struct bkey_packed *k, *p;
80 for (p = start, k = bkey_next(start);
82 p = k, k = bkey_next(k)) {
83 struct bkey l = bkey_unpack_key(b, p);
84 struct bkey r = bkey_unpack_key(b, k);
86 BUG_ON(bpos_cmp(l.p, bkey_start_pos(&r)) >= 0);
91 static void set_needs_whiteout(struct bset *i, int v)
93 struct bkey_packed *k;
95 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
96 k->needs_whiteout = v;
99 static void btree_bounce_free(struct bch_fs *c, size_t size,
100 bool used_mempool, void *p)
103 mempool_free(p, &c->btree_bounce_pool);
108 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
111 unsigned flags = memalloc_nofs_save();
114 BUG_ON(size > btree_bytes(c));
116 *used_mempool = false;
117 p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
119 *used_mempool = true;
120 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
122 memalloc_nofs_restore(flags);
126 static void sort_bkey_ptrs(const struct btree *bt,
127 struct bkey_packed **ptrs, unsigned nr)
129 unsigned n = nr, a = nr / 2, b, c, d;
134 /* Heap sort: see lib/sort.c: */
139 swap(ptrs[0], ptrs[n]);
143 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
144 b = bch2_bkey_cmp_packed(bt,
146 ptrs[d]) >= 0 ? c : d;
151 bch2_bkey_cmp_packed(bt,
158 swap(ptrs[b], ptrs[c]);
163 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
165 struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
166 bool used_mempool = false;
167 size_t bytes = b->whiteout_u64s * sizeof(u64);
169 if (!b->whiteout_u64s)
172 new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
174 ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
176 for (k = unwritten_whiteouts_start(c, b);
177 k != unwritten_whiteouts_end(c, b);
181 sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
185 while (ptrs != ptrs_end) {
191 verify_no_dups(b, new_whiteouts,
192 (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
194 memcpy_u64s(unwritten_whiteouts_start(c, b),
195 new_whiteouts, b->whiteout_u64s);
197 btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
200 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
201 bool compacting, enum compact_mode mode)
203 if (!bset_dead_u64s(b, t))
208 return should_compact_bset_lazy(b, t) ||
209 (compacting && !bset_written(b, bset(b, t)));
217 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
222 for_each_bset(b, t) {
223 struct bset *i = bset(b, t);
224 struct bkey_packed *k, *n, *out, *start, *end;
225 struct btree_node_entry *src = NULL, *dst = NULL;
227 if (t != b->set && !bset_written(b, i)) {
228 src = container_of(i, struct btree_node_entry, keys);
229 dst = max(write_block(b),
230 (void *) btree_bkey_last(b, t - 1));
236 if (!should_compact_bset(b, t, ret, mode)) {
238 memmove(dst, src, sizeof(*src) +
239 le16_to_cpu(src->keys.u64s) *
242 set_btree_bset(b, t, i);
247 start = btree_bkey_first(b, t);
248 end = btree_bkey_last(b, t);
251 memmove(dst, src, sizeof(*src));
253 set_btree_bset(b, t, i);
258 for (k = start; k != end; k = n) {
261 if (!bkey_deleted(k)) {
263 out = bkey_next(out);
265 BUG_ON(k->needs_whiteout);
269 i->u64s = cpu_to_le16((u64 *) out - i->_data);
270 set_btree_bset_end(b, t);
271 bch2_bset_set_no_aux_tree(b, t);
275 bch2_verify_btree_nr_keys(b);
277 bch2_btree_build_aux_trees(b);
282 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
283 enum compact_mode mode)
285 return bch2_drop_whiteouts(b, mode);
288 static void btree_node_sort(struct bch_fs *c, struct btree *b,
291 bool filter_whiteouts)
293 struct btree_node *out;
294 struct sort_iter sort_iter;
296 struct bset *start_bset = bset(b, &b->set[start_idx]);
297 bool used_mempool = false;
298 u64 start_time, seq = 0;
299 unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
300 bool sorting_entire_node = start_idx == 0 &&
303 sort_iter_init(&sort_iter, b);
305 for (t = b->set + start_idx;
306 t < b->set + end_idx;
308 u64s += le16_to_cpu(bset(b, t)->u64s);
309 sort_iter_add(&sort_iter,
310 btree_bkey_first(b, t),
311 btree_bkey_last(b, t));
314 bytes = sorting_entire_node
316 : __vstruct_bytes(struct btree_node, u64s);
318 out = btree_bounce_alloc(c, bytes, &used_mempool);
320 start_time = local_clock();
322 u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
324 out->keys.u64s = cpu_to_le16(u64s);
326 BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
328 if (sorting_entire_node)
329 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
332 /* Make sure we preserve bset journal_seq: */
333 for (t = b->set + start_idx; t < b->set + end_idx; t++)
334 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
335 start_bset->journal_seq = cpu_to_le64(seq);
337 if (sorting_entire_node) {
338 unsigned u64s = le16_to_cpu(out->keys.u64s);
340 BUG_ON(bytes != btree_bytes(c));
343 * Our temporary buffer is the same size as the btree node's
344 * buffer, we can just swap buffers instead of doing a big
348 out->keys.u64s = cpu_to_le16(u64s);
350 set_btree_bset(b, b->set, &b->data->keys);
352 start_bset->u64s = out->keys.u64s;
353 memcpy_u64s(start_bset->start,
355 le16_to_cpu(out->keys.u64s));
358 for (i = start_idx + 1; i < end_idx; i++)
359 b->nr.bset_u64s[start_idx] +=
364 for (i = start_idx + 1; i < b->nsets; i++) {
365 b->nr.bset_u64s[i] = b->nr.bset_u64s[i + shift];
366 b->set[i] = b->set[i + shift];
369 for (i = b->nsets; i < MAX_BSETS; i++)
370 b->nr.bset_u64s[i] = 0;
372 set_btree_bset_end(b, &b->set[start_idx]);
373 bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
375 btree_bounce_free(c, bytes, used_mempool, out);
377 bch2_verify_btree_nr_keys(b);
380 void bch2_btree_sort_into(struct bch_fs *c,
384 struct btree_nr_keys nr;
385 struct btree_node_iter src_iter;
386 u64 start_time = local_clock();
388 BUG_ON(dst->nsets != 1);
390 bch2_bset_set_no_aux_tree(dst, dst->set);
392 bch2_btree_node_iter_init_from_start(&src_iter, src);
394 nr = bch2_sort_repack(btree_bset_first(dst),
