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, 0);
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, ...) \
543 struct printbuf out = PRINTBUF; \
545 btree_err_msg(&out, c, ca, b, i, b->written, write); \
546 pr_buf(&out, ": " msg, ##__VA_ARGS__); \
548 if (type == BTREE_ERR_FIXABLE && \
550 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) { \
551 mustfix_fsck_err(c, "%s", out.buf); \
557 bch_err(c, "%s", out.buf); \
560 case BTREE_ERR_FIXABLE: \
561 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
563 case BTREE_ERR_WANT_RETRY: \
565 ret = BTREE_RETRY_READ; \
569 case BTREE_ERR_MUST_RETRY: \
570 ret = BTREE_RETRY_READ; \
572 case BTREE_ERR_FATAL: \
573 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
578 bch_err(c, "corrupt metadata before write: %s", out.buf);\
580 if (bch2_fs_inconsistent(c)) { \
581 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
587 printbuf_exit(&out); \
591 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
594 * When btree topology repair changes the start or end of a node, that might
595 * mean we have to drop keys that are no longer inside the node:
597 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
601 struct bkey unpacked;
602 struct btree_node_iter iter;
604 for_each_bset(b, t) {
605 struct bset *i = bset(b, t);
606 struct bkey_packed *k;
608 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
609 if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
613 unsigned shift = (u64 *) k - (u64 *) i->start;
615 memmove_u64s_down(i->start, k,
616 (u64 *) vstruct_end(i) - (u64 *) k);
617 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
618 set_btree_bset_end(b, t);
619 bch2_bset_set_no_aux_tree(b, t);
622 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
623 if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
626 if (k != vstruct_last(i)) {
627 i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
628 set_btree_bset_end(b, t);
629 bch2_bset_set_no_aux_tree(b, t);
633 bch2_btree_build_aux_trees(b);
635 for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
636 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
637 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
641 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
642 struct btree *b, struct bset *i,
643 unsigned offset, unsigned sectors,
644 int write, bool have_retry)
646 unsigned version = le16_to_cpu(i->version);
648 struct printbuf buf1 = PRINTBUF;
649 struct printbuf buf2 = PRINTBUF;
652 btree_err_on((version != BCH_BSET_VERSION_OLD &&
653 version < bcachefs_metadata_version_min) ||
654 version >= bcachefs_metadata_version_max,
655 BTREE_ERR_FATAL, c, ca, b, i,
656 "unsupported bset version");
658 if (btree_err_on(version < c->sb.version_min,
659 BTREE_ERR_FIXABLE, c, NULL, b, i,
660 "bset version %u older than superblock version_min %u",
661 version, c->sb.version_min)) {
662 mutex_lock(&c->sb_lock);
663 c->disk_sb.sb->version_min = cpu_to_le16(version);
665 mutex_unlock(&c->sb_lock);
668 if (btree_err_on(version > c->sb.version,
669 BTREE_ERR_FIXABLE, c, NULL, b, i,
670 "bset version %u newer than superblock version %u",
671 version, c->sb.version)) {
672 mutex_lock(&c->sb_lock);
673 c->disk_sb.sb->version = cpu_to_le16(version);
675 mutex_unlock(&c->sb_lock);
678 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
679 BTREE_ERR_FATAL, c, ca, b, i,
680 "BSET_SEPARATE_WHITEOUTS no longer supported");
682 if (btree_err_on(offset + sectors > btree_sectors(c),
683 BTREE_ERR_FIXABLE, c, ca, b, i,
684 "bset past end of btree node")) {
690 btree_err_on(offset && !i->u64s,
691 BTREE_ERR_FIXABLE, c, ca, b, i,
694 btree_err_on(BSET_OFFSET(i) &&
695 BSET_OFFSET(i) != offset,
696 BTREE_ERR_WANT_RETRY, c, ca, b, i,
697 "bset at wrong sector offset");
700 struct btree_node *bn =
701 container_of(i, struct btree_node, keys);
702 /* These indicate that we read the wrong btree node: */
704 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
705 struct bch_btree_ptr_v2 *bp =
706 &bkey_i_to_btree_ptr_v2(&b->key)->v;
709 btree_err_on(bp->seq != bn->keys.seq,
710 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
711 "incorrect sequence number (wrong btree node)");
714 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
715 BTREE_ERR_MUST_RETRY, c, ca, b, i,
716 "incorrect btree id");
718 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
719 BTREE_ERR_MUST_RETRY, c, ca, b, i,
723 compat_btree_node(b->c.level, b->c.btree_id, version,
724 BSET_BIG_ENDIAN(i), write, bn);
726 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
727 struct bch_btree_ptr_v2 *bp =
728 &bkey_i_to_btree_ptr_v2(&b->key)->v;
730 if (BTREE_PTR_RANGE_UPDATED(bp)) {
731 b->data->min_key = bp->min_key;
