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_ge(l.p, bkey_start_pos(&r)));
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 * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
456 * The first bset is going to be of similar order to the size of the node, the
457 * last bset is bounded by btree_write_set_buffer(), which is set to keep the
458 * memmove on insert from being too expensive: the middle bset should, ideally,
459 * be the geometric mean of the first and the last.
461 * Returns true if the middle bset is greater than that geometric mean:
463 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
465 unsigned mid_u64s_bits =
466 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
468 return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
472 * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
475 * Safe to call if there already is an unwritten bset - will only add a new bset
476 * if @b doesn't already have one.
478 * Returns true if we sorted (i.e. invalidated iterators
480 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
482 struct bch_fs *c = trans->c;
483 struct btree_node_entry *bne;
484 bool reinit_iter = false;
486 EBUG_ON(!(b->c.lock.state.seq & 1));
487 BUG_ON(bset_written(b, bset(b, &b->set[1])));
488 BUG_ON(btree_node_just_written(b));
490 if (b->nsets == MAX_BSETS &&
491 !btree_node_write_in_flight(b) &&
492 should_compact_all(c, b)) {
493 bch2_btree_node_write(c, b, SIX_LOCK_write,
494 BTREE_WRITE_init_next_bset);
498 if (b->nsets == MAX_BSETS &&
499 btree_node_compact(c, b))
502 BUG_ON(b->nsets >= MAX_BSETS);
504 bne = want_new_bset(c, b);
506 bch2_bset_init_next(c, b, bne);
508 bch2_btree_build_aux_trees(b);
511 bch2_trans_node_reinit_iter(trans, b);
514 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
517 prt_printf(out, "%s level %u/%u\n ",
518 bch2_btree_ids[b->c.btree_id],
520 c->btree_roots[b->c.btree_id].level);
521 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
524 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
526 struct btree *b, struct bset *i,
527 unsigned offset, int write)
529 prt_printf(out, bch2_log_msg(c, "%s"),
531 ? "error validating btree node "
532 : "corrupt btree node before write ");
534 prt_printf(out, "on %s ", ca->name);
535 prt_printf(out, "at btree ");
536 btree_pos_to_text(out, c, b);
538 prt_printf(out, "\n node offset %u", b->written);
540 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
544 enum btree_err_type {
546 * We can repair this locally, and we're after the checksum check so
547 * there's no need to try another replica:
551 * We can repair this if we have to, but we should try reading another
554 BTREE_ERR_WANT_RETRY,
556 * Read another replica if we have one, otherwise consider the whole
559 BTREE_ERR_MUST_RETRY,
561 BTREE_ERR_INCOMPATIBLE,
564 enum btree_validate_ret {
565 BTREE_RETRY_READ = 64,
568 static int __btree_err(enum btree_err_type type,
575 const char *fmt, ...)
577 struct printbuf out = PRINTBUF;
579 int ret = -BCH_ERR_fsck_fix;
581 btree_err_msg(&out, c, ca, b, i, b->written, write);
584 prt_vprintf(&out, fmt, args);
587 if (write == WRITE) {
588 bch2_print_string_as_lines(KERN_ERR, out.buf);
589 ret = c->opts.errors == BCH_ON_ERROR_continue
591 : -BCH_ERR_fsck_errors_not_fixed;
595 if (!have_retry && type == BTREE_ERR_WANT_RETRY)
596 type = BTREE_ERR_FIXABLE;
597 if (!have_retry && type == BTREE_ERR_MUST_RETRY)
598 type = BTREE_ERR_BAD_NODE;
601 case BTREE_ERR_FIXABLE:
602 mustfix_fsck_err(c, "%s", out.buf);
603 ret = -BCH_ERR_fsck_fix;
605 case BTREE_ERR_WANT_RETRY:
606 case BTREE_ERR_MUST_RETRY:
607 bch2_print_string_as_lines(KERN_ERR, out.buf);
608 ret = BTREE_RETRY_READ;
610 case BTREE_ERR_BAD_NODE:
611 bch2_print_string_as_lines(KERN_ERR, out.buf);
612 bch2_topology_error(c);
613 ret = -BCH_ERR_need_topology_repair;
615 case BTREE_ERR_INCOMPATIBLE:
616 bch2_print_string_as_lines(KERN_ERR, out.buf);
617 ret = -BCH_ERR_fsck_errors_not_fixed;
628 #define btree_err(type, c, ca, b, i, msg, ...) \
630 int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
632 if (_ret != -BCH_ERR_fsck_fix) \
637 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
640 * When btree topology repair changes the start or end of a node, that might
641 * mean we have to drop keys that are no longer inside the node:
644 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
648 struct bkey unpacked;
649 struct btree_node_iter iter;
651 for_each_bset(b, t) {
652 struct bset *i = bset(b, t);
653 struct bkey_packed *k;
655 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
656 if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
660 unsigned shift = (u64 *) k - (u64 *) i->start;
662 memmove_u64s_down(i->start, k,
663 (u64 *) vstruct_end(i) - (u64 *) k);
664 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
665 set_btree_bset_end(b, t);
668 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
669 if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
672 if (k != vstruct_last(i)) {
673 i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
674 set_btree_bset_end(b, t);
679 * Always rebuild search trees: eytzinger search tree nodes directly
680 * depend on the values of min/max key:
682 bch2_bset_set_no_aux_tree(b, b->set);
683 bch2_btree_build_aux_trees(b);
685 for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
686 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
687 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