399 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
402 set_btree_bset_end(dst, dst->set);
404 dst->nr.live_u64s += nr.live_u64s;
405 dst->nr.bset_u64s[0] += nr.bset_u64s[0];
406 dst->nr.packed_keys += nr.packed_keys;
407 dst->nr.unpacked_keys += nr.unpacked_keys;
409 bch2_verify_btree_nr_keys(dst);
412 #define SORT_CRIT (4096 / sizeof(u64))
415 * We're about to add another bset to the btree node, so if there's currently
416 * too many bsets - sort some of them together:
418 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
420 unsigned unwritten_idx;
423 for (unwritten_idx = 0;
424 unwritten_idx < b->nsets;
426 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
429 if (b->nsets - unwritten_idx > 1) {
430 btree_node_sort(c, b, unwritten_idx,
435 if (unwritten_idx > 1) {
436 btree_node_sort(c, b, 0, unwritten_idx, false);
443 void bch2_btree_build_aux_trees(struct btree *b)
448 bch2_bset_build_aux_tree(b, t,
449 !bset_written(b, bset(b, t)) &&
450 t == bset_tree_last(b));
454 * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
457 * Safe to call if there already is an unwritten bset - will only add a new bset
458 * if @b doesn't already have one.
460 * Returns true if we sorted (i.e. invalidated iterators
462 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
464 struct bch_fs *c = trans->c;
465 struct btree_node_entry *bne;
466 bool reinit_iter = false;
468 EBUG_ON(!(b->c.lock.state.seq & 1));
469 BUG_ON(bset_written(b, bset(b, &b->set[1])));
471 if (b->nsets == MAX_BSETS &&
472 !btree_node_write_in_flight(b)) {
473 unsigned log_u64s[] = {
474 ilog2(bset_u64s(&b->set[0])),
475 ilog2(bset_u64s(&b->set[1])),
476 ilog2(bset_u64s(&b->set[2])),
479 if (log_u64s[1] >= (log_u64s[0] + log_u64s[2]) / 2) {
480 bch2_btree_node_write(c, b, SIX_LOCK_write);
485 if (b->nsets == MAX_BSETS &&
486 btree_node_compact(c, b))
489 BUG_ON(b->nsets >= MAX_BSETS);
491 bne = want_new_bset(c, b);
493 bch2_bset_init_next(c, b, bne);
495 bch2_btree_build_aux_trees(b);
498 bch2_trans_node_reinit_iter(trans, b);
501 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
504 pr_buf(out, "%s level %u/%u\n ",
505 bch2_btree_ids[b->c.btree_id],
507 c->btree_roots[b->c.btree_id].level);
508 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
511 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
513 struct btree *b, struct bset *i,
514 unsigned offset, int write)
516 pr_buf(out, "error validating btree node ");
518 pr_buf(out, "before write ");
520 pr_buf(out, "on %s ", ca->name);
521 pr_buf(out, "at btree ");
522 btree_pos_to_text(out, c, b);
524 pr_buf(out, "\n node offset %u", b->written);
526 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
529 enum btree_err_type {
531 BTREE_ERR_WANT_RETRY,
532 BTREE_ERR_MUST_RETRY,
536 enum btree_validate_ret {
537 BTREE_RETRY_READ = 64,
540 #define btree_err(type, c, ca, b, i, msg, ...) \
544 char *_buf2 = _buf; \
545 struct printbuf out = PBUF(_buf); \
547 _buf2 = kmalloc(4096, GFP_ATOMIC); \
549 out = _PBUF(_buf2, 4986); \
551 btree_err_msg(&out, c, ca, b, i, b->written, write); \
552 pr_buf(&out, ": " msg, ##__VA_ARGS__); \
554 if (type == BTREE_ERR_FIXABLE && \
556 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) { \
557 mustfix_fsck_err(c, "%s", _buf2); \
564 bch_err(c, "%s", _buf2); \
567 case BTREE_ERR_FIXABLE: \
568 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
570 case BTREE_ERR_WANT_RETRY: \
572 ret = BTREE_RETRY_READ; \
576 case BTREE_ERR_MUST_RETRY: \
577 ret = BTREE_RETRY_READ; \
579 case BTREE_ERR_FATAL: \
580 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
585 bch_err(c, "corrupt metadata before write: %s", _buf2); \
587 if (bch2_fs_inconsistent(c)) { \
588 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
599 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
602 * When btree topology repair changes the start or end of a node, that might
603 * mean we have to drop keys that are no longer inside the node:
605 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
609 struct bkey unpacked;
610 struct btree_node_iter iter;
612 for_each_bset(b, t) {
613 struct bset *i = bset(b, t);
614 struct bkey_packed *k;
616 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
617 if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
621 unsigned shift = (u64 *) k - (u64 *) i->start;
623 memmove_u64s_down(i->start, k,
624 (u64 *) vstruct_end(i) - (u64 *) k);
625 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
626 set_btree_bset_end(b, t);
627 bch2_bset_set_no_aux_tree(b, t);
630 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
631 if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
634 if (k != vstruct_last(i)) {
635 i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
636 set_btree_bset_end(b, t);
637 bch2_bset_set_no_aux_tree(b, t);
641 bch2_btree_build_aux_trees(b);
643 for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
644 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
645 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
649 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
650 struct btree *b, struct bset *i,
651 unsigned offset, unsigned sectors,
652 int write, bool have_retry)
654 unsigned version = le16_to_cpu(i->version);
660 btree_err_on((version != BCH_BSET_VERSION_OLD &&
661 version < bcachefs_metadata_version_min) ||
662 version >= bcachefs_metadata_version_max,
663 BTREE_ERR_FATAL, c, ca, b, i,
664 "unsupported bset version");
666 if (btree_err_on(version < c->sb.version_min,
667 BTREE_ERR_FIXABLE, c, NULL, b, i,
668 "bset version %u older than superblock version_min %u",
669 version, c->sb.version_min)) {
670 mutex_lock(&c->sb_lock);
671 c->disk_sb.sb->version_min = cpu_to_le16(version);
673 mutex_unlock(&c->sb_lock);
676 if (btree_err_on(version > c->sb.version,
677 BTREE_ERR_FIXABLE, c, NULL, b, i,
