3 #include "bkey_methods.h"
4 #include "btree_cache.h"
6 #include "btree_iter.h"
7 #include "btree_locking.h"
8 #include "btree_update.h"
9 #include "btree_update_interior.h"
16 #include "journal_reclaim.h"
17 #include "journal_seq_blacklist.h"
20 #include <trace/events/bcachefs.h>
22 /* btree_node_iter_large: */
24 #define btree_node_iter_cmp_heap(h, _l, _r) \
25 __btree_node_iter_cmp((iter)->is_extents, b, \
26 __btree_node_offset_to_key(b, (_l).k), \
27 __btree_node_offset_to_key(b, (_r).k))
29 void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
31 const struct bkey_packed *k,
32 const struct bkey_packed *end)
35 struct btree_node_iter_set n =
36 ((struct btree_node_iter_set) {
37 __btree_node_key_to_offset(b, k),
38 __btree_node_key_to_offset(b, end)
41 __heap_add(iter, n, btree_node_iter_cmp_heap);
45 void bch2_btree_node_iter_large_advance(struct btree_node_iter_large *iter,
48 iter->data->k += __btree_node_offset_to_key(b, iter->data->k)->u64s;
51 EBUG_ON(iter->data->k > iter->data->end);
53 if (iter->data->k == iter->data->end)
54 heap_del(iter, 0, btree_node_iter_cmp_heap);
56 heap_sift_down(iter, 0, btree_node_iter_cmp_heap);
59 static void verify_no_dups(struct btree *b,
60 struct bkey_packed *start,
61 struct bkey_packed *end)
63 #ifdef CONFIG_BCACHEFS_DEBUG
64 struct bkey_packed *k;
66 for (k = start; k != end && bkey_next(k) != end; k = bkey_next(k)) {
67 struct bkey l = bkey_unpack_key(b, k);
68 struct bkey r = bkey_unpack_key(b, bkey_next(k));
70 BUG_ON(btree_node_is_extents(b)
71 ? bkey_cmp(l.p, bkey_start_pos(&r)) > 0
72 : bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
73 //BUG_ON(bkey_cmp_packed(&b->format, k, bkey_next(k)) >= 0);
78 static void clear_needs_whiteout(struct bset *i)
80 struct bkey_packed *k;
82 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
83 k->needs_whiteout = false;
86 static void set_needs_whiteout(struct bset *i)
88 struct bkey_packed *k;
90 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
91 k->needs_whiteout = true;
94 static void btree_bounce_free(struct bch_fs *c, unsigned order,
95 bool used_mempool, void *p)
98 mempool_free(p, &c->btree_bounce_pool);
100 vpfree(p, PAGE_SIZE << order);
103 static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
108 BUG_ON(order > btree_page_order(c));
110 *used_mempool = false;
111 p = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, order);
115 *used_mempool = true;
116 return mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
119 typedef int (*sort_cmp_fn)(struct btree *,
120 struct bkey_packed *,
121 struct bkey_packed *);
127 struct sort_iter_set {
128 struct bkey_packed *k, *end;
129 } data[MAX_BSETS + 1];
132 static void sort_iter_init(struct sort_iter *iter, struct btree *b)
134 memset(iter, 0, sizeof(*iter));
138 static inline void __sort_iter_sift(struct sort_iter *iter,
145 i + 1 < iter->used &&
146 cmp(iter->b, iter->data[i].k, iter->data[i + 1].k) > 0;
148 swap(iter->data[i], iter->data[i + 1]);
151 static inline void sort_iter_sift(struct sort_iter *iter, sort_cmp_fn cmp)
154 __sort_iter_sift(iter, 0, cmp);
157 static inline void sort_iter_sort(struct sort_iter *iter, sort_cmp_fn cmp)
159 unsigned i = iter->used;
162 __sort_iter_sift(iter, i, cmp);
165 static void sort_iter_add(struct sort_iter *iter,
166 struct bkey_packed *k,
167 struct bkey_packed *end)
169 BUG_ON(iter->used >= ARRAY_SIZE(iter->data));
172 iter->data[iter->used++] = (struct sort_iter_set) { k, end };
175 static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
177 return iter->used ? iter->data->k : NULL;
180 static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
182 iter->data->k = bkey_next(iter->data->k);
184 BUG_ON(iter->data->k > iter->data->end);
186 if (iter->data->k == iter->data->end)
187 array_remove_item(iter->data, iter->used, 0);
189 sort_iter_sift(iter, cmp);
192 static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
195 struct bkey_packed *ret = sort_iter_peek(iter);
198 sort_iter_advance(iter, cmp);
203 static inline int sort_key_whiteouts_cmp(struct btree *b,
204 struct bkey_packed *l,
205 struct bkey_packed *r)
207 return bkey_cmp_packed(b, l, r);
210 static unsigned sort_key_whiteouts(struct bkey_packed *dst,
211 struct sort_iter *iter)
213 struct bkey_packed *in, *out = dst;
215 sort_iter_sort(iter, sort_key_whiteouts_cmp);
217 while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
219 out = bkey_next(out);
222 return (u64 *) out - (u64 *) dst;
225 static inline int sort_extent_whiteouts_cmp(struct btree *b,
226 struct bkey_packed *l,
227 struct bkey_packed *r)
229 struct bkey ul = bkey_unpack_key(b, l);
230 struct bkey ur = bkey_unpack_key(b, r);
232 return bkey_cmp(bkey_start_pos(&ul), bkey_start_pos(&ur));
235 static unsigned sort_extent_whiteouts(struct bkey_packed *dst,
236 struct sort_iter *iter)
238 const struct bkey_format *f = &iter->b->format;
239 struct bkey_packed *in, *out = dst;
241 bool prev = false, l_packed = false;
242 u64 max_packed_size = bkey_field_max(f, BKEY_FIELD_SIZE);
243 u64 max_packed_offset = bkey_field_max(f, BKEY_FIELD_OFFSET);
246 max_packed_size = min_t(u64, max_packed_size, KEY_SIZE_MAX);
248 sort_iter_sort(iter, sort_extent_whiteouts_cmp);
250 while ((in = sort_iter_next(iter, sort_extent_whiteouts_cmp))) {
251 EBUG_ON(bkeyp_val_u64s(f, in));
252 EBUG_ON(in->type != KEY_TYPE_DISCARD);
254 r.k = bkey_unpack_key(iter->b, in);
257 bkey_cmp(l.k.p, bkey_start_pos(&r.k)) >= 0) {
258 if (bkey_cmp(l.k.p, r.k.p) >= 0)
262 ? min(max_packed_size, max_packed_offset -
263 bkey_start_offset(&l.k))
266 new_size = min(new_size, r.k.p.offset -
267 bkey_start_offset(&l.k));
269 BUG_ON(new_size < l.k.size);
271 bch2_key_resize(&l.k, new_size);
273 if (bkey_cmp(l.k.p, r.k.p) >= 0)
276 bch2_cut_front(l.k.p, &r);
280 if (!bch2_bkey_pack(out, &l, f)) {
284 out = bkey_next(out);
289 l_packed = bkey_packed(in);
293 if (!bch2_bkey_pack(out, &l, f)) {
297 out = bkey_next(out);
300 return (u64 *) out - (u64 *) dst;
303 static unsigned should_compact_bset(struct btree *b, struct bset_tree *t,