691 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
692 struct btree *b, struct bset *i,
693 unsigned offset, unsigned sectors,
694 int write, bool have_retry, bool *saw_error)
696 unsigned version = le16_to_cpu(i->version);
698 struct printbuf buf1 = PRINTBUF;
699 struct printbuf buf2 = PRINTBUF;
702 btree_err_on((version != BCH_BSET_VERSION_OLD &&
703 version < bcachefs_metadata_version_min) ||
704 version >= bcachefs_metadata_version_max,
705 BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
706 "unsupported bset version");
708 if (btree_err_on(version < c->sb.version_min,
709 BTREE_ERR_FIXABLE, c, NULL, b, i,
710 "bset version %u older than superblock version_min %u",
711 version, c->sb.version_min)) {
712 mutex_lock(&c->sb_lock);
713 c->disk_sb.sb->version_min = cpu_to_le16(version);
715 mutex_unlock(&c->sb_lock);
718 if (btree_err_on(version > c->sb.version,
719 BTREE_ERR_FIXABLE, c, NULL, b, i,
720 "bset version %u newer than superblock version %u",
721 version, c->sb.version)) {
722 mutex_lock(&c->sb_lock);
723 c->disk_sb.sb->version = cpu_to_le16(version);
725 mutex_unlock(&c->sb_lock);
728 btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
729 BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
730 "BSET_SEPARATE_WHITEOUTS no longer supported");
732 if (btree_err_on(offset + sectors > btree_sectors(c),
733 BTREE_ERR_FIXABLE, c, ca, b, i,
734 "bset past end of btree node")) {
740 btree_err_on(offset && !i->u64s,
741 BTREE_ERR_FIXABLE, c, ca, b, i,
744 btree_err_on(BSET_OFFSET(i) &&
745 BSET_OFFSET(i) != offset,
746 BTREE_ERR_WANT_RETRY, c, ca, b, i,
747 "bset at wrong sector offset");
750 struct btree_node *bn =
751 container_of(i, struct btree_node, keys);
752 /* These indicate that we read the wrong btree node: */
754 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
755 struct bch_btree_ptr_v2 *bp =
756 &bkey_i_to_btree_ptr_v2(&b->key)->v;
759 btree_err_on(bp->seq != bn->keys.seq,
760 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
761 "incorrect sequence number (wrong btree node)");
764 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
765 BTREE_ERR_MUST_RETRY, c, ca, b, i,
766 "incorrect btree id");
768 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
769 BTREE_ERR_MUST_RETRY, c, ca, b, i,
773 compat_btree_node(b->c.level, b->c.btree_id, version,
774 BSET_BIG_ENDIAN(i), write, bn);
776 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
777 struct bch_btree_ptr_v2 *bp =
778 &bkey_i_to_btree_ptr_v2(&b->key)->v;
780 if (BTREE_PTR_RANGE_UPDATED(bp)) {
781 b->data->min_key = bp->min_key;
782 b->data->max_key = b->key.k.p;
785 btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
786 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
787 "incorrect min_key: got %s should be %s",
788 (printbuf_reset(&buf1),
789 bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
790 (printbuf_reset(&buf2),
791 bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
794 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
795 BTREE_ERR_MUST_RETRY, c, ca, b, i,
796 "incorrect max key %s",
797 (printbuf_reset(&buf1),
798 bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
801 compat_btree_node(b->c.level, b->c.btree_id, version,
802 BSET_BIG_ENDIAN(i), write, bn);
804 err = bch2_bkey_format_validate(&bn->format);
806 BTREE_ERR_BAD_NODE, c, ca, b, i,
807 "invalid bkey format: %s", err);
809 compat_bformat(b->c.level, b->c.btree_id, version,
810 BSET_BIG_ENDIAN(i), write,
815 printbuf_exit(&buf2);
816 printbuf_exit(&buf1);
820 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
822 bool updated_range, int rw,
823 struct printbuf *err)
825 return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
826 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
827 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
830 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
831 struct bset *i, int write,
832 bool have_retry, bool *saw_error)
834 unsigned version = le16_to_cpu(i->version);
835 struct bkey_packed *k, *prev = NULL;
836 struct printbuf buf = PRINTBUF;
837 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
838 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
842 k != vstruct_last(i);) {
846 if (btree_err_on(bkey_next(k) > vstruct_last(i),
847 BTREE_ERR_FIXABLE, c, NULL, b, i,
848 "key extends past end of bset")) {
849 i->u64s = cpu_to_le16((u64 *) k - i->_data);
853 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
854 BTREE_ERR_FIXABLE, c, NULL, b, i,
855 "invalid bkey format %u", k->format)) {
856 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
857 memmove_u64s_down(k, bkey_next(k),
858 (u64 *) vstruct_end(i) - (u64 *) k);
862 /* XXX: validate k->u64s */
864 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
865 BSET_BIG_ENDIAN(i), write,
868 u = __bkey_disassemble(b, k, &tmp);
870 printbuf_reset(&buf);
871 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
872 printbuf_reset(&buf);
873 prt_printf(&buf, "invalid bkey: ");
874 bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
875 prt_printf(&buf, "\n ");
876 bch2_bkey_val_to_text(&buf, c, u.s_c);
878 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
880 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
881 memmove_u64s_down(k, bkey_next(k),
882 (u64 *) vstruct_end(i) - (u64 *) k);
887 bch2_bkey_compat(b->c.level, b->c.btree_id, version,
888 BSET_BIG_ENDIAN(i), write,
891 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
892 struct bkey up = bkey_unpack_key(b, prev);