678 "bset version %u newer than superblock version %u",
679 version, c->sb.version)) {
680 mutex_lock(&c->sb_lock);
681 c->disk_sb.sb->version = cpu_to_le16(version);
683 mutex_unlock(&c->sb_lock);
686 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
687 BTREE_ERR_FATAL, c, ca, b, i,
688 "BSET_SEPARATE_WHITEOUTS no longer supported");
690 if (btree_err_on(offset + sectors > btree_sectors(c),
691 BTREE_ERR_FIXABLE, c, ca, b, i,
692 "bset past end of btree node")) {
697 btree_err_on(offset && !i->u64s,
698 BTREE_ERR_FIXABLE, c, ca, b, i,
701 btree_err_on(BSET_OFFSET(i) &&
702 BSET_OFFSET(i) != offset,
703 BTREE_ERR_WANT_RETRY, c, ca, b, i,
704 "bset at wrong sector offset");
707 struct btree_node *bn =
708 container_of(i, struct btree_node, keys);
709 /* These indicate that we read the wrong btree node: */
711 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
712 struct bch_btree_ptr_v2 *bp =
713 &bkey_i_to_btree_ptr_v2(&b->key)->v;
716 btree_err_on(bp->seq != bn->keys.seq,
717 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
718 "incorrect sequence number (wrong btree node)");
721 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
722 BTREE_ERR_MUST_RETRY, c, ca, b, i,
723 "incorrect btree id");
725 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
726 BTREE_ERR_MUST_RETRY, c, ca, b, i,
730 compat_btree_node(b->c.level, b->c.btree_id, version,
731 BSET_BIG_ENDIAN(i), write, bn);
733 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
734 struct bch_btree_ptr_v2 *bp =
735 &bkey_i_to_btree_ptr_v2(&b->key)->v;
737 if (BTREE_PTR_RANGE_UPDATED(bp)) {
738 b->data->min_key = bp->min_key;
739 b->data->max_key = b->key.k.p;
742 btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
743 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
744 "incorrect min_key: got %s should be %s",
745 (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
746 (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
749 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
750 BTREE_ERR_MUST_RETRY, c, ca, b, i,
751 "incorrect max key %s",
752 (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
755 compat_btree_node(b->c.level, b->c.btree_id, version,
756 BSET_BIG_ENDIAN(i), write, bn);
758 err = bch2_bkey_format_validate(&bn->format);
760 BTREE_ERR_FATAL, c, ca, b, i,
761 "invalid bkey format: %s", err);
763 compat_bformat(b->c.level, b->c.btree_id, version,
764 BSET_BIG_ENDIAN(i), write,
771 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
772 struct bset *i, unsigned *whiteout_u64s,
773 int write, bool have_retry)
775 unsigned version = le16_to_cpu(i->version);
776 struct bkey_packed *k, *prev = NULL;
777 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
778 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
782 k != vstruct_last(i);) {
787 if (btree_err_on(bkey_next(k) > vstruct_last(i),
788 BTREE_ERR_FIXABLE, c, NULL, b, i,
789 "key extends past end of bset")) {
790 i->u64s = cpu_to_le16((u64 *) k - i->_data);
794 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
795 BTREE_ERR_FIXABLE, c, NULL, b, i,
796 "invalid bkey format %u", k->format)) {
797 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
798 memmove_u64s_down(k, bkey_next(k),
799 (u64 *) vstruct_end(i) - (u64 *) k);
803 /* XXX: validate k->u64s */
805 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
806 BSET_BIG_ENDIAN(i), write,
809 u = __bkey_disassemble(b, k, &tmp);
811 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
812 (!updated_range ? bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
813 (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
817 bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
818 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
819 "invalid bkey: %s\n%s", invalid, buf);
821 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
822 memmove_u64s_down(k, bkey_next(k),
823 (u64 *) vstruct_end(i) - (u64 *) k);
828 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
829 BSET_BIG_ENDIAN(i), write,
832 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
835 struct bkey up = bkey_unpack_key(b, prev);
837 bch2_bkey_to_text(&PBUF(buf1), &up);
838 bch2_bkey_to_text(&PBUF(buf2), u.k);
840 bch2_dump_bset(c, b, i, 0);
842 if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
843 "keys out of order: %s > %s",
845 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
846 memmove_u64s_down(k, bkey_next(k),
847 (u64 *) vstruct_end(i) - (u64 *) k);
859 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
860 struct btree *b, bool have_retry)
862 struct btree_node_entry *bne;
863 struct sort_iter *iter;
864 struct btree_node *sorted;
865 struct bkey_packed *k;
866 struct bch_extent_ptr *ptr;
868 bool used_mempool, blacklisted;
869 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
870 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
872 unsigned blacklisted_written, nonblacklisted_written = 0;
873 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
874 int ret, retry_read = 0, write = READ;
876 b->version_ondisk = U16_MAX;
878 iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
879 sort_iter_init(iter, b);
880 iter->size = (btree_blocks(c) + 1) * 2;
882 if (bch2_meta_read_fault("btree"))
883 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
886 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
887 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
890 btree_err_on(!b->data->keys.seq,
891 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
894 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
895 struct bch_btree_ptr_v2 *bp =
896 &bkey_i_to_btree_ptr_v2(&b->key)->v;
898 btree_err_on(b->data->keys.seq != bp->seq,
899 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
900 "got wrong btree node (seq %llx want %llx)",
901 b->data->keys.seq, bp->seq);
904 while (b->written < (ptr_written ?: btree_sectors(c))) {