305 enum compact_mode mode)
307 unsigned bset_u64s = le16_to_cpu(bset(b, t)->u64s);
308 unsigned dead_u64s = bset_u64s - b->nr.bset_u64s[t - b->set];
310 if (mode == COMPACT_LAZY) {
311 if (should_compact_bset_lazy(b, t) ||
312 (compacting && bset_unwritten(b, bset(b, t))))
315 if (bset_written(b, bset(b, t)))
322 bool __bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
323 enum compact_mode mode)
325 const struct bkey_format *f = &b->format;
327 struct bkey_packed *whiteouts = NULL;
328 struct bkey_packed *u_start, *u_pos;
329 struct sort_iter sort_iter;
330 unsigned order, whiteout_u64s = 0, u64s;
331 bool used_mempool, compacting = false;
334 whiteout_u64s += should_compact_bset(b, t,
335 whiteout_u64s != 0, mode);
340 sort_iter_init(&sort_iter, b);
342 whiteout_u64s += b->whiteout_u64s;
343 order = get_order(whiteout_u64s * sizeof(u64));
345 whiteouts = btree_bounce_alloc(c, order, &used_mempool);
346 u_start = u_pos = whiteouts;
348 memcpy_u64s(u_pos, unwritten_whiteouts_start(c, b),
350 u_pos = (void *) u_pos + b->whiteout_u64s * sizeof(u64);
352 sort_iter_add(&sort_iter, u_start, u_pos);
354 for_each_bset(b, t) {
355 struct bset *i = bset(b, t);
356 struct bkey_packed *k, *n, *out, *start, *end;
357 struct btree_node_entry *src = NULL, *dst = NULL;
359 if (t != b->set && bset_unwritten(b, i)) {
360 src = container_of(i, struct btree_node_entry, keys);
361 dst = max(write_block(b),
362 (void *) btree_bkey_last(b, t -1));
365 if (!should_compact_bset(b, t, compacting, mode)) {
367 memmove(dst, src, sizeof(*src) +
368 le16_to_cpu(src->keys.u64s) *
371 set_btree_bset(b, t, i);
379 end = vstruct_last(i);
382 memmove(dst, src, sizeof(*src));
384 set_btree_bset(b, t, i);
389 for (k = start; k != end; k = n) {
392 if (bkey_deleted(k) && btree_node_is_extents(b))
395 if (bkey_whiteout(k) && !k->needs_whiteout)
398 if (bkey_whiteout(k)) {
399 unreserve_whiteout(b, t, k);
400 memcpy_u64s(u_pos, k, bkeyp_key_u64s(f, k));
401 set_bkeyp_val_u64s(f, u_pos, 0);
402 u_pos = bkey_next(u_pos);
403 } else if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
405 out = bkey_next(out);
409 sort_iter_add(&sort_iter, u_start, u_pos);
411 if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK) {
412 i->u64s = cpu_to_le16((u64 *) out - i->_data);
413 set_btree_bset_end(b, t);
414 bch2_bset_set_no_aux_tree(b, t);
418 b->whiteout_u64s = (u64 *) u_pos - (u64 *) whiteouts;
420 BUG_ON((void *) unwritten_whiteouts_start(c, b) <
421 (void *) btree_bkey_last(b, bset_tree_last(b)));
423 u64s = btree_node_is_extents(b)
424 ? sort_extent_whiteouts(unwritten_whiteouts_start(c, b),
426 : sort_key_whiteouts(unwritten_whiteouts_start(c, b),
429 BUG_ON(u64s > b->whiteout_u64s);
430 BUG_ON(u64s != b->whiteout_u64s && !btree_node_is_extents(b));
431 BUG_ON(u_pos != whiteouts && !u64s);
433 if (u64s != b->whiteout_u64s) {
434 void *src = unwritten_whiteouts_start(c, b);
436 b->whiteout_u64s = u64s;
437 memmove_u64s_up(unwritten_whiteouts_start(c, b), src, u64s);
441 unwritten_whiteouts_start(c, b),
442 unwritten_whiteouts_end(c, b));
444 btree_bounce_free(c, order, used_mempool, whiteouts);
446 if (mode != COMPACT_WRITTEN_NO_WRITE_LOCK)
447 bch2_btree_build_aux_trees(b);
449 bch_btree_keys_u64s_remaining(c, b);
450 bch2_verify_btree_nr_keys(b);
455 static bool bch2_drop_whiteouts(struct btree *b)
460 for_each_bset(b, t) {
461 struct bset *i = bset(b, t);
462 struct bkey_packed *k, *n, *out, *start, *end;
464 if (!should_compact_bset(b, t, true, COMPACT_WRITTEN))
467 start = btree_bkey_first(b, t);
468 end = btree_bkey_last(b, t);
470 if (bset_unwritten(b, i) &&
473 max_t(struct bset *, write_block(b),
474 (void *) btree_bkey_last(b, t -1));
476 memmove(dst, i, sizeof(struct bset));
478 set_btree_bset(b, t, i);
483 for (k = start; k != end; k = n) {
486 if (!bkey_whiteout(k)) {
488 out = bkey_next(out);
492 i->u64s = cpu_to_le16((u64 *) out - i->_data);
493 bch2_bset_set_no_aux_tree(b, t);
497 bch2_verify_btree_nr_keys(b);
502 static inline int sort_keys_cmp(struct btree *b,
503 struct bkey_packed *l,
504 struct bkey_packed *r)
506 return bkey_cmp_packed(b, l, r) ?:
507 (int) bkey_whiteout(r) - (int) bkey_whiteout(l) ?:
508 (int) l->needs_whiteout - (int) r->needs_whiteout;
511 static unsigned sort_keys(struct bkey_packed *dst,
512 struct sort_iter *iter,
513 bool filter_whiteouts)
515 const struct bkey_format *f = &iter->b->format;
516 struct bkey_packed *in, *next, *out = dst;
518 sort_iter_sort(iter, sort_keys_cmp);
520 while ((in = sort_iter_next(iter, sort_keys_cmp))) {
521 if (bkey_whiteout(in) &&
522 (filter_whiteouts || !in->needs_whiteout))
525 if (bkey_whiteout(in) &&
526 (next = sort_iter_peek(iter)) &&
527 !bkey_cmp_packed(iter->b, in, next)) {
528 BUG_ON(in->needs_whiteout &&
529 next->needs_whiteout);
531 * XXX racy, called with read lock from write path
533 * leads to spurious BUG_ON() in bkey_unpack_key() in
536 next->needs_whiteout |= in->needs_whiteout;
540 if (bkey_whiteout(in)) {
541 memcpy_u64s(out, in, bkeyp_key_u64s(f, in));
542 set_bkeyp_val_u64s(f, out, 0);
546 out = bkey_next(out);
549 return (u64 *) out - (u64 *) dst;
552 static inline int sort_extents_cmp(struct btree *b,
553 struct bkey_packed *l,
554 struct bkey_packed *r)
556 return bkey_cmp_packed(b, l, r) ?:
557 (int) bkey_deleted(l) - (int) bkey_deleted(r);
560 static unsigned sort_extents(struct bkey_packed *dst,
561 struct sort_iter *iter,
562 bool filter_whiteouts)
564 struct bkey_packed *in, *out = dst;
566 sort_iter_sort(iter, sort_extents_cmp);
568 while ((in = sort_iter_next(iter, sort_extents_cmp))) {
569 if (bkey_deleted(in))
572 if (bkey_whiteout(in) &&
573 (filter_whiteouts || !in->needs_whiteout))
577 out = bkey_next(out);
580 return (u64 *) out - (u64 *) dst;
583 static void btree_node_sort(struct bch_fs *c, struct btree *b,
584 struct btree_iter *iter,
587 bool filter_whiteouts)
589 struct btree_node *out;
590 struct sort_iter sort_iter;
592 struct bset *start_bset = bset(b, &b->set[start_idx]);
593 bool used_mempool = false;