894 printbuf_reset(&buf);
895 prt_printf(&buf, "keys out of order: ");
896 bch2_bkey_to_text(&buf, &up);
897 prt_printf(&buf, " > ");
898 bch2_bkey_to_text(&buf, u.k);
900 bch2_dump_bset(c, b, i, 0);
902 if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
903 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
904 memmove_u64s_down(k, bkey_next(k),
905 (u64 *) vstruct_end(i) - (u64 *) k);
918 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
919 struct btree *b, bool have_retry, bool *saw_error)
921 struct btree_node_entry *bne;
922 struct sort_iter *iter;
923 struct btree_node *sorted;
924 struct bkey_packed *k;
925 struct bch_extent_ptr *ptr;
927 bool used_mempool, blacklisted;
928 bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
929 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
931 unsigned blacklisted_written, nonblacklisted_written = 0;
932 unsigned ptr_written = btree_ptr_sectors_written(&b->key);
933 struct printbuf buf = PRINTBUF;
934 int ret = 0, retry_read = 0, write = READ;
936 b->version_ondisk = U16_MAX;
937 /* We might get called multiple times on read retry: */
940 iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
941 sort_iter_init(iter, b);
942 iter->size = (btree_blocks(c) + 1) * 2;
944 if (bch2_meta_read_fault("btree"))
945 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
948 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
949 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
950 "bad magic: want %llx, got %llx",
951 bset_magic(c), le64_to_cpu(b->data->magic));
953 btree_err_on(!b->data->keys.seq,
954 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
955 "bad btree header: seq 0");
957 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
958 struct bch_btree_ptr_v2 *bp =
959 &bkey_i_to_btree_ptr_v2(&b->key)->v;
961 btree_err_on(b->data->keys.seq != bp->seq,
962 BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
963 "got wrong btree node (seq %llx want %llx)",
964 b->data->keys.seq, bp->seq);
967 while (b->written < (ptr_written ?: btree_sectors(c))) {
970 struct bch_csum csum;
971 bool first = !b->written;
976 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
977 BTREE_ERR_WANT_RETRY, c, ca, b, i,
978 "unknown checksum type %llu",
981 nonce = btree_nonce(i, b->written << 9);
982 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
984 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
985 BTREE_ERR_WANT_RETRY, c, ca, b, i,
988 ret = bset_encrypt(c, i, b->written << 9);
989 if (bch2_fs_fatal_err_on(ret, c,
990 "error decrypting btree node: %i", ret))
993 btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
994 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
995 BTREE_ERR_INCOMPATIBLE, c, NULL, b, NULL,
996 "btree node does not have NEW_EXTENT_OVERWRITE set");
998 sectors = vstruct_sectors(b->data, c->block_bits);
1000 bne = write_block(b);
1003 if (i->seq != b->data->keys.seq)
1006 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1007 BTREE_ERR_WANT_RETRY, c, ca, b, i,
1008 "unknown checksum type %llu",
1011 nonce = btree_nonce(i, b->written << 9);
1012 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1014 btree_err_on(bch2_crc_cmp(csum, bne->csum),
1015 BTREE_ERR_WANT_RETRY, c, ca, b, i,
1016 "invalid checksum");
1018 ret = bset_encrypt(c, i, b->written << 9);
1019 if (bch2_fs_fatal_err_on(ret, c,
1020 "error decrypting btree node: %i\n", ret))
1023 sectors = vstruct_sectors(bne, c->block_bits);
1026 b->version_ondisk = min(b->version_ondisk,
1027 le16_to_cpu(i->version));
1029 ret = validate_bset(c, ca, b, i, b->written, sectors,
1030 READ, have_retry, saw_error);
1035 btree_node_set_format(b, b->data->format);
1037 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1041 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1043 blacklisted = bch2_journal_seq_is_blacklisted(c,
1044 le64_to_cpu(i->journal_seq),
1047 btree_err_on(blacklisted && first,
1048 BTREE_ERR_FIXABLE, c, ca, b, i,
1049 "first btree node bset has blacklisted journal seq (%llu)",
1050 le64_to_cpu(i->journal_seq));
1052 btree_err_on(blacklisted && ptr_written,
1053 BTREE_ERR_FIXABLE, c, ca, b, i,
1054 "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1055 le64_to_cpu(i->journal_seq),
1056 b->written, b->written + sectors, ptr_written);
1058 b->written += sectors;
1060 if (blacklisted && !first)
1067 nonblacklisted_written = b->written;
1071 btree_err_on(b->written < ptr_written,
1072 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1073 "btree node data missing: expected %u sectors, found %u",
1074 ptr_written, b->written);
1076 for (bne = write_block(b);
1077 bset_byte_offset(b, bne) < btree_bytes(c);
1078 bne = (void *) bne + block_bytes(c))
1079 btree_err_on(bne->keys.seq == b->data->keys.seq &&
1080 !bch2_journal_seq_is_blacklisted(c,
1081 le64_to_cpu(bne->keys.journal_seq),
1083 BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1084 "found bset signature after last bset");
1087 * Blacklisted bsets are those that were written after the most recent
1088 * (flush) journal write. Since there wasn't a flush, they may not have
1089 * made it to all devices - which means we shouldn't write new bsets
1090 * after them, as that could leave a gap and then reads from that device
1091 * wouldn't find all the bsets in that btree node - which means it's
1092 * important that we start writing new bsets after the most recent _non_
1095 blacklisted_written = b->written;