905 unsigned sectors, whiteout_u64s = 0;
907 struct bch_csum csum;
908 bool first = !b->written;
913 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
914 BTREE_ERR_WANT_RETRY, c, ca, b, i,
915 "unknown checksum type %llu",
918 nonce = btree_nonce(i, b->written << 9);
919 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
921 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
922 BTREE_ERR_WANT_RETRY, c, ca, b, i,
925 bset_encrypt(c, i, b->written << 9);
927 btree_err_on(btree_node_is_extents(b) &&
928 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
929 BTREE_ERR_FATAL, c, NULL, b, NULL,
930 "btree node does not have NEW_EXTENT_OVERWRITE set");
932 sectors = vstruct_sectors(b->data, c->block_bits);
934 bne = write_block(b);
937 if (i->seq != b->data->keys.seq)
940 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
941 BTREE_ERR_WANT_RETRY, c, ca, b, i,
942 "unknown checksum type %llu",
945 nonce = btree_nonce(i, b->written << 9);
946 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
948 btree_err_on(bch2_crc_cmp(csum, bne->csum),
949 BTREE_ERR_WANT_RETRY, c, ca, b, i,
952 bset_encrypt(c, i, b->written << 9);
954 sectors = vstruct_sectors(bne, c->block_bits);
957 b->version_ondisk = min(b->version_ondisk,
958 le16_to_cpu(i->version));
960 ret = validate_bset(c, ca, b, i, b->written, sectors,
966 btree_node_set_format(b, b->data->format);
968 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
973 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
975 b->written += sectors;
977 blacklisted = bch2_journal_seq_is_blacklisted(c,
978 le64_to_cpu(i->journal_seq),
981 btree_err_on(blacklisted && first,
982 BTREE_ERR_FIXABLE, c, ca, b, i,
983 "first btree node bset has blacklisted journal seq");
985 btree_err_on(blacklisted && ptr_written,
986 BTREE_ERR_FIXABLE, c, ca, b, i,
987 "found blacklisted bset in btree node with sectors_written");
988 if (blacklisted && !first)
991 sort_iter_add(iter, i->start,
992 vstruct_idx(i, whiteout_u64s));
995 vstruct_idx(i, whiteout_u64s),
998 nonblacklisted_written = b->written;
1002 btree_err_on(b->written < ptr_written,
1003 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1004 "btree node data missing: expected %u sectors, found %u",
1005 ptr_written, b->written);
1007 for (bne = write_block(b);
1008 bset_byte_offset(b, bne) < btree_bytes(c);
1009 bne = (void *) bne + block_bytes(c))
1010 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1011 !bch2_journal_seq_is_blacklisted(c,
1012 le64_to_cpu(bne->keys.journal_seq),
1014 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1015 "found bset signature after last bset");
1018 * Blacklisted bsets are those that were written after the most recent
1019 * (flush) journal write. Since there wasn't a flush, they may not have
1020 * made it to all devices - which means we shouldn't write new bsets
1021 * after them, as that could leave a gap and then reads from that device
1022 * wouldn't find all the bsets in that btree node - which means it's
1023 * important that we start writing new bsets after the most recent _non_
1026 blacklisted_written = b->written;
1027 b->written = nonblacklisted_written;
1030 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1031 sorted->keys.u64s = 0;
1033 set_btree_bset(b, b->set, &b->data->keys);
1035 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1037 u64s = le16_to_cpu(sorted->keys.u64s);
1039 sorted->keys.u64s = cpu_to_le16(u64s);
1040 swap(sorted, b->data);
1041 set_btree_bset(b, b->set, &b->data->keys);
1044 BUG_ON(b->nr.live_u64s != u64s);
1046 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1049 bch2_btree_node_drop_keys_outside_node(b);
1052 for (k = i->start; k != vstruct_last(i);) {
1054 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1055 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1058 (bch2_inject_invalid_keys &&
1059 !bversion_cmp(u.k->version, MAX_VERSION))) {
1062 bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1063 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1064 "invalid bkey %s: %s", buf, invalid);
1066 btree_keys_account_key_drop(&b->nr, 0, k);
1068 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1069 memmove_u64s_down(k, bkey_next(k),
1070 (u64 *) vstruct_end(i) - (u64 *) k);
1071 set_btree_bset_end(b, b->set);
1075 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1076 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1084 bch2_bset_build_aux_tree(b, b->set, false);
1086 set_needs_whiteout(btree_bset_first(b), true);
1088 btree_node_reset_sib_u64s(b);
1090 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1091 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1093 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1094 set_btree_node_need_rewrite(b);
1098 set_btree_node_need_rewrite(b);
1100 mempool_free(iter, &c->fill_iter);
1103 if (ret == BTREE_RETRY_READ) {
1106 bch2_inconsistent_error(c);
1107 set_btree_node_read_error(b);
1112 static void btree_node_read_work(struct work_struct *work)
1114 struct btree_read_bio *rb =
1115 container_of(work, struct btree_read_bio, work);
1116 struct bch_fs *c = rb->c;
1117 struct btree *b = rb->b;
1118 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1119 struct bio *bio = &rb->bio;
1120 struct bch_io_failures failed = { .nr = 0 };
1122 struct printbuf out;
1123 bool saw_error = false;
1128 bch_info(c, "retrying read");
1129 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1130 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1132 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1133 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1134 bio->bi_iter.bi_size = btree_bytes(c);
1136 if (rb->have_ioref) {
1137 bio_set_dev(bio, ca->disk_sb.bdev);
1138 submit_bio_wait(bio);
1140 bio->bi_status = BLK_STS_REMOVED;
1144 btree_pos_to_text(&out, c, b);