594 u64 start_time, seq = 0;
595 unsigned i, u64s = 0, order, shift = end_idx - start_idx - 1;
596 bool sorting_entire_node = start_idx == 0 &&
599 sort_iter_init(&sort_iter, b);
601 for (t = b->set + start_idx;
602 t < b->set + end_idx;
604 u64s += le16_to_cpu(bset(b, t)->u64s);
605 sort_iter_add(&sort_iter,
606 btree_bkey_first(b, t),
607 btree_bkey_last(b, t));
610 order = sorting_entire_node
611 ? btree_page_order(c)
612 : get_order(__vstruct_bytes(struct btree_node, u64s));
614 out = btree_bounce_alloc(c, order, &used_mempool);
616 start_time = local_clock();
618 if (btree_node_is_extents(b))
619 filter_whiteouts = bset_written(b, start_bset);
621 u64s = btree_node_is_extents(b)
622 ? sort_extents(out->keys.start, &sort_iter, filter_whiteouts)
623 : sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
625 out->keys.u64s = cpu_to_le16(u64s);
627 BUG_ON(vstruct_end(&out->keys) > (void *) out + (PAGE_SIZE << order));
629 if (sorting_entire_node)
630 bch2_time_stats_update(&c->times[BCH_TIME_btree_sort],
633 /* Make sure we preserve bset journal_seq: */
634 for (t = b->set + start_idx; t < b->set + end_idx; t++)
635 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
636 start_bset->journal_seq = cpu_to_le64(seq);
638 if (sorting_entire_node) {
639 unsigned u64s = le16_to_cpu(out->keys.u64s);
641 BUG_ON(order != btree_page_order(c));
644 * Our temporary buffer is the same size as the btree node's
645 * buffer, we can just swap buffers instead of doing a big
649 out->keys.u64s = cpu_to_le16(u64s);
651 set_btree_bset(b, b->set, &b->data->keys);
653 start_bset->u64s = out->keys.u64s;
654 memcpy_u64s(start_bset->start,
656 le16_to_cpu(out->keys.u64s));
659 for (i = start_idx + 1; i < end_idx; i++)
660 b->nr.bset_u64s[start_idx] +=
665 for (i = start_idx + 1; i < b->nsets; i++) {
666 b->nr.bset_u64s[i] = b->nr.bset_u64s[i + shift];
667 b->set[i] = b->set[i + shift];
670 for (i = b->nsets; i < MAX_BSETS; i++)
671 b->nr.bset_u64s[i] = 0;
673 set_btree_bset_end(b, &b->set[start_idx]);
674 bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
676 btree_bounce_free(c, order, used_mempool, out);
678 bch2_verify_btree_nr_keys(b);
681 /* Sort + repack in a new format: */
682 static struct btree_nr_keys sort_repack(struct bset *dst,
684 struct btree_node_iter *src_iter,
685 struct bkey_format *out_f,
686 bool filter_whiteouts)
688 struct bkey_format *in_f = &src->format;
689 struct bkey_packed *in, *out = vstruct_last(dst);
690 struct btree_nr_keys nr;
692 memset(&nr, 0, sizeof(nr));
694 while ((in = bch2_btree_node_iter_next_all(src_iter, src))) {
695 if (filter_whiteouts && bkey_whiteout(in))
698 if (bch2_bkey_transform(out_f, out, bkey_packed(in)
699 ? in_f : &bch2_bkey_format_current, in))
700 out->format = KEY_FORMAT_LOCAL_BTREE;
702 bch2_bkey_unpack(src, (void *) out, in);
704 btree_keys_account_key_add(&nr, 0, out);
705 out = bkey_next(out);
708 dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
712 /* Sort, repack, and merge: */
713 static struct btree_nr_keys sort_repack_merge(struct bch_fs *c,
716 struct btree_node_iter *iter,
717 struct bkey_format *out_f,
718 bool filter_whiteouts,
719 key_filter_fn filter,
722 struct bkey_packed *k, *prev = NULL, *out;
723 struct btree_nr_keys nr;
726 memset(&nr, 0, sizeof(nr));
728 while ((k = bch2_btree_node_iter_next_all(iter, src))) {
729 if (filter_whiteouts && bkey_whiteout(k))
733 * The filter might modify pointers, so we have to unpack the
734 * key and values to &tmp.k:
736 bch2_bkey_unpack(src, &tmp.k, k);
738 if (filter && filter(c, src, bkey_i_to_s(&tmp.k)))
741 /* prev is always unpacked, for key merging: */
745 merge(c, src, (void *) prev, &tmp.k) == BCH_MERGE_MERGE)
749 * the current key becomes the new prev: advance prev, then
750 * copy the current key - but first pack prev (in place):
753 bch2_bkey_pack(prev, (void *) prev, out_f);
755 btree_keys_account_key_add(&nr, 0, prev);
756 prev = bkey_next(prev);
758 prev = vstruct_last(dst);
761 bkey_copy(prev, &tmp.k);
765 bch2_bkey_pack(prev, (void *) prev, out_f);
766 btree_keys_account_key_add(&nr, 0, prev);
767 out = bkey_next(prev);
769 out = vstruct_last(dst);
772 dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
776 void bch2_btree_sort_into(struct bch_fs *c,
780 struct btree_nr_keys nr;
781 struct btree_node_iter src_iter;
782 u64 start_time = local_clock();
784 BUG_ON(dst->nsets != 1);
786 bch2_bset_set_no_aux_tree(dst, dst->set);
788 bch2_btree_node_iter_init_from_start(&src_iter, src,
789 btree_node_is_extents(src));
791 if (btree_node_ops(src)->key_normalize ||
792 btree_node_ops(src)->key_merge)
793 nr = sort_repack_merge(c, btree_bset_first(dst),
797 btree_node_ops(src)->key_normalize,
798 btree_node_ops(src)->key_merge);
800 nr = sort_repack(btree_bset_first(dst),
805 bch2_time_stats_update(&c->times[BCH_TIME_btree_sort], start_time);
807 set_btree_bset_end(dst, dst->set);
809 dst->nr.live_u64s += nr.live_u64s;
810 dst->nr.bset_u64s[0] += nr.bset_u64s[0];
811 dst->nr.packed_keys += nr.packed_keys;
812 dst->nr.unpacked_keys += nr.unpacked_keys;
814 bch2_verify_btree_nr_keys(dst);
817 #define SORT_CRIT (4096 / sizeof(u64))
820 * We're about to add another bset to the btree node, so if there's currently
821 * too many bsets - sort some of them together:
823 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
824 struct btree_iter *iter)
826 unsigned unwritten_idx;
829 for (unwritten_idx = 0;
830 unwritten_idx < b->nsets;
832 if (bset_unwritten(b, bset(b, &b->set[unwritten_idx])))
835 if (b->nsets - unwritten_idx > 1) {
836 btree_node_sort(c, b, iter, unwritten_idx,
841 if (unwritten_idx > 1) {
842 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
849 void bch2_btree_build_aux_trees(struct btree *b)
854 bch2_bset_build_aux_tree(b, t,
855 bset_unwritten(b, bset(b, t)) &&
856 t == bset_tree_last(b));
860 * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
863 * Safe to call if there already is an unwritten bset - will only add a new bset
864 * if @b doesn't already have one.