732 b->data->max_key = b->key.k.p;
735 btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
736 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
737 "incorrect min_key: got %s should be %s",
738 (printbuf_reset(&buf1),
739 bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
740 (printbuf_reset(&buf2),
741 bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
744 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
745 BTREE_ERR_MUST_RETRY, c, ca, b, i,
746 "incorrect max key %s",
747 (printbuf_reset(&buf1),
748 bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
751 compat_btree_node(b->c.level, b->c.btree_id, version,
752 BSET_BIG_ENDIAN(i), write, bn);
754 err = bch2_bkey_format_validate(&bn->format);
756 BTREE_ERR_FATAL, c, ca, b, i,
757 "invalid bkey format: %s", err);
759 compat_bformat(b->c.level, b->c.btree_id, version,
760 BSET_BIG_ENDIAN(i), write,
765 printbuf_exit(&buf2);
766 printbuf_exit(&buf1);
770 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
771 struct bset *i, unsigned *whiteout_u64s,
772 int write, bool have_retry)
774 unsigned version = le16_to_cpu(i->version);
775 struct bkey_packed *k, *prev = NULL;
776 struct printbuf buf1 = PRINTBUF;
777 struct printbuf buf2 = PRINTBUF;
778 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
779 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
783 k != vstruct_last(i);) {
788 if (btree_err_on(bkey_next(k) > vstruct_last(i),
789 BTREE_ERR_FIXABLE, c, NULL, b, i,
790 "key extends past end of bset")) {
791 i->u64s = cpu_to_le16((u64 *) k - i->_data);
795 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
796 BTREE_ERR_FIXABLE, c, NULL, b, i,
797 "invalid bkey format %u", k->format)) {
798 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
799 memmove_u64s_down(k, bkey_next(k),
800 (u64 *) vstruct_end(i) - (u64 *) k);
804 /* XXX: validate k->u64s */
806 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
807 BSET_BIG_ENDIAN(i), write,
810 u = __bkey_disassemble(b, k, &tmp);
812 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
813 (!updated_range ? bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
814 (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
816 printbuf_reset(&buf1);
817 bch2_bkey_val_to_text(&buf1, c, u.s_c);
818 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
819 "invalid bkey: %s\n%s", invalid, buf1.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) {
833 struct bkey up = bkey_unpack_key(b, prev);
835 printbuf_reset(&buf1);
836 bch2_bkey_to_text(&buf1, &up);
837 printbuf_reset(&buf2);
838 bch2_bkey_to_text(&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",
844 buf1.buf, buf2.buf)) {
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);
856 printbuf_exit(&buf2);
857 printbuf_exit(&buf1);
861 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
862 struct btree *b, bool have_retry)
864 struct btree_node_entry *bne;
865 struct sort_iter *iter;
866 struct btree_node *sorted;
867 struct bkey_packed *k;
868 struct bch_extent_ptr *ptr;
870 bool used_mempool, blacklisted;
871 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
872 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
874 unsigned blacklisted_written, nonblacklisted_written = 0;
875 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
876 int ret, retry_read = 0, write = READ;
878 b->version_ondisk = U16_MAX;
880 iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
881 sort_iter_init(iter, b);
882 iter->size = (btree_blocks(c) + 1) * 2;
884 if (bch2_meta_read_fault("btree"))
885 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
888 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
889 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
890 "bad magic: want %llx, got %llx",
891 bset_magic(c), le64_to_cpu(b->data->magic));
893 btree_err_on(!b->data->keys.seq,
894 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
895 "bad btree header: seq 0");
897 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
898 struct bch_btree_ptr_v2 *bp =
899 &bkey_i_to_btree_ptr_v2(&b->key)->v;
901 btree_err_on(b->data->keys.seq != bp->seq,
902 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
903 "got wrong btree node (seq %llx want %llx)",
904 b->data->keys.seq, bp->seq);
907 while (b->written < (ptr_written ?: btree_sectors(c))) {
908 unsigned sectors, whiteout_u64s = 0;
910 struct bch_csum csum;
911 bool first = !b->written;
916 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
917 BTREE_ERR_WANT_RETRY, c, ca, b, i,
918 "unknown checksum type %llu",
921 nonce = btree_nonce(i, b->written << 9);
922 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
924 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
925 BTREE_ERR_WANT_RETRY, c, ca, b, i,