1096 b->written = nonblacklisted_written;
1099 sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1100 sorted->keys.u64s = 0;
1102 set_btree_bset(b, b->set, &b->data->keys);
1104 b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1106 u64s = le16_to_cpu(sorted->keys.u64s);
1108 sorted->keys.u64s = cpu_to_le16(u64s);
1109 swap(sorted, b->data);
1110 set_btree_bset(b, b->set, &b->data->keys);
1113 BUG_ON(b->nr.live_u64s != u64s);
1115 btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1118 bch2_btree_node_drop_keys_outside_node(b);
1121 for (k = i->start; k != vstruct_last(i);) {
1123 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1125 printbuf_reset(&buf);
1127 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1128 (bch2_inject_invalid_keys &&
1129 !bversion_cmp(u.k->version, MAX_VERSION))) {
1130 printbuf_reset(&buf);
1132 prt_printf(&buf, "invalid bkey: ");
1133 bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1134 prt_printf(&buf, "\n ");
1135 bch2_bkey_val_to_text(&buf, c, u.s_c);
1137 btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1139 btree_keys_account_key_drop(&b->nr, 0, k);
1141 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1142 memmove_u64s_down(k, bkey_next(k),
1143 (u64 *) vstruct_end(i) - (u64 *) k);
1144 set_btree_bset_end(b, b->set);
1148 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1149 struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1157 bch2_bset_build_aux_tree(b, b->set, false);
1159 set_needs_whiteout(btree_bset_first(b), true);
1161 btree_node_reset_sib_u64s(b);
1163 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1164 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1166 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1167 set_btree_node_need_rewrite(b);
1171 set_btree_node_need_rewrite(b);
1173 mempool_free(iter, &c->fill_iter);
1174 printbuf_exit(&buf);
1177 if (ret == BTREE_RETRY_READ)
1180 set_btree_node_read_error(b);
1184 static void btree_node_read_work(struct work_struct *work)
1186 struct btree_read_bio *rb =
1187 container_of(work, struct btree_read_bio, work);
1188 struct bch_fs *c = rb->c;
1189 struct btree *b = rb->b;
1190 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1191 struct bio *bio = &rb->bio;
1192 struct bch_io_failures failed = { .nr = 0 };
1193 struct printbuf buf = PRINTBUF;
1194 bool saw_error = false;
1201 bch_info(c, "retrying read");
1202 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1203 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1204 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1205 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1206 bio->bi_iter.bi_size = btree_bytes(c);
1208 if (rb->have_ioref) {
1209 bio_set_dev(bio, ca->disk_sb.bdev);
1210 submit_bio_wait(bio);
1212 bio->bi_status = BLK_STS_REMOVED;
1215 printbuf_reset(&buf);
1216 btree_pos_to_text(&buf, c, b);
1217 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1218 bch2_blk_status_to_str(bio->bi_status), buf.buf);
1220 percpu_ref_put(&ca->io_ref);
1221 rb->have_ioref = false;
1223 bch2_mark_io_failure(&failed, &rb->pick);
1225 can_retry = bch2_bkey_pick_read_device(c,
1226 bkey_i_to_s_c(&b->key),
1227 &failed, &rb->pick) > 0;
1229 if (!bio->bi_status &&
1230 !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1232 bch_info(c, "retry success");
1239 set_btree_node_read_error(b);
1244 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1247 printbuf_exit(&buf);
1249 if (saw_error && !btree_node_read_error(b))
1250 bch2_btree_node_rewrite_async(c, b);
1252 clear_btree_node_read_in_flight(b);
1253 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1256 static void btree_node_read_endio(struct bio *bio)
1258 struct btree_read_bio *rb =
1259 container_of(bio, struct btree_read_bio, bio);
1260 struct bch_fs *c = rb->c;
1262 if (rb->have_ioref) {
1263 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1265 bch2_latency_acct(ca, rb->start_time, READ);
1268 queue_work(c->io_complete_wq, &rb->work);
1271 struct btree_node_read_all {
1276 void *buf[BCH_REPLICAS_MAX];
1277 struct bio *bio[BCH_REPLICAS_MAX];
1278 int err[BCH_REPLICAS_MAX];
1281 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1283 struct btree_node *bn = data;
1284 struct btree_node_entry *bne;
1285 unsigned offset = 0;
1287 if (le64_to_cpu(bn->magic) != bset_magic(c))
1290 while (offset < btree_sectors(c)) {
1292 offset += vstruct_sectors(bn, c->block_bits);
1294 bne = data + (offset << 9);
1295 if (bne->keys.seq != bn->keys.seq)
1297 offset += vstruct_sectors(bne, c->block_bits);
1304 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1306 struct btree_node *bn = data;
1307 struct btree_node_entry *bne;
1312 while (offset < btree_sectors(c)) {
1313 bne = data + (offset << 9);
1314 if (bne->keys.seq == bn->keys.seq)
1323 static void btree_node_read_all_replicas_done(struct closure *cl)
1325 struct btree_node_read_all *ra =
1326 container_of(cl, struct btree_node_read_all, cl);
1327 struct bch_fs *c = ra->c;
1328 struct btree *b = ra->b;
1329 struct printbuf buf = PRINTBUF;
1330 bool dump_bset_maps = false;
1331 bool have_retry = false;
1332 int ret = 0, best = -1, write = READ;
1333 unsigned i, written = 0, written2 = 0;
1334 __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1335 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1336 bool _saw_error = false, *saw_error = &_saw_error;
1338 for (i = 0; i < ra->nr; i++) {
1339 struct btree_node *bn = ra->buf[i];