866 * Returns true if we sorted (i.e. invalidated iterators
868 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
869 struct btree_iter *iter)
871 struct btree_node_entry *bne;
874 EBUG_ON(!(b->lock.state.seq & 1));
875 EBUG_ON(iter && iter->l[b->level].b != b);
877 did_sort = btree_node_compact(c, b, iter);
879 bne = want_new_bset(c, b);
881 bch2_bset_init_next(c, b, bne);
883 bch2_btree_build_aux_trees(b);
885 if (iter && did_sort)
886 bch2_btree_iter_reinit_node(iter, b);
889 static struct nonce btree_nonce(struct bset *i, unsigned offset)
891 return (struct nonce) {{
892 [0] = cpu_to_le32(offset),
893 [1] = ((__le32 *) &i->seq)[0],
894 [2] = ((__le32 *) &i->seq)[1],
895 [3] = ((__le32 *) &i->journal_seq)[0]^BCH_NONCE_BTREE,
899 static void bset_encrypt(struct bch_fs *c, struct bset *i, unsigned offset)
901 struct nonce nonce = btree_nonce(i, offset);
904 struct btree_node *bn = container_of(i, struct btree_node, keys);
905 unsigned bytes = (void *) &bn->keys - (void *) &bn->flags;
907 bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, &bn->flags,
910 nonce = nonce_add(nonce, round_up(bytes, CHACHA20_BLOCK_SIZE));
913 bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, i->_data,
914 vstruct_end(i) - (void *) i->_data);
917 static int btree_err_msg(struct bch_fs *c, struct btree *b, struct bset *i,
918 unsigned offset, int write, char *buf, size_t len)
920 char *out = buf, *end = buf + len;
922 out += scnprintf(out, end - out,
923 "error validating btree node %s "
924 "at btree %u level %u/%u\n"
925 "pos %llu:%llu node offset %u",
926 write ? "before write " : "",
927 b->btree_id, b->level,
928 c->btree_roots[b->btree_id].level,
929 b->key.k.p.inode, b->key.k.p.offset,
932 out += scnprintf(out, end - out,
934 le16_to_cpu(i->u64s));
939 enum btree_err_type {
941 BTREE_ERR_WANT_RETRY,
942 BTREE_ERR_MUST_RETRY,
946 enum btree_validate_ret {
947 BTREE_RETRY_READ = 64,
950 #define btree_err(type, c, b, i, msg, ...) \
953 char _buf[300], *out = _buf, *end = out + sizeof(_buf); \
955 out += btree_err_msg(c, b, i, b->written, write, out, end - out);\
956 out += scnprintf(out, end - out, ": " msg, ##__VA_ARGS__); \
958 if (type == BTREE_ERR_FIXABLE && \
960 !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) { \
961 mustfix_fsck_err(c, "%s", _buf); \
967 bch_err(c, "%s", _buf); \
970 case BTREE_ERR_FIXABLE: \
971 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
973 case BTREE_ERR_WANT_RETRY: \
975 ret = BTREE_RETRY_READ; \
979 case BTREE_ERR_MUST_RETRY: \
980 ret = BTREE_RETRY_READ; \
982 case BTREE_ERR_FATAL: \
983 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
988 bch_err(c, "corrupt metadata before write: %s", _buf); \
990 if (bch2_fs_inconsistent(c)) { \
991 ret = BCH_FSCK_ERRORS_NOT_FIXED; \
1000 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
1002 static int validate_bset(struct bch_fs *c, struct btree *b,
1003 struct bset *i, unsigned sectors,
1004 unsigned *whiteout_u64s, int write,
1007 struct bkey_packed *k, *prev = NULL;
1008 struct bpos prev_pos = POS_MIN;
1009 enum bkey_type type = btree_node_type(b);
1010 bool seen_non_whiteout = false;
1014 if (i == &b->data->keys) {
1015 /* These indicate that we read the wrong btree node: */
1016 btree_err_on(BTREE_NODE_ID(b->data) != b->btree_id,
1017 BTREE_ERR_MUST_RETRY, c, b, i,
1018 "incorrect btree id");
1020 btree_err_on(BTREE_NODE_LEVEL(b->data) != b->level,
1021 BTREE_ERR_MUST_RETRY, c, b, i,
1024 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
1025 u64 *p = (u64 *) &b->data->ptr;
1028 bch2_bpos_swab(&b->data->min_key);
1029 bch2_bpos_swab(&b->data->max_key);
1032 btree_err_on(bkey_cmp(b->data->max_key, b->key.k.p),
1033 BTREE_ERR_MUST_RETRY, c, b, i,
1034 "incorrect max key");
1036 /* XXX: ideally we would be validating min_key too */
1039 * not correct anymore, due to btree node write error
1042 * need to add b->data->seq to btree keys and verify
1045 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
1047 BTREE_ERR_FATAL, c, b, i,
1048 "incorrect backpointer");
1050 err = bch2_bkey_format_validate(&b->data->format);
1052 BTREE_ERR_FATAL, c, b, i,
1053 "invalid bkey format: %s", err);
1056 if (btree_err_on(le16_to_cpu(i->version) != BCACHE_BSET_VERSION,
1057 BTREE_ERR_FIXABLE, c, b, i,
1058 "unsupported bset version")) {
1059 i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1064 if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
1065 BTREE_ERR_FIXABLE, c, b, i,
1066 "bset past end of btree node")) {
1071 btree_err_on(b->written && !i->u64s,
1072 BTREE_ERR_FIXABLE, c, b, i,
1075 if (!BSET_SEPARATE_WHITEOUTS(i)) {
1076 seen_non_whiteout = true;
1081 k != vstruct_last(i);) {
1084 const char *invalid;
1086 if (btree_err_on(!k->u64s,
1087 BTREE_ERR_FIXABLE, c, b, i,
1088 "KEY_U64s 0: %zu bytes of metadata lost",
1089 vstruct_end(i) - (void *) k)) {
1090 i->u64s = cpu_to_le16((u64 *) k - i->_data);
1094 if (btree_err_on(bkey_next(k) > vstruct_last(i),
1095 BTREE_ERR_FIXABLE, c, b, i,
1096 "key extends past end of bset")) {
1097 i->u64s = cpu_to_le16((u64 *) k - i->_data);
1101 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
1102 BTREE_ERR_FIXABLE, c, b, i,
1103 "invalid bkey format %u", k->format)) {
1104 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1105 memmove_u64s_down(k, bkey_next(k),
1106 (u64 *) vstruct_end(i) - (u64 *) k);
1110 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
1111 bch2_bkey_swab(type, &b->format, k);
1113 u = bkey_disassemble(b, k, &tmp);
1115 invalid = __bch2_bkey_invalid(c, type, u) ?:
1116 bch2_bkey_in_btree_node(b, u) ?:
1117 (write ? bch2_bkey_val_invalid(c, type, u) : NULL);
1121 bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1122 btree_err(BTREE_ERR_FIXABLE, c, b, i,
1123 "invalid bkey:\n%s\n%s", buf, invalid);