928 ret = bset_encrypt(c, i, b->written << 9);
929 if (bch2_fs_fatal_err_on(ret, c,
930 "error decrypting btree node: %i", ret))
933 btree_err_on(btree_node_is_extents(b) &&
934 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
935 BTREE_ERR_FATAL, c, NULL, b, NULL,
936 "btree node does not have NEW_EXTENT_OVERWRITE set");
938 sectors = vstruct_sectors(b->data, c->block_bits);
940 bne = write_block(b);
943 if (i->seq != b->data->keys.seq)
946 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
947 BTREE_ERR_WANT_RETRY, c, ca, b, i,
948 "unknown checksum type %llu",
951 nonce = btree_nonce(i, b->written << 9);
952 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
954 btree_err_on(bch2_crc_cmp(csum, bne->csum),
955 BTREE_ERR_WANT_RETRY, c, ca, b, i,
958 ret = bset_encrypt(c, i, b->written << 9);
959 if (bch2_fs_fatal_err_on(ret, c,
960 "error decrypting btree node: %i\n", ret))
963 sectors = vstruct_sectors(bne, c->block_bits);
966 b->version_ondisk = min(b->version_ondisk,
967 le16_to_cpu(i->version));
969 ret = validate_bset(c, ca, b, i, b->written, sectors,
975 btree_node_set_format(b, b->data->format);
977 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
982 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
984 blacklisted = bch2_journal_seq_is_blacklisted(c,
985 le64_to_cpu(i->journal_seq),
988 btree_err_on(blacklisted && first,
989 BTREE_ERR_FIXABLE, c, ca, b, i,
990 "first btree node bset has blacklisted journal seq (%llu)",
991 le64_to_cpu(i->journal_seq));
993 btree_err_on(blacklisted && ptr_written,
994 BTREE_ERR_FIXABLE, c, ca, b, i,
995 "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
996 le64_to_cpu(i->journal_seq),
997 b->written, b->written + sectors, ptr_written);
999 b->written += sectors;
1001 if (blacklisted && !first)
1004 sort_iter_add(iter, i->start,
1005 vstruct_idx(i, whiteout_u64s));
1008 vstruct_idx(i, whiteout_u64s),
1011 nonblacklisted_written = b->written;
1015 btree_err_on(b->written < ptr_written,
1016 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1017 "btree node data missing: expected %u sectors, found %u",
1018 ptr_written, b->written);
1020 for (bne = write_block(b);
1021 bset_byte_offset(b, bne) < btree_bytes(c);
1022 bne = (void *) bne + block_bytes(c))
1023 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1024 !bch2_journal_seq_is_blacklisted(c,
1025 le64_to_cpu(bne->keys.journal_seq),
1027 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1028 "found bset signature after last bset");
1031 * Blacklisted bsets are those that were written after the most recent
1032 * (flush) journal write. Since there wasn't a flush, they may not have
1033 * made it to all devices - which means we shouldn't write new bsets
1034 * after them, as that could leave a gap and then reads from that device
1035 * wouldn't find all the bsets in that btree node - which means it's
1036 * important that we start writing new bsets after the most recent _non_
1039 blacklisted_written = b->written;
1040 b->written = nonblacklisted_written;
1043 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1044 sorted->keys.u64s = 0;
1046 set_btree_bset(b, b->set, &b->data->keys);
1048 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1050 u64s = le16_to_cpu(sorted->keys.u64s);
1052 sorted->keys.u64s = cpu_to_le16(u64s);
1053 swap(sorted, b->data);
1054 set_btree_bset(b, b->set, &b->data->keys);
1057 BUG_ON(b->nr.live_u64s != u64s);
1059 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1062 bch2_btree_node_drop_keys_outside_node(b);
1065 for (k = i->start; k != vstruct_last(i);) {
1067 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1068 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1071 (bch2_inject_invalid_keys &&
1072 !bversion_cmp(u.k->version, MAX_VERSION))) {
1073 struct printbuf buf = PRINTBUF;
1075 bch2_bkey_val_to_text(&buf, c, u.s_c);
1076 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1077 "invalid bkey %s: %s", buf, invalid);
1078 printbuf_exit(&buf);
1080 btree_keys_account_key_drop(&b->nr, 0, k);
1082 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1083 memmove_u64s_down(k, bkey_next(k),
1084 (u64 *) vstruct_end(i) - (u64 *) k);
1085 set_btree_bset_end(b, b->set);
1089 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1090 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1098 bch2_bset_build_aux_tree(b, b->set, false);
1100 set_needs_whiteout(btree_bset_first(b), true);
1102 btree_node_reset_sib_u64s(b);
1104 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1105 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1107 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1108 set_btree_node_need_rewrite(b);
1112 set_btree_node_need_rewrite(b);