1145 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1146 bch2_blk_status_to_str(bio->bi_status), buf);
1148 percpu_ref_put(&ca->io_ref);
1149 rb->have_ioref = false;
1151 bch2_mark_io_failure(&failed, &rb->pick);
1153 can_retry = bch2_bkey_pick_read_device(c,
1154 bkey_i_to_s_c(&b->key),
1155 &failed, &rb->pick) > 0;
1157 if (!bio->bi_status &&
1158 !bch2_btree_node_read_done(c, ca, b, can_retry))
1164 set_btree_node_read_error(b);
1169 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1173 if (saw_error && !btree_node_read_error(b))
1174 bch2_btree_node_rewrite_async(c, b);
1176 clear_btree_node_read_in_flight(b);
1177 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1180 static void btree_node_read_endio(struct bio *bio)
1182 struct btree_read_bio *rb =
1183 container_of(bio, struct btree_read_bio, bio);
1184 struct bch_fs *c = rb->c;
1186 if (rb->have_ioref) {
1187 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1188 bch2_latency_acct(ca, rb->start_time, READ);
1191 queue_work(c->io_complete_wq, &rb->work);
1194 struct btree_node_read_all {
1199 void *buf[BCH_REPLICAS_MAX];
1200 struct bio *bio[BCH_REPLICAS_MAX];
1201 int err[BCH_REPLICAS_MAX];
1204 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1206 struct btree_node *bn = data;
1207 struct btree_node_entry *bne;
1208 unsigned offset = 0;
1210 if (le64_to_cpu(bn->magic) != bset_magic(c))
1213 while (offset < btree_sectors(c)) {
1215 offset += vstruct_sectors(bn, c->block_bits);
1217 bne = data + (offset << 9);
1218 if (bne->keys.seq != bn->keys.seq)
1220 offset += vstruct_sectors(bne, c->block_bits);
1227 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1229 struct btree_node *bn = data;
1230 struct btree_node_entry *bne;
1235 while (offset < btree_sectors(c)) {
1236 bne = data + (offset << 9);
1237 if (bne->keys.seq == bn->keys.seq)
1246 static void btree_node_read_all_replicas_done(struct closure *cl)
1248 struct btree_node_read_all *ra =
1249 container_of(cl, struct btree_node_read_all, cl);
1250 struct bch_fs *c = ra->c;
1251 struct btree *b = ra->b;
1252 bool dump_bset_maps = false;
1253 bool have_retry = false;
1254 int ret = 0, best = -1, write = READ;
1255 unsigned i, written = 0, written2 = 0;
1256 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1257 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1259 for (i = 0; i < ra->nr; i++) {
1260 struct btree_node *bn = ra->buf[i];
1265 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1266 (seq && seq != bn->keys.seq))
1271 written = btree_node_sectors_written(c, bn);
1275 written2 = btree_node_sectors_written(c, ra->buf[i]);
1276 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1277 "btree node sectors written mismatch: %u != %u",
1278 written, written2) ||
1279 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1280 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1281 "found bset signature after last bset") ||
1282 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1283 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1284 "btree node replicas content mismatch"))
1285 dump_bset_maps = true;
1287 if (written2 > written) {
1293 if (dump_bset_maps) {
1294 for (i = 0; i < ra->nr; i++) {
1296 struct printbuf out = PBUF(buf);
1297 struct btree_node *bn = ra->buf[i];
1298 struct btree_node_entry *bne = NULL;
1299 unsigned offset = 0, sectors;
1305 while (offset < btree_sectors(c)) {
1307 sectors = vstruct_sectors(bn, c->block_bits);
1309 bne = ra->buf[i] + (offset << 9);
1310 if (bne->keys.seq != bn->keys.seq)
1312 sectors = vstruct_sectors(bne, c->block_bits);
1315 pr_buf(&out, " %u-%u", offset, offset + sectors);
1316 if (bne && bch2_journal_seq_is_blacklisted(c,
1317 le64_to_cpu(bne->keys.journal_seq), false))
1322 while (offset < btree_sectors(c)) {
1323 bne = ra->buf[i] + (offset << 9);
1324 if (bne->keys.seq == bn->keys.seq) {
1326 pr_buf(&out, " GAP");
1329 sectors = vstruct_sectors(bne, c->block_bits);
1330 pr_buf(&out, " %u-%u", offset, offset + sectors);
1331 if (bch2_journal_seq_is_blacklisted(c,
1332 le64_to_cpu(bne->keys.journal_seq), false))
1338 bch_err(c, "replica %u:%s", i, buf);
1343 memcpy(b->data, ra->buf[best], btree_bytes(c));
1344 ret = bch2_btree_node_read_done(c, NULL, b, false);
1350 set_btree_node_read_error(b);
1352 for (i = 0; i < ra->nr; i++) {
1353 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1354 bio_put(ra->bio[i]);
1357 closure_debug_destroy(&ra->cl);
1360 clear_btree_node_read_in_flight(b);
1361 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1364 static void btree_node_read_all_replicas_endio(struct bio *bio)
1366 struct btree_read_bio *rb =
1367 container_of(bio, struct btree_read_bio, bio);
1368 struct bch_fs *c = rb->c;
1369 struct btree_node_read_all *ra = rb->ra;
1371 if (rb->have_ioref) {
1372 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1373 bch2_latency_acct(ca, rb->start_time, READ);
1376 ra->err[rb->idx] = bio->bi_status;
1377 closure_put(&ra->cl);
1381 * XXX This allocates multiple times from the same mempools, and can deadlock
1382 * under sufficient memory pressure (but is only a debug path)
1384 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1386 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1387 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1388 const union bch_extent_entry *entry;
1389 struct extent_ptr_decoded pick;
1390 struct btree_node_read_all *ra;
1393 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1397 closure_init(&ra->cl, NULL);
1400 ra->nr = bch2_bkey_nr_ptrs(k);
1402 for (i = 0; i < ra->nr; i++) {
1403 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1404 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1410 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1411 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1412 struct btree_read_bio *rb =