1125 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1126 memmove_u64s_down(k, bkey_next(k),
1127 (u64 *) vstruct_end(i) - (u64 *) k);
1132 * with the separate whiteouts thing (used for extents), the
1133 * second set of keys actually can have whiteouts too, so we
1134 * can't solely go off bkey_whiteout()...
1137 if (!seen_non_whiteout &&
1138 (!bkey_whiteout(k) ||
1139 (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0))) {
1140 *whiteout_u64s = k->_data - i->_data;
1141 seen_non_whiteout = true;
1142 } else if (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0) {
1143 btree_err(BTREE_ERR_FATAL, c, b, i,
1144 "keys out of order: %llu:%llu > %llu:%llu",
1148 bkey_start_offset(u.k));
1149 /* XXX: repair this */
1157 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1162 int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry)
1164 struct btree_node_entry *bne;
1165 struct btree_node_iter_large *iter;
1166 struct btree_node *sorted;
1167 struct bkey_packed *k;
1171 int ret, retry_read = 0, write = READ;
1173 iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
1174 __bch2_btree_node_iter_large_init(iter, btree_node_is_extents(b));
1176 if (bch2_meta_read_fault("btree"))
1177 btree_err(BTREE_ERR_MUST_RETRY, c, b, NULL,
1180 btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
1181 BTREE_ERR_MUST_RETRY, c, b, NULL,
1184 btree_err_on(!b->data->keys.seq,
1185 BTREE_ERR_MUST_RETRY, c, b, NULL,
1186 "bad btree header");
1188 while (b->written < c->opts.btree_node_size) {
1189 unsigned sectors, whiteout_u64s = 0;
1191 struct bch_csum csum;
1192 bool first = !b->written;
1197 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1198 BTREE_ERR_WANT_RETRY, c, b, i,
1199 "unknown checksum type");
1201 nonce = btree_nonce(i, b->written << 9);
1202 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1204 btree_err_on(bch2_crc_cmp(csum, b->data->csum),
1205 BTREE_ERR_WANT_RETRY, c, b, i,
1206 "invalid checksum");
1208 bset_encrypt(c, i, b->written << 9);
1210 sectors = vstruct_sectors(b->data, c->block_bits);
1212 btree_node_set_format(b, b->data->format);
1214 bne = write_block(b);
1217 if (i->seq != b->data->keys.seq)
1220 btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1221 BTREE_ERR_WANT_RETRY, c, b, i,
1222 "unknown checksum type");
1224 nonce = btree_nonce(i, b->written << 9);
1225 csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1227 btree_err_on(bch2_crc_cmp(csum, bne->csum),
1228 BTREE_ERR_WANT_RETRY, c, b, i,
1229 "invalid checksum");
1231 bset_encrypt(c, i, b->written << 9);
1233 sectors = vstruct_sectors(bne, c->block_bits);
1236 ret = validate_bset(c, b, i, sectors, &whiteout_u64s,
1241 b->written += sectors;
1243 ret = bch2_journal_seq_should_ignore(c, le64_to_cpu(i->journal_seq), b);
1245 btree_err(BTREE_ERR_FATAL, c, b, i,
1246 "insufficient memory");
1252 BTREE_ERR_FIXABLE, c, b, i,
1253 "first btree node bset has blacklisted journal seq");
1258 bch2_btree_node_iter_large_push(iter, b,
1260 vstruct_idx(i, whiteout_u64s));
1262 bch2_btree_node_iter_large_push(iter, b,
1263 vstruct_idx(i, whiteout_u64s),
1267 for (bne = write_block(b);
1268 bset_byte_offset(b, bne) < btree_bytes(c);
1269 bne = (void *) bne + block_bytes(c))
1270 btree_err_on(bne->keys.seq == b->data->keys.seq,
1271 BTREE_ERR_WANT_RETRY, c, b, NULL,
1272 "found bset signature after last bset");
1274 sorted = btree_bounce_alloc(c, btree_page_order(c), &used_mempool);
1275 sorted->keys.u64s = 0;
1277 set_btree_bset(b, b->set, &b->data->keys);
1279 b->nr = btree_node_is_extents(b)
1280 ? bch2_extent_sort_fix_overlapping(c, &sorted->keys, b, iter)
1281 : bch2_key_sort_fix_overlapping(&sorted->keys, b, iter);
1283 u64s = le16_to_cpu(sorted->keys.u64s);
1285 sorted->keys.u64s = cpu_to_le16(u64s);
1286 swap(sorted, b->data);
1287 set_btree_bset(b, b->set, &b->data->keys);
1290 BUG_ON(b->nr.live_u64s != u64s);
1292 btree_bounce_free(c, btree_page_order(c), used_mempool, sorted);
1295 for (k = i->start; k != vstruct_last(i);) {
1296 enum bkey_type type = btree_node_type(b);
1298 struct bkey_s_c u = bkey_disassemble(b, k, &tmp);
1299 const char *invalid = bch2_bkey_val_invalid(c, type, u);
1304 bch2_bkey_val_to_text(c, type, buf, sizeof(buf), u);
1305 btree_err(BTREE_ERR_FIXABLE, c, b, i,
1306 "invalid bkey %s: %s", buf, invalid);
1308 btree_keys_account_key_drop(&b->nr, 0, k);
1310 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1311 memmove_u64s_down(k, bkey_next(k),
1312 (u64 *) vstruct_end(i) - (u64 *) k);
1319 bch2_bset_build_aux_tree(b, b->set, false);
1321 set_needs_whiteout(btree_bset_first(b));
1323 btree_node_reset_sib_u64s(b);
1325 mempool_free(iter, &c->fill_iter);
1329 if (ret == BTREE_RETRY_READ) {
1332 bch2_inconsistent_error(c);
1333 set_btree_node_read_error(b);
1338 static void btree_node_read_work(struct work_struct *work)
1340 struct btree_read_bio *rb =
1341 container_of(work, struct btree_read_bio, work);
1342 struct bch_fs *c = rb->c;
1343 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1344 struct btree *b = rb->bio.bi_private;
1345 struct bio *bio = &rb->bio;
1346 struct bch_devs_mask avoid;
1349 memset(&avoid, 0, sizeof(avoid));
1353 bch_info(c, "retrying read");
1354 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1355 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1357 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1358 bio->bi_iter.bi_sector = rb->pick.ptr.offset;
1359 bio->bi_iter.bi_size = btree_bytes(c);
1361 if (rb->have_ioref) {
1362 bio_set_dev(bio, ca->disk_sb.bdev);
1363 submit_bio_wait(bio);
1365 bio->bi_status = BLK_STS_REMOVED;
1368 bch2_dev_io_err_on(bio->bi_status, ca, "btree read");
1370 percpu_ref_put(&ca->io_ref);
1371 rb->have_ioref = false;
1373 __set_bit(rb->pick.ptr.dev, avoid.d);