1114 mempool_free(iter, &c->fill_iter);
1117 if (ret == BTREE_RETRY_READ) {
1120 bch2_inconsistent_error(c);
1121 set_btree_node_read_error(b);
1126 static void btree_node_read_work(struct work_struct *work)
1128 struct btree_read_bio *rb =
1129 container_of(work, struct btree_read_bio, work);
1130 struct bch_fs *c = rb->c;
1131 struct btree *b = rb->b;
1132 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1133 struct bio *bio = &rb->bio;
1134 struct bch_io_failures failed = { .nr = 0 };
1135 struct printbuf buf = PRINTBUF;
1136 bool saw_error = false;
1141 bch_info(c, "retrying read");
1142 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1143 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1145 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1146 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1147 bio->bi_iter.bi_size = btree_bytes(c);
1149 if (rb->have_ioref) {
1150 bio_set_dev(bio, ca->disk_sb.bdev);
1151 submit_bio_wait(bio);
1153 bio->bi_status = BLK_STS_REMOVED;
1156 printbuf_reset(&buf);
1157 btree_pos_to_text(&buf, c, b);
1158 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1159 bch2_blk_status_to_str(bio->bi_status), buf.buf);
1161 percpu_ref_put(&ca->io_ref);
1162 rb->have_ioref = false;
1164 bch2_mark_io_failure(&failed, &rb->pick);
1166 can_retry = bch2_bkey_pick_read_device(c,
1167 bkey_i_to_s_c(&b->key),
1168 &failed, &rb->pick) > 0;
1170 if (!bio->bi_status &&
1171 !bch2_btree_node_read_done(c, ca, b, can_retry))
1177 set_btree_node_read_error(b);
1182 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1185 printbuf_exit(&buf);
1187 if (saw_error && !btree_node_read_error(b))
1188 bch2_btree_node_rewrite_async(c, b);
1190 clear_btree_node_read_in_flight(b);
1191 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1194 static void btree_node_read_endio(struct bio *bio)
1196 struct btree_read_bio *rb =
1197 container_of(bio, struct btree_read_bio, bio);
1198 struct bch_fs *c = rb->c;
1200 if (rb->have_ioref) {
1201 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1202 bch2_latency_acct(ca, rb->start_time, READ);
1205 queue_work(c->io_complete_wq, &rb->work);
1208 struct btree_node_read_all {
1213 void *buf[BCH_REPLICAS_MAX];
1214 struct bio *bio[BCH_REPLICAS_MAX];
1215 int err[BCH_REPLICAS_MAX];
1218 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1220 struct btree_node *bn = data;
1221 struct btree_node_entry *bne;
1222 unsigned offset = 0;
1224 if (le64_to_cpu(bn->magic) != bset_magic(c))
1227 while (offset < btree_sectors(c)) {
1229 offset += vstruct_sectors(bn, c->block_bits);
1231 bne = data + (offset << 9);
1232 if (bne->keys.seq != bn->keys.seq)
1234 offset += vstruct_sectors(bne, c->block_bits);
1241 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1243 struct btree_node *bn = data;
1244 struct btree_node_entry *bne;
1249 while (offset < btree_sectors(c)) {
1250 bne = data + (offset << 9);
1251 if (bne->keys.seq == bn->keys.seq)
1260 static void btree_node_read_all_replicas_done(struct closure *cl)
1262 struct btree_node_read_all *ra =
1263 container_of(cl, struct btree_node_read_all, cl);
1264 struct bch_fs *c = ra->c;
1265 struct btree *b = ra->b;
1266 struct printbuf buf = PRINTBUF;
1267 bool dump_bset_maps = false;
1268 bool have_retry = false;
1269 int ret = 0, best = -1, write = READ;
1270 unsigned i, written = 0, written2 = 0;
1271 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1272 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1274 for (i = 0; i < ra->nr; i++) {
1275 struct btree_node *bn = ra->buf[i];
1280 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1281 (seq && seq != bn->keys.seq))
1286 written = btree_node_sectors_written(c, bn);
1290 written2 = btree_node_sectors_written(c, ra->buf[i]);
1291 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1292 "btree node sectors written mismatch: %u != %u",
1293 written, written2) ||
1294 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1295 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1296 "found bset signature after last bset") ||
1297 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1298 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1299 "btree node replicas content mismatch"))
1300 dump_bset_maps = true;
1302 if (written2 > written) {
1308 if (dump_bset_maps) {
1309 for (i = 0; i < ra->nr; i++) {
1310 struct btree_node *bn = ra->buf[i];
1311 struct btree_node_entry *bne = NULL;
1312 unsigned offset = 0, sectors;
1318 printbuf_reset(&buf);
1320 while (offset < btree_sectors(c)) {
1322 sectors = vstruct_sectors(bn, c->block_bits);
1324 bne = ra->buf[i] + (offset << 9);
1325 if (bne->keys.seq != bn->keys.seq)
1327 sectors = vstruct_sectors(bne, c->block_bits);