1413 container_of(ra->bio[i], struct btree_read_bio, bio);
1417 rb->start_time = local_clock();
1418 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1421 rb->bio.bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1422 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1423 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1424 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1426 if (rb->have_ioref) {
1427 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1428 bio_sectors(&rb->bio));
1429 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1431 closure_get(&ra->cl);
1432 submit_bio(&rb->bio);
1434 ra->err[i] = BLK_STS_REMOVED;
1441 closure_sync(&ra->cl);
1442 btree_node_read_all_replicas_done(&ra->cl);
1444 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1451 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1454 struct extent_ptr_decoded pick;
1455 struct btree_read_bio *rb;
1461 btree_pos_to_text(&PBUF(buf), c, b);
1462 trace_btree_read(c, b);
1464 if (bch2_verify_all_btree_replicas &&
1465 !btree_node_read_all_replicas(c, b, sync))
1468 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1470 if (bch2_fs_fatal_err_on(ret <= 0, c,
1471 "btree node read error: no device to read from\n"
1473 set_btree_node_read_error(b);
1477 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1479 bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1482 rb = container_of(bio, struct btree_read_bio, bio);
1486 rb->start_time = local_clock();
1487 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1489 INIT_WORK(&rb->work, btree_node_read_work);
1490 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1491 bio->bi_iter.bi_sector = pick.ptr.offset;
1492 bio->bi_end_io = btree_node_read_endio;
1493 bch2_bio_map(bio, b->data, btree_bytes(c));
1495 if (rb->have_ioref) {
1496 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1498 bio_set_dev(bio, ca->disk_sb.bdev);
1501 submit_bio_wait(bio);
1503 btree_node_read_work(&rb->work);
1508 bio->bi_status = BLK_STS_REMOVED;
1511 btree_node_read_work(&rb->work);
1513 queue_work(c->io_complete_wq, &rb->work);
1517 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1518 const struct bkey_i *k, unsigned level)
1524 closure_init_stack(&cl);
1527 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1531 b = bch2_btree_node_mem_alloc(c);
1532 bch2_btree_cache_cannibalize_unlock(c);
1536 bkey_copy(&b->key, k);
1537 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1539 set_btree_node_read_in_flight(b);
1541 bch2_btree_node_read(c, b, true);
1543 if (btree_node_read_error(b)) {
1544 bch2_btree_node_hash_remove(&c->btree_cache, b);
1546 mutex_lock(&c->btree_cache.lock);
1547 list_move(&b->list, &c->btree_cache.freeable);
1548 mutex_unlock(&c->btree_cache.lock);
1554 bch2_btree_set_root_for_read(c, b);
1556 six_unlock_write(&b->c.lock);
1557 six_unlock_intent(&b->c.lock);
1562 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1563 struct btree_write *w)
1565 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1573 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1576 closure_put(&((struct btree_update *) new)->cl);
1578 bch2_journal_pin_drop(&c->journal, &w->journal);
1581 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1583 struct btree_write *w = btree_prev_write(b);
1584 unsigned long old, new, v;
1586 bch2_btree_complete_write(c, b, w);
1588 v = READ_ONCE(b->flags);
1592 if (old & (1U << BTREE_NODE_need_write))
1595 new &= ~(1U << BTREE_NODE_write_in_flight);
1596 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1597 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1599 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1603 six_lock_read(&b->c.lock, NULL, NULL);
1604 v = READ_ONCE(b->flags);
1608 if ((old & (1U << BTREE_NODE_dirty)) &&
1609 (old & (1U << BTREE_NODE_need_write)) &&
1610 !(old & (1U << BTREE_NODE_never_write)) &&
1611 btree_node_may_write(b)) {
1612 new &= ~(1U << BTREE_NODE_dirty);
1613 new &= ~(1U << BTREE_NODE_need_write);
1614 new |= (1U << BTREE_NODE_write_in_flight);
1615 new |= (1U << BTREE_NODE_write_in_flight_inner);
1616 new |= (1U << BTREE_NODE_just_written);
1617 new ^= (1U << BTREE_NODE_write_idx);
1619 new &= ~(1U << BTREE_NODE_write_in_flight);
1620 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1622 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1624 if (new & (1U << BTREE_NODE_write_in_flight))
1625 __bch2_btree_node_write(c, b, true);
1627 six_unlock_read(&b->c.lock);
1630 static void btree_node_write_work(struct work_struct *work)
1632 struct btree_write_bio *wbio =
1633 container_of(work, struct btree_write_bio, work);
1634 struct bch_fs *c = wbio->wbio.c;
1635 struct btree *b = wbio->wbio.bio.bi_private;
1636 struct bch_extent_ptr *ptr;
1639 btree_bounce_free(c,
1641 wbio->wbio.used_mempool,
1644 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1645 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1647 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1650 if (wbio->wbio.first_btree_write) {
1651 if (wbio->wbio.failed.nr) {
1655 ret = bch2_trans_do(c, NULL, NULL, 0,
1656 bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1657 !wbio->wbio.failed.nr));
1662 bio_put(&wbio->wbio.bio);
1663 btree_node_write_done(c, b);
1666 set_btree_node_noevict(b);
1667 bch2_fs_fatal_error(c, "fatal error writing btree node");
1671 static void btree_node_write_endio(struct bio *bio)
1673 struct bch_write_bio *wbio = to_wbio(bio);
1674 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1675 struct bch_write_bio *orig = parent ?: wbio;
1676 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1677 struct bch_fs *c = wbio->c;
1678 struct btree *b = wbio->bio.bi_private;
1679 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1680 unsigned long flags;
1682 if (wbio->have_ioref)