1374 can_retry = bch2_btree_pick_ptr(c, b, &avoid, &rb->pick) > 0;
1376 if (!bio->bi_status &&
1377 !bch2_btree_node_read_done(c, b, can_retry))
1381 set_btree_node_read_error(b);
1386 bch2_time_stats_update(&c->times[BCH_TIME_btree_read], rb->start_time);
1388 clear_btree_node_read_in_flight(b);
1389 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1392 static void btree_node_read_endio(struct bio *bio)
1394 struct btree_read_bio *rb =
1395 container_of(bio, struct btree_read_bio, bio);
1396 struct bch_fs *c = rb->c;
1398 if (rb->have_ioref) {
1399 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1400 bch2_latency_acct(ca, rb->start_time, READ);
1403 queue_work(system_unbound_wq, &rb->work);
1406 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1409 struct extent_pick_ptr pick;
1410 struct btree_read_bio *rb;
1415 trace_btree_read(c, b);
1417 ret = bch2_btree_pick_ptr(c, b, NULL, &pick);
1418 if (bch2_fs_fatal_err_on(ret <= 0, c,
1419 "btree node read error: no device to read from")) {
1420 set_btree_node_read_error(b);
1424 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1426 bio = bio_alloc_bioset(GFP_NOIO, btree_pages(c), &c->btree_bio);
1427 rb = container_of(bio, struct btree_read_bio, bio);
1429 rb->start_time = local_clock();
1430 rb->have_ioref = bch2_dev_get_ioref(ca, READ);
1432 INIT_WORK(&rb->work, btree_node_read_work);
1433 bio->bi_opf = REQ_OP_READ|REQ_SYNC|REQ_META;
1434 bio->bi_iter.bi_sector = pick.ptr.offset;
1435 bio->bi_iter.bi_size = btree_bytes(c);
1436 bio->bi_end_io = btree_node_read_endio;
1437 bio->bi_private = b;
1438 bch2_bio_map(bio, b->data);
1440 set_btree_node_read_in_flight(b);
1442 if (rb->have_ioref) {
1443 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_BTREE],
1445 bio_set_dev(bio, ca->disk_sb.bdev);
1448 submit_bio_wait(bio);
1450 bio->bi_private = b;
1451 btree_node_read_work(&rb->work);
1456 bio->bi_status = BLK_STS_REMOVED;
1459 btree_node_read_work(&rb->work);
1461 queue_work(system_unbound_wq, &rb->work);
1466 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1467 const struct bkey_i *k, unsigned level)
1473 closure_init_stack(&cl);
1476 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1480 b = bch2_btree_node_mem_alloc(c);
1481 bch2_btree_cache_cannibalize_unlock(c);
1485 bkey_copy(&b->key, k);
1486 BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1488 bch2_btree_node_read(c, b, true);
1490 if (btree_node_read_error(b)) {
1491 bch2_btree_node_hash_remove(&c->btree_cache, b);
1493 mutex_lock(&c->btree_cache.lock);
1494 list_move(&b->list, &c->btree_cache.freeable);
1495 mutex_unlock(&c->btree_cache.lock);
1501 bch2_btree_set_root_for_read(c, b);
1503 six_unlock_write(&b->lock);
1504 six_unlock_intent(&b->lock);
1509 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1510 struct btree_write *w)
1512 unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1520 } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1523 closure_put(&((struct btree_update *) new)->cl);
1525 bch2_journal_pin_drop(&c->journal, &w->journal);
1526 closure_wake_up(&w->wait);
1529 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1531 struct btree_write *w = btree_prev_write(b);
1533 bch2_btree_complete_write(c, b, w);
1534 btree_node_io_unlock(b);
1537 static void bch2_btree_node_write_error(struct bch_fs *c,
1538 struct btree_write_bio *wbio)
1540 struct btree *b = wbio->wbio.bio.bi_private;
1541 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1542 struct bkey_i_extent *new_key;
1543 struct bkey_s_extent e;
1544 struct bch_extent_ptr *ptr;
1545 struct btree_iter iter;
1548 __bch2_btree_iter_init(&iter, c, b->btree_id, b->key.k.p,
1552 ret = bch2_btree_iter_traverse(&iter);
1556 /* has node been freed? */
1557 if (iter.l[b->level].b != b) {
1558 /* node has been freed: */
1559 BUG_ON(!btree_node_dying(b));
1563 BUG_ON(!btree_node_hashed(b));
1565 bkey_copy(&tmp.k, &b->key);
1567 new_key = bkey_i_to_extent(&tmp.k);
1568 e = extent_i_to_s(new_key);
1569 extent_for_each_ptr_backwards(e, ptr)
1570 if (bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev))
1571 bch2_extent_drop_ptr(e, ptr);
1573 if (!bch2_extent_nr_ptrs(e.c))
1576 ret = bch2_btree_node_update_key(c, &iter, b, new_key);
1582 bch2_btree_iter_unlock(&iter);
1583 bio_put(&wbio->wbio.bio);
1584 btree_node_write_done(c, b);
1587 set_btree_node_noevict(b);
1588 bch2_fs_fatal_error(c, "fatal error writing btree node");
1592 void bch2_btree_write_error_work(struct work_struct *work)
1594 struct bch_fs *c = container_of(work, struct bch_fs,
1595 btree_write_error_work);
1599 spin_lock_irq(&c->btree_write_error_lock);
1600 bio = bio_list_pop(&c->btree_write_error_list);
1601 spin_unlock_irq(&c->btree_write_error_lock);
1606 bch2_btree_node_write_error(c,
1607 container_of(bio, struct btree_write_bio, wbio.bio));
1611 static void btree_node_write_work(struct work_struct *work)
1613 struct btree_write_bio *wbio =
1614 container_of(work, struct btree_write_bio, work);
1615 struct bch_fs *c = wbio->wbio.c;
1616 struct btree *b = wbio->wbio.bio.bi_private;
1618 btree_bounce_free(c,
1620 wbio->wbio.used_mempool,
1623 if (wbio->wbio.failed.nr) {
1624 unsigned long flags;
1626 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1627 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1628 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1630 queue_work(c->wq, &c->btree_write_error_work);
1634 bio_put(&wbio->wbio.bio);
1635 btree_node_write_done(c, b);
1638 static void btree_node_write_endio(struct bio *bio)
1640 struct bch_write_bio *wbio = to_wbio(bio);
1641 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
1642 struct bch_write_bio *orig = parent ?: wbio;
1643 struct bch_fs *c = wbio->c;
1644 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