1330 pr_buf(&buf, " %u-%u", offset, offset + sectors);
1331 if (bne && bch2_journal_seq_is_blacklisted(c,
1332 le64_to_cpu(bne->keys.journal_seq), false))
1337 while (offset < btree_sectors(c)) {
1338 bne = ra->buf[i] + (offset << 9);
1339 if (bne->keys.seq == bn->keys.seq) {
1341 pr_buf(&buf, " GAP");
1344 sectors = vstruct_sectors(bne, c->block_bits);
1345 pr_buf(&buf, " %u-%u", offset, offset + sectors);
1346 if (bch2_journal_seq_is_blacklisted(c,
1347 le64_to_cpu(bne->keys.journal_seq), false))
1353 bch_err(c, "replica %u:%s", i, buf.buf);
1358 memcpy(b->data, ra->buf[best], btree_bytes(c));
1359 ret = bch2_btree_node_read_done(c, NULL, b, false);
1365 set_btree_node_read_error(b);
1367 for (i = 0; i < ra->nr; i++) {
1368 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1369 bio_put(ra->bio[i]);
1372 closure_debug_destroy(&ra->cl);
1374 printbuf_exit(&buf);
1376 clear_btree_node_read_in_flight(b);
1377 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1380 static void btree_node_read_all_replicas_endio(struct bio *bio)
1382 struct btree_read_bio *rb =
1383 container_of(bio, struct btree_read_bio, bio);
1384 struct bch_fs *c = rb->c;
1385 struct btree_node_read_all *ra = rb->ra;
1387 if (rb->have_ioref) {
1388 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1389 bch2_latency_acct(ca, rb->start_time, READ);
1392 ra->err[rb->idx] = bio->bi_status;
1393 closure_put(&ra->cl);
1397 * XXX This allocates multiple times from the same mempools, and can deadlock
1398 * under sufficient memory pressure (but is only a debug path)
1400 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1402 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1403 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1404 const union bch_extent_entry *entry;
1405 struct extent_ptr_decoded pick;
1406 struct btree_node_read_all *ra;
1409 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1413 closure_init(&ra->cl, NULL);
1416 ra->nr = bch2_bkey_nr_ptrs(k);
1418 for (i = 0; i < ra->nr; i++) {
1419 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1420 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1426 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1427 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1428 struct btree_read_bio *rb =
1429 container_of(ra->bio[i], struct btree_read_bio, bio);
1433 rb->start_time = local_clock();
1434 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1437 rb->bio.bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1438 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1439 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1440 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1442 if (rb->have_ioref) {
1443 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1444 bio_sectors(&rb->bio));
1445 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1447 closure_get(&ra->cl);
1448 submit_bio(&rb->bio);
1450 ra->err[i] = BLK_STS_REMOVED;
1457 closure_sync(&ra->cl);
1458 btree_node_read_all_replicas_done(&ra->cl);
1460 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1467 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1470 struct extent_ptr_decoded pick;
1471 struct btree_read_bio *rb;
1474 struct printbuf buf = PRINTBUF;
1477 btree_pos_to_text(&buf, c, b);
1478 trace_btree_read(c, b);
1480 if (bch2_verify_all_btree_replicas &&
1481 !btree_node_read_all_replicas(c, b, sync))
1484 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1486 if (bch2_fs_fatal_err_on(ret <= 0, c,
1487 "btree node read error: no device to read from\n"
1488 " at %s", buf.buf)) {
1489 set_btree_node_read_error(b);
1493 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1495 bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1498 rb = container_of(bio, struct btree_read_bio, bio);
1502 rb->start_time = local_clock();
1503 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1505 INIT_WORK(&rb->work, btree_node_read_work);
1506 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1507 bio->bi_iter.bi_sector = pick.ptr.offset;
1508 bio->bi_end_io = btree_node_read_endio;
1509 bch2_bio_map(bio, b->data, btree_bytes(c));
1511 if (rb->have_ioref) {
1512 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1514 bio_set_dev(bio, ca->disk_sb.bdev);
1517 submit_bio_wait(bio);
1519 btree_node_read_work(&rb->work);
1524 bio->bi_status = BLK_STS_REMOVED;
1527 btree_node_read_work(&rb->work);
1529 queue_work(c->io_complete_wq, &rb->work);
1532 printbuf_exit(&buf);
1535 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1536 const struct bkey_i *k, unsigned level)
1542 closure_init_stack(&cl);
1545 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1549 b = bch2_btree_node_mem_alloc(c);