1344 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1345 (seq && seq != bn->keys.seq))
1350 written = btree_node_sectors_written(c, bn);
1354 written2 = btree_node_sectors_written(c, ra->buf[i]);
1355 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1356 "btree node sectors written mismatch: %u != %u",
1357 written, written2) ||
1358 btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1359 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1360 "found bset signature after last bset") ||
1361 btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1362 BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1363 "btree node replicas content mismatch"))
1364 dump_bset_maps = true;
1366 if (written2 > written) {
1372 if (dump_bset_maps) {
1373 for (i = 0; i < ra->nr; i++) {
1374 struct btree_node *bn = ra->buf[i];
1375 struct btree_node_entry *bne = NULL;
1376 unsigned offset = 0, sectors;
1382 printbuf_reset(&buf);
1384 while (offset < btree_sectors(c)) {
1386 sectors = vstruct_sectors(bn, c->block_bits);
1388 bne = ra->buf[i] + (offset << 9);
1389 if (bne->keys.seq != bn->keys.seq)
1391 sectors = vstruct_sectors(bne, c->block_bits);
1394 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1395 if (bne && bch2_journal_seq_is_blacklisted(c,
1396 le64_to_cpu(bne->keys.journal_seq), false))
1397 prt_printf(&buf, "*");
1401 while (offset < btree_sectors(c)) {
1402 bne = ra->buf[i] + (offset << 9);
1403 if (bne->keys.seq == bn->keys.seq) {
1405 prt_printf(&buf, " GAP");
1408 sectors = vstruct_sectors(bne, c->block_bits);
1409 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1410 if (bch2_journal_seq_is_blacklisted(c,
1411 le64_to_cpu(bne->keys.journal_seq), false))
1412 prt_printf(&buf, "*");
1417 bch_err(c, "replica %u:%s", i, buf.buf);
1422 memcpy(b->data, ra->buf[best], btree_bytes(c));
1423 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1429 set_btree_node_read_error(b);
1430 else if (*saw_error)
1431 bch2_btree_node_rewrite_async(c, b);
1433 for (i = 0; i < ra->nr; i++) {
1434 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1435 bio_put(ra->bio[i]);
1438 closure_debug_destroy(&ra->cl);
1440 printbuf_exit(&buf);
1442 clear_btree_node_read_in_flight(b);
1443 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1446 static void btree_node_read_all_replicas_endio(struct bio *bio)
1448 struct btree_read_bio *rb =
1449 container_of(bio, struct btree_read_bio, bio);
1450 struct bch_fs *c = rb->c;
1451 struct btree_node_read_all *ra = rb->ra;
1453 if (rb->have_ioref) {
1454 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1456 bch2_latency_acct(ca, rb->start_time, READ);
1459 ra->err[rb->idx] = bio->bi_status;
1460 closure_put(&ra->cl);
1464 * XXX This allocates multiple times from the same mempools, and can deadlock
1465 * under sufficient memory pressure (but is only a debug path)
1467 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1469 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1470 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1471 const union bch_extent_entry *entry;
1472 struct extent_ptr_decoded pick;
1473 struct btree_node_read_all *ra;
1476 ra = kzalloc(sizeof(*ra), GFP_NOFS);
1480 closure_init(&ra->cl, NULL);
1483 ra->nr = bch2_bkey_nr_ptrs(k);
1485 for (i = 0; i < ra->nr; i++) {
1486 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1487 ra->bio[i] = bio_alloc_bioset(NULL,
1488 buf_pages(ra->buf[i], btree_bytes(c)),
1489 REQ_OP_READ|REQ_SYNC|REQ_META,
1495 bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1496 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1497 struct btree_read_bio *rb =
1498 container_of(ra->bio[i], struct btree_read_bio, bio);
1502 rb->start_time = local_clock();
1503 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1506 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1507 rb->bio.bi_end_io = btree_node_read_all_replicas_endio;
1508 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1510 if (rb->have_ioref) {
1511 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1512 bio_sectors(&rb->bio));
1513 bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1515 closure_get(&ra->cl);
1516 submit_bio(&rb->bio);
1518 ra->err[i] = BLK_STS_REMOVED;
1525 closure_sync(&ra->cl);
1526 btree_node_read_all_replicas_done(&ra->cl);
1528 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1535 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1538 struct extent_ptr_decoded pick;
1539 struct btree_read_bio *rb;
1544 trace_and_count(c, btree_node_read, c, b);
1546 if (bch2_verify_all_btree_replicas &&
1547 !btree_node_read_all_replicas(c, b, sync))
1550 ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1554 struct printbuf buf = PRINTBUF;
1556 prt_str(&buf, "btree node read error: no device to read from\n at ");
1557 btree_pos_to_text(&buf, c, b);
1558 bch_err(c, "%s", buf.buf);
1560 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1561 bch2_fatal_error(c);
1563 set_btree_node_read_error(b);
1564 clear_btree_node_read_in_flight(b);
1565 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1566 printbuf_exit(&buf);
1570 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1572 bio = bio_alloc_bioset(NULL,
1573 buf_pages(b->data, btree_bytes(c)),
1574 REQ_OP_READ|REQ_SYNC|REQ_META,
1577 rb = container_of(bio, struct btree_read_bio, bio);
1581 rb->start_time = local_clock();
1582 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1584 INIT_WORK(&rb->work, btree_node_read_work);
1585 bio->bi_iter.bi_sector = pick.ptr.offset;