1645 unsigned long flags;
1647 if (wbio->have_ioref)
1648 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1650 if (bio->bi_status == BLK_STS_REMOVED ||
1651 bch2_dev_io_err_on(bio->bi_status, ca, "btree write") ||
1652 bch2_meta_write_fault("btree")) {
1653 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1654 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1655 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1658 if (wbio->have_ioref)
1659 percpu_ref_put(&ca->io_ref);
1663 bio_endio(&parent->bio);
1665 struct btree_write_bio *wb =
1666 container_of(orig, struct btree_write_bio, wbio);
1668 INIT_WORK(&wb->work, btree_node_write_work);
1669 queue_work(system_unbound_wq, &wb->work);
1673 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1674 struct bset *i, unsigned sectors)
1676 const struct bch_extent_ptr *ptr;
1677 unsigned whiteout_u64s = 0;
1680 extent_for_each_ptr(bkey_i_to_s_c_extent(&b->key), ptr)
1683 ret = validate_bset(c, b, i, sectors, &whiteout_u64s, WRITE, false);
1685 bch2_inconsistent_error(c);
1690 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1691 enum six_lock_type lock_type_held)
1693 struct btree_write_bio *wbio;
1694 struct bset_tree *t;
1696 struct btree_node *bn = NULL;
1697 struct btree_node_entry *bne = NULL;
1699 struct bkey_s_extent e;
1700 struct bch_extent_ptr *ptr;
1701 struct sort_iter sort_iter;
1703 unsigned bytes_to_write, sectors_to_write, order, bytes, u64s;
1706 unsigned long old, new;
1709 if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1713 * We may only have a read lock on the btree node - the dirty bit is our
1714 * "lock" against racing with other threads that may be trying to start
1715 * a write, we do a write iff we clear the dirty bit. Since setting the
1716 * dirty bit requires a write lock, we can't race with other threads
1720 old = new = READ_ONCE(b->flags);
1722 if (!(old & (1 << BTREE_NODE_dirty)))
1726 !btree_node_may_write(b))
1729 if (old & (1 << BTREE_NODE_write_in_flight)) {
1730 btree_node_wait_on_io(b);
1734 new &= ~(1 << BTREE_NODE_dirty);
1735 new &= ~(1 << BTREE_NODE_need_write);
1736 new |= (1 << BTREE_NODE_write_in_flight);
1737 new |= (1 << BTREE_NODE_just_written);
1738 new ^= (1 << BTREE_NODE_write_idx);
1739 } while (cmpxchg_acquire(&b->flags, old, new) != old);
1741 BUG_ON(btree_node_fake(b));
1742 BUG_ON(!list_empty(&b->write_blocked));
1743 BUG_ON((b->will_make_reachable != 0) != !b->written);
1745 BUG_ON(b->written >= c->opts.btree_node_size);
1746 BUG_ON(b->written & (c->opts.block_size - 1));
1747 BUG_ON(bset_written(b, btree_bset_last(b)));
1748 BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1749 BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1752 * We can't block on six_lock_write() here; another thread might be
1753 * trying to get a journal reservation with read locks held, and getting
1754 * a journal reservation might be blocked on flushing the journal and
1755 * doing btree writes:
1757 if (lock_type_held == SIX_LOCK_intent &&
1758 six_trylock_write(&b->lock)) {
1759 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN);
1760 six_unlock_write(&b->lock);
1762 __bch2_compact_whiteouts(c, b, COMPACT_WRITTEN_NO_WRITE_LOCK);
1765 BUG_ON(b->uncompacted_whiteout_u64s);
1767 sort_iter_init(&sort_iter, b);
1770 ? sizeof(struct btree_node)
1771 : sizeof(struct btree_node_entry);
1773 bytes += b->whiteout_u64s * sizeof(u64);
1775 for_each_bset(b, t) {
1778 if (bset_written(b, i))
1781 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1782 sort_iter_add(&sort_iter,
1783 btree_bkey_first(b, t),
1784 btree_bkey_last(b, t));
1785 seq = max(seq, le64_to_cpu(i->journal_seq));
1788 order = get_order(bytes);
1789 data = btree_bounce_alloc(c, order, &used_mempool);
1797 bne->keys = b->data->keys;
1801 i->journal_seq = cpu_to_le64(seq);
1804 if (!btree_node_is_extents(b)) {
1805 sort_iter_add(&sort_iter,
1806 unwritten_whiteouts_start(c, b),
1807 unwritten_whiteouts_end(c, b));
1808 SET_BSET_SEPARATE_WHITEOUTS(i, false);
1810 memcpy_u64s(i->start,
1811 unwritten_whiteouts_start(c, b),
1813 i->u64s = cpu_to_le16(b->whiteout_u64s);
1814 SET_BSET_SEPARATE_WHITEOUTS(i, true);
1817 b->whiteout_u64s = 0;
1819 u64s = btree_node_is_extents(b)
1820 ? sort_extents(vstruct_last(i), &sort_iter, false)
1821 : sort_keys(i->start, &sort_iter, false);
1822 le16_add_cpu(&i->u64s, u64s);
1824 clear_needs_whiteout(i);
1826 /* do we have data to write? */
1827 if (b->written && !i->u64s)
1830 bytes_to_write = vstruct_end(i) - data;
1831 sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1833 memset(data + bytes_to_write, 0,
1834 (sectors_to_write << 9) - bytes_to_write);
1836 BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1837 BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1838 BUG_ON(i->seq != b->data->keys.seq);
1840 i->version = cpu_to_le16(BCACHE_BSET_VERSION);
1841 SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1843 /* if we're going to be encrypting, check metadata validity first: */
1844 if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1845 validate_bset_for_write(c, b, i, sectors_to_write))
1848 bset_encrypt(c, i, b->written << 9);
1850 nonce = btree_nonce(i, b->written << 9);
1853 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1855 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1857 /* if we're not encrypting, check metadata after checksumming: */
1858 if (!bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)) &&
1859 validate_bset_for_write(c, b, i, sectors_to_write))
1863 * We handle btree write errors by immediately halting the journal -
1864 * after we've done that, we can't issue any subsequent btree writes
1865 * because they might have pointers to new nodes that failed to write.