1550 bch2_btree_cache_cannibalize_unlock(c);
1554 bkey_copy(&b->key, k);
1555 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1557 set_btree_node_read_in_flight(b);
1559 bch2_btree_node_read(c, b, true);
1561 if (btree_node_read_error(b)) {
1562 bch2_btree_node_hash_remove(&c->btree_cache, b);
1564 mutex_lock(&c->btree_cache.lock);
1565 list_move(&b->list, &c->btree_cache.freeable);
1566 mutex_unlock(&c->btree_cache.lock);
1572 bch2_btree_set_root_for_read(c, b);
1574 six_unlock_write(&b->c.lock);
1575 six_unlock_intent(&b->c.lock);
1580 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1581 struct btree_write *w)
1583 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1591 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1594 closure_put(&((struct btree_update *) new)->cl);
1596 bch2_journal_pin_drop(&c->journal, &w->journal);
1599 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1601 struct btree_write *w = btree_prev_write(b);
1602 unsigned long old, new, v;
1604 bch2_btree_complete_write(c, b, w);
1606 v = READ_ONCE(b->flags);
1610 if ((old & (1U << BTREE_NODE_dirty)) &&
1611 (old & (1U << BTREE_NODE_need_write)) &&
1612 !(old & (1U << BTREE_NODE_never_write)) &&
1613 !(old & (1U << BTREE_NODE_write_blocked)) &&
1614 !(old & (1U << BTREE_NODE_will_make_reachable))) {
1615 new &= ~(1U << BTREE_NODE_dirty);
1616 new &= ~(1U << BTREE_NODE_need_write);
1617 new |= (1U << BTREE_NODE_write_in_flight);
1618 new |= (1U << BTREE_NODE_write_in_flight_inner);
1619 new |= (1U << BTREE_NODE_just_written);
1620 new ^= (1U << BTREE_NODE_write_idx);
1622 new &= ~(1U << BTREE_NODE_write_in_flight);
1623 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1625 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1627 if (new & (1U << BTREE_NODE_write_in_flight))
1628 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED);
1631 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1633 six_lock_read(&b->c.lock, NULL, NULL);
1634 __btree_node_write_done(c, b);
1635 six_unlock_read(&b->c.lock);
1638 static void btree_node_write_work(struct work_struct *work)
1640 struct btree_write_bio *wbio =
1641 container_of(work, struct btree_write_bio, work);
1642 struct bch_fs *c = wbio->wbio.c;
1643 struct btree *b = wbio->wbio.bio.bi_private;
1644 struct bch_extent_ptr *ptr;
1647 btree_bounce_free(c,
1649 wbio->wbio.used_mempool,
1652 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1653 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1655 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1658 if (wbio->wbio.first_btree_write) {
1659 if (wbio->wbio.failed.nr) {
1663 ret = bch2_trans_do(c, NULL, NULL, 0,
1664 bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1665 !wbio->wbio.failed.nr));
1670 bio_put(&wbio->wbio.bio);
1671 btree_node_write_done(c, b);
1674 set_btree_node_noevict(b);
1675 bch2_fs_fatal_error(c, "fatal error writing btree node");
1679 static void btree_node_write_endio(struct bio *bio)
1681 struct bch_write_bio *wbio = to_wbio(bio);
1682 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1683 struct bch_write_bio *orig = parent ?: wbio;
1684 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1685 struct bch_fs *c = wbio->c;
1686 struct btree *b = wbio->bio.bi_private;
1687 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1688 unsigned long flags;
1690 if (wbio->have_ioref)
1691 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1693 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1694 bch2_blk_status_to_str(bio->bi_status)) ||
1695 bch2_meta_write_fault("btree")) {
1696 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1697 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1698 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1701 if (wbio->have_ioref)
1702 percpu_ref_put(&ca->io_ref);
1706 bio_endio(&parent->bio);
1710 clear_btree_node_write_in_flight_inner(b);
1711 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1712 INIT_WORK(&wb->work, btree_node_write_work);
1713 queue_work(c->btree_io_complete_wq, &wb->work);
1716 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1717 struct bset *i, unsigned sectors)
1719 unsigned whiteout_u64s = 0;
1722 if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1725 ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1726 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1728 bch2_inconsistent_error(c);
1735 static void btree_write_submit(struct work_struct *work)
1737 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1738 struct bch_extent_ptr *ptr;
1739 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1741 bkey_copy(&tmp.k, &wbio->key);
1743 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1744 ptr->offset += wbio->sector_offset;