1586 bio->bi_end_io = btree_node_read_endio;
1587 bch2_bio_map(bio, b->data, btree_bytes(c));
1589 if (rb->have_ioref) {
1590 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1592 bio_set_dev(bio, ca->disk_sb.bdev);
1595 submit_bio_wait(bio);
1597 btree_node_read_work(&rb->work);
1602 bio->bi_status = BLK_STS_REMOVED;
1605 btree_node_read_work(&rb->work);
1607 queue_work(c->io_complete_wq, &rb->work);
1611 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1612 const struct bkey_i *k, unsigned level)
1618 closure_init_stack(&cl);
1621 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1625 b = bch2_btree_node_mem_alloc(c, level != 0);
1626 bch2_btree_cache_cannibalize_unlock(c);
1630 bkey_copy(&b->key, k);
1631 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1633 set_btree_node_read_in_flight(b);
1635 bch2_btree_node_read(c, b, true);
1637 if (btree_node_read_error(b)) {
1638 bch2_btree_node_hash_remove(&c->btree_cache, b);
1640 mutex_lock(&c->btree_cache.lock);
1641 list_move(&b->list, &c->btree_cache.freeable);
1642 mutex_unlock(&c->btree_cache.lock);
1648 bch2_btree_set_root_for_read(c, b);
1650 six_unlock_write(&b->c.lock);
1651 six_unlock_intent(&b->c.lock);
1656 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1657 struct btree_write *w)
1659 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1667 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1670 closure_put(&((struct btree_update *) new)->cl);
1672 bch2_journal_pin_drop(&c->journal, &w->journal);
1675 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1677 struct btree_write *w = btree_prev_write(b);
1678 unsigned long old, new, v;
1681 bch2_btree_complete_write(c, b, w);
1683 v = READ_ONCE(b->flags);
1687 if ((old & (1U << BTREE_NODE_dirty)) &&
1688 (old & (1U << BTREE_NODE_need_write)) &&
1689 !(old & (1U << BTREE_NODE_never_write)) &&
1690 !(old & (1U << BTREE_NODE_write_blocked)) &&
1691 !(old & (1U << BTREE_NODE_will_make_reachable))) {
1692 new &= ~(1U << BTREE_NODE_dirty);
1693 new &= ~(1U << BTREE_NODE_need_write);
1694 new |= (1U << BTREE_NODE_write_in_flight);
1695 new |= (1U << BTREE_NODE_write_in_flight_inner);
1696 new |= (1U << BTREE_NODE_just_written);
1697 new ^= (1U << BTREE_NODE_write_idx);
1699 type = new & BTREE_WRITE_TYPE_MASK;
1700 new &= ~BTREE_WRITE_TYPE_MASK;
1702 new &= ~(1U << BTREE_NODE_write_in_flight);
1703 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1705 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1707 if (new & (1U << BTREE_NODE_write_in_flight))
1708 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1710 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1713 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1715 struct btree_trans trans;
1717 bch2_trans_init(&trans, c, 0, 0);
1719 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1720 __btree_node_write_done(c, b);
1721 six_unlock_read(&b->c.lock);
1723 bch2_trans_exit(&trans);
1726 static void btree_node_write_work(struct work_struct *work)
1728 struct btree_write_bio *wbio =
1729 container_of(work, struct btree_write_bio, work);
1730 struct bch_fs *c = wbio->wbio.c;
1731 struct btree *b = wbio->wbio.bio.bi_private;
1732 struct bch_extent_ptr *ptr;
1735 btree_bounce_free(c,
1737 wbio->wbio.used_mempool,
1740 bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1741 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1743 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1746 if (wbio->wbio.first_btree_write) {
1747 if (wbio->wbio.failed.nr) {
1751 ret = bch2_trans_do(c, NULL, NULL, 0,
1752 bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1753 !wbio->wbio.failed.nr));
1758 bio_put(&wbio->wbio.bio);
1759 btree_node_write_done(c, b);
1762 set_btree_node_noevict(b);
1763 bch2_fs_fatal_error(c, "fatal error writing btree node");
1767 static void btree_node_write_endio(struct bio *bio)
1769 struct bch_write_bio *wbio = to_wbio(bio);
1770 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1771 struct bch_write_bio *orig = parent ?: wbio;
1772 struct btree_write_bio *wb = container_of(orig, struct btree_write_bio, wbio);
1773 struct bch_fs *c = wbio->c;
1774 struct btree *b = wbio->bio.bi_private;
1775 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1776 unsigned long flags;
1778 if (wbio->have_ioref)
1779 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1781 if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1782 bch2_blk_status_to_str(bio->bi_status)) ||
1783 bch2_meta_write_fault("btree")) {
1784 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1785 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1786 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1789 if (wbio->have_ioref)
1790 percpu_ref_put(&ca->io_ref);
1794 bio_endio(&parent->bio);
1798 clear_btree_node_write_in_flight_inner(b);
1799 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1800 INIT_WORK(&wb->work, btree_node_write_work);
1801 queue_work(c->btree_io_complete_wq, &wb->work);
1804 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1805 struct bset *i, unsigned sectors)
1807 struct printbuf buf = PRINTBUF;
1811 ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1812 BKEY_TYPE_btree, WRITE, &buf);
1815 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1816 printbuf_exit(&buf);
1820 ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1821 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1823 bch2_inconsistent_error(c);