1867 * Furthermore, there's no point in doing any more btree writes because
1868 * with the journal stopped, we're never going to update the journal to
1869 * reflect that those writes were done and the data flushed from the
1872 * Make sure to update b->written so bch2_btree_init_next() doesn't
1875 if (bch2_journal_error(&c->journal) ||
1879 trace_btree_write(b, bytes_to_write, sectors_to_write);
1881 wbio = container_of(bio_alloc_bioset(GFP_NOIO, 1 << order, &c->btree_bio),
1882 struct btree_write_bio, wbio.bio);
1883 wbio_init(&wbio->wbio.bio);
1885 wbio->wbio.order = order;
1886 wbio->wbio.used_mempool = used_mempool;
1887 wbio->wbio.bio.bi_opf = REQ_OP_WRITE|REQ_META|REQ_FUA;
1888 wbio->wbio.bio.bi_iter.bi_size = sectors_to_write << 9;
1889 wbio->wbio.bio.bi_end_io = btree_node_write_endio;
1890 wbio->wbio.bio.bi_private = b;
1892 bch2_bio_map(&wbio->wbio.bio, data);
1895 * If we're appending to a leaf node, we don't technically need FUA -
1896 * this write just needs to be persisted before the next journal write,
1897 * which will be marked FLUSH|FUA.
1899 * Similarly if we're writing a new btree root - the pointer is going to
1900 * be in the next journal entry.
1902 * But if we're writing a new btree node (that isn't a root) or
1903 * appending to a non leaf btree node, we need either FUA or a flush
1904 * when we write the parent with the new pointer. FUA is cheaper than a
1905 * flush, and writes appending to leaf nodes aren't blocking anything so
1906 * just make all btree node writes FUA to keep things sane.
1909 bkey_copy(&k.key, &b->key);
1910 e = bkey_i_to_s_extent(&k.key);
1912 extent_for_each_ptr(e, ptr)
1913 ptr->offset += b->written;
1915 b->written += sectors_to_write;
1917 bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_BTREE, &k.key);
1920 set_btree_node_noevict(b);
1921 b->written += sectors_to_write;
1923 btree_bounce_free(c, order, used_mempool, data);
1924 btree_node_write_done(c, b);
1928 * Work that must be done with write lock held:
1930 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1932 bool invalidated_iter = false;
1933 struct btree_node_entry *bne;
1934 struct bset_tree *t;
1936 if (!btree_node_just_written(b))
1939 BUG_ON(b->whiteout_u64s);
1940 BUG_ON(b->uncompacted_whiteout_u64s);
1942 clear_btree_node_just_written(b);
1945 * Note: immediately after write, bset_unwritten()/bset_written() don't
1946 * work - the amount of data we had to write after compaction might have
1947 * been smaller than the offset of the last bset.
1949 * However, we know that all bsets have been written here, as long as
1950 * we're still holding the write lock:
1954 * XXX: decide if we really want to unconditionally sort down to a
1958 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1959 invalidated_iter = true;
1961 invalidated_iter = bch2_drop_whiteouts(b);
1965 set_needs_whiteout(bset(b, t));
1967 bch2_btree_verify(c, b);
1970 * If later we don't unconditionally sort down to a single bset, we have
1971 * to ensure this is still true:
1973 BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1975 bne = want_new_bset(c, b);
1977 bch2_bset_init_next(c, b, bne);
1979 bch2_btree_build_aux_trees(b);
1981 return invalidated_iter;
1985 * Use this one if the node is intent locked:
1987 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1988 enum six_lock_type lock_type_held)
1990 BUG_ON(lock_type_held == SIX_LOCK_write);
1992 if (lock_type_held == SIX_LOCK_intent ||
1993 six_lock_tryupgrade(&b->lock)) {
1994 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1996 /* don't cycle lock unnecessarily: */
1997 if (btree_node_just_written(b) &&
1998 six_trylock_write(&b->lock)) {
1999 bch2_btree_post_write_cleanup(c, b);
2000 six_unlock_write(&b->lock);
2003 if (lock_type_held == SIX_LOCK_read)
2004 six_lock_downgrade(&b->lock);
2006 __bch2_btree_node_write(c, b, SIX_LOCK_read);
2010 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2012 struct bucket_table *tbl;
2013 struct rhash_head *pos;
2018 for_each_cached_btree(b, c, tbl, i, pos)
2019 if (test_bit(flag, &b->flags)) {
2021 wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2028 void bch2_btree_flush_all_reads(struct bch_fs *c)
2030 __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2033 void bch2_btree_flush_all_writes(struct bch_fs *c)
2035 __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2038 void bch2_btree_verify_flushed(struct bch_fs *c)
2040 struct bucket_table *tbl;
2041 struct rhash_head *pos;
2046 for_each_cached_btree(b, c, tbl, i, pos) {
2047 unsigned long flags = READ_ONCE(b->flags);
2049 BUG_ON((flags & (1 << BTREE_NODE_dirty)) ||
2050 (flags & (1 << BTREE_NODE_write_in_flight)));
2055 ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
2057 char *out = buf, *end = buf + PAGE_SIZE;
2058 struct bucket_table *tbl;
2059 struct rhash_head *pos;
2064 for_each_cached_btree(b, c, tbl, i, pos) {
2065 unsigned long flags = READ_ONCE(b->flags);
2066 unsigned idx = (flags & (1 << BTREE_NODE_write_idx)) != 0;
2068 if (//!(flags & (1 << BTREE_NODE_dirty)) &&
2069 !b->writes[0].wait.list.first &&
2070 !b->writes[1].wait.list.first &&
2071 !(b->will_make_reachable & 1))
2074 out += scnprintf(out, end - out, "%p d %u l %u w %u b %u r %u:%lu c %u p %u\n",
2076 (flags & (1 << BTREE_NODE_dirty)) != 0,
2079 !list_empty_careful(&b->write_blocked),
2080 b->will_make_reachable != 0,
2081 b->will_make_reachable & 1,
2082 b->writes[ idx].wait.list.first != NULL,
2083 b->writes[!idx].wait.list.first != NULL);