1746 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1749 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1751 struct btree_write_bio *wbio;
1752 struct bset_tree *t;
1754 struct btree_node *bn = NULL;
1755 struct btree_node_entry *bne = NULL;
1756 struct sort_iter sort_iter;
1758 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1761 unsigned long old, new;
1762 bool validate_before_checksum = false;
1766 if (flags & BTREE_WRITE_ALREADY_STARTED)
1770 * We may only have a read lock on the btree node - the dirty bit is our
1771 * "lock" against racing with other threads that may be trying to start
1772 * a write, we do a write iff we clear the dirty bit. Since setting the
1773 * dirty bit requires a write lock, we can't race with other threads
1777 old = new = READ_ONCE(b->flags);
1779 if (!(old & (1 << BTREE_NODE_dirty)))
1782 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1783 !(old & (1 << BTREE_NODE_need_write)))
1787 ((1 << BTREE_NODE_never_write)|
1788 (1 << BTREE_NODE_write_blocked)))
1792 (old & (1 << BTREE_NODE_will_make_reachable)))
1795 if (old & (1 << BTREE_NODE_write_in_flight))
1798 new &= ~(1 << BTREE_NODE_dirty);
1799 new &= ~(1 << BTREE_NODE_need_write);
1800 new |= (1 << BTREE_NODE_write_in_flight);
1801 new |= (1 << BTREE_NODE_write_in_flight_inner);
1802 new |= (1 << BTREE_NODE_just_written);
1803 new ^= (1 << BTREE_NODE_write_idx);
1804 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1806 if (new & (1U << BTREE_NODE_need_write))
1809 atomic_dec(&c->btree_cache.dirty);
1811 BUG_ON(btree_node_fake(b));
1812 BUG_ON((b->will_make_reachable != 0) != !b->written);
1814 BUG_ON(b->written >= btree_sectors(c));
1815 BUG_ON(b->written & (block_sectors(c) - 1));
1816 BUG_ON(bset_written(b, btree_bset_last(b)));
1817 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1818 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1820 bch2_sort_whiteouts(c, b);
1822 sort_iter_init(&sort_iter, b);
1825 ? sizeof(struct btree_node)
1826 : sizeof(struct btree_node_entry);
1828 bytes += b->whiteout_u64s * sizeof(u64);
1830 for_each_bset(b, t) {
1833 if (bset_written(b, i))
1836 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1837 sort_iter_add(&sort_iter,
1838 btree_bkey_first(b, t),
1839 btree_bkey_last(b, t));
1840 seq = max(seq, le64_to_cpu(i->journal_seq));
1843 BUG_ON(b->written && !seq);
1845 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1848 /* buffer must be a multiple of the block size */
1849 bytes = round_up(bytes, block_bytes(c));
1851 data = btree_bounce_alloc(c, bytes, &used_mempool);
1859 bne->keys = b->data->keys;
1863 i->journal_seq = cpu_to_le64(seq);
1866 sort_iter_add(&sort_iter,
1867 unwritten_whiteouts_start(c, b),
1868 unwritten_whiteouts_end(c, b));
1869 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1871 b->whiteout_u64s = 0;
1873 u64s = bch2_sort_keys(i->start, &sort_iter, false);
1874 le16_add_cpu(&i->u64s, u64s);
1876 set_needs_whiteout(i, false);
1878 /* do we have data to write? */
1879 if (b->written && !i->u64s)
1882 bytes_to_write = vstruct_end(i) - data;
1883 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1885 memset(data + bytes_to_write, 0,
1886 (sectors_to_write << 9) - bytes_to_write);
1888 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1889 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1890 BUG_ON(i->seq != b->data->keys.seq);
1892 i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1893 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1894 : cpu_to_le16(c->sb.version);
1895 SET_BSET_OFFSET(i, b->written);
1896 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1898 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1899 validate_before_checksum = true;
1901 /* validate_bset will be modifying: */
1902 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1903 validate_before_checksum = true;
1905 /* if we're going to be encrypting, check metadata validity first: */
1906 if (validate_before_checksum &&
1907 validate_bset_for_write(c, b, i, sectors_to_write))
1910 ret = bset_encrypt(c, i, b->written << 9);
1911 if (bch2_fs_fatal_err_on(ret, c,
1912 "error encrypting btree node: %i\n", ret))
1915 nonce = btree_nonce(i, b->written << 9);
1918 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1920 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1922 /* if we're not encrypting, check metadata after checksumming: */
1923 if (!validate_before_checksum &&
1924 validate_bset_for_write(c, b, i, sectors_to_write))
1928 * We handle btree write errors by immediately halting the journal -
1929 * after we've done that, we can't issue any subsequent btree writes
1930 * because they might have pointers to new nodes that failed to write.