1830 static void btree_write_submit(struct work_struct *work)
1832 struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1833 struct bch_extent_ptr *ptr;
1834 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1836 bkey_copy(&tmp.k, &wbio->key);
1838 bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1839 ptr->offset += wbio->sector_offset;
1841 bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1845 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1847 struct btree_write_bio *wbio;
1848 struct bset_tree *t;
1850 struct btree_node *bn = NULL;
1851 struct btree_node_entry *bne = NULL;
1852 struct sort_iter sort_iter;
1854 unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1857 unsigned long old, new;
1858 bool validate_before_checksum = false;
1859 enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1863 if (flags & BTREE_WRITE_ALREADY_STARTED)
1867 * We may only have a read lock on the btree node - the dirty bit is our
1868 * "lock" against racing with other threads that may be trying to start
1869 * a write, we do a write iff we clear the dirty bit. Since setting the
1870 * dirty bit requires a write lock, we can't race with other threads
1874 old = new = READ_ONCE(b->flags);
1876 if (!(old & (1 << BTREE_NODE_dirty)))
1879 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1880 !(old & (1 << BTREE_NODE_need_write)))
1884 ((1 << BTREE_NODE_never_write)|
1885 (1 << BTREE_NODE_write_blocked)))
1889 (old & (1 << BTREE_NODE_will_make_reachable)))
1892 if (old & (1 << BTREE_NODE_write_in_flight))
1895 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1896 type = new & BTREE_WRITE_TYPE_MASK;
1897 new &= ~BTREE_WRITE_TYPE_MASK;
1899 new &= ~(1 << BTREE_NODE_dirty);
1900 new &= ~(1 << BTREE_NODE_need_write);
1901 new |= (1 << BTREE_NODE_write_in_flight);
1902 new |= (1 << BTREE_NODE_write_in_flight_inner);
1903 new |= (1 << BTREE_NODE_just_written);
1904 new ^= (1 << BTREE_NODE_write_idx);
1905 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1907 if (new & (1U << BTREE_NODE_need_write))
1910 BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1912 atomic_dec(&c->btree_cache.dirty);
1914 BUG_ON(btree_node_fake(b));
1915 BUG_ON((b->will_make_reachable != 0) != !b->written);
1917 BUG_ON(b->written >= btree_sectors(c));
1918 BUG_ON(b->written & (block_sectors(c) - 1));
1919 BUG_ON(bset_written(b, btree_bset_last(b)));
1920 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1921 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1923 bch2_sort_whiteouts(c, b);
1925 sort_iter_init(&sort_iter, b);
1928 ? sizeof(struct btree_node)
1929 : sizeof(struct btree_node_entry);
1931 bytes += b->whiteout_u64s * sizeof(u64);
1933 for_each_bset(b, t) {
1936 if (bset_written(b, i))
1939 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1940 sort_iter_add(&sort_iter,
1941 btree_bkey_first(b, t),
1942 btree_bkey_last(b, t));
1943 seq = max(seq, le64_to_cpu(i->journal_seq));
1946 BUG_ON(b->written && !seq);
1948 /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1951 /* buffer must be a multiple of the block size */
1952 bytes = round_up(bytes, block_bytes(c));
1954 data = btree_bounce_alloc(c, bytes, &used_mempool);
1962 bne->keys = b->data->keys;
1966 i->journal_seq = cpu_to_le64(seq);
1969 sort_iter_add(&sort_iter,
1970 unwritten_whiteouts_start(c, b),
1971 unwritten_whiteouts_end(c, b));
1972 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1974 b->whiteout_u64s = 0;
1976 u64s = bch2_sort_keys(i->start, &sort_iter, false);
1977 le16_add_cpu(&i->u64s, u64s);
1979 BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1981 set_needs_whiteout(i, false);
1983 /* do we have data to write? */
1984 if (b->written && !i->u64s)
1987 bytes_to_write = vstruct_end(i) - data;
1988 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1991 b->key.k.type == KEY_TYPE_btree_ptr_v2)
1992 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1994 memset(data + bytes_to_write, 0,
1995 (sectors_to_write << 9) - bytes_to_write);
1997 BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1998 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1999 BUG_ON(i->seq != b->data->keys.seq);
2001 i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
2002 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
2003 : cpu_to_le16(c->sb.version);
2004 SET_BSET_OFFSET(i, b->written);
2005 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2007 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2008 validate_before_checksum = true;
2010 /* validate_bset will be modifying: */
2011 if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2012 validate_before_checksum = true;
2014 /* if we're going to be encrypting, check metadata validity first: */
2015 if (validate_before_checksum &&
2016 validate_bset_for_write(c, b, i, sectors_to_write))
2019 ret = bset_encrypt(c, i, b->written << 9);
2020 if (bch2_fs_fatal_err_on(ret, c,
2021 "error encrypting btree node: %i\n", ret))
2024 nonce = btree_nonce(i, b->written << 9);
2027 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2029 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2031 /* if we're not encrypting, check metadata after checksumming: */
2032 if (!validate_before_checksum &&
2033 validate_bset_for_write(c, b, i, sectors_to_write))
2037 * We handle btree write errors by immediately halting the journal -
2038 * after we've done that, we can't issue any subsequent btree writes
2039 * because they might have pointers to new nodes that failed to write.