1683 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1685 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1686 bch2_blk_status_to_str(bio->bi_status)) ||
1687 bch2_meta_write_fault("btree")) {
1688 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1689 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1690 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1693 if (wbio->have_ioref)
1694 percpu_ref_put(&ca->io_ref);
1698 bio_endio(&parent->bio);
1702 clear_btree_node_write_in_flight_inner(b);
1703 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1704 INIT_WORK(&wb->work, btree_node_write_work);
1705 queue_work(c->btree_io_complete_wq, &wb->work);
1708 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1709 struct bset *i, unsigned sectors)
1711 unsigned whiteout_u64s = 0;
1714 if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1717 ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1718 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1720 bch2_inconsistent_error(c);
1727 static void btree_write_submit(struct work_struct *work)
1729 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1730 struct bch_extent_ptr *ptr;
1731 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1733 bkey_copy(&tmp.k, &wbio->key);
1735 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1736 ptr->offset += wbio->sector_offset;
1738 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1741 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, bool already_started)
1743 struct btree_write_bio *wbio;
1744 struct bset_tree *t;
1746 struct btree_node *bn = NULL;
1747 struct btree_node_entry *bne = NULL;
1748 struct sort_iter sort_iter;
1750 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1753 unsigned long old, new;
1754 bool validate_before_checksum = false;
1757 if (already_started)
1760 if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1764 * We may only have a read lock on the btree node - the dirty bit is our
1765 * "lock" against racing with other threads that may be trying to start
1766 * a write, we do a write iff we clear the dirty bit. Since setting the
1767 * dirty bit requires a write lock, we can't race with other threads
1771 old = new = READ_ONCE(b->flags);
1773 if (!(old & (1 << BTREE_NODE_dirty)))
1776 if (!btree_node_may_write(b))
1779 if (old & (1 << BTREE_NODE_never_write))
1782 BUG_ON(old & (1 << BTREE_NODE_write_in_flight));
1784 new &= ~(1 << BTREE_NODE_dirty);
1785 new &= ~(1 << BTREE_NODE_need_write);
1786 new |= (1 << BTREE_NODE_write_in_flight);
1787 new |= (1 << BTREE_NODE_write_in_flight_inner);
1788 new |= (1 << BTREE_NODE_just_written);
1789 new ^= (1 << BTREE_NODE_write_idx);
1790 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1792 if (new & (1U << BTREE_NODE_need_write))
1795 atomic_dec(&c->btree_cache.dirty);
1797 BUG_ON(btree_node_fake(b));
1798 BUG_ON((b->will_make_reachable != 0) != !b->written);
1800 BUG_ON(b->written >= btree_sectors(c));
1801 BUG_ON(b->written & (block_sectors(c) - 1));
1802 BUG_ON(bset_written(b, btree_bset_last(b)));
1803 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1804 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1806 bch2_sort_whiteouts(c, b);
1808 sort_iter_init(&sort_iter, b);
1811 ? sizeof(struct btree_node)
1812 : sizeof(struct btree_node_entry);
1814 bytes += b->whiteout_u64s * sizeof(u64);
1816 for_each_bset(b, t) {
1819 if (bset_written(b, i))
1822 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1823 sort_iter_add(&sort_iter,
1824 btree_bkey_first(b, t),
1825 btree_bkey_last(b, t));
1826 seq = max(seq, le64_to_cpu(i->journal_seq));
1829 BUG_ON(b->written && !seq);
1831 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1834 /* buffer must be a multiple of the block size */
1835 bytes = round_up(bytes, block_bytes(c));
1837 data = btree_bounce_alloc(c, bytes, &used_mempool);
1845 bne->keys = b->data->keys;
1849 i->journal_seq = cpu_to_le64(seq);
1852 sort_iter_add(&sort_iter,
1853 unwritten_whiteouts_start(c, b),
1854 unwritten_whiteouts_end(c, b));
1855 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1857 b->whiteout_u64s = 0;
1859 u64s = bch2_sort_keys(i->start, &sort_iter, false);
1860 le16_add_cpu(&i->u64s, u64s);
1862 set_needs_whiteout(i, false);
1864 /* do we have data to write? */
1865 if (b->written && !i->u64s)
1868 bytes_to_write = vstruct_end(i) - data;
1869 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1871 memset(data + bytes_to_write, 0,
1872 (sectors_to_write << 9) - bytes_to_write);
1874 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1875 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1876 BUG_ON(i->seq != b->data->keys.seq);
1878 i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1879 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1880 : cpu_to_le16(c->sb.version);
1881 SET_BSET_OFFSET(i, b->written);
1882 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1884 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1885 validate_before_checksum = true;
1887 /* validate_bset will be modifying: */
1888 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1889 validate_before_checksum = true;
1891 /* if we're going to be encrypting, check metadata validity first: */
1892 if (validate_before_checksum &&
1893 validate_bset_for_write(c, b, i, sectors_to_write))
1896 bset_encrypt(c, i, b->written << 9);
1898 nonce = btree_nonce(i, b->written << 9);
1901 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1903 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1905 /* if we're not encrypting, check metadata after checksumming: */
1906 if (!validate_before_checksum &&
1907 validate_bset_for_write(c, b, i, sectors_to_write))
1911 * We handle btree write errors by immediately halting the journal -
1912 * after we've done that, we can't issue any subsequent btree writes
1913 * because they might have pointers to new nodes that failed to write.