1932 * Furthermore, there's no point in doing any more btree writes because
1933 * with the journal stopped, we're never going to update the journal to
1934 * reflect that those writes were done and the data flushed from the
1937 * Also on journal error, the pending write may have updates that were
1938 * never journalled (interior nodes, see btree_update_nodes_written()) -
1939 * it's critical that we don't do the write in that case otherwise we
1940 * will have updates visible that weren't in the journal:
1942 * Make sure to update b->written so bch2_btree_init_next() doesn't
1945 if (bch2_journal_error(&c->journal) ||
1949 trace_btree_write(b, bytes_to_write, sectors_to_write);
1951 wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1952 buf_pages(data, sectors_to_write << 9),
1954 struct btree_write_bio, wbio.bio);
1955 wbio_init(&wbio->wbio.bio);
1957 wbio->data_bytes = bytes;
1958 wbio->sector_offset = b->written;
1960 wbio->wbio.used_mempool = used_mempool;
1961 wbio->wbio.first_btree_write = !b->written;
1962 wbio->wbio.bio.bi_opf = REQ_OP_WRITE|REQ_META;
1963 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
1964 wbio->wbio.bio.bi_private = b;
1966 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1968 bkey_copy(&wbio->key, &b->key);
1970 b->written += sectors_to_write;
1972 if (wbio->wbio.first_btree_write &&
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(b->written);
1977 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
1978 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
1979 cpu_to_le16(b->written);
1981 atomic64_inc(&c->btree_writes_nr);
1982 atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1984 INIT_WORK(&wbio->work, btree_write_submit);
1985 queue_work(c->io_complete_wq, &wbio->work);
1988 set_btree_node_noevict(b);
1990 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1991 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1992 cpu_to_le16(sectors_to_write);
1993 b->written += sectors_to_write;
1995 btree_bounce_free(c, bytes, used_mempool, data);
1996 __btree_node_write_done(c, b);
2000 * Work that must be done with write lock held:
2002 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2004 bool invalidated_iter = false;
2005 struct btree_node_entry *bne;
2006 struct bset_tree *t;
2008 if (!btree_node_just_written(b))
2011 BUG_ON(b->whiteout_u64s);
2013 clear_btree_node_just_written(b);
2016 * Note: immediately after write, bset_written() doesn't work - the
2017 * amount of data we had to write after compaction might have been
2018 * smaller than the offset of the last bset.
2020 * However, we know that all bsets have been written here, as long as
2021 * we're still holding the write lock:
2025 * XXX: decide if we really want to unconditionally sort down to a
2029 btree_node_sort(c, b, 0, b->nsets, true);
2030 invalidated_iter = true;
2032 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2036 set_needs_whiteout(bset(b, t), true);
2038 bch2_btree_verify(c, b);
2041 * If later we don't unconditionally sort down to a single bset, we have
2042 * to ensure this is still true:
2044 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2046 bne = want_new_bset(c, b);
2048 bch2_bset_init_next(c, b, bne);
2050 bch2_btree_build_aux_trees(b);
2052 return invalidated_iter;
2056 * Use this one if the node is intent locked:
2058 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2059 enum six_lock_type lock_type_held,
2062 if (lock_type_held == SIX_LOCK_intent ||
2063 (lock_type_held == SIX_LOCK_read &&
2064 six_lock_tryupgrade(&b->c.lock))) {
2065 __bch2_btree_node_write(c, b, flags);
2067 /* don't cycle lock unnecessarily: */
2068 if (btree_node_just_written(b) &&
2069 six_trylock_write(&b->c.lock)) {
2070 bch2_btree_post_write_cleanup(c, b);
2071 six_unlock_write(&b->c.lock);
2074 if (lock_type_held == SIX_LOCK_read)
2075 six_lock_downgrade(&b->c.lock);
2077 __bch2_btree_node_write(c, b, flags);
2078 if (lock_type_held == SIX_LOCK_write &&
2079 btree_node_just_written(b))
2080 bch2_btree_post_write_cleanup(c, b);
2084 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2086 struct bucket_table *tbl;
2087 struct rhash_head *pos;
2092 for_each_cached_btree(b, c, tbl, i, pos)
2093 if (test_bit(flag, &b->flags)) {
2095 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2102 void bch2_btree_flush_all_reads(struct bch_fs *c)
2104 __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2107 void bch2_btree_flush_all_writes(struct bch_fs *c)
2109 __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);