2041 * Furthermore, there's no point in doing any more btree writes because
2042 * with the journal stopped, we're never going to update the journal to
2043 * reflect that those writes were done and the data flushed from the
2046 * Also on journal error, the pending write may have updates that were
2047 * never journalled (interior nodes, see btree_update_nodes_written()) -
2048 * it's critical that we don't do the write in that case otherwise we
2049 * will have updates visible that weren't in the journal:
2051 * Make sure to update b->written so bch2_btree_init_next() doesn't
2054 if (bch2_journal_error(&c->journal) ||
2058 trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2060 wbio = container_of(bio_alloc_bioset(NULL,
2061 buf_pages(data, sectors_to_write << 9),
2062 REQ_OP_WRITE|REQ_META,
2065 struct btree_write_bio, wbio.bio);
2066 wbio_init(&wbio->wbio.bio);
2068 wbio->data_bytes = bytes;
2069 wbio->sector_offset = b->written;
2071 wbio->wbio.used_mempool = used_mempool;
2072 wbio->wbio.first_btree_write = !b->written;
2073 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
2074 wbio->wbio.bio.bi_private = b;
2076 bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2078 bkey_copy(&wbio->key, &b->key);
2080 b->written += sectors_to_write;
2082 if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2083 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2084 cpu_to_le16(b->written);
2086 atomic64_inc(&c->btree_write_stats[type].nr);
2087 atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2089 INIT_WORK(&wbio->work, btree_write_submit);
2090 queue_work(c->io_complete_wq, &wbio->work);
2093 set_btree_node_noevict(b);
2094 b->written += sectors_to_write;
2096 btree_bounce_free(c, bytes, used_mempool, data);
2097 __btree_node_write_done(c, b);
2101 * Work that must be done with write lock held:
2103 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2105 bool invalidated_iter = false;
2106 struct btree_node_entry *bne;
2107 struct bset_tree *t;
2109 if (!btree_node_just_written(b))
2112 BUG_ON(b->whiteout_u64s);
2114 clear_btree_node_just_written(b);
2117 * Note: immediately after write, bset_written() doesn't work - the
2118 * amount of data we had to write after compaction might have been
2119 * smaller than the offset of the last bset.
2121 * However, we know that all bsets have been written here, as long as
2122 * we're still holding the write lock:
2126 * XXX: decide if we really want to unconditionally sort down to a
2130 btree_node_sort(c, b, 0, b->nsets, true);
2131 invalidated_iter = true;
2133 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2137 set_needs_whiteout(bset(b, t), true);
2139 bch2_btree_verify(c, b);
2142 * If later we don't unconditionally sort down to a single bset, we have
2143 * to ensure this is still true:
2145 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2147 bne = want_new_bset(c, b);
2149 bch2_bset_init_next(c, b, bne);
2151 bch2_btree_build_aux_trees(b);
2153 return invalidated_iter;
2157 * Use this one if the node is intent locked:
2159 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2160 enum six_lock_type lock_type_held,
2163 if (lock_type_held == SIX_LOCK_intent ||
2164 (lock_type_held == SIX_LOCK_read &&
2165 six_lock_tryupgrade(&b->c.lock))) {
2166 __bch2_btree_node_write(c, b, flags);
2168 /* don't cycle lock unnecessarily: */
2169 if (btree_node_just_written(b) &&
2170 six_trylock_write(&b->c.lock)) {
2171 bch2_btree_post_write_cleanup(c, b);
2172 six_unlock_write(&b->c.lock);
2175 if (lock_type_held == SIX_LOCK_read)
2176 six_lock_downgrade(&b->c.lock);
2178 __bch2_btree_node_write(c, b, flags);
2179 if (lock_type_held == SIX_LOCK_write &&
2180 btree_node_just_written(b))
2181 bch2_btree_post_write_cleanup(c, b);
2185 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2187 struct bucket_table *tbl;
2188 struct rhash_head *pos;
2194 for_each_cached_btree(b, c, tbl, i, pos)
2195 if (test_bit(flag, &b->flags)) {
2197 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2206 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2208 return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2211 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2213 return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2216 const char * const bch2_btree_write_types[] = {
2217 #define x(t, n) [n] = #t,
2218 BCH_BTREE_WRITE_TYPES()
2222 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2224 printbuf_tabstop_push(out, 20);
2225 printbuf_tabstop_push(out, 10);
2230 prt_str(out, "size");
2233 for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2234 u64 nr = atomic64_read(&c->btree_write_stats[i].nr);
2235 u64 bytes = atomic64_read(&c->btree_write_stats[i].bytes);
2237 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2241 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);