1915 * Furthermore, there's no point in doing any more btree writes because
1916 * with the journal stopped, we're never going to update the journal to
1917 * reflect that those writes were done and the data flushed from the
1920 * Also on journal error, the pending write may have updates that were
1921 * never journalled (interior nodes, see btree_update_nodes_written()) -
1922 * it's critical that we don't do the write in that case otherwise we
1923 * will have updates visible that weren't in the journal:
1925 * Make sure to update b->written so bch2_btree_init_next() doesn't
1928 if (bch2_journal_error(&c->journal) ||
1932 trace_btree_write(b, bytes_to_write, sectors_to_write);
1934 wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1935 buf_pages(data, sectors_to_write << 9),
1937 struct btree_write_bio, wbio.bio);
1938 wbio_init(&wbio->wbio.bio);
1940 wbio->data_bytes = bytes;
1941 wbio->sector_offset = b->written;
1943 wbio->wbio.used_mempool = used_mempool;
1944 wbio->wbio.first_btree_write = !b->written;
1945 wbio->wbio.bio.bi_opf = REQ_OP_WRITE|REQ_META;
1946 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
1947 wbio->wbio.bio.bi_private = b;
1949 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1951 bkey_copy(&wbio->key, &b->key);
1953 b->written += sectors_to_write;
1955 if (wbio->wbio.first_btree_write &&
1956 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1957 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1958 cpu_to_le16(b->written);
1960 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
1961 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
1962 cpu_to_le16(b->written);
1964 atomic64_inc(&c->btree_writes_nr);
1965 atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1967 INIT_WORK(&wbio->work, btree_write_submit);
1968 queue_work(c->io_complete_wq, &wbio->work);
1971 set_btree_node_noevict(b);
1973 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1974 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1975 cpu_to_le16(sectors_to_write);
1976 b->written += sectors_to_write;
1978 btree_bounce_free(c, bytes, used_mempool, data);
1979 btree_node_write_done(c, b);
1983 * Work that must be done with write lock held:
1985 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1987 bool invalidated_iter = false;
1988 struct btree_node_entry *bne;
1989 struct bset_tree *t;
1991 if (!btree_node_just_written(b))
1994 BUG_ON(b->whiteout_u64s);
1996 clear_btree_node_just_written(b);
1999 * Note: immediately after write, bset_written() doesn't work - the
2000 * amount of data we had to write after compaction might have been
2001 * smaller than the offset of the last bset.
2003 * However, we know that all bsets have been written here, as long as
2004 * we're still holding the write lock:
2008 * XXX: decide if we really want to unconditionally sort down to a
2012 btree_node_sort(c, b, 0, b->nsets, true);
2013 invalidated_iter = true;
2015 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2019 set_needs_whiteout(bset(b, t), true);
2021 bch2_btree_verify(c, b);
2024 * If later we don't unconditionally sort down to a single bset, we have
2025 * to ensure this is still true:
2027 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2029 bne = want_new_bset(c, b);
2031 bch2_bset_init_next(c, b, bne);
2033 bch2_btree_build_aux_trees(b);
2035 return invalidated_iter;
2039 * Use this one if the node is intent locked:
2041 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2042 enum six_lock_type lock_type_held)
2044 if (lock_type_held == SIX_LOCK_intent ||
2045 (lock_type_held == SIX_LOCK_read &&
2046 six_lock_tryupgrade(&b->c.lock))) {
2047 __bch2_btree_node_write(c, b, false);
2049 /* don't cycle lock unnecessarily: */
2050 if (btree_node_just_written(b) &&
2051 six_trylock_write(&b->c.lock)) {
2052 bch2_btree_post_write_cleanup(c, b);
2053 six_unlock_write(&b->c.lock);
2056 if (lock_type_held == SIX_LOCK_read)
2057 six_lock_downgrade(&b->c.lock);
2059 __bch2_btree_node_write(c, b, false);
2060 if (lock_type_held == SIX_LOCK_write &&
2061 btree_node_just_written(b))
2062 bch2_btree_post_write_cleanup(c, b);
2066 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2068 struct bucket_table *tbl;
2069 struct rhash_head *pos;
2074 for_each_cached_btree(b, c, tbl, i, pos)
2075 if (test_bit(flag, &b->flags)) {
2077 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2084 void bch2_btree_flush_all_reads(struct bch_fs *c)
2086 __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2089 void bch2_btree_flush_all_writes(struct bch_fs *c)
2091 __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2094 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2096 struct bucket_table *tbl;
2097 struct rhash_head *pos;
2102 for_each_cached_btree(b, c, tbl, i, pos) {
2103 unsigned long flags = READ_ONCE(b->flags);
2105 if (!(flags & (1 << BTREE_NODE_dirty)))
2108 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2110 (flags & (1 << BTREE_NODE_dirty)) != 0,
2111 (flags & (1 << BTREE_NODE_need_write)) != 0,
2114 !list_empty_careful(&b->write_blocked),
2115 b->will_make_reachable != 0,
2116 b->will_make_reachable & 1);